Ann. Henri Poincar´e 3 (2002) 1 – 17 c Birkh¨ auser Verlag, Basel, 2002 1424-0637/02/010001-17 $ 1.50+0.20/0
Annales Henri Poincar´ e
On the Gap in the Spectrum of the Translations H. J. Borchers Abstract. The spectrum of the translations in local quantum field theory will be analyzed in order to show that in a positive energy representation without vacuum vector and with lowest mass m1 there is no gap in the spectrum which is larger than 2m1 . In particular in a zero mass representation there is no hole at all. These results are obtained with methods of analytic functions of several complex variables.
1 Introduction Let π be a representation of quantum field theory of local observables in the sense of Araki, Haag and Kastler [Ha92], which fulfils the spectrum condition. We assume that the representation of the translation is the minimal one, which implies that its spectrum is located on a Lorentz invariant set of the closed forward light cone V + (see [BB85] and [Bch85]). If the representation is generated from a vacuum state then Wightman’s result [Wi64] implies that the spectrum is an additive set, i.e. it fulfils S + S ⊂ S, (1.1) where S denotes the support of the spectrum. If the representation space does not contain a vacuum vector then I conjectured in [Bch96] that the spectrum fulfils the relation S + S + S ⊂ S. (1.2) In particle physics factor representations are also called super selection sectors because a superposition of vectors in two different sectors does not lead to observable consequences. Therefore, the charges, describing the different sectors, must commute with every observable. Since the time development belongs to the weak closure of the observable algebra it acts separately in every sector. In such a situation one can characterize the factor representations of the observable algebra by charge quantum numbers. Dealing with vectors carrying a charge and adding to this a particle with a non-zero charge, then the total charge changes. This implies that we obtain another sector. Therefore, one stays in the same sector only if we add particles carrying the total charge zero. Let us associate to every sector a charge quantum number Q, then the structure of the spectrum of the translation is known, provided the representation in a charged sector is generated from the vacuum sector by a charged field, in the sense of Doplicher, Haag and Roberts [DHR69a,b]. In this situation one has in the case of Abelian charges the relation SQ1 + SQ2 = SQ1 +Q2 ,
2
H. J. Borchers
Ann. Henri Poincar´e
which has been shown in [Bch65]. SQ denotes the spectrum of the sector characterized by Q. We are interested in a statement concerning only one sector. If the above picture with the charged fields is correct, then the adjoint of a charged field carries the opposite charge. Therefore, one must find to every sector a sector with opposite charge. Since the spectrum does not change by this charge conjugation one has: S−Q = SQ . More generally this relation holds for particles and anti-particles. Since a particle together with its anti-particle form a neutral pair we remain in the same sector if we add to a vector in this sector such a pair. If one takes the particle from the given sector, one obtains Eq. (1.2). In general theories one cannot expect that the representations obtained from a vacuum sector are all representations. In the presence of zero-mass particles the investigations of Doplicher and Spera [DS82], Borek [Bk82] in the non-interacting, and of Buchholz and Doplicher [BD84] in the interacting situation show, that most positive energy representations are not connected with vacuum representations. A similar situation might hold also in a gauge field theory. Therefore, Eq. (1.2) needs a separate investigation. One special result is known for arbitrary sectors [Bch86], which indicates that the conjecture is probably correct. Theorem. Let π be a covariant factor representation of a theory of local observables fulfilling the spectrum condition, and let T (a) be the unique minimal representation of the translations. Assume the spectrum of T (a) consists of two parts: (a) Isolated hyperboloids with the masses m1 < m2 · · · mi < · · · . (b) The rest starting at mc > mi where mc denotes the beginning of the continuous spectrum or the first accumulation point of the mi s. Then 3m1 ≥ mc . One would like to show that above 3m1 there is no gap in the spectrum at all. Unfortunately techniques with which one can derive such result are missing. Since we know that there can be a gap between m1 and 3m1 , such method must distinguish between the regions below 3m1 and above 3m1 . Our procedure does not do it, therefore, we only obtain that there is no gap larger than 2m1 . The technique we are using is that of analytic functions of several complex variables, in particular the edge of the wedge theorem (Brehmermann, Oehme and Taylor [BOT56]). In this case one has two analytic functions holomorphic in the forwardand backward tube respectively. These functions coincide on some real coincidence domain and the edge of the wedge theorem implies that both functions are different branches of one analytic function. The problem consists in calculating the envelope of holomorphy or of parts of it.
Vol. 3, 2002
On the Gap in the Spectrum of the Translations
3
In the next section the details of the problem and the technique to be used will be presented. In our situation the coincidence domain consists of three parts each of which can be handled with help of the Jost-Lehmann-Dyson representation [Dy58]. We will compute the holomorphy domains of the three parts in section 3. In the last section the influence of the different parts on each other will be discussed with help of the double cone theorem. In favorable situations there is a simpler proof of the main result. It is using the ”reentrant nose theorem”, which is a consequence of the Jost-Lehmann-Dyson representation. The method will be described at the end of section 2.
2 Representation of the problem Quantum field theory of local observables in the sense of Araki, Haag and Kastler [Ha92] is concerned with C ∗ -algebras A(O), associated with bounded open regions O ⊂ Rd , which have a common unit. These algebras shall fulfil isotony, i.e. O1 ⊂ O2 implies A(O1 ) ⊂ A(O2 ), and locality, i.e. if O1 and O2 are space-like separated, then the algebras A(O1 ) and A(O2 ) commute element-wise. In order that this statement makes sense the space Rd must be furnished with the Minkowski metric. The global algebra A is defined as the C ∗ -inductive limit of {A(O)}. We also assume that the net {A(O)} is covariant under a representation of the translation group of Rd by automorphisms, i.e. to a ∈ Rd exists an automorphism αa with αa A(O) = A(O + a). A covariant representation π of A in the representation space H is such that there exists a continuous unitary representation T (a) of the translation group of Rd implementing αa T (a)π(A)T (−a) = π(αa (A)),
A ∈ A.
In addition T (a) shall fulfil the spectrum condition, i.e. spectrum T (Rd ) is contained in the closed forward light-cone. As a consequence of the spectrum condition one can choose the representation T (a) of the translation as elements of the von Neumann algebra generated by π(A) [Bch96]. Since T (a) leaves the center of π(A)” point-wise invariant [Ar64] it is no restriction assuming that π is a factor representation. The spectrum of the translations shall have the following structure: It starts at m1 > 0 and in addition there shall be a gap in the spectrum between m2 and m3 , m2 < m3 . We want to show that these assumptions lead to a contradiction if m3 − m2 > 2m1 holds. Since the statement we are looking for is scale invariant we may assume m1 = 1. For the investigation one looks at matrix elements of the form F + (a) = (ψ, BT (a)Aψ)
(2.1)
4
H. J. Borchers
Ann. Henri Poincar´e
where ψ belongs to the representation space Hπ and A, B are operators belonging to π(A(D)). D denotes a double cone with center at the origin. One also looks at the expression (2.2) F − (a) = (ψ, T (a)AT (−a)BT (a)ψ). Locality implies that F + (a) and F − (a) coincide for a ∈ (2D) , where the prime denotes the space-like complement. We are interested in the Fourier transform of (2.1) and (2.2). The support of the Fourier transform of (2.1) is contained in the support of the spectrum S described above. In order that the Fourier transform of (2.2) has good support property we must choose ψ in a suitable manner. The best way is to choose the spectral support of ψ close to the bottom of the spectrum. Let t = (1 + , 0) then we choose the support of ψ in S ∩ (t + V − ). Because of the curvature of the hyperboloid the support of the vector is contained in the double 1 , 0). This choice allows to describe the support of the cone Dt ,t with t = ( 1+ − Fourier transform of F (a). If we denote this support by S then we get S = 2t + (V1− \ {crescent moon}), − − crescent moon = set below 2t + Hm and above 2t + Hm . 2 3
(2.3)
+ In this formula Vm+ characterizes the set {p ∈ V + ; p2 ≥ m2 }, Vm− = −Vm+ . Hm 2 2 − denotes the upper sheet of the hyperboloid p = m and Hm its lower sheet. For the future investigations we choose so small that the crescent moon domain is sufficiently big. The details will be discussed later. About the spectrum of the translations we know definitely only the values mi , i = 1, 2, 3. Whether or not there exists other gaps in the spectrum than that between m2 and m3 is not determined by the formulation of the result Thm. 4.1. Therefore, one knows only vectors with energy-momentum support in the neighbourhood of the masses mi . This motivates the choice of the vector ψ. If m1 = 0, then one has to choose for the support of ψ a small double cone such that its closure is in the interior of the forward light cone. Such vector exists because the theorem in the introduction implies that m1 = 0 is not an isolated hyperboloid. Let F (a) = F + (a)− F − (a) be the commutator function. It vanishes in (2D) . Therefore, we can write F (a) = G+ (a) − G− (a) with
supp G+ (a) ⊂ −b + V + ,
supp G− (a) ⊂ b + V − ,
(2.4)
where b is the upper tip of the double cone 2D. Looking at the Fourier transform we obtain F˜ (p) = F˜ + (p) − F˜ − (p) and
supp F˜ (p) ⊂ S ∪ S .
(2.5)
On the other hand one can write ˜ + (p) − G ˜ − (p), F˜ (p) = G
(2.6)
Vol. 3, 2002
On the Gap in the Spectrum of the Translations
5
˜ + (p) is the boundary value of a function holomorphic in the forward where G + tube T = {z = x + iy ∈ Cd ; x ∈ Rd , y ∈ V + }. Here V + denotes the interior ˜ − (p) is the boundary value of an of the forward light cone. Correspondingly G − + analytic function holomorphic in T = −T . Since the two functions coincide on Rd \ {S ∪ S } we have to deal with an edge of the wedge problem. If one has an edge of the wedge problem where the coincidence domain is bounded by two space-like hyper-surfaces then the domain of holomorphy can be computed with help of the Jost-Lehmann-Dyson formula [Dy58] (see also Bros, Messiah and Stora [BMS61] [Bch96]), i.e. the complement of the holomorphy domain is the union of hyperboloids not entering the coincidence domain. If we assume that the same method would also work in our situation then the coincidence domain could be enlarged provided m3 > 3 holds. But one does not know whether or not also for time-like disconnected coincidence domains a formula like the Jost-Lehmann-Dyson representation holds. Therefore, we have to use other methods of enlarging the domain of holomorphy. If the domain of holomorphy can be enlarged into S \ {2t − V1 } then one can conclude that the spectrum is smaller than assumed. The reason is the following: Let ∆ be the set Dt ,t and E(∆) the corresponding spectral projection of the translation, then we can choose an arbitrary vector in E(∆)H. Moreover, the matrix elements are formed with operators in π(A(D)). Since the sets D are arbitrary, except for the requirement that their center is the origin, any operator in ∪A(D) can be chosen. These are norm-dense in A. Therefore, the conclusion for the spectrum holds for any vector in the closure of π(A)E(∆)H. The projection F on this space belongs to π(A) . Let Z be the central carrier of F then F π(A) and Zπ(A) are isomorphic, and hence we obtain a change of ZS. In the coming investigations we are dealing with factor representations. So one has Z = 1l and the analyticity methods show directly the change of the spectrum.
2.a A simple plausibility argument If one disregards the problem of the support of the vector ψ then there is a direct argument showing that there is no gap of width larger than 2m1 . This method uses the reentrant nose theorem, which follows from the Jost-Lehmann-Dyson representation. Since in this subsection we do not present a correct proof we look at the enlargement of the domain of holomorphy, which cuts into the support of −S convoluted with the double of the support of ψ. For a correct proof one would have to interchange the gap with the crescent moon domain. At the end of this subsection an example will be given showing that this method does not work in all cases, and that one has to use the more complicated methods described in section 3 and 4. Let τ be the vector (τ, 0) and Dτ − ,τ be the double cone {τ − + V + } ∩ {τ + V − } and assume that for τ > m1 the spectral projection E(Dτ − ,τ ) of the translations does not vanish. Choosing τ then the support of F − (p) is (2τ +
6
H. J. Borchers
Ann. Henri Poincar´e
Vm−1 ) \ crescent moon. (The crescent moon part is unimportant for the coming considerations.) Therefore, for m2 < 2τ − m1 < m3
(2.7)
The support of F − (p) cuts into the gap between m2 and m3 . The gap itself is a Jost-Lehmann-Dyson domain. Therefore, one can ask for conditions such that the part of (2τ + Vm−1 ) which cuts into the gap is a reentrant nose such that part of it can be erased by holomorphic completion. This means (gap) \ (2τ + Vm−1 ) is not the complement of hyperboloids not entering this domain. If this is the case then holomorphic completion shows that (2τ + Vm−1 ) is not a stable set, contradicting the assumptions that m1 is the lower boundary of the spectrum. The critical hyperboloid is Hm1 centered at 2τ . Its upper sheet shall not cut the hyperboloid Hm3 . The upper sheets of the two hyperboloids hit each-other if m3 − 2τ > m1
or
m3 − m1 > 2τ
holds. Together with Eq. (2.7) this implies m3 − m1 > m2 + m1 ,
(2.8)
which shows that there is no gap with width larger than 2m1 . The difficulty of this approach is the existence of suitable vectors ψτ which we have to assume in the above approach. That this assumption is not always satisfied shows the following Example: Look at the situation: (i) There is a gap between m1 and 3m1 . (ii) There is a gap between 3m1 and 6m1 . Trying to apply the above method we find that ψτ must be located at the + . Therefore, 2τ is a little above 6m1 and the reentrant nose hyperboloid H3m 1 appears only if m3 > 7m1 . Therefore, the existence of the second gap can not be excluded with this method.
3 The holomorphy domain for subsets of the coincidence region The coincidence domain of our problem consists of three parts. The envelope of holomorphy of each part can be computed separately with help of the JostLehmann-Dyson representation. The union of the three domains will be the starting point for the investigation of the whole problem. The three domains in question are: 1. The complement of {V1+ ∪ (2t + V1− )}.
Vol. 3, 2002
On the Gap in the Spectrum of the Translations
7
2. The gap, i.e. the domain between the two upper sheets of the hyperboloids + + Hm and Hm . 2 3 3. The crescent moon shaped domain, i.e. the domain between the hyperboloids − − − 2t + Hm and 2t + Hm . This means the points below 2t + Hm and above 2 3 2 − 2t + Hm3 . We start with the first coincidence domain and notice that it is invariant under rotations. The complement of the domain of holomorphy is given by the union of d ; (p − u)2 = µ2 } the real or complex points of the hyperboloids H(µ, √ u) = {p ∈ C √ which have their center u at the surface u0 = t2 and u = eλ t2 , where e is a real unit vector in the plane p0 = 0 and 0 ≤ λ ≤ 1. The possible values of µ are determined by the condition that the real part of H(µ, u) is contained in {V1+ ∪ (2t + V1− )}. Fixing u then the possible values of µ are larger or equal than a minimal value µ0 (u). Since the zero-component of u is fixed the function µ0 (u) depends only on the norm of the space-like components of u. √ √ First let us compute the function µ0 (u). To this end set τ = t2 , τ = t2 , which implies the representation u = τ (1, λe). The hyperboloids are given by the formula (p0 − τ )2 − (p − λτ e)2 = µ2 or p0 = τ ± µ2 + (p − λτ e)2 . In order that the upper sheet of this hyperboloid belongs to V1+ there must hold τ + µ2 + (p − λτ e)2 − 1 + p2 ≥ 0. (3.1) Since we want to look at the minimum we can choose p and e parallel and set p = πe. To obtain an equation for minimum one has to differentiate (3.1) with respect to π. π π − λτ −√ . 0= 2 2 1 + π2 µ + (π − λτ ) This leads to a second degree equation for π (π − λτ )2 (1 + π 2 ) (π − λτ )2 (π −
2 λτ 1−µ2 )
with the solution π=
= π 2 (µ2 + (π − λτ )2 ), = π 2 µ2 , =
λ2 τ 2 µ2 (1−µ2 )2 ,
λτ λτ , (1 ± µ) = 2 1−µ 1∓µ
which leads to the expressions 1 + π2 µ2 + (π − λτ )2
= =
1 (1 ∓ µ)2 + λ2 τ 2 , 1∓µ µ (1 ∓ µ)2 + λ2 τ 2 . 1∓µ
8
H. J. Borchers
Ann. Henri Poincar´e
The minimal value of (3.1) shall be zero, which implies τ+
µ − 1 (1 ∓ µ)2 + λ2 τ 2 = 0. 1∓µ
Hence we have to take the − sign. This implies for the minimal µ0 (λ) µ0 (λ) = max{1 − τ 1 − λ2 , 0}.
(3.2)
Since µ has to be non-negative one must choose µ = 0 if the first expression becomes negative. This first expression is the equation of a semi-circle with radius τ. For the holomorphy domain one must find the points p + iq such that the equation (p − u + iq)2 = µ2 ≥ µ20 (3.3) cannot be fulfilled. Eq. (3.3) reads (p − u)2 − q 2 + 2i(q, p − u) = µ2 . Therefore, points of analyticity are: Either one has (q, p − u) = 0 ∀ u, or if (q, p − u) = 0 for one u then (p − u)2 < µ20 (u) + q 2 must hold.
(3.4)
Next we look at the gap as coincidence domain. This set is invariant under Lorentz transformations. Therefore, the surface for the center of the hyperboloids is also invariant, and the same holds for the minimal µ. The centers of 2 2 ) and the minimal µ is the hyperboloids are located on the surface u2 = ( m3 +m 2 m3 −m2 µ0 = . Let (p, q) be given, then exists always a u with (q, p − u) = 0. In 2 order that p + iq belongs to the domain of holomorphy (q, p − u) = 0 and (p − u)2 < µ20 + q 2
∀ u fulfilling the first eqation
(3.5)
must hold. Since (p − u) is time-like this equation can only be fulfilled if −q 2 < 2 2 ) holds. ( m3 −m 2 Finally we look at the crescent moon problem. The interesting part of the surface for the centers of the hyperboloids must be inside this coincidence domain. The maximal value of v := u, which we denote by vm is obtained if (u0 , u) belongs to both hyperboloids, i.e. if 2 = 2τ − 2 m23 + vm u0 = 2τ − m22 + vm holds. With this value for u0 we obtain for vm a second order equation 2 2 2 )(m2 + v 2 ), 4(τ − τ )2 = (m23 + vm ) + (m22 + vm ) − 2 (m23 + vm m 2 2 2 2 (4(τ − τ )2 − 2vm − (m22 + m23 ))2 = 4(m23 + vm )(m22 + vm ),
2 vm 16(τ − τ )2 − 16(τ − τ )4 + 8(τ − τ )2 (m22 + m23 ) − (m23 − m22 )2 = 0,
Vol. 3, 2002
On the Gap in the Spectrum of the Translations
9
which implies 2 vm =
16(τ − τ )4 + (m23 − m22 )2 − 8(τ − τ )2 (m22 + m23 ) . 16(τ − τ )2
(3.6)
2
2) This formula is valid for (τ − τ )2 < (m3 −m as should be. One concludes this 4 2 because vm has to be positive. The corresponding value of u0 is restricted to 2 2 (3.7) 2τ − m3 + v < u0 < 2τ − m22 + v 2 ,
where we have set v = u. The exact value of u0 will be determined together with the minimal value µ0 of µ by the requirement that the hyperboloid H(µ0 , u) has to touch both H − (m2 , 2t ) and H − (m3 , 2t). With π = p one obtains two equations: (3.8a) u0 + µ2 + (π − v)2 − 2τ + m22 + π 2 > 0, u0 − µ2 + (π − v)2 − 2τ + m23 + π 2 < 0. (3.8b) One finds the minimal value of (3.8a) by differentiating this equation with respect to π. π−v 2 µ + (π − v)2
+
(π − v)2 (m22 + π 2 ) =
π = 0, 2 m2 + π 2
π 2 (µ2 + (π − v)2 ).
2 This implies π = mvm . The sign of the square root has to be chosen in such a 2 ±µ manner that π ≤ v holds. The minimal value of (3.8a) is 2 )2 . (3.9a) u0 − 2τ + µ2 + v 2 ( m2µ+µ )2 + m22 + v 2 ( mm 2 +µ
The maximal value of (3.8b) is obtained by π = u0 − 2τ +
vm3 m3 −µ
which leads to
3 µ2 + v 2 ( m3µ−µ )2 + m23 + v 2 ( mm )2 . 3 −µ
(3.9b)
The minimal value of µ is obtained by setting (3.9a) resp. (3.9b) equal to zero. This implies (u0 − 2τ )2
=
= =
2 (µ2 + v 2 ( m2µ+µ )2 ) + (m22 + v 2 ( mm )2 ) + 2 +µ 2 2 2 (µ2 + v 2 ( mµ2 +µ )2 )(m22 + v 2 ( mm )2 ), 2 +µ 2 2 (m22 + µ2 )(1 + (m2v+µ)2 ) + 2 µ2 m22 (1 + (m2v+µ)2 )2 ,
(m2 + µ)2 (1 +
v2 (m2 +µ)2 )
= (m2 + µ)2 + v 2 ,
10
H. J. Borchers
and hence µa =
Ann. Henri Poincar´e
(u0 − 2τ )2 − v 2 − m2 .
(3.10a)
(u0 − 2τ )2 − v 2 .
(3.10b)
In the same manner one finds µb = m3 −
Now we can determine u0 from the equation µa = µb . This leads to (m2 + m3 )2 = ((u0 − 2τ )2 − v 2 ) + ((u0 − 2τ )2 − v 2 )+ 2 ((u0 − 2τ )2 − v 2 )((u0 − 2τ )2 − v 2 ), {(m2 + m3 )2 − ((u0 − 2τ )2 − v 2 ) − ((u0 − 2τ )2 − v 2 )}2 = 4((u0 − 2τ )2 − v 2 )((u0 − 2τ )2 − v 2 ), u20 [−4(m2 + m3 )2 + 16(τ − τ )2 ] + u0 [8(m2 + m3 )2 (τ + τ ) − 32(τ + τ )(τ − τ )2 ] + (m2 + m3 )4 + 4(m2 + m3 )2 v 2 − 8(m2 + m3 )2 (τ 2 + τ 2 ) + 16(τ 2 − τ 2 )2 = 0, 0 = [u0 − (τ + τ )]2 − (τ + τ )2 − {(m2 + m3 )4 + 4(m2 + m3 )2 v 2 − 8(m2 +
1 4[(m2 +m3 )2 −4(τ −τ )2 ] m3 )2 (τ 2 + τ 2 ) + 16(τ 2
− τ 2 )2 },
and hence [u0 − (τ + τ )]2 =
(m2 +m3 )2 4[(m2 +m3 )2 −4(τ −τ )2 ] [(m2
For v = 0 we have to find u0 = (τ + τ ) − take the negative square root, i.e. u0 = (τ + τ ) −
(m2 +m3 ) 2
+ m3 )2 − 4(τ − τ )2 + 4v 2 ].
(m2 +m3 ) . 2
1+
(3.11)
This implies that one has to
4v 2 [(m2 +m3 )2 −4(τ −τ )2 ] .
(3.12)
Inserting this expression into (3.10a) we obtain µ0
= =
(τ − τ ) −
(m2 +m3 ) 2
1+
2 4v 2 [(m2 +m3 )2 −4(τ −τ )2 ]
3) )) 1 + { (m2 +m − (τ − τ 2
− v 2 − m2 ,
4v 2 2 [(m2 +m3 )2 −4(τ −τ )2 ] }
− m2 ,
which implies µ0 =
m3 −m2 2
− (τ − τ ) 1 +
4v 2 [(m2 +m3 )2 −4(τ −τ )2 ] .
(3.13)
For v larger than vm we take the surface u0 (v) = u0 (vm ), v > vm and µ0 (v) = 0.
Vol. 3, 2002
On the Gap in the Spectrum of the Translations
11
4 The case of a big gap In this section we want to treat the gap problem in a situation where techniques of enlarging the domain of holomorphy at real points apply. The method which shall be used is the double cone theorem discovered by Vladimirov [Vl60] and Borchers [Bch61]. We need the local version of this theorem explained in [Bch96]. Let the domain D ⊂ Cd be invariant under the map z → z ∗ . For x ∈ Rd we define the local cone Cx as follows: Cx = {y ∈ Rd ; ∃ ρ > 0
with x + iρy ∈ D}.
By the symmetry the set Cx is either Rd or it contains an even number of components. We are interested in the situation of two components. In the case where x belongs to D ∩ Rd one has Cx = Rd . Since the domain of holomorphy is open, the cone Cx is open too. If D contains real points, then they are called the coincidence ˜ = H(D)∩Rd . Asdomain K. Let H(D) be the envelope of holomorphy of D and K ˜ has a tangent hyper-plane at a point x ∈ ∂ K, ˜ sume that the coincidence domain K then the double cone theorem requires that this hyper-plane is not allowed to enter the local cone Cx . In other terms if for the original coincidence domain K at a point x ∈ ∂K the tangent hyper-plane enters the local cone Cx , then D can be enlarged through this point. In our situation the local cones Cx contain at least the open light-cones V + ∪ V − . With help of this double cone theorem we want to show the following result:
4.1
Theorem
Let π be a translational covariant factor representation of the theory of local observables fulfilling the spectrum condition for the translations. Assume that the representation of T (a) is the unique minimal one. Let the spectrum of the translations start at m1 , then there is no gap with the width larger than 2m1 . Before proving the theorem we draw the following conclusion:
4.2
Corollary
Use the assumptions of the last theorem and assume the spectrum starts at m1 = 0, then there is no gap in the spectrum. This is a simple consequence of the theorem because one has in this case 2m1 = 0. Proof of the Theorem. If m2 coincides with m1 , then the result follows from the theorem mentioned in the introduction. Therefore, it is no restriction if m2 > m1 is assumed in the proof. The result will be shown by contradiction, i.e. the assumption that there is a gap of width larger than 2m1 is not in accordance with the value of the lowest mass m1 . Since the support of the spectrum is Lorentz invariant
12
H. J. Borchers
Ann. Henri Poincar´e
[BB85],[Bch85], it is sufficient to show that one can enlarge the coincidence domain at some parts of the hyperboloid H + (m1 , 0). This will be done by computing the local cones at some points of H + (m1 , 0). Since the complement of Vm+1 ∪ (2t + Vm−1 ) and also the gap between H + (m2 , 0) and H + (m3 , 0) does not lead to an instability of H + (m1 , 0) the interesting contribution for the local cone must come from the crescent moon domain. Our domain is invariant under rotations, therefore, we can restrict our attention to the (0, 1)-plane. Since the domain of analyticity we are starting with is open it is sufficient to compute Cp ∩ R2 , where the plane in question is the (0, 1)-plane, because Cp is open in Rd . For the calculations we use again m1 = 1. We start with the local cone obtained from the complement of {V1+ ∪ (2t + − V1 )}, and afterwards we look at the enlargement of this cone influenced by the crescent moon domain. The local cone contains always the interior of the light cone, therefore, one is interested in light-like and space-like directions. In order to find them we recall first that the hyperboloids H(m, u), filling the complement of the holomorphy domain, are real. Therefore, one must have (p − u, q) = 0 if (p − u + iq) belongs to such hyperboloid. In addition one must have (p − u)2 − q 2 = m2 . So we have to look at points (p − u + iq) where one of the two conditions is violated for every of the hyperboloid filling the complement of the domain of holomorphy. The hyperboloids of interest are those which are centered in the hyper-surface described in the last section. Let p be the point of interest then there are two possibilities of finding directions of the local cone Cp . 1. The set of directions of p − u form a cone C. Then (p − u + iq) belongs to the domain of holomorphy if q is not perpendicular to any of these directions, ˆ i.e. if q or −q belongs to the interior of the dual cone C. 2. If (p − u) belongs to C and q ⊥ (p − u) then q belongs to the local cone Cp if the equations
(p − u)2 − q 2 < µ0 (u)2 ,
(p − u, q) = 0,
(4.1)
can be fulfilled. Altogether the local cone at p will consist of three parts. Two of them are consequences of the coincidence domain {complement of Vm+1 ∪ (2t + Vm−1 )} and the third part is the contribution of the crescent moon domain. (See Fig.1) We choose a point p ∈ H + (1, 0) with p1 > 0 and p21 > 1 − τ12 . This condition implies that the set p − V + does not meet the intersection of V1+ with 2t − V1+ . With this choice p − V + meets the surface for the centers of the hyperboloids in the interval [λ0 τ, τ ] where λ0 is obtained from the relation ( 1 + p21 , p1 ) − α(1, 1) = (τ, λ0 τ ),
Vol. 3, 2002
On the Gap in the Spectrum of the Translations
13
(2) (1)
tangent
(3)
Figure 1: Structure of the local cone (1) y ⊥ (p − u) ∀ u (2) y ⊥ (p − u) and (p − u)2 − y 2 < µ20 (u) (3) Contribution of the crescent moon domain. which implies 1 + p21 − α = τ, (1 − λ0 )τ
=
p1 − α = λ0 τ. 1 + p21 − p1 ,
and consequently
λ0 τ = τ + p1 − 1 + p21 . (4.2) If we take the vector u = √ τ 2 ( 1 + p21 , p1 ) in the u-hyper-surface, then we 1+p1
obtain (p − u)2 = (1 − √ τ 2 ). In this situation one has λ1 = √ p1 2 and √ 1 2 = 1+p1 1+p1 1+p1 2 1 − λ1 . This implies (p − u(λ1 ))2 = µ0 (λ1 ). (4.3) If λ0 ≤ λ < λ1 , then the second case can be fulfilled because one finds with p1 = √ λ1 2 the relation 1−λ1
(p − u(λ)) = eτ (λ1 − λ) + (p − u(λ1 )),
e2 = −1.
This implies with Eq. (3.2), (4.3) and the above value of p1 (p − u(λ))2
= (p − u(λ1 ))2 − τ 2 (λ1 − λ)2 − 2τ (λ1 − λ)(p1 − τ λ1 ) = µ0 (λ)2 − (1 + τ 2 (1 − λ2 ) − 2τ 1 − λ2 ))(1 + τ 2 (1 − λ21 ) − 2τ 1 − λ21 ) − τ 2 (λ21 + λ2 − 2λλ1 ) − 2τ (λ1 − λ)( √ λ1 2 − τ λ1 ) 1−λ1
14
H. J. Borchers
Ann. Henri Poincar´e
1 (λ1 −λ) = µ0 (λ)2 − 2τ ( 1 − λ21 − 1 − λ2 + λ√ ) 1−λ21 = µ0 (λ)2 − √ 2τ 2 (1 − λ1 λ − (1 − λ1 λ)2 − (λ1 − λ)2 ) 1−λ1 λ1 −λ 2 ) ) = µ0 (λ)2 − √ 2τ 2 (1 − λ1 λ − (1 − λ1 λ) 1 − ( 1−λ 1λ 1−λ1
2
< µ0 (λ) −
2 1 (λ1 −λ) √ 2τ 2 2 1−λ λ 1−λ 1
< µ0 (λ)2 .
Next we turn to the crescent moon domain. Since can be chosen arbitrarily small we take first = 0 in order to see what can be expected. In this situation the u-hyper-plane is met by every time-like direction starting from p. Looking for conditions implying that the local cone Cp contains the tangent plane at p one takes p − u0 = (1 + α)p. Since p2 = 1 and q 2 = −β 2 equation (4.1) reads (1 + α)2 + β 2 < µ20 . The quantity α is a function of p1 . In order that the tangent direction belongs to the local cone one must have µ0 > 1 + min{α(p1 )}. Later we 2 will show that α → 0 for p1 → ∞. Since µ0 = m3 −m , one obtains an enlargement 2 of the local cone for µ0 > 1. With µ0 − 1 = δ and the choice α(p1 ) < 2δ one can vary β in the interval (0, δ + 34 δ 2 ). Since (p − u) depends continuously on u one can still vary u in a small neighbourhood around u0 without violating Eq. (4.1), i.e. the local cone at p contains the tangent direction at p as an interior direction. It remains to show α(p1 ) → 0 for p1 → ∞ and to look at the variation for small 3 . With m2 +m = M , the point u0 must lie on the hyperboloid H − (M, 2t) which 2 leads to the equations −αp = 2t − ( M 2 + r2 , r), −αp1 = −r, −α 1 + p21 = 2 − M 2 + r2 ,
and hence 2=
M 2 + α2 p21 − α 1 + p21 .
This implies 4 = M 2 + α2 (1 + 2p21 ) − 4(M 2 + α2 p21 )α2 (1 + p21 ) = −α4 + α2 {2M 2 + α2 = (M 2 + 4(1 + 2p21 )) ±
2 (M 2 + α2 p21 )α2 (1 + p21 ), {(M 2 − 4) + α2 (1 + 2p21 )}2 , 8(1 + 2p21 )} = (M 2 − 4)2 , (M 2 + 4(1 + 2p21 ))2 − (M 2 − 4)2 .
Since α is bounded by 1 for all p1 we must choose the minus sign and get 1 (M 2 −4)2 2 −4)2 ≈ 2 M 2 +4(1+2p2 ) . α2 = (M 2 + 4(1 + 2p21 )) 1 − 1 − (M 2(M +4(1+2p2 ))2 1
This tends to zero for p1 → ∞ as claimed before.
1
(4.4)
Vol. 3, 2002
On the Gap in the Spectrum of the Translations
15
Finally we must look at the approximation in . For the future calculations 1 1 one sets σ = +( 1+ −1), δ = −( 1+ −1) which implies τ +τ = 2+δ, τ −τ = σ. The crescent moon domain becomes smaller for δ = 0, which simplifies the coming calculations. With these notations we obtain with (3.12) and (3.13) the following equations for α and µ0 (v). −α 1 + p21
= −
2 2 − M 2 + v 2 MM −αp1 = −v, 2 −σ 2 , 2 α 1 + p21 = 2 − M 2 + α2 p21 MM 2 −σ 2 ,
and µ0 =
m3 −m2 2
−σ
1+
α2 p21 M 2 −σ2 .
(4.5)
(4.6)
In order that the tangent at p belongs to the local cone one has to choose small enough such that m3 − m2 α2 p2 − 1 > σ 1 + M 2 −σ1 2 2 holds. To show that this is possible we have to calculate p1 as a function of α. 2 With Eq. (4.5) and with Γ = MM 2 −σ 2 = 1 − γ one obtains as equation for α 4 = M 2 + α2 (1 + p21 (1 + Γ)) − 2 (M 2 + α2 p21 Γ)α2 (1 + p21 ), 4(M 2 + α2 p21 Γ)α2 (1 + p21 ) = {(M 2 − 4) + α2 (1 + p21 (1 + Γ))}2 , 4{M 2 α2 (1 + p21 ) + α4 (p21 Γ(1 + p21 ))} = (M 2 − 4)2 + 2(M 2 − 4)α2 (1 + p21 (1 + Γ))+ α4 (1 + 2p21 (1 + Γ) + p41 (1 + Γ)2 ), p41 {−α4 γ 2 } + p21 {2α4 γ + 8α2 (1 + Γ) − 2M 2 α2 γ} + {−α4 + α2 (2M 2 + 8) − (M 2 − 4)2 } = 0, 16α2 +γ(2α4 −α2 (2M 2 −8)) (+α4 −α2 (2M 2 +8)+(M 2 −4)2 )α4 γ 2 2 1 − . (4.7) 1 − p1 = α4 γ 2 {16α2 +γ(2α4 −α2 (2M 2 −8))}2 One must choose p1 at least as big as stated in the formula. This is not too different from the value of the case = 0. With (p − u) = (1 + α)p one chooses α such that the inequality (4.1) is fulfilled. Then one can find an appropriate value for p1 using Eq. (4.7) and finally a finite value for with help of (4.6). Hence the local cone Cp contains the tangent at p, which leads to the desired contradiction. Acknowledgment. I like to thank Jacques Bros for discussions.
16
H. J. Borchers
Ann. Henri Poincar´e
References [Ar64]
H. Araki, On the algebra of all local observables, Progr. Theor. Phys. 32, 844–854 (1964).
[Bch61]
¨ H. J. Borchers, Uber die Vollst¨andigkeit lorentzinvarianter Felder in einer zeitartigen R¨ohre, Nuovo Cimento 19, 787–796 (1961).
[Bch65]
H. J. Borchers, Local Rings and the Connection of Spin with Statistics, Commun. Math. Phys. 1, 281–307 (1965).
[Bch85]
H. J. Borchers, Locality and covariance of the spectrum, Fizika 17, 289– 304 (1985).
[Bch86]
H. J. Borchers, A Remark on Antiparticles, in Lecture Notes in Physics Vol 257, 268–280 (1986).
[Bch96]
H. J. Borchers, Translation Group and Particle Representations in Quantum Field Theory, Lecture Notes in Physics m40 Springer, Heidelberg (1996).
[BB]
H. J. Borchers and D. Buchholz, The Energy-Momentum Spectrum in Local Field Theories with Broken Lorentz-Symmetry, Commun. Math. Phys. 97, 169–185 (1985).
[Bk82]
R. Borek, Kovariante Darstellungen Masseloser Fermi Felder, Dissertation, G¨ ottingen (1982).
[BOT]
H. J. Brehmermann, R. Oehme and J. G. Taylor, Proof of dispersion relation in quantized field theories, Phys. Rev. 109, 2178–2190 (1958).
[BMS]
J. Bros, A. Messiah and R. Stora, A problem of analytic completion related to the Jost-Lehmann-Dyson formula, J. Math. Phys. 2, 639–651 (1961).
[BD84]
D. Buchholz and S. Doplicher, Exotic Infrared Representations of interacting Systems, Ann. Inst. Henri Poincar´e 40, 157–184 (1994).
[DHR69a] S. Doplicher, R. Haag and J. E. Roberts, Fields, observables and gauge transformations I, Commun. Math. Phys. 13, 1 (1969). [DHR69b] S. Doplicher, R. Haag and J. E. Roberts, Fields, observables and gauge transformations II, Commun. Math. Phys. 15, 173 (1969). [DS82]
S. Doplicher and M. Spera, Representations obeying the spectrum condition, Commun. Math. Phys. 84, 505–513 (1982).
[Dy58]
F. J. Dyson, Integral representation of causal commutators, Phys. Rev. 110, 1460–1464 (1958).
Vol. 3, 2002
On the Gap in the Spectrum of the Translations
17
[Ha92]
R. Haag, Local Quantum Physics, Springer Verlag, Berlin-HeidelbergNew York (1992).
[Vl]
V. S. Vladimirov, The construction of envelopes of holomorphy for domains of special type, Doklady Akad. Nauk SSSR 134, 251 (1960).
[Wi64]
A. S. Wightman, La th´eorie quantique locale et la th´eorie quantique des champs, Ann. Inst. Henri Poincar´e A, I, 403–420 (1964).
Hans J¨ urgen Borchers Institut f¨ ur Theoretische Physik Universit¨ at G¨ ottingen Bunsenstrasse 9 D-37073 G¨ ottingen Germany Communicated by Klaus Fredenhagen submitted 19/04/01, accepted 02/10/01
Ann. Henri Poincar´e 3 (2002) 19 – 27 c Birkh¨ auser Verlag, Basel, 2002 1424-0637/02/010019-9 $ 1.50+0.20/0
Annales Henri Poincar´ e
Singular Vectors of the N = 1 Superconformal Algebra K. Iohara and Y. Koga Abstract. We give two singular vector formulae of the N = 1 superconformal algebra.
1 Introduction Singular vectors of Verma modules over the Virasoro algebra play important roles both in mathematics and physics. In fact, many studies on their explicit forms have been done e.g., [Ke] etc. In particular, two formulae were obtained by B. L. Feigin and D. B. Fuchs [FeFu1], and these formulae played an essential role when they determined the structure of the fusion rings of the minimal models [FeFu2]. In this paper, we will derive similar formulae for the N = 1 superconformal algebra, i.e., the Neveu-Schwarz algebra. In our case, one technical difficulty arises, viz. factorizations in a non-commutative ring are required (see Remark 3.2 and Theorem 3.5). We note that one of our formulae, i.e. Theorem, 3.1 has been partially obtained in [BS] in a special case. Namely, they obtained their formula explicitly in the case when one of a and b is one and provided an algorithm to compute the general case via fusion procedure. Our method is based on the embedding diagram given in Proposition 2.3 and the result is given for the general cases. This paper is organized as follows: In §2, we will recall the definition of the N = 1 superconformal algebra and its Verma modules. Further, we will collect some properties of them. In §3, we will derive two singular vector formulae.
2 N = 1 Superconformal Algebra Here we recall the definition of the N = 1 superconformal algebra. Further we introduce Verma modules and state their properties.
2.1
Definitions
Let V ir 12 be the N = 1 superconformal algebra, the Neveu-Schwarz algebra, i.e., V ir 12 is the Lie superalgebra V ir 12 =
i∈Z
CLi ⊕
j∈Z+ 12
CGj ⊕ CC,
20
K. Iohara and Y. Koga
Ann. Henri Poincar´e
and these generators satisfy the following commutation relations: deg Li = deg C = 0, deg Gi = 1, 1 [Li , Lj ] = (i − j)Li+j + (i3 − i)δi+j,0 C, 12 1 [Gi , Lj ] = (i − j)Gi+j , 2 1 1 [Gi , Gj ] = 2Li+j + (i2 − )δi+j,0 C, 3 4 [C, V ir ] = {0}. Note that V ir 12 admits the following triangular decomposition: − 0 V ir 12 = V ir+ 1 ⊕ V ir 1 ⊕ V ir 1 , 2
2
where
V ir± 1 := 2
For later use, we set
CLi ⊕
±i>0
V ir01 := CL0 ⊕ CC. 2
±i>0
+ 0 V ir≥ 1 := V ir 1 ⊕ V ir 1 . 2
2.2
CGi ,
2
2
2
Verma Modules
Next, we introduce Verma modules over V ir 12 and collect some basic properties of them. For (z, h) ∈ C2 , let Cz,h = C1z,h be the 1-dimensional representation of V ir≥ 1 defined by 2
x.1z,h = 0
if x ∈ V ir+ 1 , 2
C.1z,h = z1z,h , L0 .1z,h = h1z,h . The Verma module M (z, h) over V ir 12 with highest weight (z, h) is defined as follows: M (z, h) := U (V ir 12 ) ⊗U(V ir≥ ) Cz,h . 1 2
The values z and h are called central charge and conformal weight, respectively. For simplicity, we set |z, h := 1 ⊗ 1z,h . Since the choice of the parity of a highest weight vector is not essential in the sequel, we do not specify it. Note that M (z, h) has the following weight space decomposition: M (z, h) = M (z, h)n , M (z, h)n := {v ∈ M (z, h)|L0 .v = (h + n)v}. n∈ 12 Z≥0
A non-zero vector in {M (z, h)n }
Vir+ 1 2
is called a singular vector of level n.
Vol. 3, 2002
Singular Vectors of the N = 1 Superconformal Algebra
21
Here we recall two basic facts on singular vectors. The first fact is the existence of singular vectors. As a consequence of the determinant formula for M (z, h) [KW], we see the following: If there exist t ∈ C \ {0} and a, b ∈ Z>0 satisfying a ≡ b (mod 2) and 15 − 3(t + t−1 ), 2 1 1 1 h = ha,b (t) := (a2 − 1)t − (ab − 1) + (b2 − 1)t−1 , 8 4 8 z = z(t) :=
then M (z, h) has a singular vector of level 12 ab. The second fact is the uniqueness of singular vectors. Indeed, we have Proposition 2.1 ([IK1]) For any highest weight (z, h) ∈ C2 , the following hold: dim{M (z, h)n}
2.3
Vir + 1 2
≤1
(∀n ∈
1 Z≥0 ). 2
Embedding Diagrams of Verma Modules
To prove singular vector formulae, we will use embedding diagrams of Verma modules [IK1]: Here we recall them. It should be noted that non-trivial homomorphisms between Verma modules over a Lie superalgebras are not injective in general. But for the N = 1 superconformal algebra, the injectivity holds. Proposition 2.2 ([IK1]) Suppose that M (z, h) has a singular vector of level n. Then the following map is an embedding: M (z, h + n) −→ M (z, h). When z = z(t) for t ∈ Q \ {0}, we have the following embedding diagrams of Verma modules. Let p and q be integers such that z = z(± pq ), p, q ≥ 2, p − q ∈ 2Z and ( p−q 2 , q) = 1. For the integers p and q, we set ± ± ± ± ± := Kp,q (1) ∪ Kp,q (2) ∪ Kp,q (3) ∪ Kp,q (4), Kp,q ± (i) are give by where Kp,q ± (1) := {(r, s) ∈ Z2 |0 < r < q, 0 < ±s < p, rp + sq ≤ pq}, Kp,q ± (2) := {(q, s)|0 < ±s < p}, Kp,q ± (3) := {(r, −p)|0 < r < q}, Kp,q ± Kp,q (4) := {(0, ±p), (q, ±p)}.
22
K. Iohara and Y. Koga
Ann. Henri Poincar´e
± For (r, s) ∈ Kp,q , we also set h(i−1)q+r,−s ± p i ≡ 1 (mod 2) q . hi := p hiq+r,s ± i ≡ 0 (mod 2) q
Then we have the following proposition: Proposition 2.3 Let us take integers p, q, r and s as above. Then there exist the following embedding diagrams of Verma modules: + For (r, s) ∈ Kp,q (1),
/ M (z, h3 ) / M (z, h2 ) / M (z, h1 ) · · · M (z, h4 ) UUU* 9 KKK KKK MMM q8 s9 q s ss K K q MM s s q K K s s M (z, h0 ). qqMMMM ssKKKK ssKKKK qqq sss sss & % % iii4 · · · M (z, h−4 ) / M (z, h−3 ) / M (z, h−2 ) / M (z, h−1 ) + (2), For (r, s) ∈ Kp,q
· · · M (z, h4 )
/ M (z, h3 )
/ M (z, h2 )
/ M (z, h3 )
/ M (z, h−2 )
/ M (z, h1 )
/ M (z, h0 ).
+ (3), For (r, s) ∈ Kp,q
· · · M (z, h−4 )
/ M (z, h1 )
/ M (z, h0 ).
+ For (r, s) = (p, q) ∈ Kp,q (4)
· · · M (z, h8 )
/ M (z, h6 )
/ M (z, h4 )
/ M (z, h2 )
/ M (z, h4 )
/ M (z, h−2 )
/ M (z, h0 ).
+ (4) For (r, s) = (0, q) ∈ Kp,q
· · · M (z, h8 )
/ M (z, h−6 )
/ M (z, h0 ).
− (1), For (r, s) ∈ Kp,q
M (z, h3 ) o · · · M (z, h4 ) o M (z, h2 ) o M (z, h1 ) jU fMMM eKKK eKKK UU q s q s s MM KKs KKsKs s q s q K M q s s M (z, h0 ). K K M q s s KK KK MM tiii xqqq ysss ysss · · · M (z, h−4 ) o M (z, h−3 ) o M (z, h−2 ) o M (z, h−1 ) − For (r, s) ∈ Kp,q (2),
· · · M (z, h4 ) o
M (z, h3 ) o
M (z, h2 ) o
M (z, h1 ) o
M (z, h0 ).
Vol. 3, 2002
Singular Vectors of the N = 1 Superconformal Algebra
23
− For (r, s) ∈ Kp,q (3),
· · · M (z, h−4 ) o
M (z, h3 ) o
M (z, h−2 ) o
M (z, h1 ) o
M (z, h0 ).
− For (r, s) = (p, −q) ∈ Kp,q (4),
· · · M (z, h8 ) o
M (z, h6 ) o
M (z, h4 ) o
M (z, h2 ) o
M (z, h4 ) o
M (z, h−2 ) o
M (z, h0 ).
− For (r, s) = (0, −q) ∈ Kp,q (4),
· · · M (z, h8 ) o
M (z, h−6 ) o
M (z, h0 ).
3 Singular Vector Formulae In this section, we will present our main results, singular vector formulae.
3.1
Notation
Suppose that z = z(t) and h = ha,b (t) for some a, b ∈ Z>0 such that a ≡ b (mod 2). As stated before, there exists a unique element (up to a scalar multiple) −1 ] Sa,b (t) ∈ U (V ir− 1 ) ⊗C C[t, t 2
such that Sa,b (t)|z, h is a non-zero singular vector of level 12 ab. From now on, we will give two formulae on Sa,b (t).
3.2
Loop Modules
The first formula is concerned with the representations over V ir 12 called loop modules. Although, this formula has already been considered in [IK2], we have stated there without proof. Here, we will show a way to prove this formula. Let θ be the Grassmann variable i.e. θ2 = 0. For λ, µ ∈ C and σ ∈ {0, 1}, we set σ Fλ,µ = CFi ⊕ CFi θ, i∈Z+σ 12
i∈Z+(1−σ) 12
and regard it as a V ir -module via C.Fi θγ = 0, 1 Ln .Fi θγ = {−i + µ + (n − 1)λ − nγ}Fi+n θγ , 2 Gm .Fi θγ = δγ,1 Fi+m + δγ,0 {−i + µ + (2m − 1)λ}Fi+m θ, where γ = 0, 1, n ∈ Z and m ∈ Z + 12 . σ Here we describe the action of Sa,b (t) on Fλ,µ explicitly. To do it, we set 1 if i ≡ j (mod 2), (2) δi,j = 0 if i ≡ j (mod 2).
24
K. Iohara and Y. Koga
Ann. Henri Poincar´e
σ Then Sa,b (t) acts on F0 θγ ∈ Fλ,µ as follows:
Theorem 3.1 ([IK2]) Suppose that a, b ∈ Z>0 such that a ≡ b (mod 2). Then we have (2) γ Sa,b (t).F0 θγ = Pa,b (λ, µ, t)F− 12 ab θδab,1−γ , γ where Pa,b (λ, µ, t)2 is, up to a scalar depending on the normalization, given by
Qk,l a,b (λ, µ, t)
0≤k≤a−1 0≤l≤b−1 k−l≡γ (mod 2)
and Qk,l a,b (λ, µ, t)
1 1 1 1 1 = (µ − 2λ) − (kt 2 − lt− 2 ){(a − k)t 2 − (b − l)t− 2 } 2
1 1 1 1 1 × (µ − 2λ) − {(k + 1)t 2 − (l + 1)t− 2 }{(a − k − 1)t 2 − (b − l − 1)t− 2 } 2 2 1 1 1 + (a − 1 − 2k)t 2 − (b − 1 − 2l)t− 2 λ. 2 (1)
σ Proof. Since every weight space of the loop module Fλ,µ is of dimension 1, there γ −1 exists Φa,b (λ, µ, t) ∈ C[λ, µ, t, t ] which satisfies (2)
Sa,b (t).F0 θγ = Φγa,b (λ, µ, t)F− 12 ab θδab,1−γ . γ From now on, we verify Φγa,b (λ, µ, t) = Pa,b (λ, µ, t). To do it, we first compute the γ degree of Φa,b (t) as a polynomial in t. Following [AsFu], one obtains
degΦγa,b (t) ≤
ab + 1 , 2
where x denotes the maximal integer that does not exceed x. Let us prove γ Φγa,b (λ, µ, t) = Pa,b (λ, µ, t) by induction on the level of singular vectors. It follows from Proposition 2.3 that there exists t0 ∈ Q \ {0} such that the singular vector of level 12 ab factorize as Sa1 ,b1 (t0 )Sa2 ,b2 (t0 ) · · · Sak ,bk (t0 )|z(t0 ), ha,b (t0 ) ,
(2)
k where k > 1 and ai , bi ∈ Z>0 satisfy ai ≡ bi (mod 2) and 12 ab = i=1 21 ai bi . By the uniqueness of singular vectors and induction hypothesis, we have Φγa,b (λ, µ, t0 ) ∝
k i=1
Paγi ,bi (λ, µ, t0 ).
(3)
Vol. 3, 2002
Singular Vectors of the N = 1 Superconformal Algebra
25
γ Here, one can show that the right hand side of (3) coincides with Pa,b (λ, µ, t) by direct computation. On the other hand, by Proposition 2.3, we can check that the number ({t0 } such that the singular vector of level 12 ab factorize as (2) is greater than the γ γ degree of Pa,b (λ, µ, t). This implies that Φγa,b (λ, µ, t) = Pa,b (λ, µ, t) holds for any t ∈ C \ {0}. We have completed the proof.
3.3
Projections of Singular Vectors
Next we describe an explicit form of the image of Sa,b (t) under the projection − − − − − π : U (Vir− 1 ) −→ U (Vir 1 )/U (Vir 1 )[Vir 1 , [Vir 1 , Vir 1 ]]. 2
2
2
2
2
2
(4)
Remark 3.2 In [FeFu1], a similar projection was computed for the Virasoro algebra. One of the differences between their case and our case is whether the quotient space is a commutative ring or not. Indeed, the quotient space in (4) is noncommutative, and, to derive the formula, we have to consider factorizations in this non-commutative space. We introduce Xp,q , Yr ∈ U (Vir− 1 ) for p, q, r ∈ C as follows: 2
Xp,q :=
L2−1
− pG− 12 G− 32 − qG− 32 G− 12 ,
Yr := L−1 G− 12 − rG− 32 . Then we have the following lemma: Lemma 3.3
1. π(Xp,r )π(Xq,s ) = π(Xp,s )π(Xq,r ) = π(Xq,r )π(Xp,s ).
2. π(Xp,q )π(Yr ) = π(Xp,r )π(Yq ) = π(Yr )π(Xp,q ). To describe the image of π(Sa,b (t)), we introduce some notation. Let us first introduce an index set Pσa,b for σ ∈ {0, 1} and a, b ∈ Z>0 such that a − b ≡ 0 (mod 2). Set (i) (r, s) ∈ (Z>0 × Z) ∪ ({0} × Z≥σ ), (ii) 0 ≤ r ≤ a − 1 ∧ −(b − 1) ≤ s ≤ b − 1, σ , Pa,b := (r, s) (iii) a − 1 − r ≡ 0 (mod 2) ∧ b − 1 − s ≡ 0 (mod 2), (iv) a + b − (r + s) ≡ 2(1 − σ) (mod 4) and nσa,b := (Pσa,b . We remark that n0a,b − n1a,b =
0 1
if a ≡ b ≡ 0 mod 2 . if a ≡ b ≡ 1 mod 2
Moreover, let us take a sequence {(r1σ , sσ1 ), (r2σ , sσ2 ), · · · , (rnσ , sσn )}
(n := nσa,b )
of elements Pσa,b , which satisfies the following conditions:
26
K. Iohara and Y. Koga
Ann. Henri Poincar´e
1. (riσ , sσi ) = (rjσ , sσj ) for any i, j (1 ≤ i < j ≤ n), 2. if (0, 0) ∈ Pσa,b , i.e., σ = 0, a ≡ b ≡ 1 (mod 2) and a + b ≡ 2 (mod 4), then (rnσ , sσn ) = (0, 0). Next, we define a subgroup Ga,b of the symmetric group Sn , where n := nσa,b , by Ga,b :=
{τ ∈ Sn |τ (n0a,b ) = n0a,b } Sn
if (0, 0) ∈ Pσa,b . otherwise
For i ∈ Z such that 1 ≤ i ≤ n0a,b and τ ∈ Ga,b , we define an element mτi ∈ P0a,b × (P1a,b ∪ {vac}) as follows: mτi := Further, we set
[(rτ0 −1 (i) , s0τ −1 (i) ), (ri1 , s1i )]
if 1 ≤ i ≤ n1a,b
[(rτ0 −1 (n0 ) , s0τ −1 (n0 ) ), vac] a,b a,b
if n0a,b = n1a,b ∧ i = n1a,b
.
Ωτa,b := {mτi |1 ≤ i ≤ n0a,b }.
Finally, for m ∈ Ωτa,b , we set
Sm a,b
and
1 1 1 1 X 14 (r t 2 +s t− 2 )2 , 14 (r t 2 +s t− 2 )2 := Y 1 12 − 12 2 (rt +st ) 4 1 G− 2 mτ
if m = [(r, s), (r , s )] if m = [(r, s), vac] for (r, s) = (0, 0) , if m = [(0, 0), vac] mτ
mτ
Sa,b;τ := Sa,b1 Sa,b2 · · · Sa,bn , where n := n0a,b . By Lemma 3.3, we have Lemma 3.4 Sa,b;τ does not depend on the choice of τ ∈ Ga,b . Consequently, the image of Sa,b (t) under the projection π in (4) is described as follows: Theorem 3.5 Suppose that a, b ∈ Z>0 such that a ≡ b (mod 2). Then we have π(Sa,b (t)) ∝ Sa,b;τ up to scalar multiplication depending on the normalization. Proof. The proof of this formula is essentially the same as the proof of Theorem 3.1. Hence we omit it. Acknowledgment. We would like to thank the referee for pointing out the reference [BS], where a similar result was obtained.
Vol. 3, 2002
Singular Vectors of the N = 1 Superconformal Algebra
27
References [AsFu]
A. Astashkevich and D. B. Fuchs, Asymptotics for singular vectors in Verma modules over the Virasoro algebra, Pacific J. Math. 177, no. 2, 201–209 (1997).
[BS]
L. Benoit and Y. Saint-Aubin, Fusion and the Neveu-Schwarz Singular Vectors, Int. Jour. Mod. Phys. A 9, 547–566 (1994).
[FeFu1]
B. L. Feigin and D. B. Fuchs, Verma modules over the Virasoro algebra, Lecture Notes in Math., 1060, 230–245 (1982).
[FeFu2]
B. L. Feigin and D. B. Fuchs, Cohomology of some nilpotent subalgebras of the Virasoro and Kac-Moody Lie algebras, J. Geom. Phys. 5, 209–235 (1988).
[IK1]
K. Iohara and Y. Koga, Representation Theory of Neveu-Schwarz and Ramond Algebras I: Verma Modules, preprint.
[IK2]
K. Iohara and Y. Koga, Fusion algebras for N=1 superconformal field theories through coinvariants II: N = 1 super Virasoro symmetry, J. Lie Theory 11, 305–337 (2001).
[Ke]
A. Kent, Singular vectors of the Virasoro algebra, Phys. Lett. B 273, 56–62 (1991).
[KW]
V. G. Kac and M. Wakimoto, Unitarizable highest weight representation of the Virasoro, Neveu-Schwarz and Ramond algebras, Lect. Notes in Phys., 261, 345–372, (1986).
Kenji Iohara Department of Mathematics Faculty of Science Kobe University Kobe 657-8501 Japan email:
[email protected] Communicated by Tetsuji Miwa submitted 26/01/01, accepted 08/09/01
Yoshiyuki Koga Department of Mathematics Faculty of Science and Technology Sophia University Tokyo 102-8554 Japan email:
[email protected]
Ann. Henri Poincar´e 3 (2002) 29 – 86 c Birkh¨ auser Verlag, Basel, 2002 1424-0637/02/010029-58 $ 1.50+0.20/0
Annales Henri Poincar´ e
Front Fluctuations in One Dimensional Stochastic Phase Field Equations L. Bertini∗, S. Brassesco†, P. Butt` a‡ and E. Presutti‡
Abstract. We consider a conservative system of stochastic PDE’s, namely a weakly coupled, one dimensional phase field model with additive noise. We study the fluctuations of the front proving that, in a suitable scaling limit, the front evolves according to a non–Markov process, solution of a linear stochastic equation with long memory drift.
Part I. Introduction 1 General setting, model, and results The term “sharp interface limit” denotes a scaling procedure aimed at the derivation of interfaces as geometric objects, e.g. surfaces of codimension one with bounded variation, that is, enough regular for the area measure to be well defined. Of course this makes only sense in the context of systems which undergo phase transitions and of states where different phases coexist. In the limit the other degrees of freedom are lost and we are left with the interface alone. Rigorous proofs are hard, yet a great variety of models has been successfully worked out. The mathematics involved is correspondingly rich, e.g. the theory of Γ–convergence (to study the sharp interface limit of Ginzburg–Landau like free energy functionals in relation with the equilibrium shape of the interface, as in the Wulff problem) and correspondingly the theory of Gibbsian large deviations (to study the same problems at the more microscopic level of statistical mechanics); singular limit in PDE’s, like in the Allen–Cahn, Cahn–Hilliard and phase field equations, and correspondingly, at the microscopic level, hydrodynamic limits of spin or particle systems. This paper deals with fluctuations. Here again the questions are, first, whether in a sharp interface limit the system is described by a [fluctuating] interface with closed equations of motion and, secondly, the nature of such equations. The problem greatly simplifies in one space dimension where the limit interface is represented by a point which separates the two phases (one to its left and the other one ∗ Partially
supported by Cofinanziamento MURST 1999. supported by agreement CONICIT–CNR, and Proyecto F–97 000004 (CONICIT). ‡ Partially supported by Cofinanziamento MURST 1999 and by NATO Grant PST. CLG. 976552. † Partially
30
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
to its right). Indeed, until now, most fluctuation results are restricted to d = 1. Such a state is marginally stable (w.r.t. displacements of the interfaces, see Section 2 for a precise statement) while it becomes stable when there is a conservation law (which imposes that the mass of each phase is preserved). By the conservation law, each displacement of the interface must then necessarily come with a deformation of the profile, hence the intuition that other degrees of freedom may then enter into play. This is confirmed, but in a sense also infirmed, by our results: indeed we will see that extra degrees of freedom become relevant, but their effect can be represented in the limit by terms which depends on the previous history of the interface evolution. The system we consider is a phase field equation with additive stochastic noise, see (1.1)–(1.2) below. Without noise the interface is given by a special instantonic profile connecting the two phases; in the presence of noise, after suitable rescalings, the limit state is still represented by the same instantonic profile which is however randomly displaced. The displacement obeys an ordinary stochastic equation driven by a white noise forcing term and with a long memory drift, whose effect is to force the interface back toward its initial position, thus restoring the equilibrium of the two coexisting phases. The evolution is defined by the following stochastic equations √ 1 ∆m(t) − V (m(t)) + λh(t) dt + γ dW (a) (t) dm(t) = 2 1 d [h(t) + m(t)] = ∆h(t)dt 2
(1.1) (1.2)
where the unknowns, m(t) = m(t, x), h(t) = h(t, x), (t, x) ∈ R+ × R, are two scalar random fields. In (1.1), λ and γ are positive parameters; ∆ is the Laplacian on R and W (a) (t) is a white noise with a cutoff in the spatial covariance. This means W (a) (t) is the canonical process in the filtered probability space (Ω, F , Ft , P) where Ω := C(R+ ; S (R)) (S (R) the space of tempered distributions), F its Borel σ– algebra, Ft the canonical filtration, and P the Gaussian measure with mean zero and covariance aγ (x) := a γ β/2 x (1.3) E W (a) (t), ϕW (a) (t ), ϕ = t ∧ t ϕ, a2γ ϕ , where t ∧ t := min{t, t }, ·, · denotes the inner product in L2 (R, dx) as well as the canonical pairing between S and S . Above β > 0 and a ∈ C02 (R), i.e. a twice differentiable function with compact support. We assume the normalization a(0) = 1. Finally V (m) is the derivative of a symmetric, smooth double well potential V (m); for simplicity we assume m2 m4 − . (1.4) 4 2 Note we are omitting to write explicitly the dependence on the randomness ω ∈ Ω. This will be done throughout the whole paper without further mention. V (m) =
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
31
In a companion paper, [2], we prove global existence and uniqueness in a space of H¨older continuous functions for the system t t √ m(t) = pt m(0) + ds pt−s [−V (m(s)) + λh(s)] + γ pt−s dW (a) (s) (1.5) 0 0 1 t ds (∆pt−s )m(s) (1.6) h(t) + m(t) = pt [h(0) + m(0)] − 2 0 where pt = et∆/2 is the heat semigroup, namely the integral operator with kernel 2
pt (x, y) =
e−(x−y) /2t √ . 2πt
(1.7)
Observe that the integral on the r.h.s. of (1.6) is well defined because m(s) is H¨ older continuous. The system (1.1)–(1.2) is defined in terms of the integral equations (1.5)–(1.6) and called “stochastic phase field equations”. General background and physical interpretation are discussed in the sequel, here we proceed by stating our main result, presented in the next theorem. We consider the initial condition m(0) = m, ¯
h(0) = 0
(1.8)
where m ¯ ξ (x) := tanh(x − ξ),
m ¯ := m ¯0
(1.9)
is a standing wave (that we call “instanton”) with “center” ξ ∈ R. We suppose that √ (1.10) λ= γ (λ) (λ) and denote by m (t), h (t) the solution of (1.1)–(1.2) with (1.10) and initial condition (1.8). Our main result is
Theorem 1.1 There exists a process x(λ) (t) in C(R+ ), adapted to Ft , such that for each τ, ε > 0
(λ) sup m (t) − m ¯ x(λ) (t) > ε = 0 . (1.11) lim P λ↓0
t≤λ−2 τ
∞
Furthermore, denoting weak convergence in C(R+ ) by =⇒ and defining, after scaling, xλ (τ ) := x(λ) λ−2 τ , we have that xλ =⇒ x as λ ↓ 0, where x(τ ) is the unique solution of τ x(s) x(τ ) = b(τ ) − 3 (1.12) ds 2π(τ − s) 0 in which b(τ ) is a one dimensional Brownian motion with diffusion coefficient D = 3/4.
32
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
The coefficient 3 in (1.12) and the value D = 3/4 above are related to the specific choice of the potential V . Existence and properties of the process solution of (1.12) are discussed in [3]. For the physical interpretation, we start from the equation without noise and with λ = 0, namely (1.1) with λ = γ = 0. This is the well known Allen–Cahn (AC) equation with double well potential V (m), which arises as the gradient flow associated to the Ginzburg–Landau free energy functional 1 dx (1.13) F (m) = |∇m(x)|2 + V m(x) 4 R (F decreases along the solutions of the AC equation). The minimizers of F (m) are the two constant functions m+ (x) = 1 and m− (x) = −1, therefore the values of the order parameter m = ±1 correspond to pure phases and the interface for (1.13) is (up to translations) the profile m(x) which minimizes F (m) under the condition that asymptotically as x → ±∞ it converges to ±1. The associated Euler–Lagrange equation is the stationary AC equation 1 ∆m − V (m) = 0 2
(1.14)
which, imposing the above conditions at ±∞, has the instanton m ¯ of (1.9) as its unique solution (modulo translations). Therefore m ¯ ξ is the equilibrium state which has the two phases coexisting to the right and to the left of ξ, it represents the “mesoscopic interface” with ξ its “mesoscopic location” (mesoscopic instead of macroscopic because the interface is “diffuse” and the transition from one phase to the other, even though exponentially fast, is not sharp; mesoscopic instead of microscopic because the AC equation and the Ginzburg–Landau functional can be derived by a scaling procedure from particle or spin systems, i.e. from an underlying more microscopic structure). The next step is with λ > 0, but still γ = 0. Then (1.1) is coupled to (1.2) and the two together give an example of phase field equations (PFE). Here h is a thermodynamic potential conjugated to the order parameter m: if m is a magnetization density, then h is a magnetic field, our notation is inspired by such an interpretation. More commonly however, m is a relative concentration of one species in a binary alloy and h is a relative temperature. The effective potential will then depend on the relative temperature h, our choice is simply to replace V (m) by V (m) − hm. At the critical temperature, which corresponds to h = 0, the alloy can exist in two different concentrations m = ±1, but, as the temperature h changes, the equilibrium concentrations vary, one becomes stable and the other one metastable. In the presence of a given temperature profile h, the AC rate of change of the concentration density at x at time t, i.e. dm(t, x)/dt, is given by (1.1) (with γ = 0). Due to latent heat, there is a corresponding change of temperature given (in the proper units) by dh(t, x)/dt = −dm(t, x)/dt. Simultaneously, by the
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
33
Fourier law, the temperature diffuses according to the r.h.s. of (1.2). This has a feedback in (1.1) so that (1.1) and (1.2) are coupled. In conclusion the PFE describe a change of phase including latent heat effects and, because of that, unlike for the AC equation, there is a conservation law: the integral of m + h is in fact invariant under PFE. The noise term in (1.1) takes into account some external fluctuating force and the resulting equation is known in the physical literature as model C of Hohenberg and Halperin, [14]. Since m ¯ solves (1.14), the state m = m, ¯ h = 0 is a stationary solution of PFE, which is therefore interpreted, like in AC, as the mesoscopic interface. We are then studying in Theorem 1.1 what happens to the interface when there are small (because γ → 0 in Theorem 1.1), external perturbations which produce a √ random change γdW (a) (t) of magnetization (or, in the other interpretation, of concentration), the analysis including latent heat effects. The small parameter γ in the noise term has the meaning of a ratio between mesoscopic and microscopic space units, the former referring to (1.1), the latter to some microscopic model, as for instance the Glauber + Kawasaki process introduced in [12]. A formal comparison with the microscopic model in [12] would lead to a more complex structure for the additive noise; however we stick to (1.1) (which catches the correct behavior of critical fluctuations, see [5]) to make our analysis simpler. In conclusion the scaling γ → 0 has a natural justification in terms of the microscopic origin of the noise term, the scaling of time in Theorem 1.1 is on the other hand justified a posteriori: it is the correct scaling for observing finite displacements of the interface. On the contrary, the equality (1.10) has no clear physical interpretation; it is true that a scaling with λ → 0 is widely used in the PFE literature to stress “kinetic undercooling effects”, see [11], [17], but relating λ to the noise as in (1.10) is just a matter of technical convenience. We will come back to this in the next section in the paragraph “the role of the assumptions”.
2 Heuristic analysis and outline of proofs By Theorem 1.1, with probability going to 1, the process m(λ) (t), t ≤ λ−2 τ , is always close to the manifold of instantons M = {m ¯ ξ , ξ ∈ R}, see (1.11); the theorem then identifies the motion along M, (1.12). It is then natural to describe m(λ) (t) in terms of coordinates along and transversal to M, these are the Fermi coordinates that we are going to define. Stability of instantons, Fermi coordinates. Closeness to M is a consequence of the stability of M under the AC evolution. Under AC, in fact, M attracts exponentially fast all data which are in a small neighborhood, in sup norm, · ∞ , of M: namely there are δ, c and a all positive so that if for some ξ, m − m ¯ ξ ∞ < δ, then there is a ξ for which, for all t, m(t) − m ¯ ξ ∞ ≤ ce−at , m(t) being the solution of the AC equation starting from m. This obviously fails if we add noise (thus considering (1.1) with λ = 0 and γ > 0), but the noise, in a polynomial scale, cannot drive too far away from M because of the exponential attraction of
34
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
the deterministic part of the equation: in [10] it is shown that for any τ > 0 and ζ > 0,
lim P
γ→0
sup dist(m(t), M) ≥ γ 1/2−ζ
=0
(2.1)
t≤γ −1 τ
where “dist” denotes distance in sup norm. The analysis extends to our case with √ λ = γ and (1.1)–(1.2) coupled, as stated in (1.11) and proved in Section 8. Going back to AC without noise, observe that stability of M does not mean stability of the single instanton: let m be a small deviation from m ¯ ξ , then from what we said above it will relax under AC to some m ¯ ξ , with ξ close but not necessarily equal to ξ. In the space of all profiles m, m ¯ ξ is marginally stable along the direction M while all the other directions are stable. It is then natural to associate to each m (as above) the value ξ of the center of the limit instanton. In practice, however, it is better to work with a more “geometrical” definition. Following [10, 9], we define a center ξ of m as a real number which minimizes the ¯ x (as a function of x). Then ξ is such that L2 –norm of m − m ¯ ξ = 0. m − m ¯ ξ, m
(2.2)
The center ξ has also a dynamical interpretation. Let Lξ be the operator 1 ∆v + (1 − 3m ¯ 2ξ )v (2.3) 2 obtained by linearizing the AC equation around m ¯ ξ . It is readily seen that Lξ is ¯ ξ while the rest of self-adjoint in L2 (R, dx), it has eigenvalue 0 with eigenvector m the spectrum is on the negative axis strictly away from 0. Then, if m has center ξ, the deviation v = m − m ¯ ξ by (2.2) has no component along m ¯ ξ , hence the linearized evolution starting from v decays exponentially fast and correspondingly m converges to m ¯ ξ , so that the center of m is also the center of the instanton to which the linearized AC evolution converges. The pair {ξ, m − m ¯ ξ } is known as the Fermi coordinates of m. This notion of a center of a function plays an important role also in our proofs, so we will spend a few more words, recalling Proposition 3.2 in [9], which gives a sufficient condition for m = m(x) to have a center. Lξ v =
Proposition 2.1 There is a constant δ > 0 such that if there exists x0 ∈ R so that m− m ¯ x0 ∞ ≤ δ then the following holds for some constant C = C(δ) independent of x0 and m. (i) The function m has a unique center x and ¯ x0 ∞ |x − x0 | ≤ Cm − m
(2.4)
(ii) the center x has the expansion 3 9 ¯ x0 , m − m ¯ ,m− m x = x0 − m ¯ x0 − m ¯ x0 m ¯ x0 , m − m ¯ x0 + R(m − m ¯ x0 ), 4 16 x0 ¯ x0 3∞ . (2.5) |R(m − m ¯ x0 )| ≤ Cm − m
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
35
Thus, by (2.1) and (1.11) and with the help of Proposition 2.1, we can talk unambiguously, with probability going to 1, of the center ξ(t) of m(t) both for the stochastic AC equation and PFE. This gives us a precise definition of the location of the interface even without sharp interface limits; we will prove convergence to (1.12) by studying the asymptotic law of ξ(t) as γ → 0. Heuristics of the AC equation and PFE with noise. We start from the simpler (and instructive) case of the AC equation with noise, i.e. (1.1) with λ = 0 but with γ > 0. This is well studied in the literature, [9, 10, 13], even though in slightly different contexts. The scaling procedure is the same as in Theorem 1.1 and it leads to the same limit law but without drift, i.e. a Brownian motion. To √ explain heuristically the result, let us regard the forcing term γdW (a) (t) as a “source of small kicks” which we decompose in a component along M and another one orthogonal to M. The latter fights against the AC drift which pushes back √ toward M, and because of the small factor γ, to a first order, we forget about orthogonal components. On the contrary the kicks along M are not contrasted and they sum up: thus we are approximating dm(t) ≈
3√ γm ¯ ξ(t) m ¯ ξ(t) , dW (a) (t), 4
m ¯ , m ¯ =
4 3
(2.6)
where ξ(t) is the center of m(t). Also m(t) ≈ m ¯ ξ(t) , hence dm(t) ≈ m ¯ ξ(t) dξ(t)
(2.7)
and, in conclusion,
3√ γm ¯ ξ(t) , dW (a) (t) (2.8) 4 namely ξ(γ −1 t) is a Brownian motion with diffusion 3/4, which is what is proved in [10]. The argument for the system (1.1)–(1.2) is similar, the only difference from the stochastic AC equation in (1.1) lies in the simple, innocent looking term λh, which is however the source of all problems. The same heuristics leading to (2.6) √ applies to (1.1) by simply adding λh(t) to the noise; writing γ = λ according to (1.10), we then get
3 ¯ ξ(t) m dm(λ) (t) ≈ λm ¯ ξ(t) , dW (a) (t) (2.9) ¯ ξ(t) , h(t)dt + m 4 dξ(t) ≈
and, using (2.7), dξ(t) ≈
3 λ m ¯ ξ(t) , h(t)dt + m ¯ ξ(t) , dW (a) (t) . 4
(2.10)
On the other hand, writing (1.2) in integral form and recalling that h(0) = 0, we have t pt−s dm(λ) (s) (2.11) h(t) = − 0
36
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
Using again (2.7) and taking the scalar product with m ¯ ξ(t) we get from (2.11) m ¯ ξ(t) , h(t)
≈−
0
t
m ¯ ξ(t) , pt−s m ¯ ξ(s) dξ(s) .
(2.12)
The approximate system (2.10)–(2.12) (with ≈ replaced by equality) is not as easy to study as the one which approximates the stochastic AC equation, but it can be seen to give the correct result (1.12) for the limit motion of the center, [3]. To see the relation with (1.12) we make another approximation whose validity will be justified in the course of the proof of Theorem 1.1. The approximation consists in ¯ (i.e. ξ(t) → 0) in the scalar products in (2.10) and (2.12) replacing m ¯ ξ(t) by m (the reason for its validity is that the displacements of the center are finite while the field h becomes flat because the diffusion in (1.2) acts for long times). The new system is then (forgetting about the noise cutoff) 3 λm ¯ , h(t) + λdb(t) 4 t m ¯ , pt−s m ¯ dξ(s) m ¯ , h(t) = −
(2.13)
dξ(t) =
(2.14)
0
where b(t) is the Brownian motion with diffusion 3/4 of Theorem 1.1. Using (2.14) to rewrite the first term on the r.h.s. of (2.13) we get t s 3 ξ(t) = λb(t) − λ ds m ¯ , ps−s m ¯ dξ(s ) . 4 0 0 Integrating by parts, after some simple algebra, t 3 ξ(t) = λb(t) − λ ds ξ(s )m ¯ , pt−s m ¯ . 4 0
(2.15)
Approximating
¯ (y) m ¯ (x)m dy 2π(t − s ) (2.16) and recalling that m ¯ = 2, (2.15) becomes (1.12), in the above approximation, which can be made rigorous in the limit λ → 0, having set t = λ−2 τ . ¯ = m ¯ , pt−s m
dx
dy m ¯ (x)pt−s (x, y)m ¯ (y) ≈
dx
Main difficulties and outline of proof. The heuristic arguments outlined above are essentially based on a linear approximation, their validity therefore rests on a rigorous proof that the non linear effects are negligible. Since the strength of √ the noise is γ, we cannot hope to improve the a priori bounds beyond m(t) − √ m ¯ ξ(t) ∞ ≈ γ. Then the non linear terms which are, to lowest order, quadratic have order γ; since they act for a time λ−2 τ (that is γ −1 τ ), a naive estimate gives a non–negligible contribution. The fact that they are indeed negligible must then
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
37
come from cancellations, and the true problem is to find them and take them into proper account. This is more clearly seen in the stochastic AC equation. Following [10], we split the time axis into intervals of length T , taking T long, yet very small as compared to the macroscopic time γ −1 τ , for instance T = γ −1/10 . The crucial point is an iterative procedure for which the problem reduces to the analysis of only one of these intervals, a clear advantage, because in such a “short time” the non linear effects are under control. Couplings are used for this crucial reduction. We compare in fact, in the generic interval [nT, (n + 1)T ], the true process m(t) which starts from m(nT ) and the new process m(t) ˆ which starts from m ¯ ξ(nT ) , ξ(nT ) the center of m(nT ), the coupling is simply to use for both processes the √ ¯ ξ(nT ) ∞ ≤ γ 1/2−ζ , same noise γdW (a) (t). Under the assumption that m(nT )−m see (2.1), it can be seen that, with probability going to 1 as γ → 0, m((n + 1)T ) − m((n ˆ + 1)T )∞ ≤ Cγ 1−2ζ , (C a constant). By Proposition 2.1, the displacements of the center in the interval [nT, (n + 1)T ], as computed with the two evolutions, differ proportionally to γ 1−2ζ : since the number of intervals grows proportionally to γ −1 τ /T , the sum of all these differences goes like γ 1−2ζ γ −1 τ /T , which vanishes, after choosing 2ζ < 1/10. We can then study in each interval the process which starts from an instanton. Neglecting for simplicity the cutoff on the noise (with the cutoff some extra computations are needed), then the displacement of the center in a time interval T does not depend on the initial center and it is therefore independent of the past. The displacements of the centers (each time restarting from an instanton) are thus independent variables with mean 0 (by the symmetry between right and left) and, using classical arguments on convergence to Brownian motion, in the end, we need to sum their variances; since we are already with squares, it turns out that the linear approximation is sufficiently accurate and this explains the validity of the linear approximation in the previous heuristic analysis. While the above approach works well in the stochastic AC equation, Theorem 1.1 tell us that it fails, as there are long memory effects in the limit law. More bad news: the last term in (1.12), responsible for these effects, according to the heuristic analysis of the previous paragraph, comes from the term λh(t) and since it produces a finite drift in a time λ−2 , the order of magnitude of h(t) must be ≈ λ. Therefore we need an accuracy of order λ, which is comparable with the √ deviations of m(λ) from M (recall λ = γ) that has been neglected so far. Our approach to the problem, since when we began the present work, has been “to trust” [10] and to consider the non linear terms that are left as being negligible when we linearize (1.1) around m ¯ ξ(nT ) , ξ(nT ) the center of m(λ) (nT ). It is evidently not possible to use the coupling argument of [10], yet in some maybe more complicated way, in the end we “must” see that they are not relevant. The extra term λh(t) is new w.r.t. [10] and has to be dealt anew. According to the final result and for what said before, we expect h(t) ≈ λ, but let us even suppose, pessimistically, that h(t) is of the order of unity. Its effect for a time T will then be of order λT , hence still infinitesimal. Moreover, by (2.11), h(t) can be large only if dm(λ) (s) is large, but dm(λ) (s) is under control except for the term λh(s). Due
38
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
to the presence of the small factor λ this gives a virtuous feedback which allows to control the magnitude of h(t). This is done in Sections 3 and 4. We first write down an integral equation, ¯ ξ(nT ) , the superscript n recalls that we are con(3.7), for v (n) (t) = m(λ) (t) − m sidering t ∈ [nT, (n + 1)T ]. v (n) (t) is written as the sum of 8 terms, two of them, (n) (n) called Γ2 (t) and Λ3 (t) are the important ones, which give contributions to the limit, respectively the Brownian motion and the drift. All the others vanish, but at this stage this is not yet established as they depend on the unknown m(λ) (t). In Section 4 we study by iterations the integral equation (3.7) to prove bounds on v (n) (t) and on the displacements of the center. We derive in this way (1.11) and establish even sharper bounds that will be used in the successive proofs. The proof of the convergence to (1.12) is reported in the remaining sections. (n) We begin Section 5 by splitting the term Λ3 (t) into the sum of 3 other terms, see (5.1)–(5.5), which foresees the approximation done in (2.16). We then use (2.5) to deduce an equation for the displacements of the center, which are called ψn , see (5.7). The equation is (5.9), which is a sort of linear integral equation in the ψn with kernel An,k , k < n, and known data ηn : “sort of” because the ηn still depend on the unknowns v (n) (·). The elements An,k decay as (n − k)−1/2 . It is then convenient to reduce to the matrix A2 , so that we iterate once (5.9) obtaining (5.17), where the kernel is now A2n,k and the “known terms” are ηn and (Aη)n . In Section 6 we study these “known terms” which are splitted into four groups. The first one consists of truly known terms which survive in the limit (n) (they come from Γ2 (t)). The terms in the second group, which instead may de(n) pend on v (t), are all directly proved to be negligible using the a priori bounds of Section 4. For those in the third group we cannot proceed in this way, but we need to use the integral equation for v (n) (t) and only after sufficiently many iterations, we can show that they are negligible. Finally, the last group collects terms which become negligible because of stochastic cancellations. The latter are studied in Section 7, the others in Section 6. We draw the conclusions of our analysis in Section 8 where we complete the proof of Theorem 1.1. Role of assumptions. We start from the assumption (1.10) which is conceptually the most important one, the others are of a more technical nature. As already remarked, there are several studies of sharp interface limits on PFE where λ is scaled to 0. This describes an intermediate regime (called kinetic undercooling) where thermodynamical equilibrium is not fully reached. Thus our model should be regarded as kinetic undercooling in the presence of stochastic perturbations. As said, the relation between λ and γ stated in (1.10) does not have a straight physical interpretation, it is just the right way to scale (1.1)–(1.2) and have a nice limit law. One may however wonder what would happen if we took a different relation than (1.10). We have not worked out the details, but we can at least present some educated guess. If we multiply h(t) in (1.1) by a constant θ, we would then derive a limit law with such a factor multiplying the last term in (1.12). Let us then
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
39
consider the scaling behavior of x(t) = b(t) − 3θ
t
0
x(s) ds 2π(t − s)
(2.17)
both when θ vanishes and when it diverges. Set y(t) = θa x(θc t), then a c/2
y(t) = θ θ
b(t) − 3θθ
c/2
0
t
y(s) . ds 2π(t − s)
(2.18)
Imposing a + c/2 = 0,
1 + c/2 = 0;
a = 1,
c = −2
(2.19)
we have (2.18) equal to (1.12). Thus if we take (1.1) with λ → λ1+α and call θ = λα , we believe that our analysis extends, at least for small |α|, and that the center ξ(t) of the solution of the corresponding equation is such that λα ξ(λ−2−2α t)
(2.20)
converges in law to (1.12). Concluding remarks and perspectives. A forthcoming paper, [3], is devoted to the analysis of the limit process (1.12). This can be characterized in terms of a Brownian motion with absorption at the origin, which in turn is reduced to the well studied one dimensional Schr¨ odinger equation with Dirac’s delta potentials. A quite explicit expression for the solution of (1.12), is then available, in √ particular it shows that the displacements ξ(τ ) of the front have typical size log τ for τ large; “aging phenomena” are also present. The “cluster fluctuations” have instead the usual Brownian structure. Consider an initial state with h0 = 0 and m0 (x) the symmetric function which coincides with m ¯ −ξ (x), ξ > 0, for x ≤ 0. We interpret it as a “plus cluster” in the region (−ξ, ξ) with the minus phase outside. To make it sharp we set ξ = λ−1 -, - > 0, and √ consider the process (1.1)–(1.2) (always assuming λ = γ). Preliminary results (see also [3]) indicate that, in proper units, the coordinates ξ1 (τ ) and ξ2 (τ ) of the two centers, evolve in the limit according to a system of two stochastic equations. The system can be diagonalized into the variables ξG (τ ) := [ξ1 (τ ) + ξ2 (τ )]/2 and ξ(τ ) := ξ1 (τ ) − ξ2 (τ ). The variable ξG (τ ) has the law of a Brownian motion, ξ(τ ) is independent of ξG (τ ) and obeys an equation like (1.12). We hope to be able to accomplish the above program including the analysis of the case with several clusters (and suitable scalings), the ultimate goal being the study of interface fluctuations in many dimensions.
40
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
Part II. Stability of fronts 3 The iterative scheme Before starting with the proofs we introduce some notation that will be used throughout the paper: Notation. C will denote a generic constant whose numerical value may change from line to line (from the statements it will appear clear on which parameters it depends). For p ∈ [1, ∞] we denote by · p the norm in Lp (R, dx). We will study the problem (1.1)–(1.10) by an iterative procedure. To this end, we divide the microscopic time–line R+ into intervals [Tn , Tn+1 ], Tn = nT , where n ∈ N and T = λ−(1∧β)/20 (β as in (1.3)). We then associate to any macroscopic time interval [0, τ ], τ > 0, the set of microscopic time intervals [Tn , Tn+1 ], with n ≤ nλ (τ ), where nλ (τ ) is the integer part of (λ2 T )−1 τ , namely nλ (τ ) = [(λ2 T )−1 τ ]. To simplify notation, we write m(t) = {m(t, x), x ∈ R} and h(t) = {h(t, x), x ∈ R} for the solution of (1.1)–(1.10) (omitting the dependence on λ). We next define, by induction on n ≥ 0, the numbers xn and the functions v (n) (t) = {v (n) (t, x), x ∈ R}, t ∈ [Tn , Tn+1 ]. They will have the property that for each t ∈ [Tk , Tk+1 ] ¯ xk v (k) (t) = m(t) − m
if v (h) (Th+1 )∞ ≤ δ for all 0 ≤ h ≤ k − 1
(3.1)
with δ as in Proposition 2.1. For n = 0 we set x0 = 0 and v (0) (t) = m(t) − m ¯ x0 , 0 ≤ t ≤ T , so that (3.1) holds. Suppose that, by the induction hypothesis, we have already defined for all k ≤ n − 1 both xk and v (k) (t) and that (3.1) holds for such k’s. If there is k ≤ n − 1 such that v (k) (Tk+1 )∞ > δ, we set xn = 0 and v (n) (t) = 0, ¯ xn−1 + v (n−1) (Tn ), which t ∈ [Tn , Tn+1 ]. Otherwise we define xn as the center of m (n−1) by Proposition 2.1 is well defined, as v (Tn )∞ ≤ δ. We then set, according to ¯ xn , t ∈ [Tn , Tn+1 ], and have (always under the assumption (3.1), v (n) (t) = m(t)− m that v (k) (Tk+1 )∞ ≤ δ for all k ≤ n − 1) ¯ xn−1 + v (n−1) (Tn ) = m ¯ xn + v (n) (Tn ), m(Tn ) = m Let
h(n) (t) = h(t),
m ¯ xn , v (n) (Tn ) = 0 . (3.2)
t ∈ [Tn , Tn+1 ],
introduce the stopping time w.r.t. the discrete filtration FTn+1
Nδ := inf k : v (k) (Tk+1 )∞ > δ (n)
(δ as in Proposition 2.1) and define gt (0) (2.3). We abbreviate gt = gt .
(3.3)
(3.4)
:= exp{tLxn }, t ∈ R+ , with Lxn as in
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
41
After expanding V (m(t)) in (1.8) and (1.10) about m ¯ xn , for t ∈ [Tn , Tn+1 ] as long as n ≤ Nδ , v (n) (t) and h(n) (t) are given as the solution of dv (n) (t) = Lxn v (n) (t) + λh(n) (t) − 3m ¯ xn v (n) (t)2 − v (n) (t)3 dt + λdW (a) (t)
h(n) (t) = pt−Tn h(n−1) (Tn ) − v (n) (t) + pt−Tn v (n) (Tn ) +
(3.5) t
ds Tn
∂pt−s (n) v (s) ∂s (3.6)
with initial condition v (n) (Tn ) for the first equation. Recall that v (n) (Tn ) is determined from v (n−1) (Tn ) via (3.2), and that from (3.3) we have h(n−1) (Tn ) = h(n) (Tn ). We will write in the next section v (n) (t) as a sum of 10 terms, v (n) (t) = 10 (n) i=1 Γi (t), but for the moment it is more convenient to keep some of the terms together. We are going to prove that for n < Nδ , (n)
(n)
(n)
(n)
(n)
(n)
(n)
(n)
v (n) (t) = Γ1 (t) + Γ2 (t) + Λ3 (t) + Γ4 (t) + Γ5 (t) + Γ6 (t) + Γ9 (t) + Γ10 (t) (3.7) where (n)
Γ1 (t) (n)
Γ2 (t) (n)
Λ3 (t) (n) Γ4 (t) (n)
Γ5 (t) (n) Γ6 (t) (n)
Γ9 (t) (n)
Γ10 (t)
(n)
:= gt−Tn v (n) (Tn ) (n) := λz (t) := λ n
:= −λ
k=1 t
:= λ
Tn
:= −λ
k=1 t
:= −
t
Tn
Tn t
Tn
gt−s dW (a) (s)
(3.9)
(n)
ds gt−s ps−Tk [m ¯ xk − m ¯ xk−1 ]
s
ds
Tn
∂ps−s (n) v (s ) ∂s
(n)
Tn
(n) ds gt−s
Tk
(3.12) ds
Tk−1
∂ps−s (k−1) v (s ) ∂s
(n) ¯ xn v (n) (s)2 ds gt−s m
(3.15)
−1
to (3.5) we get (n) (n) (n) (n) v (n) (t) = Γ1 (t) + Γ2 (t) + Γ9 (t) + Γ10 (t) + λ
(3.13) (3.14)
(n) ds gt−s v (n) (s)3 .
Proof of (3.7). By applying (∂t − Lxn )
(3.10) (3.11)
ds gt−s v (n) (s)
Tn n t
:= −λ := −3
t
(n)
Tn
(n) ds gt−s
(3.8) t
t Tn
(n)
ds gt−s h(n) (s) .
42
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
We use for h(n) (s) the expression given by (3.6), and get v (n) (t) =
(n)
(n)
(n)
(n)
(n)
(n)
Γ1 (t) + Γ2 (t) + Γ4 (t) + Γ5 (t) + Γ9 (t) + Γ10 (t) t (n) ds gt−s ps−Tn h(n−1) (Tn ) + v (n) (Tn ) . +λ Tn
To complete the proof of (3.7) we thus need to show
t
λ Tn
(n) (n) (n) ds gt−s ps−Tn h(n−1) (Tn ) + v (n) (Tn ) = Λ3 (t) + Γ6 (t) .
(3.16)
From (3.2), we have ¯ xn − m ¯ xn−1 ) h(n−1) (Tn ) + v (n) (Tn ) = h(n−1) (Tn ) + v (n−1) (Tn ) − (m
(3.17)
¯ xn−1 ) reproduces We use this identity in the l.h.s. of (3.16). The difference (m ¯ xn − m (n) the last term of the sum in (3.10) (i.e. in the definition of Λ3 (t)). For the term with h(n−1) (Tn ) + v (n−1) (Tn ), we use (3.6) to write h(n−1) (Tn ) + v (n−1) (Tn )
(n−1) (n−1) (Tn−1 ) + v (Tn−1 ) + = pTn −Tn−1 h
Tn
Tn−1
ds
∂pTn −s (n−1) v (s ). ∂s
The contribution of the last integral to the l.h.s. of (3.16) gives the last term of (n) the sum in (3.13), i.e. the definition of Γ6 (t). By iteration we then get (3.16) and (3.7) is therefore proved.
4 A priori bounds We will use the representation (3.7) to prove in Proposition 4.1 below some a priori bounds on v (n) (t) and other quantities. We need first some more notation; recalling Proposition 2.1, we set ξn := −
n−1 3 m ¯ xk , v (k) (Tk+1 ) 4
(4.1)
k=0
We should think of ξn as a linear approximation to xn , since the increment ξk −ξk−1 is, according to (2.5), the linear approximation to the displacement xk − xk−1 of the center in the time interval [Tk−1 , Tk ]. We set (n) sup Vn,∗ := sup Vk , V∗ (τ ) := Vnλ (τ ),∗ (4.2) Vn := v (t) , t∈[Tn ,Tn+1 ]
∞
k≤n
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
and
δ ∗ (τ ) :=
sup |xk+1 − xk |
43
(4.3)
k≤nλ (τ )
Let also Zn :=
sup s∈[Tn ,Tn+1 ]
(n) z (s)
∞
,
Zn,∗ := sup Zk k≤n
(4.4)
(1)
and Bλ,τ ⊂ Ω be the event (recall the definition of (Ω, F , Ft , P) given in Section 1) √ (1) (4.5) Bλ,τ := ω ∈ Ω : Znλ (τ ),∗ ≤ λ−ζ T . We will prove in Appendix B the following Gaussian estimate: for each τ, ζ, q > 0 there is a constant C = C(τ, ζ, q) such that for any λ > 0 (1) (4.6) P Bλ,τ ≥ 1 − Cλq . The next proposition is the key ingredient for the bound (1.11), see the beginning of §8 for the conclusion of the proof. Proposition 4.1 For each τ, ζ > 0 there is a constant C = C(τ, ζ) such that, for any λ > 0, √ √ (1) V∗ (τ ) ≤ Cλ1−ζ T , δ ∗ (τ ) ≤ Cλ1−ζ T on the set Bλ,τ . (4.7) In particular, for λ > 0 small enough, Nδ > nλ (τ )
(1)
on the set Bλ,τ .
(4.8)
Proof. Let (mλ , tλ ), tλ ∈ [Tmλ , Tmλ +1 ], mλ ≤ nλ (τ ), be the stopping times (mλ w.r.t. FTn+1 , tλ w.r.t. Ft ) so that (mλ , tλ ) = (nλ (τ ), Tnλ (τ )+1 ) if V∗ (τ ) < λT , otherwise (mλ , tλ ) = (n, t), where n is the first index such that Vn ≥ λT and t the first time in [Tn , Tn+1 ] for which v (n) (t)∞ ≥ λT . In the following we may assume λ is small enough, otherwise (4.7) holds trivially. We claim there is a constant C so that v (mλ ) (tλ )∞ ≤ CλT .
(4.9)
In fact, if tλ ∈ (Tmλ , Tmλ +1 ], (4.9) follows from the continuity of v (mλ ) (t)∞ ; otherwise, i.e. if tλ = Tmλ , v (mλ ) (Tmλ )∞ ≤ v (mλ −1) (Tmλ )∞ + m ¯ mλ − m ¯ mλ −1 ≤ λT + |xmλ − xmλ −1 | ≤ CλT from (3.2) and (2.4), since m ¯ ∞ = 1 and v (mλ −1) (Tmλ )∞ ≤ λT .
44
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
We will study the evolution till time tλ , so that our next considerations are tacitly restricted to n ≤ mλ and t ≤ tλ . By (4.9), Nδ > mλ (for λ small enough) so that we can use (3.7), regarding it as an integral equation in v (n) (t). We are going to bound one after the other all eight terms on the r.h.s. of (3.7). We will use in the sequel the following bounds (see e.g. [9] and references therein). There is C < ∞ so that, for any measurable function ϕ, gt ϕ∞ ≤ Cϕ∞ .
(4.10)
Moreover, by the Perron–Frobenius Theorem, there are α > 0 and C < ∞ so that for any ϕ orthogonal to m ¯ , m ¯ , ϕ = 0, gt ϕ∞ ≤ Ce−αt ϕ∞ .
(4.11)
Then, by the last identity in (3.2), for t ∈ [Tn , Tn+1 ] (n) ≤ Ce−α(t−Tn ) v (n) (Tn )∞ Γ1 (t) ∞ ≤ Ce−α(t−Tn ) v (n−1) (Tn )∞ + m ¯ xn − m ¯ xn−1 ∞ ≤
Ce−α(t−Tn ) v (n−1) (Tn )∞ .
(4.12)
The second inequality follows from (3.2) and the last one follows from (2.4), proceeding as in the proof of (4.9). By definition (4.4) we get (n) sup (4.13) Γ2 (t) ≤ CλZn . ∞
t∈[Tn ,Tn+1 ]
(n)
We next bound Λ3 . We have 2 1 pt m ¯ z ∞ ≤ √ ≤ √ t 2πt √ ¯ z 1 = 2 for any z ∈ R. Then because pt (x, y) ≤ 1/ 2πt and m |x − x xk k k−1 | pt (m √ ¯ xk − m ¯ xk−1 ) ∞ ≤ dz pt m ¯ z ∞ ≤ xk−1 t so that, by (4.10), we get (n) Λ3 (t)
∞
≤ Cλ
n k=1
t
|xk − xk−1 | ds √ . s − Tk Tn
(4.14)
(4.15)
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
45
By (2.4), we then get (n) Λ3 (t)
n √ Vk−1 √ ≤ Cλ T . n −k+1 k=1
∞
(4.16)
(n)
We will next bound Γ4 (t). We write
(n)
gt−s = pt−s −
t
ds
s
d (n) gt−s ps −s = pt−s + ds
t
s
(n)
ds gt−s [1 − 3m ¯ 2xn ]ps −s
having used (2.3) in the second equality. Then (n)
Γ4 (t)
t
= λ
s
ds
ds Tn
Tn
t
+λ
∂pt−s (n) v (s ) ∂s
t
ds
ds
Tn
s
(n) gt−s [1
−
3m ¯ 2xn ]
s
ds
Tn
∂ps −s (n) v (s ) . ∂s
By using (4.10) and (see Appendix A) ∂ pt ϕ ≤ 1 ϕ∞ ∂t t ∞ since v (n) (s )∞ ≤ Vn , we get (n) Γ4 (t)
∞
≤ CλT 2 Vn .
This is not optimal but good enough for our purposes. Analogously t (n) ds v (n) (s)∞ ≤ CλT Vn Γ5 (t) ≤ Cλ ∞
(n) Γ6 (t)
∞
≤ Cλ
n k=1
Tn
(4.18)
(4.19)
Tn
t
(4.17)
Tk
ds
ds
Tk−1
Finally, using again (4.10), we have (n) Γ9 (t) ∞ (n) Γ10 (t) ∞
n v (k−1) (s )∞ Vk−1 . ≤ CλT s − s n−k+1 k=1 (4.20)
≤
CT Vn2
(4.21)
≤
CT Vn3
(4.22)
hence, recalling the definition of (mλ , tλ ) and (4.9), (n) (n) Γ9 (t) + Γ10 (t) ≤ CλT 2 Vn ∞
∞
(4.23)
46
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
After collecting all these bounds, we have, for t ∈ [Tn , Tn+1 ], n ≤ mλ , t ≤ tλ , (n) ≤ C e−α(t−Tn ) v (n−1) (Tn ) + λZn + λT 2 Vn v (t) ∞
∞
n n √ V Vk−1 √ k−1 . + λT +λ T n−k+1 n−k+1 k=1 k=1
(4.24)
By iterating once the above inequality we get Vn ≤ C e−αT Vn−2 + λ(Zn + Zn−1 ) + λT 2 (Vn + Vn−1 ) n n √ Vk−1 Vk−1 √ +λ T + λT . n−k+1 n−k+1 k=1 k=1 Recalling (4.4), we have Vn
≤ C e−αT Vn,∗ + λZn,∗ + λT 2 Vn,∗ n n √ Vk−1,∗ Vk−1,∗ √ +λ T + λT n−k+1 n−k+1 k=1 k=1
and since the r.h.s. is an increasing function of n, Vn,∗ ≤ C e−αT Vn,∗ + λZn,∗ + λT 2 Vn,∗ n n √ V Vk−1,∗ √ k−1,∗ + λT . +λ T n−k+1 n−k+1 k=1 k=1
(4.25)
Inequality (4.25) yields, for n ≤ mλ ≤ nλ (τ ), √ n−1 Cλ T 2 −αT √ 1 − C λT + e + λT log nλ (τ ) Vn,∗ ≤ CλZn,∗ + Vk,∗ . n−k k=0 For λ small enough the square bracket term is larger than 1/2, so that (provided we double the value of the constant C) Vn,∗ ≤ CλZn,∗ +
n−1 k=0
√ Cλ T √ Vk,∗ . n−k
(4.26)
By iteration of (4.26) we get Vn,∗ ≤ CλZn,∗ +
n−1 k=0
√ n−2 C 2 λ2 T 2 2 √ α(n, k)Vk,∗ Zk,∗ + C λ T n−k k=0
(4.27)
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
where α(n, k) =
n−1
1 1 √ √ n−h h−k h=k+1
47
(4.28)
is a bounded function of n and k. We bound the second term on the r.h.s. of (4.27) by Zn,∗
n−1 k=0
√ √ C 2 λ2 T √ ≤ CZn,∗ λ2 T n ≤ CλZn,∗ n−k
since n ≤ mλ ≤ (λ2 T )−1 τ . We then get from (4.27) Vn,∗ ≤ CλZn,∗ + Cλ2 T
n−2
Vk,∗
(4.29)
k=0
from which, by Gronwall Lemma, there is C = C(τ ) such that Vn,∗ ≤ CλZn,∗ ,
for all n ≤ mλ .
(4.30)
By (4.5), choosing λ small enough, we have, for n ≤ mλ , √ (1) on the set Bλ,τ Vn,∗ ≤ Cλ1−ζ T (1)
which implies mλ = nλ (τ ) on the set Bλ,τ . Thus the first bound in (4.7) follows from (4.5) and (4.30). Since Nδ > mλ , (4.8) also follows. Since m ¯ ∞ = 1, the second estimate in (4.7) follows directly from the first one and Proposition 2.1. ¯ xn , is We are going to prove that the component of v (n) (Tn+1 ) orthogonal to m 1−ζ (n) bounded by Cλ , thus considerably improving the bound on the full v (Tn+1 ). Let 3 (n,⊥) (n) ¯ m gt := gt ¯ xn | 1 − |m 4 xn the operator whose kernel is (n,⊥)
gt
3 (n) ¯ (x)m (x, y) = gt (x, y) − m ¯ xn (y) . 4 xn
(4.31) (n)
The superscript ⊥ recalls L2 –orthogonality w.r.t. the eigenvector m ¯ xn of gt , i.e. (n) ¯ xn = m ¯ xn . It follows from (4.11) that there are constants α > 0 and C < ∞ gt m so that, for any ϕ, (n,⊥) ϕ ≤ Ce−αt ϕ∞ . (4.32) gt ∞
Let also
3 ¯ , z (n) (t)m z (n,⊥) (t) := z (n) (t) − m ¯ xn 4 xn
(4.33)
48
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
be the component of z (n) (t) orthogonal to m ¯ xn and introduce the event (2)
Bλ,τ :=
ω∈Ω :
sup
sup
n≤nλ (τ ) t∈[Tn ,Tn+1 ]
z (n,⊥)(t)∞ ≤ λ−ζ
.
(4.34)
In Appendix B we will prove that for each τ, ζ, q > 0 there exists a constant C = C(τ, ζ, q) such that for any λ > 0 (2) P Bλ,τ ≥ 1 − Cλq . (4.35) Define now
V∗⊥ (τ ) :=
sup v (n,⊥) (Tn+1 )
n≤nλ (τ )
∞
,
3 ¯ , v (n) (Tn+1 )m v (n,⊥) (Tn+1 ) := v (n) (Tn+1 ) − m ¯ xn . 4 xn
(4.36)
Proposition 4.2 Recalling (4.5) and (4.34), set (1,2)
(1)
(2)
Bλ,τ := Bλ,τ ∩ Bλ,τ .
(4.37)
Then, for each τ, ζ > 0 there is a constant C = C(τ, ζ) such that, for any λ > 0, V∗⊥ (τ ) ≤ Cλ1−2ζ
(1,2)
on the set Bλ,τ
(4.38)
and, recalling (4.1) for the definition of ξn , sup n≤nλ (τ )+1
|ξn − xn | ≤ CT −1/2+ζ
(1,2)
on the set Bλ,τ
.
(4.39)
Proof. Let (n,⊥)
Λ3
(Tn+1 ) := =
3 (n) (n) ¯ , Λ (Tn+1 )m Λ3 (Tn+1 ) − m ¯ xn 4 xn 3 n Tn+1 (n,⊥) −λ dt gTn+1 −t pt−Tk [m ¯ xk − m ¯ xk−1 ] k=1
(4.40)
Tn
where we used (4.31). By using (4.14), (4.32), m ¯ ∞ ≤ 1, and recalling (4.3) we have n Tn+1 1 (n,⊥) ∗ (Tn+1 ) ≤ Cδ (τ )λ dt e−α(Tn+1 −t) √ Λ3 ∞ t − Tk k=1 Tn λ ≤ Cδ ∗ (τ ) √ nλ (τ ) ≤ Cλ1−ζ T −1/2 (4.41) T (1)
the last inequality being true, by (4.7), on the set Bλ,τ .
Vol. 3, 2002
for
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
49
Shorthanding by R(n) the sum of all the terms on the r.h.s. of (3.7), except (n) and Λ3 , and calling R(n,⊥) its orthogonal projection,
(n) Γ2
(n,⊥)
v (n,⊥) (Tn+1 ) = λz (n,⊥) (Tn+1 ) + Λ3
(Tn+1 ) + R(n,⊥) .
(4.42)
The last term is bounded by |R(n,⊥) | ≤ C e−αT + λT 2 + λT log nλ (τ ) V∗ (τ )
(4.43)
as it follows from (4.12) and (4.18)–(4.20), and (4.23). The bound (4.38) now follows from the definition (4.34), and equations (4.41)–(4.43). It remains to prove (4.39). Recalling that x0 = 0, from (2.5) and (4.1), xn =
n−1
(xk+1 − xk ) = ξn −
k=0
n−1 9 m ¯ xk , v (k) (Tk+1 )m ¯ xk , v (k) (Tk+1 ) 16 k=0
+
n−1
R(v (k) (Tk+1 )) (4.44)
k=0
then, for n ≤ nλ (τ ), since m ¯ 1 = 2,
¯ xn , v (n) (Tn+1 ) + |ξn − xn | ≤ nλ (τ ) 2V∗ (τ ) sup m n≤nλ (τ )
sup R(v (n) (Tn+1 ))
∞
n≤nλ (τ )
By (2.5)
sup R(v (n) (Tn+1 ))
∞
n≤nλ (τ )
≤ CV∗ (τ )3 .
.
(4.45)
(4.46)
¯ = 0, |m ¯ xn , v (n) (Tn+1 )| ≤ CV∗⊥ (τ ), (4.39) follows from (4.7), Since m ¯ , m (4.38), (4.45), and (4.46). ¯ xn , v (n) (Tn ) = 0, The bound in (4.38) holds also for v (n) (Tn ). Indeed, since m we have the following proposition. (1,2)
Proposition 4.3 Let Bλ,τ be as in (4.37). Then, for each τ, ζ > 0 there is a constant C = C(τ, ζ) such that, for any λ > 0 (1,2) on the set Bλ,τ . (4.47) sup v (n) (Tn ) ≤ Cλ1−2ζ n≤nλ (τ )
Proof. By (3.2)
∞
v (n) (Tn ) = v (n−1) (Tn ) + m ¯ xn−1 − m ¯ xn .
(4.48)
50
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
By Taylor expansion up to fourth order, we have ¯ xn−1 m ¯ xn − m
1 ¯ −m ¯ xn−1 (xn − xn−1 ) + m (xn − xn−1 )2 2 xn−1 1 − m (xn − xn−1 )3 + an−1 (xn − xn−1 )4 ¯ 6 xn−1
=
(4.49)
where an−1 is bounded. Thus, by Proposition 2.1, we have m ¯ xn − m ¯ xn−1 =
3 m ¯ , v (n−1) (Tn )m ¯ xn−1 + αn 4 xn−1
(4.50)
where 1 1 m ¯ ¯ (xn − xn−1 )2 − m (xn − xn−1 )3 + an−1 (xn − xn−1 )4 2 xn−1 6 xn−1 9 (n−1) (n−1) (n−1) m ¯ +m ¯ xn−1 ,v (Tn )m ¯ xn−1 , v (Tn ) − R(v (Tn )) . (4.51) 16 xn−1 αn =
From (4.48) and (4.50) we get v (n) (Tn ) = v (n−1,⊥) (Tn ) − αn .
(4.52)
√ (1,2) Note that, from (4.7), on the set Bλ,τ we have supn≤nλ (τ ) |xn −xn−1 | ≤ Cλ1−ζ T ; by (4.38) and (4.46) we have |αn | ≤ Cλ2−2ζ T (a better bound is proved in Section 6). The bound (4.47) follows.
Part III. Limit motion 5
A new integral equation (1,2)
Unless otherwise stated, we will work in the set Bλ,τ which appears in Propositions 4.2 and 4.3. In particular we can use the integral representation (3.7) for all n ≤ nλ (τ ) (if λ is small enough) and the bounds of Section 4. It is now convenient (n) to decompose the term Λ3 (t) into the sum of three new terms; we thus use (4.50) to write (n) (n) (n) (5.1) Λ3 (t) = Γ3+8 (t) + Γ7 (t) where (n)
Γ3+8 (t) := (n)
Γ7 (t) :=
n
3λ m ¯ xk−1 , v (k−1) (Tk ) 4 k=1 n t (n) −λ ds gt−s ps−Tk αk −
k=1
Tn
t
Tn
(n)
ds gt−s ps−Tk m ¯ xk−1
(5.2) (5.3)
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations (n)
(n)
51
(n)
and next we decompose Γ3+8 (t) = Γ3 (t) + Γ8 (t) where (n)
Γ3 (t) := −
t n−1 3λ 2 (n) m ¯ xk , v (k) (Tk+1 ) ds gt−s 4 2π(s − Tk+1 ) Tn
(5.4)
k=0
t n−1 3λ 2 (n) (k) := − m ¯ xk , v (Tk+1 ) ds gt−s ps−Tk+1 m ¯ xk − . 4 2π(s − Tk+1 ) Tn k=0 (5.5) (n) (n) Note that Γ3 (t) is obtained from Γ3+8 (t) replacing pt (x, y) by (2πt)−1/2 (recall that dx m ¯ (x) = 2). In conclusion we have (n) Γ8 (t)
v (n) (t) =
10 i=1
Let us define
(n)
Γi (t) .
3 ψn := − m ¯ , v (n) (Tn+1 ) 4 xn
(5.6)
(5.7)
so that, recalling (4.1), ξn =
n−1
ψk .
(5.8)
k=0
We then set t = Tn+1 in (5.6) and project it on m ¯ xn , getting ψn = ηn −
n−1
An,k ψk
(5.9)
3 (n) ¯ , Γ (Tn+1 ), i = 1, . . . , 10 ηn (i) = − m 4 xn i
(5.10)
k=0
where ηn :=
10 i=1 i=3
and
ηn (i),
√ 2 T 3 √ 1k
Note that ηn (3) ≡ −
n−1
An,k ψk .
(5.11)
(5.12)
k=0
Despite its appearance, (5.9) is not a linear equation in the variables ψn , because the term ηn still depends on the unknowns v (n) (t). We shall however see, using the a priori bounds of Section 4, that all contributions to ηn which contain the unknowns vanish as λ ↓ 0 (this is not exactly true, as some terms will vanish
52
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
only after averaging over n). As already seen in Section 4 the matrix A := ((An,k )) improves after iterations. Calling ψ and η the vectors whose components are ψn and ηn , (5.9) becomes ψ = η − Aψ . (5.13) In Rnλ (τ )+1 we introduce the seminorms |y|n := sup0≤j≤n |yj | and the corresponding seminorms on matrices as M n := sup |M y|n = sup |y|n =1
n
0≤i≤n j=0
|Mi,j | .
(5.14)
By the triangular structure of A, see (5.11), we easily get the following lemma.
Lemma 5.1 For each τ > 0 there is a constant C = C(τ ) such that for any λ > 0 Cj . sup Aj n ≤ j! n≤nλ (τ )
(5.15)
Hence, from (5.13), ψ=
∞
(−A)j η = (1 + A)−1 η,
sup (1 + A)−1 n ≤ C .
(5.16)
n≤nλ (τ )
j=0
It will be convenient to consider also one iteration of (5.13), i.e. ψ = η − Aη + A2 ψ .
(5.17)
Explicitly we have (A2 )n,k =
n−1
An,j Aj,k 10≤k
(5.18)
j=k+1
We conclude the section by showing that A2 has a scaling limit. Proposition 5.2 For each τ > 0 there is a constant C = C(τ ) such that 2 sup A n,k ≤ Cλ2 T .
(5.19)
0≤k
Moreover, for each δ, τ > 0, lim
sup
λ↓0 0≤τ2 <τ1 ≤τ τ1 −τ2 >δ
2 −1 2 (λ T ) A n (τ1 ),n (τ2 ) − λ λ
9 =0. 2
(5.20)
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
Proof. By (5.11)
53
n−1 1 2 √ √ A n,k ≤ Cλ2 T n − j j−k j=k+1
which proves (5.19). By (5.11) we have 9 9 (λ2 T )−1 A2 n (τ ),n (τ ) − = λ 1 λ 2 2 2π 1 )−1 nλ (τ 2 × nλ (τ1 ) − j + nλ (τ1 ) − j − 1 j=n (τ )+1 λ
2
2 −π × j − nλ (τ2 ) + j − nλ (τ2 ) − 1
.
(5.21)
1 Since π = 0 dx [x(1 − x)]−1/2 , by letting σ = τ1 − τ2 and changing the index of summation the r.h.s. of (5.21) equals 1 nλ (σ)−2 2 2 1 9 √ √ − dx 2π j=0 nλ (σ) − j + nλ (σ) − j − 1 j + j + 1 x(1 − x) 0 (5.22) which converges, as λ ↓ 0, to 0 uniformly for σ ∈ [δ, τ ], as can be easily checked. Proposition 5.2 is proved.
6 Bounds on η Recalling (5.8)–(5.10), a term η(i) (as in the previous section, we are here using vectorial notation) contributing to ψ whose seminorm is |η(i)|n = o(λ2 T ) (i.e. such that (λ2 T )−1 |η(i)|n vanishes as λ ↓ 0) does not contribute to the limiting equation for ξn since n ≤ (λ2 T )−1 τ . This is the case for some of the η(i)’s, i.e. η(1), η(7), and η(10), as we shall see. √ Clearly ηn (2) is not negligible because its typical magnitude is λ T . It will be examined in the next section, where we shall see that , summed over n it gives a finite contribution because of cancellations related to its martingale nature.The other terms, i.e. ηn (i), with i from 4 to√9, live on an intermediate stage: they are smaller than the a priori bound λ1−ζ T , yet not small enough to be directly negligible. We shall rewrite the factors v (n) (t) via (5.6), with the idea that if we get two η(i), i = 4, . . . , 10 then the corresponding terms become negligible. We will use the following notation. Definition 6.1 Let ηn (i1 , . . . , ik ), k > 1, ij ∈ {3, 4, 5, 6, 8} when j < k, ik ∈ {1, . . . , 10}, be the term which is obtained from ηn (i1 ) by replacing the v () (·)– () function in its expression by Γi2 (·); then, the new v ( ) (·) function which appears
54
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
( )
is replaced by Γi3 (·) and so forth till the last one, which is not changed. Instead of an index ij may appear the symbol 3 + 8 or 4 + 5, it means that, at that stage () () of the iteration, we replace the v () (·)–function by Γ3 (·) + Γ8 (·), or respectively () () by Γ4 (·) + Γ5 (·). We will also consider i1 = 9, in which case there is a product of two v– functions. Then, ηn (9|i2 , . . . , ik ; i2 , . . . , ih ) is defined by doing the branching i2 , . . . , ik (with the previous rules) for the first v function and the branching i2 , . . . , ih for the second one (the second branching may be absent). We define σn := ηn (2) + ηn (4 + 5, 2) + ηn (6, 2) + ηn (9|2; 2)
(6.1)
which will be studied in the next section where we will bound its sum over n using probabilistic arguments, σn is in fact a truly stochastic term. The difference between ηn and σn is negligible, this being the main result in this section. Let (3) ω ∈ Ω : sup An,k ηk (2) ≤ λ2−ζ T (6.2) Bλ,τ := n≤nλ (τ ) k
where ηk (2) and An,k are defined in (5.10) and (5.11) respectively. We will prove in Appendix B that for each τ, ζ, q > 0 there is C = C(τ, ζ, q) so that (3) (4) (6.4) P Bλ,τ ∩ Bλ,τ ≥ 1 − Cλq . Finally define xn,∗ := sup |xn |, k≤n
x∗ (τ ) = xnλ (τ ),∗ .
(6.5)
Proposition 6.2 Recalling (4.37), let (1,2)
(3)
(4)
Bλ,τ := Bλ,τ ∩ Bλ,τ ∩ Bλ,τ .
(6.6)
Then for each τ > 0, there is a constant C = C(τ ) so that for any λ > 0 and ζ small enough |ηn − σn | ≤ Cλ2 T 3/4 (1 + x2n+1,∗ )
on the set Bλ,τ .
(6.7)
Proof. We will call negligible a term which is bounded by the r.h.s. of (6.7). We will next examine one by one all the terms which contribute to ηn − σn and show that they are all negligible, thus proving Proposition 6.2.
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
55
Step 1. The terms ηn (1) and ηn (10). We have that ηn (1) = 0; indeed by (3.8), (n) (n) since gt is self–adjoint and gt m ¯ xn = m ¯ xn , 3 3 (n) ¯ , g v (n) (Tn ) = − m ¯ , v (n) (Tn ) = 0 ηn (1) = − m 4 xn T 4 xn
(6.8)
by (3.2). To show ηn (10) is negligible we use the bound (4.22); since we are in the set Bλ,τ , we get, from (4.7) (n)
sup t∈[Tn ,Tn+1 ]
Γ10 (t)∞ ≤ CT 5/2 λ3−3ζ
hence
|ηn (10)| ≤ CT 5/2 λ3−3ζ . (6.9)
Step 2. The term ηn (7). By (5.3) and (4.51), (n)
Γ7 (t) = I1 + I2 + I3 + I4
(6.10)
where t n λ (n) 2 I1 := − (xk − xk−1 ) ds gt−s ps−Tk m ¯ xk−1 2 T n k=1 n " (xk − xk−1 )3 t (n) ds gt−s ps−Tk m ¯ I2 := −λ xk−1 6 T n k=1 # t (n) + (xk − xk−1 )4 ds gt−s ps−Tk ak−1 n
I3 := −
Tn
9λ m ¯ xk−1 , v (k−1) (Tk )m ¯ xk−1 , v (k−1) (Tk ) 16 k=1
I4 := λ
n t k=1
Tn
(6.11)
(6.12)
t Tn
(n)
ds gt−s ps−Tk m ¯ xk−1 (6.13)
(n) ¯ xk−1 R(v (k−1) (Tk )) . ds gt−s ps−Tk m
(6.14)
To bound I1 we first consider the term with k = n. We write, for t > Tn + 1, t t Tn +1 (n) (n) ds gt−s ps−Tn m ¯ xn−1 = + ¯ xn−1 . ds gt−s ps−Tn m Tn
Tn
Tn +1
The first integral is bounded by a constant. In the second integral, as well as in the integrals in I1 when k < n, we write ps−Tk m ¯ xk−1 = 2
∂ps−Tk m ¯ xk−1 ∂s
56
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
and then use (4.17). We thus get recalling the definition (4.3) of δ ∗ (τ ), and using (4.10), ∗ 2 n δ (τ ) 1 ∗ 2 √ ≤C I1 ∞ ≤ Cλδ (τ ) log T + λ3 T | log λ| . n−k+1 λ T k=1
−1/2 Since m ¯ ¯ and as ak−1 is bounded, using xk−1 ∈ L1 (R, dx), pt m xk−1 ∞ ≤ Ct again (4.10) n t n ∗ 3 −1/2 ∗ 4 ds (s − Tk ) + Cλδ (τ ) (t − Tn ) I2 ∞ ≤ Cλδ (τ )
≤ C
k=1
∗
δ (τ ) √ λ T
3
Tn
λ3 T 3/2 + C
δ ∗ (τ ) √ λ T
4
k=1
λ3 T 2 .
Recalling the definition (4.2) of V∗ (τ ) and the definition (4.36) of V∗⊥ (τ ), we get √ V∗ (τ )V∗⊥ (τ ) √ I3 ∞ ≤ CλV∗ (τ )V∗⊥ (τ ) T nλ (τ ) ≤ C λ2 T . 2 λ T Finally, recalling (2.5), I4 ∞ ≤ Cλ T nλ (τ )V∗ (τ )3 ≤ C
V∗ (τ ) √ λ T
3
λ3 T 3/2 .
√ Since we are √restricting our considerations to the set where δ ∗ (τ ) ≤ λ1−ζ T , V∗ (τ ) ≤ λ1−ζ T and V∗⊥ (τ ) ≤ λ1−ζ , we get √ √ (n) Γ7 (t)∞ ≤ Cλ2−2ζ T hence |ηn (7)| ≤ Cλ2−2ζ T . (6.15) sup t∈[Tn ,Tn+1 ]
Step 3. The term ηn (8). In order to show ηn (8) is negligible (as specified at the beginning of the present proof) we cannot use directly the a priori bounds, but we use equation (5.6) to get ηn (8) =
10
ηn (8, i) + ηn (8, 3 + 8)
(6.16)
i=1 i=3,8
and show each of the terms on the r.h.s. above are negligible. We have n−1 (1) ηn (8) = An,k ψk k=0
where, recalling 2 = 1, m ¯ , Tn+1xk 3λ (1) An,k := − ds 4 Tn
(6.17)
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
m ¯ xn , ps−Tk+1 m ¯ xk
−
1 1, m ¯ xk 1k
m ¯ xn , 1
Let us define (recall that m ¯ > 0)
57
(6.18)
:= λ ds dx dy − y) − ¯ xk (y) . m 2π(s − Tk+1 ) Tn (6.19) We are going to prove that for any τ > 0 there is a constant C = C(τ ) such that for any 0 ≤ k < n ≤ nλ (τ ) 1 (1) (1) 2 √ + 1k=n−1 . (6.20) An,k ≤ CBn,k ≤ Cλ 1 + xn+1,∗ T (n − k)3/2 (1) Bn,k
Tn+1
m ¯ xn (x) ps−Tk+1 (x
1
Proof of (6.20). Let xn,k := xn − xk and define, for any θ ∈ R and t > 0, e−(y−x−θ)2 /2t − 1 √ f (t, θ) = dx dy m ¯ (x)m ¯ (y) . 2πt Then, after a change of variables in (6.19), its r.h.s. becomes equal to λ f (s − Tk+1 , xn,k ).
Tn+1 Tn
ds
The case k < n − 1. After changing variables in the time integral, we get, for k < n − 1, 1 (1) (6.21) Bn,k = λT dt f T (t + n − 1 − k), xn,k . 0
Since |e
−|ξ|
− 1| ≤ |ξ| and m ¯ (x) decays exponentially to 0 as |x| → ∞, we get f (t, θ) ≤ C
1 + θ2 t3/2
(6.22)
and, from (6.21), (1) Bn,k
≤ CλT
−1/2
[1 +
x2n,k ]
1
dt 0
1 (t + n − 1 − k)3/2
which proves (6.20) when k < n − 1. The case k = n − 1. We have, after a change of variables in the time integral, T (1) Bn,n−1 = λ dt f (t, xn,n−1 ) . 0
Using the inequality f (t, θ) ≤ 4(2πt)−1/2 when t ≤ 1 and (6.22) when t > 1, we get T 1 1 + x2n,n−1 4 (1) √ + Cλ dt dt ≤ Cλ(1 + x2n,∗ ) . Bn,n−1 ≤ λ t3/2 2πt 0 1
58
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
The proof of (6.20) is concluded. Recall now that ηn (8, i) = −
n−1 3 (1) An,k m ¯ xk , Γki (Tk+1 ), 4
i = 1, 2, 3 + 8, 4, 5, 6, 7, 9, 10.
k=0
The term ηn (8, 1) in (6.16) vanishes by the same argument as in (6.8). Since (1) we are working in the set Bλ,τ , see (4.37), we have √ n−2 1 2 √ + 1 λ1−ζ T |ηn (8, 2)| ≤ Cλ 1 + xn+1,∗ T (n − k)3/2 k=0 ≤ Cλ2−ζ 1 + x2n+1,∗
(6.23)
which is negligible for ζ small. To bound ηn (8, 3 + 8) use (6.17), (5.2), and (5.7) to write ηn (8, 3 + 8) = λ
n−1 k=0
(1) An,k
k−1
Bk,h ψh ,
Bk,h :=
h=0
Tk+1
Tk
ds m ¯ xk , ps−Th+1 m ¯ xh . √
(6.24)
We note that, by (5.7), |ψ|n ≤ CVn,∗ ; it is also easy to verify Bk,h ≤ C T (k − h)−1/2 . Plugging this bounds, together with (6.20), into (6.24), we get √ n−1 |ηn (8, 3 + 8)| ≤ Cλ2 T Vn,∗ 1 + x2n+1,∗ k=0
×
n−1
k−1
1 √ ≤ CλVn,∗ 1 + x2n+1,∗ k−h h=0 k=0
1 √ + 1k=n−1 T (n − k)3/2
1 √ + 1k=n−1 T (n − k)3/2 ≤ CλVn,∗ 1 + x2n+1,∗ (6.25)
which shows ηn (8, 3 + 8) is negligible. It remains to bound ηn (8, i), i = 4, 5, 6, 7, 9, 10. The negligibility of those terms follows directly from (6.17), (6.20), and the bounds (4.18)–(4.22), (6.15). We thus conclude √ |ηn (8)| ≤ Cλ2−ζ T 1 + x2n+1,∗ . (6.26) Step 4. The term ηn (4 + 5). We claim ηn (4 + 5) =
3λ 4
Tn+1
Tn
ds m ¯ xn , pTn+1 −s v (n) (s)
(6.27)
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
59
Indeed s 3λ Tn+1 ∂ps−s (n) ηn (4) = − ds ds m ¯ xn , v (s ) 4 Tn ∂s Tn 3λ Tn+1 Tn+1 ∂ps−s (n) = v (s ) ds ds m ¯ xn , 4 Tn ∂s s 3λ Tn+1 = ds m ¯ xn , pTn+1 −s v (n) (s ) − ηn (5) 4 Tn which proves (6.27). In order to bound ηn (4 + 5), we use equation (5.6) to get ηn (4 + 5) =
ηn (4 + 5, 2) + ηn (4 + 5, 3, 2) + ηn (4 + 5, 3, 3) 10 + [ηn (4 + 5, i) + ηn (4 + 5, 3, i)] .
(6.28)
i=1 i=2,3
The term ηn (4 + 5, 2) is in σn . Postponing the analysis of ηn (4 + 5, 3, 2) let us first show that ηn (4 + 5, i), i = 2, 3 are negligible. The terms ηn (4 + 5, i), i = 2, 3. We have ηn (4 + 5, 1) =
3λ 4
Tn+1
Tn
(n)
ds m ¯ xn , pTn+1 −s gs−Tn v (n) (Tn )
¯ xn , see (3.2), we can apply (4.11) to deduce that since v (n) (Tn ) is orthogonal to m |ηn (4 + 5, 1)| ≤ Cλ
Tn+1
ds e−α(s−Tn ) λ1−2ζ
Tn
having used that v (n) (Tn )∞ ≤ Cλ1−2ζ , see (4.47). Thus |ηn (4 + 5, 1)| ≤ Cλ2−2ζ which for ζ small is negligible. We next show ηn (4 + 5, 8) is negligible. We have n−1
ηn (4 + 5, 8) =
3 2 λ ψk 4 k=0
$ ×
(n) m ¯ xn , pTn+1 −t gt−s
Tn+1
t
dt Tn
¯ xk ps−Tk+1 m
ds Tn
2 − 2π(s − Tk+1 ) (n)
% .
(6.29)
We now use the decomposition (4.31) for gt−s above. Recalling (6.19), the (n) term obtained by replacing gt−s (x, y) by (3/4)m ¯ xn (x)m ¯ xn (y) in (6.29) is bounded
60
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
by Cλ
n−1 k=0
(1) Bn,k |ψk |
×
Tn+1
Tn
n−1 k=0
√ dt m ¯ xn , pTn+1 −t m ¯ xn ≤ C T λ2 Vn,∗ 1 + x2n+1,∗
1 √ + 1k=n−1 T (n − k)3/2
√ ≤ C T λ2 Vn,∗ 1 + x2n+1,∗ (n)
where we used (6.20). We next consider the case when we replace gt−s in (6.29) (n,⊥) by gt−s . Since, by (4.14) and (4.32), (n,⊥) ¯ xk ≤ Ce−α(t−s) ps−tk+1 m ¯ xk ∞ pTn+1 −t gt−s ps−Tk+1 m ∞
1 ≤ Ce−α(t−s) s − Tk+1 (n)
(n,⊥)
we can bound (6.29) (with gt−s replaced by gt−s ) as n−1
Tn+1
t
1 ds e−α(t−s) s − Tk+1 Tn Tn k=0 n−1 √ √ 1 ≤ C T λ2 Vn,∗ ≤ C λ3−ζ T + λ2−ζ T 1k=n−1 + √ n−k k=0 Cλ2
|ψk |
dt
which is negligible for ζ small enough. It remains to show ηn (4 + 5, i), i = 4, 5, 6, 7, 9, 10 is negligible. This follows from the bounds (4.18)–(4.22), (6.15), and (6.27). The terms ηn (4 + 5, 3, i). We write ηn (4 + 5, 3) =
n−1 k=0
(2)
An,k ψk
(6.30)
where (2) An,k
3λ2 = 4
Tn+1
ds Tn
$
Tn+1
ds s
(n) m ¯ xn , pTn+1 −s gs −s
2
2π(s − Tk+1 )
% 1k
By using the bounds (4.14) and (4.10), it is easy to show λ2 T 3/2 (2) 1k
(6.32)
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
61
From (6.30) we thus get ηn (4 + 5, 3, i) =
n−1 k=0
(2)
An,k ηk (i) .
(6.33)
The first one, ηn (4 + 5, 3, 1), vanishes, by (6.8). By using the bounds (4.18)–(4.22), (6.15), (6.26), and (6.32) it is immediate to verify ηn (4 + 5, 3, i) is negligible for i = 2, 3 if ζ is chosen small enough. To bound η(4 + 5, 3, 3) we note that, by (5.12) η(4 + 5, 3, 3) = −A(2) Aψ with A = ((An,k )) as in (5.11). Hence, recalling (5.16), η(4 + 5, 3, 3) = −A(2) (1 + A)−1 Aη .
(6.34)
By (6.32), it is easy to show sup A(2) n ≤ CλT
(6.35)
n≤nλ (τ )
so that, by the bound in (5.16), |ηn (4 + 5, 3, 3)| ≤ CλT
i=3
max A,k ηk (i) . ≤n
(6.36)
k<
By definitions (6.2) and (6.6) the term with i = 2 in the r.h.s. of (6.47) is therefore negligible. For i = 1, 7, 8, 10 (resp. i = 4, 5, 6, 9) we can use the bounds already proved (6.8), (6.15), (6.26), and (6.9) (resp. the a priori bounds (4.18)–(4.21)) together with An ≤ C to conclude that ηn (4 + 5, 3, 3) is also negligible. We are thus left with the term ηn (4 + 5, 3, 2). Recalling (6.33), we write it as ηn (4 + 5, 3, 2) =
n−1 k=0
(2,L)
An,k ηk (2) +
n−1 k=0
(2,R)
An,k ηk (2)
(6.37)
where Tn+1 Tn+1 2 3λ2 (n) ds ds m ¯ xn , pTn+1 −s gs −s 11k
An,k
:=
62
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
Since, by (4.10) Tn+1 Tn+1 (n) ds ds m ¯ xn , pTn+1 −s gs −s 1 ≤ CT 2 Tn s n−1 (2,L) An,k ηk (2) ≤ CλT
we have
k=0
√ n−1 λ T √ ηk (2) n−k
(6.39)
k=0
(4)
which is negligible because we are in the set Bλ,τ , recall (6.3). On the other hand, by using again (4.10) we get " # 1 (2,R) (6.40) An,k ≤ Cλ2 T 3/2 1k=n−1 + (n − k − 1)3/2 whence, recalling (4.5) n−1 n−2 1 (2,R) 3−ζ 2 An,k ηk (2) ≤ Cλ T 1+ (n − k − 1)3/2 k=0
(6.41)
k=0
which is negligible. Step 5. The term ηn (6). To study this term we are going to use the same strategy as for ηn (4 + 5). The explicit expression of ηn (6) is Tk+1 n−1 3λ Tn+1 ∂ps−s (k) v (s ) ds ds m ¯ xn , ηn (6) = 4 ∂s Tn Tk
(6.42)
k=0
which we decompose as we did for ηn (4 + 5), i.e. as in (6.28) with 4 + 5 replaced by 6. The terms ηn (6, i), i = 2, 3. We have ηn (6, 1) =
Tk+1 n−1 3λ Tn+1 ∂ps−s (k) gs −Tk v (k) (Tk ) . ds ds m ¯ xn , 4 ∂s Tn Tk
(6.43)
k=0
√ (k) Since, by (3.2) and(4.7) gs −Tk v (k) (Tk )∞ ≤ Ce−α(s −Tk ) λ1−ζ T , by (4.17) the double integral above for k = n − 1 is less or equal than √ Tn+1 Tn−1 +T /2 Ce−α(s −Tn−1 ) λ1−ζ T ds ds s − [Tn−1 + T /2] Tn−1 Tn √ Tn+1 Tn −αT /2 1−ζ √ λ T Ce + ds ds ≤ Cλ1−ζ T . s−s Tn Tn−1 +T /2
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
63
On the other hand,√for k < n − 1 the double integral in (6.43) is bounded by C(n − k + 1)−1 λ1−ζ T , so that √ |ηn (6, 1)| ≤ Cλ2−ζ T | log λ| (6.44) which is therefore negligible. We next consider ηn (6, 8); its explicit expression is Tn+1 Tk+1 s n−1 k−1 3λ2 ψh dt ds ds ηn (6, 8) = 4 T T T n k k k=0 h=0 $ % ∂pt−s (k) 2 × m ¯ xn , ¯ xh − g ps −Th+1 m . ∂t s−s 2π(s − Th+1 )
(6.45)
(n)
We now use the decomposition (4.31) for gt−s above. Since ( ) C ∂pt−s m ¯ xk ≤ ¯ xn , ∂t m (t − s)3/2 (k)
and recalling (6.19), the term obtained replacing gs−s (x, y) in (6.45) by ¯ xk (y) can be bounded by (3/4)m ¯ xk (x)m CλVn,∗
n−1 k−1 k=0 h=0
√ T (1) B ≤ Cλ3−ζ T [1 + xn+1,∗ ] (n − k)3/2 k,h
(6.46)
(k)
which is negligible. We next consider the case when gs−s in (6.45) is replaced by
(k,⊥) gs−s .
By (4.17), (4.32), and (4.14) we have ∂p 1 2 e−α(s−s ) t−s (k,⊥) gs−s ps −Th+1 m ¯ xh − ≤C ∂t t−s s − Th+1 2π(s − Th+1 ) ∞ (k)
(k,⊥)
hence the r.h.s. of (6.45) with gt−s replaced by gt−s
can be bounded by
n−1 k−1 √ Cλ2 T Vn,∗
√ 1 √ ≤ Cλ2−ζ | log λ| T (n − k) k − h k=0 h=0
which is negligible. It remains to consider the terms ηn (6, i), i ≥ 4, i = 8. We have |ηn (6, i)| ≤ Cλ| log λ|T max
sup
k≤n t∈[Tk ,Tk+1 ]
(k)
Γi (t)∞
(6.47)
which, together with (4.18) for i = 4, (4.19) for i = 5, (4.20) for i = 6, (6.15) for i = 7, (4.21) for i = 9, (6.9) for i = 10, shows that the terms ηn (6, i), i ≥ 4, i = 8 are all negligible.
64
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
The terms ηn (6, 3, i). The term ηn (6, 3, 1) = 0. Analogously to (6.30), we write n−1
ηn (6, 3) =
k=0
(3)
An,k ψk
(6.48)
where (3)
An,k =
s Tn+1 Th n 3λ2 dt ds ds 4 Th−1 Th−1 h=k+2 Tn $ % 2 ∂pt−s (h−1) g × m ¯ xn , 1k
(6.49)
By (4.17), (4.10), and (4.14) it is easy to show λ2 T 3/2 log(n − k) (3) √ 1k
(6.50) (2)
Except for the factor log(n − k) ≤ C| log λ|, An,k has the same behaviour as An,k , compare (6.32) with (6.50). The analysis in Step 4 applies and we can therefore conclude that all the terms ηn (6, 3, i), i = 2, are negligible. We are left with ηn (6, 3, 2). As in (6.38) we define s Tn+1 Th n 2 3λ2 dt ds ds 4 2π(T − T ) T T T h+1 k+1 n h−1 h−1 h=k+2 ( ) ∂pt−s (h−1) gs−s 1 1k
(3,L)
An,k :=
(3,R) An,k
s Tn+1 Th n 3λ2 := dt ds ds 4 Th−1 Tn Th−1 h=k+2
2
(
2
− 2π(s − Tk+1 ) 2π(Th+1 − Tk+1 )
m ¯ xn ,
) ∂pt−s (h−1) gs−s 1 1k
Setting 3λ (2) Bn,h := √ 2 2πT we have
Tn+1
Tn
Th
dt
s
ds Th−1
Th−1
ds
) ( ∂pt−s (h−1) gs−s 1 m ¯ xn , ∂t
√ n−1 n h−2 (3,L) λ T (2) √ An,j ηj (2) ≤ ηj (2) Bn,h j=0 h=2 j=0 h − j
(6.52)
(6.53)
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
65
and, using (4.10) and (4.17), λT (2) . Bn,h ≤ C n−h (4)
Thus, since we are in the set Bλ,τ , see (6.3), the l.h.s. of (6.53) is negligible. On the other hand, by using again (4.10) and (4.17) it is easy to show 1 (3,R) An,k ≤ Cλ2 | log λ| T 3/2 (n − k)3/2 n−1 n−1 (3,R) 1 3−ζ 2 A η (2) ≤ Cλ | log λ| T j n,j (n − j)3/2 j=0 j=0
hence
(6.54)
which is negligible. Step 6. The term ηn (9). The iterative scheme to estimate η(9) is the following. We first branch, using (5.6), one of the v (n) (s) which appears on the r.h.s. of (3.14) and stop the iteration for the terms η(9|i), i = 2, 3. In η(9|3) we use again the equation (5.6) and stop the iteration. We then branch the other v (n) (s) on the term η(9|2) stopping the iteration with the same rule as before. We thus get η(9)
= η(9|2; 2) + η(9|2; 3, 2) + η(9|2; 3, 3) + η(9|3, 2) + η(9|3, 3) [η(9|i) + η(9|2; i) + η(9|2; 3, i) + η(9|3, i)] . (6.55) + i=2,3
We need to show that η(9) − η(9|2; 2) is negligible. We start from η(9|i), i = 2, 3. We have 9 Tn+1 (n) ηn (9|i) = dt m ¯ xn , m ¯ xn v (n) (t)Γi (t) . (6.56) 4 Tn (n)
For i = 1, by (3.2), gt−Tn v (n) (Tn )∞ ≤ Ce−α(t−Tn ) v (n) (Tn )∞ , by (4.47) v (n) (Tn )∞ ≤ λ1−ζ , hence √ |ηn (9|1)| ≤ Cλ1−ζ T λ1−ζ (6.57) (1,2)
so that it is negligible. On the set Bλ,τ , see (4.37) and (6.6), we can use the bounds (4.7), (4.18), (4.19), (4.20), (6.15), (4.21), and (6.9), which show η(9|i) is negligible for i = 4, 5, 6, 7, 9, 10. To bound η(9|8) we note that, by (5.5), (4.7), (n) ¯ (x) > 0, and gt (x, y) ≥ 0), we have (using also m ¯ ∞ = 1, m |ηn (9|8)| ≤ λCVn,∗
n−1 k=0
|ψk |
Tn+1
dt Tn
t
1 ds s − Tk+1 Tn
66
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
2 (n) ¯ xk (z) . dx dy dz m ¯ xn (x)gt−s (x, y) e−(y−z) /[2(s−Tk+1 )] − 1 m
× (n)
Since gt is self–adjoint in L2 (R, dx) we have Then, from (6.19) and (6.20) it follows that |ηn (9|8)| ≤ CV∗ (τ )T
n−1 k=0
(6.58)
(n)
dx m ¯ xn (x)gt−s (x, y) = m ¯ xn (y).
(1) Bn,k |ψk | ≤ CλV∗ (τ )2 T 1 + x2n+1,∗
which is negligible because of (4.7). We next consider η(9|3, i), i = 2, 3, Tn+1 t n−1 9λ ηn (9|3, i) = ηk (i) dt ds 4 Tn Tn
$
2
(n) m ¯ xn , m ¯ xn v (n) (s)gt−s
%
. 2π(s − Tk+1 ) (6.59) Note that, from (6.8) η(9|3, 1) = 0 and that, for i ≥ 4, from (5.11), k=0
n−1
|ηn (9|3, i)| ≤ CλVn,∗
k=0
|ηk (i)|
Tn+1
dt Tn
t
1 ds s − Tk+1 Tn
≤ T CVn,∗
n−1
An,k |ηk (i)| .
(6.60)
k=0
Since An ≤ C, by using again (4.7), (4.18), (4.19), and (4.20)–(6.9), we show η(9|3, i), i ≥ 4 is negligible. Next, using (5.12), (5.16) and (6.59), Tn+1 t n−1 9λ −1 (1 + A) Aη k dt ds η(9|3, 3) = − 4 Tn Tn k=0
$ ×
2 2π(s − Tk+1 )
%
(n) m ¯ xn , m ¯ xn v (n) (s)gt−s
(6.61)
whence, recalling (1 + A)−1 n ≤ C, An ≤ C, ηn (1) = 0,
√ 10 n−1 T √ |η(i)|n Vn,∗ |η(9|3, 3)|n ≤ CλT sup |(Aη(2))k | + n −k k≤n i=4 k=0 2−ζ ≤ CT λ T + λ2−2ζ T 2 1 + x2n+1,∗ Vn,∗ where we used the bounds (6.8), (6.15), (6.26), (6.9), and (4.18)–(4.21) to estimate |η(i)| together with the fact that we are on the set Bλ,τ (see (6.2) and (6.6)). Hence η(9|3, 3) is negligible. (n) Since on the set Bλ,τ Γ2 (t) has the same order as v (n) (t) (see (4.5)) we can bound η(9|2; i), η(9|2; 3, i), i = 2, 3, and η(9|2; 3, 3) as above.
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
67
We are left with the bounds of η(9|3, 2) and η(9|2; 3, 2). We have ηn (9|3, 2) = (L) (R) ηn (9|3, 2) + ηn (9|3, 2) where ηn(L) (9|3, 2) :=
n−1 2 9λ ηk (2) 4 2π(T n+1 − Tk+1 ) k=0
Tn+1
t
dt Tn
Tn
ηn(R) (9|3, 2) := ×
Tn+1
Tn
& ' (n) ds m ¯ xn , m ¯ xn v (n) (s)gt−s 1
n−1 9λ ηk (2) 4 k=0 t
2 2 − dt ds 2π(s − Tk+1 ) 2π(Tn+1 − Tk+1 ) Tn & ' (n) × m ¯ xn , m ¯ xn v (n) (s)gt−s 1 .
(6.62)
Since, by (4.7) and (4.10), Tn+1 t & ' (n) dt ds m ¯ xn , m ¯ xn v (n) (s)gt−s 1 ≤ Cλ1−ζ T 5/2 Tn Tn √ n−1 λ T (L) 1−ζ 3/2 √ ηk (2) ηn (9|3, 2) ≤ Cλ T n−k
we have
(6.63)
k=0
(4)
which is negligible since we are in the set Bλ,τ , see (6.3). On the other hand, again by (4.7) and (4.10), Tn+1 t & ' 2 2 (n) m ¯ xn , m dt ds ¯ xn v (n) (s)gt−s 1 − Tn 2π(s − Tk+1 ) 2π(Tn+1 − Tk+1 ) Tn " # 1 ≤ Cλ1−ζ T 2 1k=n−1 + (n − k − 1)3/2 whence, by (4.13) and (4.5) n−2 (R) 2−ζ 2 ηn (9|3, 2) ≤ Cλ T k=0
1 |ηk (2)| + |ηn−1 (2)| ≤ λ3−ζ T 5/2 (n − k − 1)3/2 (6.64)
which is negligible. (2)
Since Γn (t) has the same order as v (n) (t), it follows from (6.56) that the term η(9|2; 3, 2) can be analyzed as η(9|3, 2). Proposition 6.2 is thus proved.
68
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
7 Bounds on the stochastic terms The stochastic terms in the title are the ones left from the previous section, i.e. σn , see (6.1). As already mentioned we will obtain bounds for the sums of the components of these vectors rather than for the components themselves. We define Sn (-)
:=
n−1
- ∈ L := {2, (4 + 5, 2), (6, 2), (9|2; 2)}
ηk (-),
k=0
Sn
:=
Sn (-) =
∈L
n−1
σk
(7.1)
k=0
and call S the vector whose components are Sn . Analogously ξ is the vector whose components are ξn , see (5.8). Proposition 7.1 Recall that A is the matrix whose entries are defined in (5.11). Then ξ = S − AS + R + A2 ξ on the set Bλ,τ (7.2) where Rn :=
n−1
(ηk − σk ) −
k=0
n−1 k=0
(A[η − σ])k .
(7.3)
Moreover, for each τ > 0 there is a constant C = C(τ ) such that for any λ and ζ small enough on the set Bλ,τ . (7.4) |R|n ≤ CT −1/4 1 + x2n+1,∗ The proof is based on the following simple lemma. Lemma 7.2 Let M be a lower triangular matrix such that Mn,k = Mn ,k whenever n−1 n −k = n−k. Let also u = {uh ; h = 0, . . . , n} be a vector and define vn = h=0 uh (resp. v0 = 0). Then n−1 (M u)h = (M v)n . (7.5) h=0
Proof. The l.h.s. of (7.5) can be written as n−1 h−1
Mh,k (vk+1 − vk ) =
h=0 k=0
n−1 h−1
Mh,k vk+1 −
h=0 k=0
n−1 h−1
Mh,k vk .
h=0 k=0
We call k = k + 1 in the first sum and h = h − 1 in the second one, getting h
n−1
n−1
h=0
h=0 k =1
(M u)h =
Mh,k −1 vk −
n−2
h
h =−1 k=0
Mh +1,k vk
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
69
We may as well extend the first sum from k = 0 because v0 = 0. We then get n−1
n−1
h=0
k=1
(M u)h =
Mn−1,k−1 vk +
h n−2
[Mh,k−1 − Mh+1,k ] vk .
h=0 k=0
Note that the second term above vanishes by the assumptions on M , and this completes the proof of (7.5). Proof of Proposition 7.1. From (5.8) and (5.17), ξn =
n−1
[η − Aη + A2 ψ]k .
k=0
Next, apply Lemma 7.2 for A and A2 . From (7.1) and (5.8), we get (7.2). Since An ≤ C, the bound (7.4) follows from Proposition 6.2. Our next task is to control the stochastic term Sn . Proposition 7.3 For any τ > 0 we have lim lim P
L→∞ λ↓0
sup |Sn (2)| > L
=0
(7.6)
- ∈ L \ {2} .
(7.7)
n≤nλ (τ )
|Sn (-)| lim P sup > T −1/4 λ↓0 1 + x n,∗ n≤nλ (τ )
= 0,
Proof of (7.6). By (3.9), (5.10), and (7.1), n−1 3λ m ¯ xk , z (k) (Tk+1 ) Sn (2) = − 4
(7.8)
k=0
which is an FTn –martingale. In Appendix B it is proved there is a C so that, for any n ≤ nλ (τ ) and t ∈ [Tn + 1, Tn+1 ], 2p sup E z (n) (t, x) p = 1, 2 . (7.9) ≤ C(t − Tn )p , x∈R
Then, by Doob’s martingale inequality, we get, for some constant C,
Cnλ (τ )λ2 T 1 P sup |Sn (2)| > L ≤ 2 E Snλ (τ ) (2)2 ≤ L L2 n≤nλ (τ ) which proves (7.6).
(7.10)
70
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
Proof of (7.7). The case - = (4 + 5, 2). We have (recall (6.27)) Sn (4 + 5, 2) =
n−1 3λ2 Th+1 ds m ¯ xh , pTh+1 −s z (h) (s) . 4 Th h=0
This is again a FTn –martingale, so that, using again Doob’s inequality, and the bound 2 Th+1 E ds m ¯ xh , pTh+1 −s z (h) (s) ≤ CT 3 Th (which follows from (7.9)) we conclude that for each ζ > 0 there is a constant C = C(τ, ζ) such that
√ −1/4 P sup |Sn (4 + 5, 2)| > T (7.11) ≤ Cnλ (τ )λ4 T 3 T ≤ Cλ2 T 5/2 . n≤nλ (τ )
The case - = (6, 2). By (3.13) and (3.9) ) ( Tk n−1 h 3λ2 Th+1 ∂ps−s (k−1) z ds ds (s ) . Sn (6, 2) = m ¯ xh , 4 ∂s Th Tk−1
(7.12)
h=1 k=1
For each h and k above, we write, recalling that m ¯ ξ (x) = m(x ¯ − ξ), xh m ¯ xh (x) = m ¯ xk−1 (x) − dξ m ¯ ξ (x) xk−1
thus obtaining, by exchanging the sums in (7.12), Sn (6, 2) = −S˜n (6, 2) +
n−1
Xn,k
(7.13)
k=1
where Tk xh n−1 h 3λ2 Th+1 ∂ps−s (k−1) ˜ z Sn (6, 2) := ds ds dξ m ¯ ξ , (s ) (7.14) 4 ∂s Tk−1 xk−1 h=1 k=1 Th Tk 3λ2 Tn ∂ps−s (k−1) Xn,k := z ds ds m ¯ xk−1 , (s ) . (7.15) 4 Tk ∂s Tk−1 We define (n) Mk
:=
k−1 h=1
Xn,h ,
k = 1, . . . , n
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
(n)
71
(n)
with M1 = 0. For each fixed n, {Mk ; k = 1, . . . , n} is an FTk−1 –martingale (n) since E(Xn,h |FTh−1 ) = 0. The quadratic variation of Mk is k 2 k−1 (n) (n) 2 Xn,h . Mh − Mh−1 = M (n) = k
h=2
h=1
We have Tk Tn Tk s ∧u 9λ4 Tn E = ds ds du du ds dx dx dy 16 Tk Tk−1 Tk Tk−1 Tk−1 ∂ps−s ∂pu−u × m ¯ (x) m ¯ (x ) gs −s (x, y) gu −s (x , y)a(λβ (y + xk−1 ))2 ∂s ∂u Tn Tk Tn Tk s ∧u 2 4 ≤ Ca∞ λ ds ds du du ds Tk Tk−1 Tk Tk−1 Tk−1 ∂pu−u ∂ps−s m ¯ m ¯ × gs −s gu −s 1∞ ∂s ∂u ∞ 1 Tn Tk Tn Tk s ∧u 1 1 2 4 ≤ Ca∞ λ ds ds du du ds )3/2 u − u (s − s Tk Tk−1 Tk Tk−1 Tk−1
2 FTk−1 Xn,k
≤ Ca2∞ λ4 | log λ|T 5/2 . Since, conditionally on FTk−1 , z k−1 is Gaussian we also have 4 FT E Xn,k ≤ Ca4∞ λ8 | log λ|2 T 5 . k−1 By the BDG inequality (see [16, §6, E. 4.1]) and the above bounds it follows that
2 n−1 4 2 (n) (n) 2 Mn E Xn,h ≤ CE M =CE ≤ C
n−1
n
4 EXn,h
h=1 2 8
h=1
+2
n−1 k−1
E
2 Xn,h E
2 FTk−1 Xn,k
k=1 h=1
≤ C n λ | log λ|2 T 5 .
(7.16)
We next use the following corollary of the Chebyshev inequality, P sup |ϑn | > δ ≤ P |ϑn |p > δ p ≤ δ −p E (|ϑn |p ) n≤N
n≤N
(7.17)
n≤N
and get, by (7.16), choosing p = 4, n−1
nλ (τ ) −1/4 P sup Xn,k > T Cn2 λ8 | log λ|2 T 5 ≤ Cλ2 T 4 . ≤T n≤nλ (τ ) k=1
n=1
72
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
It remains to bound S˜n (6, 2). We split it into two terms by using the decomposition (4.33) for z (k−1) . We get Tk xh n−1 h 3λ2 Th+1 ∂ps−s (k−1,⊥) S˜n⊥ (6, 2) := z ds ds dξ m ¯ ξ , (s ) 4 ∂s Tk−1 xk−1 h=1 k=1 Th Tk n−1 h 3λ2 Th+1 S˜n (6, 2) := ds ds 4 Tk−1 h=1 k=1 Th xh ∂ps−s m ¯ xk−1 w(k−1) (s ) × dξ m ¯ ξ , (7.18) ∂s xk−1 where we set w(k) (t) := 34 m ¯ xk , z (k) (t). In Appendix A we prove for each γ ∈ (0, 1/2) there exists a constant C such that ( ) ϕ1 ∂pt ≤ C m ϕ , ∧ ϕ . (7.19) ¯ ∞ t3/2 ∂t tγ We thus get |S˜n⊥ (6, 2)| ≤ Cλ2 n≤nλ (τ ) 1 + xn,∗ sup
≤ Cλ2 ≤ CT
nλ (τ )−1 h Th+1
Tk
ds
h=1
k=1
nλ (τ )−1 h
h=1 −1/2
k=1
sup
Th
ds
Tk−1
√ T (h − k + 1)3/2 sup
n≤nλ (τ ) t∈[Tk−1 ,Tk ]
z
1 z (k−1,⊥) (s )∞ (s − s )3/2 sup
t∈[Tk−1 ,Tk ]
(k−1,⊥)
z (k−1,⊥) (t)∞
(t)∞ .
(7.20)
On the other hand, for each γ ∈ (0, 1/2), since m ¯ xk−1 1 ≤ C, we also have
|S˜n (6, 2)| ≤ Cλ2 n≤nλ (τ ) 1 + xn,∗ sup
≤ Cλ2
nλ (τ )−1 h Th+1 h=1
k=1
nλ (τ )−1 h h=1
≤ C T −γ
k=1
sup
Tk
ds
Th
ds
Tk−1
T 1/2−γ (h − k + 1)3/2+γ sup
n≤nλ (τ ) t∈[Tk−1 ,Tk ]
|w
1 |w(k−1) (s )| (s − s )3/2+γ sup
t∈[Tk−1 ,Tk ]
|w(k−1) (t)|
(k)
(t)| √ . T
(7.21)
Since |w(k) (t)| ≤ Cz (k−1) (t)∞ by using the Gaussian estimate (4.6) (resp. (4.35)) and (7.21) (resp. in (7.20)), the bound in (7.7) for S˜n (6, 2) now follows. n−1 The case - = (9|2; 2). We have Sn (9|2; 2) = k=0 ηk (9|2; 2) where 9λ2 Tk+1 ηk (9|2; 2) = dt m ¯ xk , m ¯ xk z (k) (t)2 . (7.22) 4 Tk
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
73
We define γk := E ( ηk (9|2; 2)| FTk ) and then decompose Sn (9|2; 2) = Mn +
n−1
γk
(7.23)
(7.24)
k=0
where Mn is an FTn –martingale with bracket M n =
n−1
E ηk (9|2; 2)2 FTk − γk2 .
(7.25)
k=0
Note that the difference between the bracket ·n and the quadratic variation [·]n is a martingale, see [16]. We have t β 9λ2 Tk+1 2 2 dt ds dx dy m ¯ (x)m(x)g ¯ γk = t−s (x, y) a(λ (y + xk )) − 1 4 Tk Tk (7.26) where we exploited the identity 2 dx dy m ¯ (x)m(x)g ¯ (x, y) = dx m ¯ (x)m(x)g ¯ t−s 2(t−s) (x, x) = 0 which holds because x "→ gt (x, x) is an even function of x. We claim for each q > logλ−1 T = (1 ∧ β)/20 there is C = C(q) so that
−q |γk | ≤ C(λT )2 λβ−q + λβ xk,∗ + e−λ /C . (7.27) By Taylor expansion β a(λ (y + xk ))2 − 1 ≤ ≤
/ . C λβ (|y| + |xk |)1|y|≤λ−q + 1|y|>λ−q . / C λβ−q + λβ xk,∗ + 1|y|>λ−q .
(7.28)
We use the bound (7.28) in (7.26). There is C so that, for any t ∈ [0, T ] (T > 1), t t 2 ds dx dy m ¯ (x)|m(x)|g ¯ ds dx m ¯ (x)g2s (x, x) ≤ CT s (x, y) ≤ 0
0
the first two terms on the r.h.s. of (7.28) produce the first two terms on the r.h.s. of (7.27). We estimate next the contribution of the last one. Denoting by Ex,y,t the expectation w.r.t. a Brownian bridge from x to y in time t, by the Feynman–Kac formula we get t ds V (m(ω ¯ s )) ≤ e2t pt (x, y) (7.29) gt (x, y) = pt (x, y) Ex,y,t exp − 0
74
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
whence ¯ (x) gt (x, y)2 1|y|>λ−q dx dy 1|x|≤λ−q /2 m −2q
≤
e−[λ
/(8t)]+2t
√ 2πt
dx
−2q
dy gt (x, y) m ¯ (x) = 2
e−[λ
/(8t)]+2t
√ 2πt
. (7.30)
On the other hand, since m ¯ (x) decays exponentially as x → ∞, −q ¯ (x)gt (x, y)2 1|y|>λ−q ≤ Ce−λ /C dx g2t (x, x) m ¯ (x) dx dy 1|x|>λ−q /2 m −q
e−λ /C+4t √ . ≤ C 4πt
(7.31)
Taking q > logλ−1 T , from (7.30) and (7.31) we find that the last term on the r.h.s. of (7.28) yields the last one on the r.h.s. of (7.27). From (7.27), for any λ small enough, n−1 1 sup γk ≤ CT λβ−q . (7.32) 1 + x n,∗ n≤nλ (τ ) k=0
By choosing q ∈ (logλ−1 T, β − logλ−1 T ) (recall we fixed logλ−1 T = (1 ∧ β)/20) it follows the l.h.s. of (7.32) vanishes as λ ↓ 0. We are left with the bound on the martingale part Mn . By Doob’s inequality, recalling (7.25),
λ (τ ) √ n √ −1/4 E ηk (9|2; 2)2 ≤ C T E M nλ (τ ) ≤ C T P sup |Mn | > T n≤nλ (τ )
≤
2 √ C T (λ2 T )−1 Cλ2 T 3/2
k=0
(7.33)
in the second estimate we used (recall (7.9)) Tk+1 0 2 2 E (ηk (9|2; 2) ) ≤ Cλ dt dx m ¯ (x) E z (k) (t, x)4 ≤ Cλ2 T 3/2 . Tk
From (7.32) and (7.33) (recall xn,∗ is increasing), (7.7) for - = (9|2; 2) follows.
8 Conclusion of the proof In this section we prove Theorem 1.1. As before we denote by (m(t), h(t)) the solution of (1.1)–(1.10) (omitting the dependence on λ). Let δ be as in Proposition 2.1 and define the stopping time ¯ z ∞ ≥ δ} . tδ := inf{t ≥ 0 : inf m(t) − m z∈R
(8.1)
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
75
By Proposition 2.1, for each t ∈ [0, tδ ] the profile m(t) has a well defined center x(m(t)). We prove the theorem with x(λ) (t) := x(m(t ∧ tδ )) . Observe the process x(λ) (t) is adapted to Ft . Moreover, by Proposition 2.1 and the continuity properties of m(t), the process x(λ) (t) is continuous. Proof of (1.11). Given τ, ζ > 0, from Propositions 2.1, 4.1, and recalling (3.1) there is C = C(τ, ζ) so that, for any λ small enough, √ (1) sup sup |x(λ) (t) − xn | ≤ Cλ1−ζ T on the set Bλ,τ . tδ > λ−2 τ, n≤nλ (τ ) t∈[Tn ,Tn+1 ]
On the other hand, by (3.1), (4.2), and m ¯ ∞ ≤ 1, we have m(t) − m ¯ x(λ) (t) ∞ ≤ Vn + sup sup |x(λ) (t) − xn |
t∈[Tn ,Tn+1 ]
t∈[Tn ,Tn+1 ]
(8.2) ∀ n : Tn+1 ≤ tδ .
(8.3) From (4.6), (4.7), (8.2), and (8.3), by choosing ζ small enough (1.11) follows. To complete the proof of Theorem 1.1 we need to prove the weak convergence of the scaled process xλ (τ ) := x(λ) λ−2 τ . Let ξλ (τ ), τ ∈ R+ , be the (continuous) process obtained by linearly interpolating the values ξλ (λ2 Tn ) = ξn . From (4.6), (4.35), (4.39), and (8.2), for any τ and ε positive, lim P sup |ξλ (s) − xλ (s)| > ε = 0 . λ↓0
0≤s≤τ
It is therefore enough to prove the convergence of the process ξλ . To this end we shall use that ξn solves the equation (7.2). In Sections 7 and 8 we proved bounds on R and S, see (7.4) and Proposition 7.3 which hold with probability going to 1 as λ ↓ 0. They however depend on the unknown quantities xn,∗ . The a priori bounds of Section 4 yield xn,∗ ≤ nδ ∗ (τ ), which goes like λ−1 T −1/2+ζ , according to (4.7) and with n = nλ (τ ). With such a bound, our estimates on R and S become very bad and in any case inadequate to study the equation (7.2), for which we could only tolerate a bound on xn,∗ which diverges very weakly as λ ↓ 0. As mentioned, the bounds of Section 4 on δ ∗ (τ ) are quasi optimal, yet they yield a bound on xn,∗ which is far from correct. The point is that xn is the sum of the increments xk − xk−1 , whose size has indeed the order of δ ∗ (τ ); but there are a lot of cancellations, which make the absolute value of the sum much smaller than the sum of the absolute values. Such cancellations are lost using the a priori bounds of Section 4, we thus need to go back to the equation (7.2) itself (recalling that ξn and xn are essentially the same, see (4.39)). It seems, at this point, that we are back to a non linear problem, with the unknown ξn in both the “known terms” R and S hidden through xn . Such a non linearity is however not really dangerous,
76
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
as in our bounds for R and S the dependence on xn,∗ is always multiplied by a function of λ which vanishes as λ ↓ 0. As we will see, this makes easy the proof of Proposition 8.1 below. Let (8.4) ξ ∗ (τ ) := sup |ξn | n≤nλ (τ )
and x∗ (τ ) be as in (6.5). Proposition 8.1 For any τ ∈ R+ , lim lim P (ξ ∗ (τ ) > L) = 0 ,
(8.5)
lim lim P (x∗ (τ ) > L) = 0 .
(8.6)
L→∞ λ↓0
L→∞ λ↓0
Proof. Since (8.6) follows from (8.5) and (4.39), we only need to prove (8.5). Let (recall (7.1)) Sn(R) := Sn (-) (8.7) ∈L\{2} (R)
and denote as usual by S (R) the vector whose components are Sn . Given τ > 0 and ζ > 0 small enough, let Bλ,τ be as in (6.6) and define Gλ,L := Aλ,L ∩ Bλ,τ where Aλ,L is the event Aλ,L :=
ω ∈ Ω : |S(2)|nλ (τ )
(R) S −1/4 n ≤ L ; sup ≤T . n≤nλ (τ ) 1 + xn,∗
(8.8)
(8.9)
From (4.6), (4.35), (6.6), (6.4), and Proposition 7.3 we have, for any τ > 0, lim lim P (Gλ,L ) = 1 .
L→∞ λ↓0
(8.10)
We are going to show that given L there is a constant c so that, for all λ small enough, in Gλ,L we have |ξn | ≤ c for all n ≤ nλ (τ ) + 1. Given any L1 > 0, let N (L1 , λ) be the first index n ≤ nλ (τ ) for which |ξn | ≥ L1 ; otherwise we set N (L1 , λ) = nλ (τ ) + 1. Then what we have to show is that there is an L1 so that for all λ small enough, N (L1 , λ) = nλ (τ ) + 1 in Gλ,L . Having fixed L1 , there is a λ1 so that for λ ≤ λ1 and n ≤ N (L1 , λ), by (4.39) (1,2) (recall Gλ,L ⊆ Bλ,τ ) |xn | ≤ |ξn | + 1 ≤ L1 + 1,
on the set Gλ,L .
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
77
Using this, (7.4), and (8.9) we then get there is λ2 = λ2 (λ1 , L1 ) ∈ (0, λ1 ) such that for any λ ≤ λ2 and n ≤ N (L1 , λ), |R|n ≤ 1,
|S|n ≤ L + 1,
on the set Gλ,L .
(8.11)
Then, by (7.2) and (5.19), for n ≤ N (L1 , λ), |ξn | ≤ C(L + 1) + C
n−1
λ2 T |ξk |,
on the set Gλ,L .
(8.12)
k=0
By solving this inequality, we conclude that there is a c = c(C, L, τ ), independent of L1 , so that |ξn | ≤ c, for all n ≤ N (L1 , λ). By taking L1 ≥ c, we have N (L1 , λ) > nλ (τ ). Proposition 8.1 is proved. Setting Yn := Sn (2) −
n−1
An,k Sk (2)
(8.13)
k=0
from (7.4), (8.6), (8.10), and An ≤ C, we have, for each ε > 0,
lim P λ↓0
sup |S − AS + R − Y |n > ε
=0
(8.14)
n≤nλ (τ )
so that the analysis of the limiting behavior of the “known term” S − AS + R in (7.2) reduces to that of Y , which is the content of the following proposition. Proposition 8.2 Let Yλ (τ ), τ ∈ R+ , be the process obtained by linearly interpolating the values Yλ (λ2 Tn ) = Yn , n ∈ N. Then τ 3 b(s) λ↓0 Yλ (τ ) =⇒ b(τ ) − √ ds √ (8.15) τ −s 2π 0 where the convergence is in C(R+ ), and b(τ ) is a one dimensional Brownian motion with diffusion coefficient D = 3/4. Proof. We prove first that (S(2))λ (τ ) (obtained by linear interpolation from Sn (2)) converges weakly to a Brownian motion. We first prove tightness. Boundedness has already been proved, see (7.6). Since Sn (2) is a FTn –martingale, by Doob’s inequality, for each ε > 0, we have lim lim P δ↓0 λ↓0
sup
0≤σ<τ ≤T τ −σ≤δ
|(S(2))λ (τ ) − (S(2))λ (σ)| > ε
≤ lim lim δ↓0 λ↓0
C[nλ (δ) + 1]λ2 T =0. ε2
78
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
To identify the limit we compute the bracket of Snλ (τ ) (2) as follows nλ (τ )
k=1
λ (τ ) 9λ2 n 2 E (Sk (2) − Sk−1 (2)) FTk−1 = E m ¯ xk−1 , z (k−1) (Tk )2 FTk−1 16
k=1
nλ (τ ) 9λ T dx m ¯ (x)2 a(λβ (x + xk−1 ))2 16 2
=
k=1
which converges to Dτ on the set {x∗ (τ ) ≤ L} for any L > 0. From (8.6) and routine manipulation, (S(2))λ converges weakly to a Brownian motion with diffusion D by Levy’s characterization theorem. To complete the proof, we introduce the family of linear mappings Jλ : C(R+ ) → C(R+ ) defined by nλ (τ )−1
Jλ (ϕ)(τ ) :=
Anλ (τ )−1,k−1 ϕ(λ2 T k)
k=0
when (λ2 T )−1 τ is an integer and by linear interpolation otherwise. It is easy to verify that τ 3 ϕ(s) lim Jλ (ϕ)(τ ) = J(ϕ)(τ ) := √ ds √ λ↓0 τ −s 2π 0 uniformly for ϕ in a compact set. Since J is continuous, we can apply [6, Thm. 5.3] and get 3 Jλ ((S(2))λ )(τ ) =⇒ √ 2π λ↓0
τ 0
b(s) ds √ τ −s
in C(R+ ) .
Proposition 8.2 is proved.
Conclusion of the proof of Theorem 1.1. We first show that ξλ converges by subsequences to a continuous process, and then that any limit point solves the integral equation (1.12). By the uniqueness (in law) of the latter, Theorem 1.1 follows. The boundedness of ξλ follows from (8.5). In order to prove equicontinuity, we note that from (7.2) and (8.14) we have, for each ε > 0,
n−1 2 lim P sup ξn − Yn − (A )n,k ξk > ε = 0 . λ↓0 n≤nλ (τ )
(8.16)
k=0
Setting nλ (s)−1
Ξλ (s) :=
k=0
(A2 )nλ (s),k ξk
(8.17)
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
we postpone the proof that, for each ε > 0, lim lim P sup δ↓0 λ↓0
|Ξλ (s) − Ξλ (s )| > ε = 0
s,s ∈[0,τ ] |s−s |≤δ
79
(8.18)
and observe that, by (8.16), Proposition 8.2, and (8.18), for each ε > 0, lim lim P sup δ↓0 λ↓0
|ξλ (s) − ξλ (s )| > ε = 0
s,s ∈[0,τ ] |s−s |≤δ
(8.19)
which shows ξλ is tight in C(R+ ). It remains to prove (8.18). Let us fix 0 ≤ s ≤ s ≤ τ , s − s ≤ δ. We have nλ (s −δ)−1
Ξλ (s) − Ξλ (s ) =
2 (A )nλ (s),k − (A2 )nλ (s ),k ξk
k=0 nλ (s )−1
+
2 (A )nλ (s),k − (A2 )nλ (s ),k ξk
k=nλ (s −δ) nλ (s)−1
+
(A2 )nλ (s),k ξk
k=nλ (s )
so that
|Ξλ (s) − Ξλ (s )| ≤
∗
2
2 ξ (τ ) nλ (τ ) λ T + 3 ξ ∗ (τ ) nλ (δ)
sup τ1 ,τ2 ∈[0,τ ] τ1 −τ2 >δ
sup 0≤k
2 −1 2 (λ T ) (A )n (τ ),n (τ ) − λ 1 λ 2
9 2
2 (A )n,k
by using (5.19), (5.20), and (8.5) the equation (8.18) follows. We have concluded the proof that ξλ is tight. Finally, by (5.20), Proposition 8.2 and (8.16), we conclude that any limit ξ solves τ 9 τ 3 b(s) + ξ(τ ) = b(τ ) − √ ds ds ξ(s) . (8.20) 2π 0 2π(τ − s) 2 0 It is now easy to verify (1.12) and (8.20) are equivalent; indeed the latter is obtained by an iteration of the former.
80
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
Appendices A
Some technical estimates
We have pt (x, y) = pt (x − y, 0) and # " 2 2 ∂pt (x, 0) x 1 = √ − 1 e−x /2t . ∂t π(2t)3/2 t
(A.1)
Proof of (4.17). " # 2 ∂ (pt ϕ)(x) ≤ ϕ∞ 1 dy pt (y, 0) 1 + y = 1 ϕ∞ . ∂t 2t t t Proof of (7.19). ) ( ) ( 2 ∂pt ∂ pt ϕ =− m ¯ , ϕ m ¯ , ∂t ∂x∂t " # (x − y)3 1 3(x − y) −(x−y)2 /2t = dx m ¯ (x) dy √ − ϕ(y) e t2 t π(2t)3/2 " # (x − y)3 1 3(x − y) √ = 3/2 dx m − ¯ (x) dy pt (x, y) ϕ(y) 3/2 2t t t hence, by the Young and H¨ older inequalities, (7.19) follows.
B Gaussian estimates Proof of (7.9). From (1.3) and (3.9) we have, for t ∈ [Tn + 1, Tn+1 ] E z (n) (t, x)2 |xn =
t
Tn
ds
(n)
2
dy gt−s (x, y) a(λβ y)2 ≤ C
0
t−Tn
(n)
du g2u (x, x) .
(B.1) (n) Next, since sup{gu (x, y) : x, y ∈ R; 2u ≥ 1} < ∞ (see [7, Lemma A.9]), and using also (7.29) the desired estimate holds for p = 1. Since the process z (n) conditioned on xn is Gaussian, the estimate also holds for p = 2. Proof of (4.6). Let us denote by Pn the probability P conditioned on the center xn , and by En the corresponding expectation. From (B.1) and (7.29) to treat the case Tn ≤ t ≤ Tn + 1, it follows there exists C > 0 such that, for any n ≤ nλ (τ ), sup (B.2) sup En z (n) (t, x)2 ≤ CT . t∈[Tn ,Tn+1 ] x∈R
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
81
Proceeding as in the proof of [7, Thm. 2.3], (taking ε = 1 and µ(y) = a(λβ y)), it follows that given α ∈ (0, 1) there exists C such that, for any h, λ ∈ (0, 1], t ∈ [Tn , Tn+1 ], and x ∈ R, 2 z (n) (t, x + h) − z (n) (t, x) (B.3) ≤ CT hα . En It is not difficult to prove also H¨older continuity in t: 2 √ En z (n) (t + h, x) − z (n) (t, x) ≤ CT 2 λ−β h .
(B.4)
Indeed recall that, from (1.3), 2 (n) (n) z (t + h, x) − z (t, x) = En
t
ds Tn
h
+
2 (n) (n) dy gt−s+h (x, y) − gt−s (x, y) a(λβ y)2
ds 0
dygs(n) (x, y)2 a(λβ y)2 .
(B.5)
Using (7.29) it is easy to see that h √ ds dy gs(n) (x, y)2 a(λβ y)2 ≤ C h . 0
To bound the first term on the r.h.s. of (B.5), we use the following formula for (n) gt : t (n) ds dz pt−s (x, z)[1 − 3m ¯ 2xn (z)]gs(n) (z, y) . (B.6) gt (x, y) = pt (x, y) + 0
(n)
From the properties of the heat kernel pt , the estimate sup{gt (x, y); x, y ∈ R, t ≥ 1} < ∞, the fact that a has compact support and (7.29) for t ∈ (0, 1], one gets (B.4). Recall that zn conditioned on xn is Gaussian, and let us now define (B.7) sup sup En z (n) (t, x)2 σn2 = t∈[Tn ,Tn+1 ] x∈R
µn
=
En
sup t∈[Tn ,Tn+1 ]
z
(n)
(t)∞
.
(B.8)
Borell’s inequality (see for instance [1, Thm. 2.1]) tells us that, if µn < √ −ζ T λ , then
√ −ζ 2 √ T − µ ) (λ n z (n) (t)∞ > λ−ζ T ≤ 4 exp − sup . (B.9) Pn 2σn2 t∈[Tn ,Tn+1 ]
82
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
To estimate µn , we use (see [1, Cor. 4.15]), ∞ µn ≤ K dr log N (r),
Ann. Henri Poincar´e
(B.10)
0
where K is a universal constant and N (r) is the minimal number of balls of radius r needed to cover [Tn , Tn+1 ] × R when considering in that set the metric 3 2 . d (t, x), (s, y) = En z (n) (t, x) − z (n) (s, y) Let us then estimate N (r). Recall that, from (B.2), we only need to consider √ r ≤ CT , since for larger r, N (r) = 1 and the integrand in (B.10) equals zero. For each r > 0, consider the sets A1 = [Tn , Tn+1 ] × [−(λ2β r)−1 , (λ2β r)−1 ] and A2 = ([Tn , Tn+1 ] × R) \ A1 . Let us estimate first the number of balls needed to cover A2 . From now on, to avoid introducing new constants in the notation we suppose that the support of a is included in [−1, 1], and that√λ is sufficiently small. From (7.29), and since Tn+1 − Tn = T ≤ λ−β/20 and r ≤ CT , it is easy to see that if (t, x) ∈ A2 t ≤ En z (n) (t, x)2 ds dy e4(t−s) pt−s (x, y)2 a(λβ y)2 Tn 2 t (λ2β r)−1 − λ−β 1 4T ≤ Ce ds √ exp − 2(t − s) t−s Tn ≤
exp [−λ−β /(2r2 )] ≤
r2 . 4
(B.11)
From (B.11), it follows that we may cover A2 with just one ball. From (B.3) and (B.4), it follows that A1 may be covered by (λ7β r9 )−1 balls of radius r. Then, from (B.10) it follows that √CT 0 K dr log N (r) ≤ K dr log (1 + (λ7β r9 )−1 ) 0 0 ∞ ∞ log (1 + u9 ) 9 log u −7β/9 −7β/9 du ≤ Kλ du Kλ 2 −7β/9 −7β/9 λ√ λ√ u u2 CT CT √ C(β) T | log λ| (B.12)
µn
≤ ≤ ≤
∞
where we have done the change of variables u = λ−7β/9 r−1 , and C(β) is a constant that depends on β. Then, we may apply (B.9), to obtain
√ Pn z (n) (t)∞ > λ−ζ T ≤ 4 exp (−Cλ−ζ/2 ) . (B.13) sup t∈[Tn ,Tn+1 ]
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
Finally, recall that, given xn , the random variables sup t∈[Tn ,Tn+1 ]
83
z (n) (t)∞ ; n ≤ nλτ
are independent, and the estimate (B.13) does not depend on n (in particular, it does not depend on xn ). Then nλ (τ ) (1) P Bλ,τ ≥ 1 − 4 exp (−Cλ−ζ/2 )
(B.14)
and (4.6) follows. Proof of (4.35). The estimate (4.35) follows by the same procedure used to prove (4.6). We use the notation introduced in that proof. Recall that z
n,⊥
t
(t) = Tn
and define 2 σ(n,⊥)
=
sup
n,⊥ gt−s dW (a) (s)
(B.15)
sup En z (n,⊥) (t, x)2
(B.16)
t∈[Tn ,Tn+1 ] x∈R
µ(n,⊥)
En
=
sup t∈[Tn ,Tn+1 ]
z
(n,⊥)
(t)∞
.
(B.17)
Then, = En z (n,⊥)(t, x)2 ≤ =
2 3 (n) ¯ xn (x)m gt−s (x, y) − m ¯ xn (y) a(λβ y)2 4 Tn 2 ∞ 3 ¯ (x)m a∞ ds dy gt−s (x, y) − m ¯ (y) 4 0 ∞ 3 2 ¯ (x) ≤ C a∞ ds g2t (x, x) − m (B.18) 4 0
t
ds
dy
(see [8] for the last inequality). It is easy to see that the estimates (B.3) and (B.4) also hold for z (n,⊥) . Moreover, ¯ (x)2 (B.19) En z (n,⊥) (t, x)2 ≤ CEn z (n) (t, x)2 + Ctm (see (4.33)), hence from (B.11) and recalling that m ¯ is exponentially small as |x| → ∞, it follows that, for λ small enough, r2 En z (n,⊥) (t, x)2 ≤ 4
for (t, x) ∈ A2 .
(B.20)
84
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
Then N (r) ≤ 1 + (λ7β r9 )−1 , so proceeding as in (B.12), but now with T = 1, we obtain µn,⊥ ≤ C| log λ| (B.21) and the proof finishes as that of (4.6). Proof of (6.4). Note that, conditionally on the xk−1 ’s, the random variables {ηk (2); k ≤ nλ (τ )} are independent and Gaussian with variance bounded by CT . Therefore (6.4) is easily deduced from the following elementary lemma. LemmaB.1 Let {ωh , h = 1, . . . , N } be mean zero i.i.d. random variables and set n Yn := h=1 (n − h + 1)−1/2 ωh . Assume that for each p ∈ [1, ∞) we have E|ωi |p < ∞. Then for each ζ > 0 and q < ∞ there exists a constant C = C(ζ, q) such that ζ P sup |Yn | > N (B.22) ≤ CN −q . n≤N
k (n) Proof. Let Yk := h=1 (n − h + 1)−1/2 ωh , k = 1, . . . , n, which is a martingale with quadratic variation [Y (n) ]k =
k h=1
1 ω2 n−h+1 h
hence, by the BDG inequality (see [16, §6, E. 4.1]), for each p ∈ [1, ∞) there exists C = C(p) such that
p/2 1 2 [Y ]n ≤ CE ω E (|Yn | ) = E ≤ CE n−h+1 h h=1
p/2 n n−1 p/2 1 1 p 2/p (E|ωh | ) ≤C ≤C ≤ C (log n)p/2 . n−h+1 k+1
p
|Yn(n) |p
(n)
p/2
h=1
n
k=0
By using (7.17) we thus get P
sup |Yn | > N −ζ
n≤N
≤ N −ζp
N
p/2
E (|Yn |p ) ≤ CN 1−ζp (log N )
n=1
taking p large enough the lemma follows.
Acknowledgments. S. Brassesco acknowledges the very kind hospitality of the Department of Mathematics of the University of Rome Tor Vergata. Part of this work was done while P. Butt` a was visiting that Department with a Fellowship in Mathematics of the “Istituto Nazionale di Alta Matematica Francesco Severi”.
Vol. 3, 2002
Front Fluctuations in One Dimensional Stochastic Phase Field Equations
85
References [1] R. J. Adler, An Introduction to Continuity, Extrema and Related Topics for General Gaussian Processes, Lecture Notes Monograph series 12, Institute of Mathematical Statistics 1990. [2] L. Bertini, S. Brassesco, P. Butt` a and E. Presutti, Stochastic phase field equations: existence and uniqueness, Ann. Henri Poincar´e 3, 87–98 (2002). [3] L. Bertini, P. Butt` a, E. Presutti and E. Saada, Interface fluctuations in a conserved system: derivation and long time behavior, Preprint. [4] L. Bertini, P. Butt` a and B. R¨ udiger, Interface dynamics and Stefan problem from a microscopic conservative model, Rend. Mat. Appl. (7) 19, 547–581 (1999). [5] L. Bertini, E. Presutti, B. R¨ udiger and E. Saada, Dynamical fluctuations at the critical point: convergence to a nonlinear stochastic PDE, Teor. Veroyatnost. i Primenen 38, no. 4, 689–741 (1993); Translation in Theory Probab. Appl. 38, no. 4, 586–629 (1993). [6] P. Billingsley, Convergence of Probability Measures, New York: Wiley 1968. [7] S. Brassesco, Stability of the Instanton under small random perturbations, Stoch. Proc. Appl. 54, 309–330 (1994). [8] S. Brassesco and P. Butt`a, Interface fluctuations for the D=1 Stochastic Ginzburg–Landau equation with non–symmetric reaction term, J. Statist. Phys. 93, 1111–1142 (1998). [9] S. Brassesco, P. Butt` a, A. De Masi and E. Presutti, Interface fluctuations and couplings in the d = 1 Ginzburg–Landau equation with noise, J. Theoret. Probab. 11, 25–80 (1998). [10] S. Brassesco, A. De Masi and E. Presutti, Brownian fluctuations of the interface in the d = 1 Ginzburg–Landau equation with noise, Annal. Inst. H. Poincar´e 31, 81–118 (1995). [11] G. Caginalp and X. Chen, Convergence of the phase field model to its sharp interface limits, European J. Appl. Math. 9, 417–445 (1998). [12] A. De Masi, P. Ferrari and J. L. Lebowitz, Reaction–diffusion equations for interacting particle systems, J. Statist. Phys. 44, 589–644 (1986). [13] T. Funaki, The scaling limit for a stochastic PDE and the separation of phases, Prob. Theory Relat. Fields 102, 221–288 (1995) . [14] P. C. Hohenberg and B. I. Halperin, Theory of dynamic critical phenomena, Rev. Mod. Phys. 49, 435–479 (1977).
86
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
[15] P. L. Lions and P. E. Souganidis, Fully nonlinear stochastic partial differential equations: non–smooth equations and applications, C. R. Acad. Sci. Paris S´er. I Math. 327, 735–741 (1998). [16] M. M´etivier, Semimartingales: A Course on Stochastic Processes, Berlin: de Gruyter 1982. [17] H. M. Soner, Convergence of the phase–field equations to the Mullins–Sekerka problem with kinetic undercooling, Arch. Rational Mech. Anal. 131, 139–197 (1995). [18] F. G. Tricomi, Integral Equations, New York: Dover 1985. Lorenzo Bertini Dipartimento di Matematica Universit` a di Roma La Sapienza Piazzale Aldo Moro 2 00185 Roma Italy email:
[email protected] Stella Brassesco Departamento de Matem´aticas Instituto Venezolano de Investigaciones Cient´ıficas Apartado Postal 21827 Caracas 1020–A Venezuela email:
[email protected] Paolo Butt` a Dipartimento di Matematica Universit` a di Roma La Sapienza Piazzale Aldo Moro 2 00185 Roma Italy email:
[email protected] Errico Presutti Dipartimento di Matematica Universit` a di Roma Tor Vergata Via della Ricerca Scientifica 00133 Roma Italy email:
[email protected] Communicated by Jean-Pierre Eckmann submitted 11/01/01, accepted 13/06/01
Ann. Henri Poincar´e 3 (2002) 87 – 98 c Birkh¨ auser Verlag, Basel, 2002 1424-0637/02/010087-12 $ 1.50+0.20/0
Annales Henri Poincar´ e
Stochastic Phase Field Equations: Existence and Uniqueness L. Bertini∗, S. Brassesco, P. Butt` a† and E. Presutti†
Abstract. We consider a conservative system of stochastic PDE’s, namely a one dimensional phase field model perturbed by an additive space–time white noise. We prove a global existence and uniqueness result in a space of continuous functions on R+ × R. This result is obtained by extending previous results of Doering [3] on the stochastic Allen–Cahn equation.
1 Introduction and results We consider a phase field system with additive stochastic noise, which is formally written as 1 ∂t m(t) = ∆m(t) − V (m(t)) + λh(t) + aη(t) 2 (1.1) 1 ∂t [h(t) + m(t)] = ∆h(t) 2 where (t, x) ∈ R+ × R, m(t) = m(t, x), h(t) = h(t, x) are two scalar random fields (we omit to write explicitly the dependence on the randomness), λ is a positive parameter, ∆ is the Laplacian on R, V (m) = m4 /4 − m2 /2 is a double well potential, a = a(x) is a bounded and continuous function, and η(t) = η(t, x) is a space–time white noise, i.e. E (η(t, x) η(s, y)) = δ(t − s) δ(x − y). In particular a translationally covariant noise is obtained for a = 1. The deterministic system obtained by setting a = 0 in (1.1), usually referred to as phase field equations, describes the kinetic of phase segregation when the presence of the latent heat is taken into account. The first equation describes in fact the evolution of the order parameter m which is coupled to the external field h (which can be thought as the excess temperature measured from the melting temperature); h is however itself a dynamic variable which diffuses and, via the coupling λ, introduces a feedback into m whose effect is to slow down the phase segregation process. We stress that q = m + h is locally conserved as apparent from the second equation in (1.1). Scaling limits of the deterministic phase field equations as λ → 0 have been considered in [2] and [6]. ∗ Partially
† Partially
976552.
supported by Cofinanziamento MURST 1999. supported by Cofinanziamento MURST 1999 and by NATO Grant PST.CLG.
88
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
Due to the phenomenological character of the phase field model, the introduction of a random forcing term appears natural and makes possible to discuss statistical properties of solutions. We also mention that the stochastic system (1.1) is very similar to the so–called, in the physical literature on critical phenomena, model C of Hohenberg and Halperin [5]. The specific choice of an additive white noise only in the first equation has been done to keep the model as simple as possible: the random forcing term is non–conservative, whereas the conservation law is still linear and not perturbed. Moreover, for λ small, m and h are weakly coupled so that we may refer to (1.1) as a weakly conservative system. This simplifying feature might help in developing a mathematical theory for phase segregation in conservative models. Indeed, in the companion paper [1], front fluctuations for (1.1) are analyzed in a suitable scaling limit as λ → 0 and a = O(λ). The need of an existence and uniqueness result for the stochastic system (1.1) in [1] is the main motivation for the present paper. Referring to [1] for a more exhaustive discussion on the stochastic phase field equations, we next state precisely our results. For α > 0 and γ ∈ (0, 1], let us define the following norms on C(R): ϕCα (R)
:= sup e−α|x| |ϕ(x)|
ϕCαγ (R)
:= ϕCα (R) + sup e−α(|x|+|y|)
x∈R
x=y
|ϕ(x) − ϕ(y)| |x − y|γ
and also the following ones on C(R+ × R) f Cα (R+ ×R)
:=
f Cαγ (R+ ×R)
:=
2
sup e−α
t/2
f (t)Cα(R)
t/2
f (t)Cαγ (R)
t∈R+ 2
sup e−α t∈R+
we shall denote by Cα (R), . . . , the corresponding Banach spaces. Let pt = et∆/2 be the heat semigroup, namely the integral operator with kernel pt (x, y) = (2πt)−1/2 exp{−(x − y)2 /2t}. We introduce the process Z(t) = t Z(t, x) given by Z(t) = 0 ds pt−s [aη(s)]. Then Z is the mean zero Gaussian process with covariance t∧s E (Z(t, x)Z(s, y)) = Γ(t, s; x, y) := du dz pt−u (x−z)ps−u (y−z)a(z)2 (1.2) 0
where t ∧ s := min{t, s}. In the next lemma we state some properties of the paths of Z, which follow, by standard Gaussian arguments, from the properties of the covariance Γ, see e.g. [3] and references therein. Lemma 1.1 For each α > 0 and γ ∈ (0, 1/2) we have that Z ∈ Cαγ (R+ × R), P–a.s. We shall denote by Ft the filtration given by Ft := σ{Z(s, x); (s, x) ∈ [0, t] × R}. In the following we consider a fixed realization of Z ∈ Cαγ (R+ × R).
Vol. 3, 2002
Stochastic Phase Field Equations: Existence and Uniqueness
89
Let q := m + h and U (m) := V (m) − λm; we formulate the problem (1.1) in the integral form t ds pt−s [−U (m(s)) + λq(s)] + Z(t) m(t) = pt m(0) + 0
1 q(t) = pt q(0) − 2
0
(1.3) t
ds ∆pt−s m(s)
note that the integral on the r.h.s. of the second equation is well defined (the result is a continuous function) provided x → m(s, x) is H¨older continuous. Our main result is the following existence and uniqueness result for the system (1.3) on the space of H¨ older continuous functions. Theorem 1.2 Let m(0), q(0) ∈ Cαγ (R) for any α > 0 and γ ∈ (0, 1/2). Then there exists a unique Ft –adapted process (m, q) ∈ Cαγ (R+ × R) × Cαγ (R+ × R) for any α > 0 and γ ∈ (0, 1/2) which solves (1.3). ˜ on Cα (R+ × R) given by We introduce the linear operators G, G t G[F ] (t) := ds pt−s F (s) 0 t 1 ˜ ] (t) := G[F ds (t − s) ∆pt−s F (s) 2 0
(1.4)
and set R(t) := pt m(0)+ λt pt q(0)+ Z(t). By plugging the second equation in (1.3) into the first one we get that m solves the problem ˜ m = R − G[U (m)] − λ G[m] .
(1.5)
By standard estimates on the heat kernel, it is easy to verify that for each α > 0 and γ ∈ (0, 1] there exist an α > 0 and a constant C = C(α, γ) such that Gf Cαγ (R+ ×R) ≤ Cf Cα (R+ ×R) , ˜ C γ (R ×R) ≤ Cf C (R ×R) . Gf + + α α Furthermore, for each α > 0 and γ ∈ (0, 1), there exists an α > 0, γ ∈ (0, 1) and a constant C = C(α, γ) such that t ds ∆pt−s f (s) ≤ Cf C γ (R ×R) . 0
γ Cα (R+ ×R)
α
+
Therefore Theorem 1.2 can be easily deduced from the following existence and uniqueness result for the problem (1.5) on the space of continuous functions. Theorem 1.3 Let m(0), q(0) ∈ Cα (R) for any α > 0. Then there exists a unique Ft –adapted process m ∈ Cα (R+ × R) for any α > 0 which solves (1.5).
90
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
In the rest of the paper we prove Theorem 1.3. A uniqueness and existence result for equation 1.5 with λ = 0 is given in [3]. See also [4] for the analogous result in a bounded domain. We shall follow closely the proof in [3] for the one dimensional case, referring to that paper for some technical Lemmata. The term ˜ coming from λh in (1.1), is the source of the difficulties. Since we cannot λG, estimate the Lp norm of h in terms of the Lp norm m, the necessary a priori bounds are not a straightforward extension of that in [3]. To overcame this problem we shall estimate an appropriate negative Sobolev norm of h in terms of the Lp norm of m, see Lemmata 2.4 and 2.5 below. We finally remark that the quartic double well potential, V (m) = m4 /4 − 2 m /2 has been chosen only for notation simplicity; the proof works for any polynomial of even degree with positive leading coefficient. Since the parameter λ will be kept fixed throughout all the paper we omit to indicate the dependence on it. We shall denote by C a generic positive constant whose numerical value may change from line to line.
2 Finite volume approximations ∞ Let CK (R) be the space of infinitely differentiable functions on R with compact ∞ support and introduce L = {Λ ∈ CK (R) : 0 ≤ Λ ≤ 1}. For Λ ∈ L, we introduce the following finite volume approximations of the problem (1.5)
˜ mΛ = ΛR − G[ΛU (mΛ )] − λ G[Λm Λ] .
(2.1)
In this section we establish a global existence result for (2.1) together with some bounds uniform for Λ ∈ L. To this end we need to introduce some more notation. For α > 0, let %(x) := e−α|x| . We shall denote by % also the measure %(x)dx on R. We introduce the following finite measures on R+ × R, omitting the dependence on α > 0 from the notation 2
µ(dt, dx) µT (dt, dx)
:= :=
e−α t/2 %(x) dx dt χ[0,T ] (t) µ(dt, dx)
µΛ (dt, dx) µT,Λ (dt, dx)
:= :=
Λ(x) µ(dt, dx) χ[0,T ] (t) Λ(x) µ(dt, dx)
where χ[0,T ] denotes the characteristic function of [0, T ]. For ν a measure and f a function use the notation ν(f ) = dν f . For p ∈ (1, ∞), we introduce the Sobolev space H1p (%) obtained by completing ∞ CK (R) with respect to the norm ϕpH p () := ∇ϕpLp () + ϕpLp () 1
where ∇ denotes the derivative with respect to x. Since H1p (%) ⊂ Lp (%), for p, q ∈ q (1, ∞) such that p−1 + q −1 = 1 we introduce the dual space H−1 (%) by completing
Vol. 3, 2002
Stochastic Phase Field Equations: Existence and Uniqueness
91
Lq (%) with respect to the norm q +H−1 () :=
sup
%(+ ϕ) .
ϕ : ϕ H p () ≤1 1
For f = f (t, x) we shall use the notation f pH p
−1 (µT )
T
:= 0
2
dt e−α
t/2
f (t)pH p
−1 ()
omitting to write T on the l.h.s. if T = ∞. From Lemma 1.1 it follows that for each α > 0 and p ∈ [1, ∞) we have Z ∈ Lp (µ). Furthermore for each T > 0 and Λ ∈ L we have sup sup |Λ(x)Z(t, x)| < ∞.
t∈[0,T ] x∈R
Moreover, by the assumptions on the initial data, the same holds for R(t) = pt m(0) + λt pt q(0) + Z(t). ˜ has a We next state a local existence result for the problem (2.1). Since G kernel which can be estimated as the one of G, the proof of the next result, which is based on Picard iterations, is the same as [3, Prop. 1] and we omit it. Lemma 2.1 For each Λ ∈ L and each realization of Z in Cα (R+ × R) there is a time T0 > 0 such that there exists a unique Ft –adapted continuous and bounded solution of (2.1) on [0, T0 ] × R. To show the existence of a global solution, it is enough to prove that if mΛ is a continuous solution of (2.1) on [0, T ∗) × R then sup sup |mΛ (t, x)| < ∞ .
(2.2)
t∈[0,T ∗ ) x∈R
The key ingredient for proving (2.2) is the following a priori bound on the Lp (µT ∗ ,Λ ) norm of solutions. Proposition 2.2 Let mΛ be a continuous solution of (2.1) on [0, T ∗ ) × R. For each α > 0 and p ∈ [1, ∞) there exists a constant C = C α, p, RLp(µ) < ∞, independent of T ∗ > 0 and of Λ ∈ L, such that mΛ Lp (µT ∗ ,Λ ) ≤ C .
(2.3)
The proof of the proposition is split in several Lemmata, the first one, which is proven integrating by parts, is [3, Lemma 7].
92
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
Lemma 2.3 If both f (t, x) and (∂t − 12 ∆)f are continuous on (0, T ) × R, f (0) = 0 and |f |2n+2 and |∇f |2n+2 are in L1 (µT ), then, for any n = 0, 1, . . . , 1 2n+1 µT f ∂t − ∆ f ≥ (2n + 1)µT f 2n |∇f |2 . 2
(2.4)
Let mΛ be a continuous solution of (2.1). We define uΛ := mΛ − ΛR, 1 ˜ qΛ := ∂t − ∆ G[Λm Λ] . 2
(2.5)
From (2.1), if mΛ is continuous and bounded on [0, T ] × R, then uΛ ∈ C ∞ ˜ Moreover, for t ∈ [0, T ], ((0, T ] × R) by the regularizing properties of G and G. uΛ (t, x) (together with its derivative) is exponentially decaying as x → ∞. Lemma 2.4 Let mΛ be a continuous solution of (2.1) on [0, T ∗ ) × R; then for each n = 0, 1, . . . and β > 0 there exists a constant C = C(n, β), independent of T ∗ > 0 and Λ ∈ L, such that 2n+2 µT ∗ ,Λ e−(2n+2)βt u2n+1 U (mΛ ) ≤ Ce−βt qΛ H 2n+2 Λ (µ −1
T∗)
.
(2.6)
Proof. We apply Lemma 2.3 with f = e−βt uΛ and T < T ∗ . From (2.1) we get 1 ∂t − ∆ e−βt uΛ = −βe−βt uΛ − e−βt [ΛU (mΛ ) + λqΛ ] 2 hence, by (2.4)
2n+2 2n+1 2 µT e−(2n+2)βt (2n + 1)u2n + uΛ ΛU (mΛ ) Λ (∇uΛ ) + βuΛ 2n+1 (2.7) qΛ . ≤ −λµT e−(2n+2)βt uΛ Let p, q ∈ (1, ∞) such that p−1 + q −1 = 1; then for each γ > 0 there exists a constant C = C(γ, p) such that for any a, b ∈ R |a b| ≤ γ|a|p + C|b|q
(2.8)
q (%), we have therefore, by the duality between H1p (%) and H−1
|%(f g)| ≤ γf pH p () + CgqH q 1
−1 ()
.
(2.9)
Vol. 3, 2002
Stochastic Phase Field Equations: Existence and Uniqueness
93
By applying (2.9) we get T 2n+1 2 q dt e−α t/2−(2n+2)βt % uΛ (t)qΛ (t) µT e−(2n+2)βt u2n+1 Λ ≤ Λ 0 T
2n+2 2n+2 2n+2 2 ≤ dt e−α t/2−(2n+2)βt γ(2n + 1) 2n+1 % uΛ (t)2n 2n+1 (∇uΛ (t)) 2n+1 0 2n+2 +γ% u2n+2 (t) + Cq (t) Λ Λ H 2n+2 ()
T
2
2n+2
−1
2
γ(2n + 1) 2n+1 c1 (n)% uΛ (t)2n (∇uΛ (t)) + uΛ (t)2n+2 0 2n+2 2n+2 +γ% uΛ (t) + CqΛ (t)H 2n+2 () −1 2n+2 2 = γ(2n + 1) 2n+1 c1 (n)µT e−(2n+2)βt u2n Λ (∇uΛ )
2n+2 2n+2 + Ce−(2n+2)βt qΛ 2n+2 + γ (2n + 1) 2n+1 c1 (n) + 1 µT e−(2n+2)βt uΛ H 2n+2 (µ ) ≤
dt e−α
t/2−(2n+2)βt
−1
T
(2.10) where we used, in the third step, H¨ older inequality in the form (as follows from (2.8)) 2n+1
2n+1 %(f g) ≤ c1 (n) % |f | n+1 + % |g| n 2n+2
2n+2
with f = [uΛ (t)n ∇uΛ (t)] 2n+1 and g = uΛ (t)n 2n+1 . Choosing γ = γ(n, λ, β) small enough and taking T ↑ T ∗ the lemma now follows from (2.7) and (2.10). Lemma 2.5 For each p ∈ (1, ∞) there exists a constant C = C(p) such that, for any α, β, T > 0 and F ∈ Lp (µT ), −βt 1 ˜ 1 e e−βt F p G[F ] ∆ − ≤ C 1 + . (2.11) ∂ t L (µT ) 2 α H p (µT ) −1
Proof. We can write
˜ ](t) = G[F 0
t
ds pt−s
s
0
1 ds ∆ps−s F (s ) 2
(2.12)
so that (in distribution sense)
1 ˜ ] (t) = ∂t − ∆ G[F 2 We thus get ∂t − 1 ∆ G[F ˜ ] (t) p 2 H
−1 ()
1 ≤ 2
t
0
0
t
1 ds ∆pt−s F (s) . 2
ds ∆pt−s F (s)H p
−1 ()
(2.13)
.
(2.14)
94
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
∞ For q −1 + p−1 = 1 and ϕ ∈ CK (R) we now have
(ϕ, ∆pt−s F (s)) = (%ϕ, ∆pt−s F (s))L2 (dx) ϕ ∇% = − % ∇ϕ + , ∇pt−s F (s) % L2 (dx) ≤ ∇ϕLq () + ϕLq () sup |∇ log %(x)| ∇pt−s F (s)Lp ()
(2.15)
x∈R
since supx∈R |∇ log %(x)| = α we conclude ∆pt−s F (s)H p
−1 ()
≤ (1 + α) ∇pt−s F (s)Lp () .
(2.16)
As exp{−α|x|} ≤ exp{−α|y|} exp{α|x − y|}, −α|x|/p ≤ dy e−α|y|/p |F (s, y)| |∇pt−s (x − y)| eα|x−y|/p , (2.17) |∇pt−s F (s) (x)| e by applying Young’s inequality for convolutions we then obtain ∇pt−s F (s)Lp () ≤ F (s)Lp () dx |∇pt−s (x)| eα|x|/p . We define ψα,p (t) :=
dx |∇pt (x)| e
α|x|/p
2 2α exp ≤√ + p 2πt
α2 t 2p2
(2.18)
.
(2.19)
Then, from (2.14), (2.16) and (2.18), 2 1 ˜ − α2pt −βt χ[0,T ] (t) e ∂t − 2 ∆ G[F ] (t) p H−1 () 2 α s 1+α ≤ ds χ[0,T ] (s)e− 2p −βs F (s)Lp () χ[0,T ] (t − s) 2 × e−
α2 (t−s) −β(t−s) 2p
ψα,p (t − s) .
Again by Young’s inequality for convolutions we get −βt 1 ˜ e ∂t − ∆ G[F ] = p 2 H−1 (µT ) 1/p p T 2 1 ˜ − α2 t −βpt dte ∂t − 2 ∆ G[F ](t) p 0 H−1 () 1/p T T 2 2 1+α p − α2pt − α2 t −βpt ≤ dt e ψα,p (t) dt e F (t)Lp () . 2 0 0
Vol. 3, 2002
Stochastic Phase Field Equations: Existence and Uniqueness
95
By using the estimate in (2.19) and recalling p > 1, it is easy to show there is a constant C = C(p) > 0 so that α2 t 1+α T 1 dt e− 2p ψα,p (t) ≤ C 1 + . 2 α 0
The lemma is proved.
3
Proof of Proposition 2.2. Recalling (2.5) and that U (m) = m − (1 − λ)m, by expanding the l.h.s. of (2.6), using Lemma 2.5, and Λ2 ≤ Λ, we get there exists a constant C = C(α, n, β) such that µT ∗ ,Λ e−(2n+2)βt u2n+4 Λ
≤ CµT ∗ ,Λ e−(2n+2)βt |uΛ |2n+1 u2Λ |ΛR| + |uΛ |(ΛR)2 + |ΛR|3 2n+2 . (2.20) + |uΛ |2n+1 |ΛR| + |ΛR|2n+2 +uΛ Let
1/(2n+4) 2n+4 M := µT ∗ ,Λ e−(2n+2)βt uΛ
by using repeatedly H¨ older inequality in (2.20) we get
M 2n+4 ≤ C M 2n+3 ΛRL2n+4(µ) + M 2n+2 1 + ΛR2L2n+4(µ)
+ M 2n+1 ΛRL2n+4(µ) 1 + ΛR2L2n+4(µ) + ΛR2n+2 L2n+4(µ) we then conclude that M is bounded by some constant C = C(α, n, β, ΛRL2n+4(µ) ). Recalling mΛ = uΛ + ΛR, by using triangular and Cauchy–Schwartz inequalities, we have mΛ Lp (µΛ,T ∗ ) ≤ ΛRLp(µΛ,T ∗ ) + eβt L2p (µΛ,T ∗ ) e−βt uΛ L2p (µΛ,T ∗ ) . Since, for each p ∈ [1, ∞), ΛRLp(µΛ,T ∗ ) ≤ RLp(µ) < ∞, by choosing β = β(α, p) small enough and n = n(p) large enough, the proposition follows. Proof of (2.2). Let Y := sup{|x| : Λ(x) > 0}; writing explicitly the kernels in the integral equation (2.1), and using Young’s inequality for convolutions, it follows sup sup |mΛ (t, x)| ≤
t∈[0.T ∗ ) x∈R
+e
α2 4
T ∗+ α 2Y
×
T∗
+λ
dt 0
sup sup |Λ(x)R(t, x)|
t∈[0,T ∗ ) x∈R
T∗
dt 0
2
12
dz pt (z)
dz pt (z)2
2
z 1 + 2t 2
2
µT ∗ ,Λ (dt, dx) U (mΛ (t, y))
2 12
µT ∗ ,Λ (dt, dx) mΛ (t, y)2
12
12 (2.21)
96
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
where we used also that Λ2 ≤ Λ. By using that, as follows from Proposition 2.2, U (mΛ ) is in L2 (µT ∗ ,Λ ), and the properties of the heat kernel, it is easy to see that the r.h.s. of (2.21) is bounded. The following lemma states an a priori uniform bound for the solutions of (2.1), which by (2.2) are defined for any t ≥ 0. Lemma 2.6 Let mΛ be a continuous solution of(2.1) on R+ × R. Then for each α > 0, p ∈ [1, ∞) there exists a constant C = C α, p, RLp(µ) < ∞ independent of Λ ∈ L such that mΛ Lp (µ) ≤ C. Proof. Since the continuous solution of equation (2.1) exist globally in time, the inequality (2.3) may be extended for T ∗ ↑ ∞; we get ΛmΛ Lp (µ) ≤ mΛ Lp (µΛ ) ≤ C < ∞ .
(2.22)
˜ are bounded operators in Lp (µ) (see [3, Note that for p > 1 both G and G ˜ From the integral equation Lemma 9] for G and the same proof also works for G). (2.1), the bound (2.22), and R ∈ Lp (µ) the lemma follows.
3 Infinite volume equation In this section we conclude the proof of Theorem 1.3 by removing the truncation Λ in (2.1); following [3], we prove first that, provided β is chosen large enough, exp (−βt) Λ mΛ converges in Lp (µ). Lemma 3.1 Let mΛ be a solution of (2.1). For each p ∈ [1, ∞) there exists a positive constant kp < ∞ such that, for any α > 0 and β ≥ kp (1 + α−p ), {exp (−βt) Λ mΛ , Λ ∈ L} is Cauchy in Lp (µ) as Λ ↑ 1. Proof. Recalling (2.5), if we consider f = e−βt [uΛ − uΛ¯ ], from equation (2.1),we have
1 ¯ (mΛ¯ ) − λe−βt [qΛ − qΛ¯ ] . ∂t − ∆ f = −βf − e−βt ΛU (mΛ ) − ΛU 2 By the same computations as in Lemma 2.4 with f as above, (see equations (2.7) and (2.10)), but choosing γ = γ(n, λ) independent of β and taking T ↑ ∞, there exists a constant c1 = c1 (n) independent of α, β such that 2n+1
¯ (mΛ¯ ) µ e−(2n+2)βt [uΛ − uΛ¯ ] ΛU (mΛ ) − ΛU . + (β − 1)µ e−(2n+2)βt [uΛ − uΛ¯ ]2n+2 ≤ c1 e−βt (qΛ − qΛ¯ ) 2n+2 H 2n+2 (µ) −1
(3.1)
Vol. 3, 2002
Stochastic Phase Field Equations: Existence and Uniqueness
97
By using Lemma 2.5 there is a constant c2 = c2 (n) such that e−βt (qΛ − qΛ¯ ) 2n+2 2n+2 H−1 (µ)
¯ Λ¯ 2n+2 . (3.2) ≤ c2 1 + α−(2n+2) µ e−(2n+2)βt ΛmΛ − Λm On the other hand, since for any a, b ∈ R we have (a3 − b3 )(a − b) ≥ (a − b)4 /4, there exists a constant c3 = c3 (λ) such that
¯ (mΛ¯ ) (uΛ − uΛ¯ ) ΛU (mΛ ) − ΛU ¯ ¯ 3¯ − (1 − λ) (uΛ − uΛ¯ ) ΛmΛ − Λm ¯ Λ¯ = mΛ − mΛ¯ − (Λ − Λ)R Λm3Λ − Λm Λ
1 ¯ Λ¯ 4 − c3 (uΛ − uΛ¯ )2 ΛmΛ − Λm ≥ 4
¯ m2 (1 + m2 ) + m2¯ (1 + m2¯ ) + R2 (1 + R2 ) − c3 |1 − Λ| + |1 − Λ| Λ Λ Λ Λ
(3.3)
By plugging the bounds (3.2) and (3.3) into (3.1) and using H¨ older inequality, we find there exists a c4 = c4 (n, α, λ) such that 1 −(2n+2)βt ¯ Λ¯ 2n+4 µ e ΛmΛ − Λm 4 2n+2 + β − 1 − c3 − c1 c2 1 + α−(2n+2) µ e−(2n+2)βt (uΛ − uΛ¯ ) ¯ L2 (µ) ≤ c4 1 − ΛL2 (µ) + 1 − Λ 2n+4 2n+4 2n+4 × 1 + mΛ L4n+8 (µ) + mΛ¯ L4n+8 (µ) + RL4n+8 (µ) . Given p ≥ 2, let n = n(p) = [p/2] − 1 and kp such that 1 + c3 + c1 c2 (1 + α−(2n+2) ) ≤ kp (1 + α−p ) for any α > 0. The lemma now follows from Lemma 2.6 and R ∈ Lp (µ) for any p ∈ [1, ∞). The proof of Theorem 1.3 can now be completed as in [3], we shall just sketch the argument. Proof of Theorem 1.3. To prove existence of a continuous solution of (1.5), we first note that, for each p ∈ [1, ∞) and α so large that α2 > kp (1 + α−p ), mΛ is Cauchy in Lp (µ). This follows from Lemmata 2.6, 3.1, and H¨older inequality; moreover the ˜ map L2 (µ) into Cα (R+ × R), see limit m satisfies equation (1.5). Since G and G [3, Lemma 12], we also have m ∈ Cα (R+ × R). In order to show m ∈ Cα (R+ × R) for any α > 0, we note that, by Lemma 2.6, mΛ is uniformly bounded in Lp (µ); we can thus find a weakly convergent subsequence mΛk → m . On the other hand, by Lemma 3.1, e−βt Λk mΛk converges strongly in Lp (µ) for β ≥ kp (1 + α−p ), hence m = m µ–a.s. Since m ∈ Lp (µ) for any α > 0 and p ∈ [1, ∞), by the same argument as above, we get m ∈ Cα (R+ ×R) and m = m . To prove uniqueness, let m1 and m2 be two continuous solutions of (1.5). By applying Lemma 2.3 to the function f = e−βt [m1 − m2 ] and repeating the same computations as in Lemma 3.1 it is easy to show m1 = m2 .
98
L. Bertini, S. Brassesco, P. Butt` a and E. Presutti
Ann. Henri Poincar´e
Acknowledgments. L. Bertini acknowledges the very kind hospitality of Departamento de Matem´aticas, Instituto Venezolano de Investigaciones Cient´ıficas.
References [1] L. Bertini, S. Brassesco, P. Butt` a and E. Presutti, Front fluctuations in one dimensional stochastic phase field equations, Ann. Henri Poincar´e 3, 29–86 (2002). [2] G. Caginalp and X. Chen, Convergence of the phase field model to its sharp interface limits, European J. Appl. Math. 9, 417–445 (1998). [3] C. R. Doering, Nonlinear parabolic stochastic differential equations with additive colored noise on Rd × R+ : a regulated stochastic quantization, Comm. Math. Phys. 109, 537–561 (1987). [4] W. G. Faris and G. Jona–Lasinio, Large fluctuations for a nonlinear heat equation with noise, J. Phys. A 15, 3025–3055 (1982). [5] P. C. Hohenberg and B. I. Halperin, Theory of dynamic critical phenomena, Rev. Mod. Phys. 49, 435–479 (1977). [6] H. M. Soner, Convergence of the phase–field equations to the Mullins–Sekerka problem with kinetic undercooling, Arch. Rational Mech. Anal. 131, 139–197 (1995). Lorenzo Bertini Dipartimento di Matematica Universit` a di Roma La Sapienza Piazzale Aldo Moro 2 00185 Roma Italy email:
[email protected]
Stella Brassesco Departamento de Matem´ aticas Instituto Venezolano de Investigaciones Cient´ıficas Apartado Postal 21827 Caracas 1020–A Venezuela email:
[email protected]
Paolo Butt` a Dipartimento di Matematica Universit` a di Roma La Sapienza Piazzale Aldo Moro 2 00185 Roma Italy email:
[email protected]
Errico Presutti Dipartimento di Matematica Universit` a di Roma Tor Vergata Via della Ricerca Scientifica 00133 Roma Italy email:
[email protected]
Communicated by Jean-Pierre Eckmann submitted 30/01/01, accepted 13/06/01
Ann. Henri Poincar´e 3 (2002) 99 – 105 c Birkh¨ auser Verlag, Basel, 2002 1424-0637/02/010099-7 $ 1.50+0.20/0
Annales Henri Poincar´ e
Some High-Complexity Hamiltonians with Purely Singular Continuous Spectrum D. Damanik and B. Solomyak Abstract. We give examples of discrete one-dimensional Schr¨ odinger operators with potential having high combinatorial complexity that have purely singular continuous spectrum. This is accomplished by constructing minimal, palindromic subshifts with entropy arbitrarily close to maximal entropy. For a dense Gδ set of elements in such a subshift, purely singular continuous spectrum can then be established using a criterion of Hof, Knill, and Simon.
1 Introduction In this paper we discuss discrete one-dimensional Schr¨odinger operators (Hφ)(n) = φ(n + 1) + φ(n − 1) + V (n)φ(n)
(1)
in 2 (Z) with potential V taking finitely many values. It is known that for these operators, any spectral type can occur. Concretely, periodic potentials lead to purely absolutely continuous spectrum [12], Sturmian potentials (to be defined below) lead to purely singular continuous spectrum [4], and random potentials lead almost surely to pure point spectrum [2]. A point of view that has been proposed and investigated recently [3] is to link spectral properties of the operators to combinatorial properties of the potentials, particularly their combinatorial complexity. Apparently, the spectra become more and more singular as combinatorial complexity is increased. For example, referring to the above classes, periodic potentials have minimal complexity, Sturmian potentials have minimal complexity among the aperiodic potentials, and, on the other extreme, random potentials have (almost surely) maximal complexity. For potentials of intermediate complexity (i.e., potentials intermediate between Sturmian and random), there are almost no definite results on the operator side, in particular for potentials with positive entropy. Our purpose here therefore is to exhibit concrete examples of potentials with high combinatorial complexity that lead to purely singular continuous spectrum. This indicates that a possible correspondence between complexity and spectral type might be quite complicated and we discuss this issue at the end of the paper. Let us be more specific. Let A ⊆ R be a finite set and consider some V ∈ AZ . A finite string w = a1 . . . ak of elements of A is called a word and its length is |w| = k. A word a1 . . . ak is called a subword of V if there is some m such that
100
D. Damanik and B. Solomyak
Ann. Henri Poincar´e
V (m + i) = ai , 1 ≤ i ≤ k. Let f (n) denote the number of different subwords of V having length n. The function f : N → N is called the complexity function of V . A classical result by Morse and Hedlund states that f is bounded if and only if V is periodic [7]. Moreover, any unbounded f satisfies the universal lower bound f (n) ≥ n + 1 for every n. There exist V ’s with f (n) = n + 1 for every n and these sequences are called Sturmian (if they satisfy the additional requirement that every subword occurs infinitely often). Sturmian sequences are therefore of minimal complexity among the aperiodic sequences. On the other hand, if we generate a random sequence in AZ (by i.i.d. random variables giving non-zero weight to every element of A), we get f (n) = |A|n with probability one. One defines the entropy of V by h(V ) = limn→∞ n1 ln f (n). By a subadditivity argument one can show that the limit exists. The maximal entropy in AZ is therefore ln |A|. Our goal is the following: Given A with |A| ≥ 2 and any number h < ln |A|, we will construct a potential V with entropy h(V ) > h such that the operator (1) has purely singular continuous spectrum. The organization is as follows. In Section 2 we use a construction that is similar to the one in Furstenberg [5] which gives a minimal, palindromic subshift Ω ⊂ AZ with high entropy. We then show in Section 3 that for an uncountable set of V ’s in Ω (which, in fact, is a dense Gδ ), the operator (1) has purely singular continuous spectrum. This will follow easily from Section 2, Hof, Knill and Simon [8], Kotani [10], and Last and Simon [11].
2 Construction of minimal, palindromic subshifts with large entropy In this section we construct subshifts Ω ⊆ AZ which are minimal, contain arbitrarily long palindromes, and have large entropy. Our approach is similar to the one used by Furstenberg in [5] who constructs minimal subshifts with positive entropy by generating suitable Toeplitz sequences. Let us put the topology of pointwise convergence on AZ and define the shift transformation T : AZ → AZ by T x(n) = x(n + 1). Any closed, T -invariant subset Ω ⊆ AZ is called a subshift. A subshift Ω is called minimal if the T -orbit of every V ∈ Ω is dense in Ω. It is relatively easy to see that if Ω is minimal, every V ∈ Ω has the same set of subwords. In this case, one may associate a complexity function f : N → N with the subshift because the complexity function is the same for each one of its elements. Similarly, we can then define the entropy of a minimal subshift as above by taking h(Ω) = limn→∞ n1 ln f (n). We shall be interested in minimal subshifts that contain palindromes. A palindrome is a word w = a1 . . . ak that is invariant under reversal, that is, w = ak . . . a1 . We define the palindrome complexity function p : N → N0 := N ∪ {0} for a sequence (resp., a minimal subshift) by p(n) = number of palindromes of length n occurring in V (resp., the elements of the subshift). For general background on palindrome complexity, we refer the reader to [1] (and references therein).
Vol. 3, 2002
High-complexity Hamiltonians
101
We call a minimal subshift palindromic if it contains arbitrarily long palindromes (equivalently, lim supn→∞ p(n) > 0). This notion was introduced by Hof, Knill and Simon [8] who also established a connection between palindromicity and spectral theory. Our main result is the following: Theorem 1 Given any A ⊆ R with k elements, where k ≥ 2, and any number h < ln k, there exists a minimal, palindromic subshift Ω ⊆ AZ with entropy larger than h. Proof. The strategy is as follows. We construct a Toeplitz sequence T by starting with the “empty sequence” and successively filling the “holes” with periodic sequences (also having holes) of rapidly increasing periods. This ensures that the limit sequence, provided that it exists, is uniformly recurrent, that is, every subword occurs infinitely often and with bounded gaps. This implies that its T -orbit closure in AZ is a minimal subshift, see, e.g., [6]. We therefore have to take care of two issues: In the generation process, we have to introduce sufficiently many palindromes (to get palindromicity) and sufficiently many subwords (to get large entropy). The first issue is dealt with by “symmetrizing” Furstenberg’s construction [5], the second issue only requires a slightly more elaborate line of argument. We first decompose Z into a suitable disjoint union of arithmetic progressions: Let c0 = 0 and let {dn } be an increasing sequence of positive integers satisfying dn |dn+1 for every n. Let J0 = (c0 + d0 Z), and then for every n ≥ 0, n cn+1 = min m ≥ 0 : m ∈ Ji i=0
and Jn+1 = (cn+1 + dn+1 Z) ∪ (−cn+1 + dn+1 Z). It is clear that Z is the disjoint union of the sets Jn , n ≥ 0. We will choose a sequence {sn } ∈ AN0 and then define the sequence V ∈ AZ by V (m) = sn if m ∈ Jn . From the construction it follows that V is uniformly recurrent and hence we obtain a minimal subshift Ω by taking its T -orbit closure. It also follows that V contains arbitrarily long palindromes since it is symmetric about the origin: V (m) = V (−m) for every m. Thus, Ω is palindromic. To arrange for entropy larger than h, we have to choose the sequences {dn } and {sn } suitably. Given any ε > 0, we can choose a rapidly increasing sequence {dn } sufficient to ensure cn < 1 + ε. (2) lim sup n→∞ n ε −1 Indeed, we can make sure that ∞ m=0 dm < 4 . Since [−cn + 1, cn − 1] ∩ Z ⊂
n−1 m=0
Jm
102
D. Damanik and B. Solomyak
and #(Jm ∩ [−cn + 1, cn − 1]) ≤ 2
we obtain 2cn − 1 < 2
Ann. Henri Poincar´e
2cn − 1 +1 , dm
(2cn − 1)ε +n 4
which implies (2) for ε > 0 sufficiently small. For every m, (2) implies that there are infinitely many n such that cn+m − cn+1 < (1 + ε)m.
(3)
From this point, we can proceed very similarly to Furstenberg. For the reader’s convenience, we provide a sketch of the subsequent arguments. Let Qm = {n ∈ N : (3) holds }. It is possible to choose a sequence {sn } such that for every m and every word a1 . . . am ∈ Am , there exists r ∈ Qm such that sr+1 . . . sr+m = a1 . . . am . Furstenberg [5] used that this holds for a random sequence in AZ . There is also a constructive way: We can just inductively define the sequence {sn } by prescribing the places in Qm for all the words in Am so that no overlaps occur, neither between words on the same m-level nor for words of different levels. It helps of course that we have in each step only finitely many words, so that there is always enough (infinite) space left to work with. With this sequence {sn } we define V as described above. We get that for every word a1 . . . am ∈ Am , there exists r such that cr+m −cr+1 < (1+ε)m and V (cr+j ) = aj , 1 ≤ j ≤ m. Thus there exists n such that V (n+1)V (n+2) . . . V (n+(1+ε)m) contains the symbols a1 , . . . , am (possibly with gaps) in precisely that order. Since every fixed block V (n + 1)V (n + 2) . . . V (n + (1 + ε)m) can contain at most (1+ε)m ordered m-tuples, we deduce that for every m, V must contain at least m |A|m (1+ε)m m
subwords of length (1 + ε)m. It follows from Stirling’s formula that, for λ > 1, m λm λλ ≤ const . m (λ − 1)λ−1 Thus, choosing ε sufficiently small, we can conclude the proof.
3 Application to Schr¨ odinger operators Given a minimal subshift (Ω, T ), certain combinatorial properties of Ω are linked to spectral properties of (1) with V ∈ Ω. Recall that f (n) counts the number of subwords of length n that occur in elements of Ω and p(n) counts the number of palindromes of length n that occur in elements of Ω.
Vol. 3, 2002
High-complexity Hamiltonians
103
Proposition 3.1 (Absence of absolutely continuous spectrum) If the complexity function obeys lim supn→∞ f (n) = ∞, then for every V ∈ Ω, the operator H in (1) has empty absolutely continuous spectrum. Remark. If lim supn→∞ f (n) < ∞, then every V ∈ Ω is periodic and hence the operator H has purely absolutely continuous spectrum. Proof. It follows from the assumption and Morse-Hedlund that every V ∈ Ω is aperiodic. Moreover, since (Ω, T ) must have a T -invariant measure, it follows that it also has an ergodic measure µ. Kotani [10] then implies that for µ-almost every V ∈ Ω, the operator H has empty absolutely continuous spectrum. Finally, it was shown by Last-Simon [11] that minimality implies independence of σac (H) of V ∈ Ω. Putting this together, we get the assertion. Proposition 3.2 (Absence of point spectrum) If the palindrome complexity function obeys lim supn→∞ p(n) > 0, then for a dense Gδ -set of V ’s in Ω, the operator H in (1) has empty point spectrum. Proof. This was shown by Hof, Knill and Simon in [8]. We remark that it is possible to construct uncountably many elements V ∈ Ω for which (1) has empty point spectrum. Combining Theorem 1, Proposition 3.1, and Proposition 3.2, we obtain the following theorem: Theorem 2 Given any A ⊆ R with k elements, where k ≥ 2, and any number h < ln k, there exists a closed, T -invariant set Ω ⊆ AZ such that 1. (Ω, T ) is minimal, 2. (Ω, T ) has entropy larger than h, 3. for a dense Gδ set of V ∈ Ω, the operator H in (1) has purely singular continuous spectrum. Let us summarize our results: For each finite alphabet with at least two elements, it is possible to construct uniformly recurrent two-sided sequences with entropy arbitrarily close to the maximal possible value such that the associated Schr¨ odinger operator has purely singular continuous spectrum. This phenomenon should be seen in the general context of the interplay between combinatorial complexity and spectral properties of Schr¨ odinger operators. It turns out that sufficiently high complexity along with uniform recurrence does not necessarily imply pure point spectrum. One may ask, however, the following: Given a uniformly recurrent sequence V , consider the associated subshift Ω and some ergodic measure µ on Ω. Is it true then that if V has high complexity, µalmost every element of Ω generates an operator with pure point spectrum? If this turns out to be true, one will have an interesting result as a byproduct – the
104
D. Damanik and B. Solomyak
Ann. Henri Poincar´e
spectral type for potentials from some minimal subshift over a finite alphabet need not be constant! Such a non-constancy result is known in the uniformly almost periodic case; see Jitomirskaya and Simon [9] who prove this for the family of almost Mathieu operators. However, in the finite alphabet case such an example has not yet been found.
References [1] J.-P. Allouche, M. Baake, J. Cassaigne and D. Damanik, Palindrome complexity, preprint (2000), to appear in Theoret. Comput. Sci. [2] R. Carmona, A. Klein and F. Martinelli, Anderson localization for Bernoulli and other singular potentials, Commun. Math. Phys. 108 (1987), 41–66 [3] D. Damanik, Gordon-type arguments in the spectral theory of one-dimensional quasicrystals, in Directions in Mathematical Quasicrystals, Eds. M. Baake and R. V. Moody, CRM Monograph Series 13, AMS, Providence, RI (2000), 277–304 [4] D. Damanik, R. Killip, and D. Lenz, Uniform spectral properties of onedimensional quasicrystals, III. α-continuity, Commun. Math. Phys. 212 (2000), 191–204 [5] H. Furstenberg, Disjointness in Ergodic Theory, minimal sets, and a problem in diophantine approximation, Math. Systems Theory 1 (1967), 1–49 [6] H. Furstenberg, Recurrence in Ergodic Theory and Combinatorial Number Theory, Princeton University Press, 1981. [7] G. A. Hedlund and M. Morse, Symbolic dynamics, Amer. J. Math. 60 (1938), 815–866 [8] A. Hof, O. Knill and B. Simon, Singular continuous spectrum for palindromic Schr¨ odinger operators, Commun. Math. Phys. 174 (1995), 149–159 [9] S. Jitomirskaya and B. Simon, Operators with singular continuous spectrum: III. Almost periodic Schr¨ odinger operators, Commun. Math. Phys. 165 (1994), 201–205 [10] S. Kotani, Jacobi matrices with random potentials taking finitely many values, Rev. Math. Phys. 1 (1989), 129–133 [11] Y. Last and B. Simon, Eigenfunctions, transfer matrices, and absolutely continuous spectrum of one-dimensional Schr¨ odinger operators, Invent. Math. 135 (1999), 329–367 [12] G. Teschl, Jacobi Operators and Completely Integrable Nonlinear Lattices, Mathematical Surveys and Monographs 72, AMS, Providence (2000)
Vol. 3, 2002
High-complexity Hamiltonians
David Damanik Department of Mathematics 253–37, California Institute of Technology Pasadena, CA 91125 USA email:
[email protected] Boris Solomyak Department of Mathematics University of Washington Seattle, WA 98195-4350 USA email:
[email protected] Communicated by Jean Bellissard submitted 05/01/01, accepted 11/09/01
105
Ann. Henri Poincar´e 3 (2002) 107 – 170 c Birkh¨ auser Verlag, Basel, 2002 1424-0637/02/010107-64 $ 1.50+0.20/0
Annales Henri Poincar´ e
Asymptotic Completeness for Rayleigh Scattering J. Fr¨ ohlich, M. Griesemer and B. Schlein
Abstract. It is expected that the state of an atom or molecule, initially put into an excited state with an energy below the ionization threshold, relaxes to a ground state by spontaneous emission of photons which propagate to spatial infinity. In this paper, this picture is established for a large class of models of non-relativistic atoms and molecules coupled to the quantized radiation field, but with the simplifying feature that an (arbitrarily tiny, but positive) infrared cutoff is imposed on the interaction Hamiltonian. This result relies on a proof of asymptotic completeness for Rayleigh scattering of light on an atom. We establish asymptotic completeness of Rayleigh scattering for a class of model Hamiltonians with the features that the atomic Hamiltonian has point spectrum coexisting with absolutely continuous spectrum, and that either an infrared cutoff is imposed on the interaction Hamiltonian or photons are treated as massive particles. We show that, for models of massless photons, the spectrum of the Hamiltonian strictly below the ionization threshold is purely continuous, except for the ground state energy.
1 Introduction Ever since the inception of the quantum theory of atoms interacting with the quantized radiation field, theoreticians have expected that when an atom (with an infinitely heavy nucleus) in a state where all electrons are bound to the nucleus is targeted by a finite number of photons in such a way that the total energy of the composed system remains below the ionization threshold of the atom the following physical processes unfold: First, some of the electrons in the shells of the atom are lifted into an excited state by absorbing incoming photons; but, since the total energy is below the ionization threshold, they remain bound to the nucleus. As time goes on, the excited state relaxes to a ground state of the atom by spontaneous emission of photons, which propagate essentially freely to spatial infinity. Thus, asymptotically, the state of the total system, atom plus quantized radiation field, describes an atom in its ground state and a cloud of photons escaping to infinity with the velocity of light. Relaxation of an excited initial state to a ground state by emission of outgoing radiation is the simplest example of an “irreversible process”, accompanied by information loss at infinity, occuring in an open quantum system with infinitely many degrees of freedom. It would seem worthwhile to attempt to understand this process mathematically precisely; (see Sect. 10).
108
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
The picture described above suggests that the scattering operator describing Rayleigh scattering of light off an atom with a static nucleus, i.e., the scattering operator restricted to the subspace of states with energies below the ionization threshold of the atom, is unitary; (see Sect. 9 and 10). If true - one says that “asymptotic completeness” (AC) is valid for Rayleigh scattering. In attempting to establish this picture mathematically, one faces the problem that, in the scattering of light at an atom, an arbitrarily large number of soft photons of arbitrarily small total energy can, in priciple, be produced (in processes of high order in the feinstructure constant). Perturbative calculations of scattering amplitudes suggest, however, that in the analysis of Rayleigh scattering of light at an atom with a static nucleus, one does not encouter a genuine infrared catastrophe of the kind first described by Bloch and Nordsieck. Yet, the mathematical problems connected with controlling very large numbers of very soft photons in a mathematically rigorous, non-perturbative way are quite substantial and have not been fully mastered, yet. In order to simplify matters to a manageable size, we propose to study Rayleigh scattering and the phenomenon of relaxation to a ground state for models of massive photons and for models of massless photons with an infrared cutoff. In this paper, results in this direction are proven. In order to avoid inessential technical complications, we consider models of “scalar photons”, bosons. But our analysis can be extended to the quantum electrodynamics of non-relativistic electrons (bound to a static nucleus) interacting with the quantized electromagnetic field, provided we work within the dipole approximation and impose an (arbitrarily tiny) infrared cutoff on the electron-field interaction. We plan to study more difficult scattering problems in similar models and their consequences for “irreversible phenomena” in future papers. Next, we describe our main results in more detail. To avoid starting with a list of assumptions, we formulate our results for a concrete, simple model, which is physically relevant and captures the main features of the problems we propose to solve. For precise assumptions and other models see Section 3. Consider N non-relativistic electrons subject to a potential V , which may also include two–body interactions. The electrons are linearly coupled to a quantized field of relativistic bosons. The Hamilton operator of this system is H = K ⊗ 1 + 1 ⊗ dΓ(ω) + φ(G) and acts on the Hilbert space H = Hel ⊗F, where Hel = ∧N L2 (R3 ; C2 ) is the antisymmetric tensor product of N copies of L2 (R3 ; C2 ), and F is the bosonic Fock space over L2 (R3 ). The operator K describes the time evolution of the electrons without radiation and is given by K = −∆ + V , where ∆ denotes the Laplacian on R3N . (In our units, and the electron mass are equal to one.) Electron spin will be neglected henceforth.
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
109
We assume that V− , the negative part of V , is infinitesimally form–bounded with respect to −∆, that K is essentially self-adjoint on C0∞ (R3N ), and that inf σess (K) > inf σ(K), i.e., that K has bound states. In particular, inf σ(K) is an isolated eigenvalue of K. The operator dΓ(ω) describes the energy of free bosons. Formally dk ω(k)a∗ (k)a(k) dΓ(ω) = R3
where a(k) and a∗ (k) are the usual annihilation- and creation operators (operator– √ valued distributions) depending on the wave vector k, and ω(k) = k 2 + m2 is the energy of a relativistic particle with momentum k and mass m ≥ 0. The operator φ(G) = dk Gx (k)a(k) + Gx (k)a∗ (k) describes the interaction between electrons and bosons. In this introduction, we −ik·xi e κ(k), and κ(k) ∈ C0∞ (R3 ). If m = 0 we cut off the choose Gx (k) = N i=1 infrared modes from the interaction by assuming that κ(k) = 0, for |k| small. Thus, in any case (1) inf ω(k) > 0. k∈supp κ
Without this assumption, it is presently not known how to control the number of soft bosons produced in the course of the time evolution. Let Σ denote the ionization threshold. This is the smallest energy the system can reach when one or several electrons have been moved to infinity. States with energy below Σ are exponentially localized w.r. to the electron coordinates. More precisely
eα|x| E∆ (H) < ∞ if ∆ ⊂ (−∞, Σ) and sup ∆ + α2 < Σ. Clearly Σ ≥ inf σ(H), and, for one-electron atoms and if the coupling is weak enough, it is known that Σ > inf σ(H) [GLL00, BFS98]. Of course, one expects Σ > inf σ(H) for all neutral atoms and molecules. Note that Σ = ∞ if V (x) → ∞, for |x| → ∞, i.e., when σ(K) is discrete. This situation is included in our analysis, but one of our main points is to prove results on Rayleigh scattering when Σ < ∞. For states ϕ ∈ H with energy below Σ, that is ϕ = E(−∞,Σ) (H)ϕ, one expects that ϕt = e−iHt ϕ is well approximated, in the distant future, by a linear combination of states of the form a∗ (h1,t ) . . . a∗ (hn,t )e−iEt ϕb
(2)
where ϕb is a bound state of H, Hϕb = Eϕb , hj,t = e−iωt hj , j = 1, . . . , n, are one– particle wave functions of freely propagating bosons, and a∗ (h) = h(k)a∗ (k) dk, a(h) := (a∗ (h))∗ . This property is called asymptotic completeness (AC) for
110
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
Rayleigh scattering. It asserts, in particular, that the asymptotic dynamics of escaping photons is well approximated by their free dynamics. This requires that the strength of the interaction between ballistically moving bosons and electrons decays at an integrable rate. In our model of non-relativistic electrons bound to static nuclei this is true, thanks to the spatial localization of the electrons. For massless bosons, it follows more generally from the fact that the propagation velocity of electrons is strictly smaller than the velocity of the bosons, i.e., the velocity of light, [FGS00]. To give a mathematically more precise formulation of AC, let us introduce asymptotic creation operators a∗+ (h). Let ϕ = Eλ (H)ϕ, hj ∈ L2 (R3 ), j = 1, . . . , n and Mj = sup{ω(k)|hj (k) = 0}. Then a∗+ (h1 ) . . . a∗+ (hn )ϕ = lim eiHt a∗ (h1,t ) . . . a∗ (hn,t )e−iHt ϕ t→∞
(3)
exists if λ+
n
Mi < Σ.
(4)
i=1
Asymptotic completeness of Rayleigh scattering is the statement that linear combinations of states of the form (3), (4), with ϕ ∈ Hpp (H), are dense in E(−∞,Σ) (H)H. Assuming (1), we prove AC for all m ≥ 0, with an infrared cutoff imposed when m = 0. If m = 0 one can show more: If the Pauli principle for the electrons is neglected then H has a unique ground state ϕ0 [BFS98, GLL00] but no other stationary states with energy below inf σess (K) − ε, ε > 0, provided g is sufficiently small and the life times of all excited states of Hg=0 , as computed by Fermi’s Golden Rule, are finite. This follows from results of Bach et al. [BFSS99] together with an argument given in the present paper, which excludes eigenvalues close to inf σ(H). As a consequence, states of the form inf σ(H) + M (hi ) ≤ inf σess (K) − ε a∗+ (h1 ) . . . a∗+ (hn )ϕ0 , are dense in Einf σess (K)−ε (H)H, for some ε > 0 depending on the coupling constant. Moreover, we show that every state ψt ∈ Einf σess (K)−ε (H)H eventually relaxes to the ground state ϕ0 , in the following sense: Let A denote the C ∗ algebra generated by the Weyl operators eiφ(h) where φ(h) = a(h) + a∗ (h), and h ∈ S(R3 ), the Schwartz space of test functions. By taking sums of tensor products of operators in A with arbitrary bounded operators acting on the N –electron Hilbert ˜ space one obtains a C ∗ algebra A. “Relaxation of ψt to the ground state ϕ0 ” is the statement that lim ψt , Aψt = ϕ0 , Aϕ0 ψ, ψ,
t→∞
(5)
˜ and for all ψ ∈ Einf σ (K)−ε (H)H. This is our second for all operators A ∈ A, ess main result. It essentially follows from asymptotic completeness and, of course, from the absence of eigenvalues in (inf σ(H), inf σess (K) − ε].
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
111
Asymptotic Completeness for massive bosons was previously established by Derezi´ nski and G´erard, for confined electrons (i.e., Σ = ∞) and under a somewhat unphysical short-range assumption [DG99]. From this important paper, and from [DG00], we have learned how to translate techniques from N -body quantum theory to quantum field theory. Before [DG99], Arai had established AC in the standard model of non-relativistic QED in the dipole–approximation and with V (x) = x2 , a model which is explicitly soluble [Ara83]. Later, Spohn extended this result, using the Dyson series, to include potentials which are small perturbations of x2 [Spo97]. Our proof of asymptotic completeness adapts methods and techniques from the scattering theory of N -particle Schr¨ odinger operators to the present situation. In particular, we use a Mourre estimate and propagation estimates, and we rely on localization techniques in bosonic configuration space. As in the more recent papers on N -body quantum scattering, we derive AC from the fact that the mean square diameter dΓ(y 2 )ψt of a given state ψt , with y the position operator in bosonic configuration space, diverges like t2 if ψ is away, in energy, from thresholds and eigenvalues. Correspondingly, a central object in our proof is an asymptotic observable W that measures the square of the asymptotic velocities of the escaping photons. That is, W = lim dΓ(y 2 /2t2 )t t→∞
= lim
t→∞
d dΓ(y 2 /2t)t . dt
(6)
Thanks to the ballistic escape property mentioned above, W is positive and thus invertible, on suitable spectral subspaces. We construct a Deift-Simon wave operator W+ with the property that W+ W −1 is a right-inverse of an extended wave operator on a dense subspace of Hcont (H) ∩ E(−∞,Σ) (H)H, where Hcont (H) is the orthogonal complement of all eigenvectors. The proof is completed with an inductive argument explained further below. Let us temporarily assume that the interaction φ(G) vanishes, in order to explain the ideas underlying the construction of W and W+ in their purest form. The main observation is that D2
1 y2 = (∇ω − y)2 ≥ 0, 2t t
(7)
where D denotes the Heisenberg derivative [iω, .] + ∂/∂t. As a consequence the time derivative of the expression (6), whose limit is W , is non-negative. Since d/dtdΓ(y 2 /2t)t is bounded uniformly in time, it follows that d2 /dt2 dΓ(y 2 /2t)t = dΓ(D2 [y 2 /2t])t is integrable. This propagation estimate, with small modifications to accommodate the interaction, proves existence of W and suffices to establish existence of W as a strong limit. Existence of W+ requires, in addition, some geometry in bosonic configuration space. Once these asymptotic operators are constructed, AC follows by induction in the number of energy intervals of length m, the smallest energy of a boson.
112
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
Assuming that AC holds on the spectral subspace of H corresponding to energies below min(Σ, m(n − 1)), we prove AC for energies below min(Σ, mn). Roughly speaking, the positivity of W on suitable spectral subspaces E∆ (H)H, where ∆ ⊂ (−∞, min(Σ, mn)), allows us to show that at least one boson of a given state ϕ ∈ E∆ (H)H escapes to infinity. It thus carries away an energy of at least m. The energy distribution of the remaining system is contained in (−∞, m(n − 1)), where asymptotic completeness holds by assumption, and hence ϕt , for large t, is of the form (2). Obviously the positivity of the boson mass, or condition (1), in the case of more general dispersion relations, is absolutely crucial in this argument. Our strategy and the constructions of W and W+ are strongly inspired by ideas and constructions developed by Graf and Schenker for N -body quantum scattering theory [GS97]. The Mourre estimate we use is essentially the one of Derezi´ nsky and G´erard [DG99]. We follow closely the notation of [DG99]; but, otherwise, there are only few similarities between our approach to AC and the one in [DG99]. Our paper is organized as follows. In Sect. 2, we consider the quantum theory of the bosons. We briefly review the standard formalism of second quantization and introduce some basic notions that are useful in scattering theory. In Sect. 3, we describe the physical systems and define the models studied in this paper. We formulate some basic assumptions on the Hamiltonians generating the dynamics in these models which will be important to gain mathematical control over Rayleigh scattering. We describe some concrete examples of models. In Sect. 4, we review and prove results on spectral properties of the Hamiltonians of our models. In particular, we recapitulate a theorem on the existence of a ground state and on the location of the essential spectrum. We state a Mourre estimate and properties of the set of thresholds and eigenvalues. In Sect. 5, we construct the Møller wave operators of our models on spectral subspaces corresponding to bound electrons, using a variant of Cook’s argument; see also [FGS00] for more detailed results. Sect. 6 contains our basic propagation estimates needed for the construction of an asymptotic observable and of a Deift-Simon wave operator. An asymptotic observable, W , is constructed in Sect. 7 and shown to be selfadjoint, positive on appropriate spectral subspaces, and to commute with the Hamiltonian H. In Sect. 8, a Deift-Simon wave operator is constructed and shown to invert, with respect to W , an extended variant of the Møller wave operator on spectral subspaces where W is positive. By combining the results of previous sections, asymptotic completeness for Rayleigh scattering is established in Sect. 9 with the help of an inductive argument in the number of asymptotic bosons. In Sect. 10, models of massless bosons with an infrared cutoff are analyzed, and the phenomenon of relaxation to a ground state is exhibited. A novel positivecommutator estimate is proven which, together with results in [BFSS99], excludes
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
113
the existence of point spectrum above the ground state energy (and below the ionization threshold). Our most difficult and innovative results appear in Sects. 6 through 10. Various technical arguments are deferred to appendices, the most important ones being Appendices E through G. Acknowledgements. We thank V. Bach, Chr. G´erard, G.-M. Graf and I. M. Sigal for many useful discussions of problems related to those studied in this paper.
2 Fock Space and Second Quantization The natural Hilbert space of states of the radiation field is the Fock space. Let h be a complex Hilbert space and let ⊗ns h denote the n-fold symmetric tensor product of h. The bosonic Fock space over h F = F (h) = ⊕n≥0 ⊗ns h is the space of sequences ϕ = (ϕn )n≥0 , with ϕ0 ∈ C, ϕn ∈ ⊗ns h, and with the scalar product given by (ϕn , ψn ), ϕ, ψ := n≥0
where (ϕn , ψn ) denotes the inner product in ⊗n h. The vector Ω = (1, 0, . . .) ∈ F is called the vacuum. By F0 ⊂ F we denote the dense subspace of vectors ϕ for which ϕn = 0, for all but finitely many n. The number operator N on F is defined by (N ϕ)n = nϕn .
2.1
Creation- and Annihilation Operators
The creation operator a∗ (h), h ∈ h, on F is defined by √ a∗ (h)ϕ = n S(h ⊗ ϕ), for ϕ ∈ ⊗n−1 h, s and extended by linearity to F0 . Here S ∈ B(⊗n h) denotes the orthogonal projection onto the subspace ⊗ns h ⊂ ⊗n h. The annihilation operator a(h) is the adjoint of a∗ (h) restricted to F0 . Creation- and annihilation operators satisfy the canonical commutation relations (CCR) [a(g), a∗ (h)] = (g, h),
[a# (g), a# (h)] = 0.
In particular [a(h), a∗ (h)] = h 2 , which implies that the graph norms associated with the closable operators a(h) and a∗ (h) are equivalent. It follows that the closures of a(h) and a∗ (h) have the same domain. On this common domain we define 1 (8) φ(h) = √ (a(h) + a∗ (h)). 2
114
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
The creation- and annihilation operators, and thus φ(h), are bounded relative to the square root of the number operator:
a# (h)(N + 1)−1/2 ≤ h
(N + 1)−1/2 a# (h) ≤ h .
(9)
More generally, for any p ∈ R and any integer n
(N + 1)p a# (h1 ) . . . a# (hn )(N + 1)−p−n/2 ≤ Cn,p h1 · . . . · hn . This follows from a∗ (h)N = (N − 1)a∗ (h), a(h)N = (N + 1)a(h), and from (9).
2.2
The Functor Γ
Let h1 and h2 be two Hilbert spaces and let b ∈ B(h1 , h2 ). We define Γ(b) : F (h1 ) → F (h2 ) Γ(b)|` ⊗ns h1 = b ⊗ . . . ⊗ b. In general Γ(b) is unbounded but if b ≤ 1 then Γ(b) ≤ 1. From the definition of a∗ (h) it easily follows that Γ(b)a∗ (h) = a∗ (bh)Γ(b), Γ(b)a(b∗ h) = a(h)Γ(b),
h ∈ h1 . h ∈ h2 .
(10) (11)
where (11) is derived by taking the adjoint of (10). If b∗ b = 1 on h1 then these equations imply that Γ(b)a(h) = a(bh)Γ(b) Γ(b)φ(h) = φ(bh)Γ(b)
2.3
h ∈ h1 h ∈ h1 .
(12) (13)
The Operator dΓ(b)
Let b be an operator on h. Then dΓ(b) : F (h) → F (h) n dΓ(b)|` ⊗ns h = (1 ⊗ . . . b ⊗ . . . 1). i=1
For example N = dΓ(1). From the definition of a∗ (h) we get [dΓ(b), a∗ (h)] = a∗ (bh) [dΓ(b), a(h)] =−a(b∗ h), where the second equation follows from the first one by taking the adjoint. If b = b∗ then i[dΓ(b), φ(h)] = φ(ibh). (14) Note that dΓ(b)(N + 1)−1 ≤ b .
Vol. 3, 2002
2.4
Asymptotic Completeness for Rayleigh Scattering
115
The Operator dΓ(a, b)
Suppose a, b ∈ B(h1 , h2 ). Then dΓ(a, b) : F (h1 ) → F (h2 ) n dΓ(a, b)|` ⊗ns h = (a . . . a ). ⊗ . . . a ⊗b ⊗ a ⊗ j=1
j−1
n−j
For a, b ∈ B(h) this definition is motivated by Γ(a)dΓ(b) = dΓ(a, ab),
and
[Γ(a), dΓ(b)] = dΓ(a, [a, b]).
If a ≤ 1 then dΓ(a, b)(N + 1)−1 ≤ b .
2.5
The Tensor Product of two Fock Spaces
Let h1 and h2 be two Hilbert spaces. We define a linear operator U : F (h1 ⊕ h2 ) → F (h1 ) ⊗ F(h2 ) by UΩ = Ω ⊗ Ω U a (h) = [a∗ (h(0) ) ⊗ 1 + 1 ⊗ a∗ (h(∞) )]U,
(15)
∗
for h = (h(0) , h(∞) ) ∈ h1 ⊕h2 . This defines U on finite linear combinations of vectors of the form a∗ (h1 ) . . . a∗ (hn )Ω. From the CCRs it follows that U is isometric. Its closure is isometric and onto, hence unitary. It follows that U a(h) = [a(h(0) ) ⊗ 1 + 1 ⊗ a(h(∞) )]U.
(16)
Furthermore we note that U dΓ(b) = [dΓ(b0 ) ⊗ 1 + 1 ⊗ dΓ(b∞ )]U
if
b=
b0 0
0 b∞
(17)
For example U N = (N0 + N∞ )U where N0 = N ⊗ 1 and N∞ = 1 ⊗ N . Let Fn = ⊗ns h and let Pn be the projection from F = ⊕n≥0 Fn onto Fn . Then the tensor product F ⊗ F is norm-isomorphic to ⊕n≥0 ⊕nk=0 Fn−k ⊗ Fk , the corresponding isomorphism being given by ϕ → (ϕn,k )n≥0, k=0..n where ϕn,k = (Pn−k ⊗ Pk )ϕ. In this representation of F ⊗ F and with pi (h(0) , h(∞) ) = h(i) , U becomes U |` ⊗ns (h ⊕ h) =
n 1/2 n k=0
k
p0 ⊗ . . . ⊗ p0 ⊗ p∞ ⊗ . . . ⊗ p∞ .
n−k factors
k factors
(18)
116
2.6
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
Factorizing Fock Space in a Tensor Product
Suppose j0 and j∞ are linear operators on h and j : h → h ⊕ h is defined by ∗ h2 and consequently j ∗ j = jh = (j0 h, j∞ h), h ∈ h. Then j ∗ (h1 , h2 ) = j0∗ h1 + j∞ ∗ ∗ j0 j0 + j∞ j∞ . On the level of Fock spaces, Γ(j) : F (h) → F (h ⊕ h), and we define ˘ Γ(j) = U Γ(j) : F → F ⊗ F. ˘ ∗ Γ(j) ˘ It follows that Γ(j) = Γ(j ∗ j) which is the identity if j ∗ j = 1. Henceforth ∗ j j = 1 is tacitly assumed in this subsection. From (10) through (13), (15) and (16) it follows that # ˘ ˘ Γ(j)a (h) = [a# (j0 h) ⊗ 1 + 1 ⊗ a# (j∞ h)]Γ(j) ˘ ˘ Γ(j)φ(h) = [φ(j0 h) ⊗ 1 + 1 ⊗ φ(j∞ h)]Γ(j).
(19) (20)
Furthermore, if ω = ω ⊕ ω on h ⊕ h, then by (17) ˘ Γ(j)dΓ(ω) = U Γ(j)dΓ(ω) = U dΓ(ω)Γ(j) − U dΓ(j, ω j − jω) ˘ ˘ ω j − jω) = [dΓ(ω) ⊗ 1 + 1 ⊗ dΓ(ω)]Γ(j) − dΓ(j,
(21)
˘ b) = U dΓ(a, b) was introduced. In particular Γ(j)N ˘ where the notation dΓ(a, = ˘ Finally we remark that, by (18), (N0 + N∞ )Γ(j). ˘ ` ⊗ns h = Γ(j)|
n 1/2 n k=0
2.7
k
j0 ⊗ . . . ⊗ j0 ⊗ j∞ ⊗ . . . ⊗ j∞ .
n−k factors
(22)
k factors
The “Scattering Identification”
An important role will be played by the scattering identification I : F ⊗ F → F defined by I(ϕ ⊗ Ω) = ϕ I(ϕ ⊗ a (h1 ) · · · a∗ (hn )Ω) = a∗ (h1 ) · · · a∗ (hn )ϕ, ∗
ϕ ∈ F0 ,
and extended by linearity to F0 ⊗ F0 . (Note that this definition is symmetric with respect to the two factors in the tensor product.) There is a second characterization of I which will often be used. Let ι : h ⊗ h → h be defined by ι(h(0) , h(∞) ) = h(0) + h(∞) . Then (23) I = Γ(ι)U ∗ with U as above. To see this consider states of the form ψ = a∗ (h1 ) · · · a∗ (hm )Ω ⊗ a∗ (g1 ) · · · a∗ (gn )Ω m n
= [a∗ (hi ) ⊗ 1] [1 ⊗ a∗ (gj )] Ω ⊗ Ω i=1
j=1
(24)
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
117
in F0 ⊗ F0 . By equations (15) and (10) Γ(ι)U ∗ [a∗ (hi ) ⊗ 1] = Γ(ι)a∗ (hi , 0)U ∗ = a∗ (hi )Γ(ι)U ∗ and similarly Γ(ι)U ∗ [1 ⊗ a∗ (gj )] = a∗ (gj )Γ(ι)U ∗ . Furthermore Γ(ι)U ∗ Ω ⊗ Ω = Ω = I(Ω ⊗ Ω). This shows that Γ(ι)U ∗ ψ = Iψ for ψ given by (24) which proves equation (23). √ Since ι = 2, the operator I is unbounded. Lemma 1 For each positive integer k, the operator I(N + 1)−k ⊗ χ(N ≤ k) is bounded. Let j : h → h ⊕ h be defined by jh = (j0 h, j∞ h) where j0 , j∞ ∈ B(h). If ˘ is a right inverse of I, that is, j0 + j∞ = 1, then Γ(j) ˘ I Γ(j) = 1. ˘ Indeed I Γ(j) = Γ(ι)U ∗ U Γ(j) = Γ(ιj) = Γ(1) = 1.
3 Definition of the System and Basic Assumptions 3.1
The Electron System
The dynamics of the electron system (atom) is given by a self-adjoint operator K on L2 (X), where X is a measure space. Typically X = (Rn , dn x) or X = {1, . . . , n} equipped with the counting measure, in which case L2 (X) = Cn . We assume that K is bounded from below and that inf σess (K) > inf σ(K).
(H0)
In other words, inf σ(K) is an isolated eigenvalue of K. We use |x| to denote the norm of x ∈ X if X is a euclidean space. Otherwise |x| := 0.
3.2
The Radiation Field
Pure states of the radiation field are described by vectors in the bosonic Fock space F (h) over h = L2 (Rd , dk). Their time evolution is generated by the Hamiltonian dΓ(ω), where ω denotes multiplication with a real-valued function ω(k) on Rd . For easy reference, we summarize all further properties of ω in the following assumption. m := inf ω(k) > 0, and ω(k) → ∞ as |k| → ∞, ω ∈ C ∞ (Rd ), and ∂ α ω is bounded for all α = 0, (H1) ∇ω(k) = 0 if k = 0. Remarks. (i) The last condition ensures positivity of the Mourre √ constant away from thresholds. (ii) Typical examples we have in mind are ω(k) = k 2 + m2 and
118
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
smooth dispersions ω that differ from |k| only for small |k|, where ω is chosen such that inf k ω(k) > 0 (see Section 10). (iii) Bosons with non-zero spin or helicity, with h = L2 (Rd , dk) ⊗ Cs , can also be handled. We then k as a pair interpret (k, λ) ∈ Rd × {1, . . . , s} and define integration over k as sλ=1 dd k. Throughout this paper, y denotes the position operator in h, i.e., y = i∇k .
3.3
The Composed System
The dynamics of the composed system of matter (electron system) interacting with the radiation field is given by the Hamilton operator H = K ⊗ 1 + 1 ⊗ dΓ(ω) + φ(G) = H0 + φ(G), where
⊕
φ(Gx ) dx
φ(G) = X
and φ has been defined in (8). It acts on the Hilbert space H = L2 (X) ⊗ F ∼ = ⊕ F dx. For each x ∈ X, G ∈ h, and we assume that x X sup Gx < ∞. x
−1/2
and hence that φ(G)(H0 + i)−1/2 are bounded It follows that φ(G)(N + 1) operators. This implies that φ(G) is infinitesimal w.r.to H0 , which shows that H is self-adjoint on D(H0 ) and bounded from below. All further assumptions are listed below and will be cited upon use. (H2) Exponential decay: There exists an ionization threshold Σ > inf σ(H) with the property that
eα|x|E∆ (H) < ∞, for some α > 0, and for any given closed interval ∆ contained in (−∞, Σ). (H3) Fall-off of the form factor: For arbitrary α > 0, sup e−α|x| (1 + y 2 )Gx < ∞. x
(H4) Dispersion of the form-factor: If h ∈ C0∞ (Rd \{0}) then, for arbitrary α > 0, sup e−α|x| |(Gx , ht )| dt < ∞. x
Here ht = e
−iωt
h.
(H5) Short range condition: There exists a µ > 1 such that sup e−α|x| χ(|y| ≥ R)y n Gx ≤ Cα R−µ+n x
for all
R≥0
for all α > 0 and n ∈ {0, 1, 2}. Further hypotheses (IR), (H6), and (H7) are introduced in Section 10.2.
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
119
Remarks. (i) These hypotheses are satisfied and, with the exception of (H2), easily verified for many concrete systems (see the next section and Sect. 10.3). (ii) Hypothesis (H2) says that the particle system is localized near the origin for small energies. This is the central physical assumption in this paper. Since (H2) can often be derived from (H0) (see [BFS98, GLL00]) it is legitimate to impose (H2). Assumption (H2) makes (H0) obsolete in all of this paper with the exception of Section 10. (iii) In (H5), χ(|y| ≥ R) stands for multiplication with the characteristic function of the set {y : |y| ≥ R}. More generally we use χ(A < c) and χ(A ≤ c) to denote the spectral projections E(−∞,c) (A) and E(−∞,c] (A) of a given self-adjoint operator A. (iv) Hypothesis (H3) follows from (H5). ˜ =H⊗F We conclude this section by defining an extended Hilbert space H and an extended Hamilton operator ˜ = H ⊗ 1 + 1 ⊗ dΓ(ω), H
(25)
which describes a time evolution where the photons in the auxiliary Fock space do not interact with the particles. With the help of (25) and the scattering identification I the time evolution (2) will be described as a unitary time evolution on the extended Hilbert space in Section 5. Moreover the extended Hamilton operator is ˜ we define invaluable in the spectral analysis of H (see Section 4). In analogy to H ˜ H0 = H0 ⊗ 1 + 1 ⊗ dΓ(ω).
3.4
Examples of Concrete Physical Systems
In this section some concrete systems are discussed for which the hypotheses (H0) through (H5) are all satisfied. See also Section 10.3 where we explain how the standard model of non-relativistic QED fits into a slightly expanded version of the general framework introduced above. The example discussed in the introduction, where K is an arbitrary atomic Schr¨ odinger operator, that is, K = −∆ + V on L2 (R3N ), √ satisfying hypothesis (H0), ω(k) = k 2 + m2 , and Gx (k) = g
N
e−ik·xj κ(k),
κ ∈ C0∞ (R3 ),
j=1
satisfies hypotheses (H1) through (H5), for g sufficiently small. In fact, (H2) is proven in [BFS98], with Σ = inf σess (H) − g supx dk |Gx (k)|2 /ω(k), (H3) follows from y 2 = −∆k , (H4) from |(Gx , ht )| ≤ Ct−3/2 , see e.g. [RS79], and (H5) even holds in the strong form sup e−α|x| χ(|y| ≥ R)y n Gx ≤ Cn,µ R−µ x
120
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
for arbitrary n, µ ∈ N. To see this, let f (y) = (1 + |y|)µ κ ˆ (y) and put N = 1 for simplicity (this means that we consider the case with only one electron). Then −2α|x| n 2 −2α|x|
χ(|y| ≥ R)y Gx = e |y n κ ˆ (x − y)|2 dy e |y|≥R −2α|x| =e (1 + |x − y|)−2µ |y n f (x − y)|2 dy. |y|≥R
This decays exponentially as R → ∞ for |x| ≥ R/2, and if |x| < R/2 it is bounded by (1 + R/2)−2µ supx e−2α|x| χ(|y| ≥ R)y n f (x − ·) 2 which is of order O(R−2µ ), because κ ∈ C0∞ (R3 ) by assumption and thus f is rapidly decreasing. The spin-boson model, where K is a hermitian n × n matrix on Cn = L2 (X), with X = {1, . . . , n}, also fits into the our general framework. Suppose ω is as above and Gx ∈ S(R3 ) for all x ∈ X. Then hypotheses (H0) through (H5) are satisfied with the convention that |x| ≡ 0. Hypotheses (H0) and (H2) are trivial, (H4) is seen as in the first example above, and (H3) and (H5) follow from the fact that Gx ∈ S(R3 ), for x = 1, . . . , n.
4 The Spectrum of Pauli-Fierz Hamiltonians 4.1
Essential Spectrum and Existence of a Ground State
Theorem 2 Assume (H1) and (H2) (Sect. 3), and let E = inf σ(H). Then inf σess (H) ≥ min{Σ, E + m}. In particular, inf σ(H) is an isolated eigenvalue of H. Proof. Given λ ∈ σess (H) with λ < Σ we need to show that λ ≥ E + m. Let ∆ be an open interval in R containing λ with sup ∆ < Σ and let (ϕn )n≥0 ⊂ E∆ (H)H with (H − λ)ϕn → 0, ϕn = 1 and ϕn 8 0. Then λ = lim ϕn , Hϕn n→∞
and we estimate the r.h.s from below. Let j0,R , j∞,R ∈ C ∞ (R3 ) be a partition of 2 2 unity defined as in Lemma 32 with j0,R + j∞,R = 1. Pick α > 0 according to (H2) α|x| such that e E∆ (H) is bounded. Then supn eα|x| ϕn < ∞ by assumption on ϕn and hence by Lemma 32 and Lemma 31 ˘ R )Hϕn ˘ R )∗ Γ(j ϕn , Hϕn = ϕn , Γ(j ˜ Γ(j ˘ R )ϕn + o(R0 ) ˘ R )∗ H = ϕn , Γ(j
(26)
uniformly in n. From H ≥ E and dΓ(ω) ≥ m − mχ(N = 0), it is clear that ˜ ≥ (E + m) − mχ(N∞ = 0). H
(27)
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
121
˘ R )∗ χ(N∞ = 0)Γ(j ˘ R ) = Γ(j 2 ) and since Since Γ(j 0,R 2 2 E∆ (H)Γ(j0,R )E∆ (H) = (E∆ (H)eα|x| ) × (e−α|x| Γ(j0,R )E∆ (H))
is compact by Lemma 34 in Appendix E, the equation (26) combined with (27) implies that λ = lim ϕn , Hϕn ≥ E + m + o(R0 ). n→∞
Letting R → ∞ this proves the theorem.
4.2
The Mourre Estimate
Next we establish a type of Mourre theorem with conjugate operator A = dΓ(a) and 1 a = [iω, y 2 /2] = (∇ω · y + y · ∇ω). 2 That is we prove positivity of i[H, A] on spectral subspaces of H away from thresholds and eigenvalues, and, as in N -body quantum theory, we obtain important spectral properties of H as a byproduct. Here the thresholds are the elements of τ := σpp (H) + Nm, (0 ∈ N). The Mourre inequality will allow us to show that dΓ(y 2 )ψt ≥ ct2 ,
as |t| → ∞,
(28)
where c > 0 for states separated in energy from thresholds and eigenvalues. This together with the above mentioned spectral properties suffices to derive AC. As (28) can only be true if the particles are spatially confined, our Mourre estimate only holds for energies below Σ. On a suitable dense subspace, i[H, A] = dΓ(|∇ω|2 ) − φ(iaG). We use this equation to define the operator i[H, A] on ∪µ<Σ Eµ (H)H. Note that φ(−iaG)ϕ makes sense for ϕ ∈ Ran Eµ (H), µ < Σ, thanks to the exponential decay and the boundedness of e−α|x| φ(−iaG)(N + 1)−1/2 . The following virial theorem is an important ingredient in the proof of Theorem 4. Furthermore, in the case of massless bosons and IR-cutoff interaction (see Section 10) it will allow us to prove absence of eigenvalues close to, but different from the ground state energy. Lemma 3 (Virial Theorem) Assume hypotheses (H1), (H2), and (H3). If Hϕ = Eϕ and E < Σ then ϕ, i[H, A]ϕ = 0. The proof of this lemma is deferred to Appendix E.
122
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
The following theorem is the main result of this section. Theorem 4 Assume (H1), (H2) and (H5). Then (i) For each λ ∈ (−∞, Σ)\τ there exists an open interval ∆ constant Cλ , and a compact operator E such that
λ, a positive
E∆ (H)[iH, A]E∆ (H) ≥ Cλ E∆ (H) + E. (ii) Non-threshold eigenvalues in (−∞, Σ) have finite multiplicity and can accumulate only at thresholds. Furthermore τ ∩ (−∞, Σ] is closed and countable. (iii) If λ ∈ (−∞, Σ)\τ is not an eigenvalue, then there exists an open interval ∆ λ and a positive constant Cλ such that E∆ (H)[iH, A]E∆ (H) ≥ Cλ E∆ (H). The proof of this theorem follows the lines of the proof in [DG99] with only minor modifications due to the presence of continuous spectrum in the particle Hamiltonian K. For the sake of completeness we have included a proof of Theorem (4) in this paper, but it is deferred to Appendix E.
5 The Wave Operator ˜ = H ⊗ 1 + 1 ⊗ dΓ(ω) on H ˜ = H ⊗ F, and let Recall from Section 3 that H PB (H) in (29) denote the orthogonal projector onto Hpp (H), the closure of the space spanned by all eigenvectors of H. The purpose of this section is to establish existence of the wave operator ˜
Ω+ = s − lim eiHt Ie−iHt PB (H) ⊗ 1 t→∞
(29)
˜ corresponding to compact intervals ∆ ⊂ (−∞, Σ). on spectral subspaces of H Furthermore we will see that Ω+ is isometric if restricted to vectors in Hpp (H)⊗F. The existence of (29) will essentially follow from the existence of asymptotic field operators (30) a'+ (h)ϕ = lim eiHt a' (ht )e−iHt ϕ t→∞
and the existence of products of such operators, which is established in the next theorem. Theorem 5 Assume hypotheses (H1), (H2) and (H4) are satisfied, and let f, h ∈ L2 (Rd ). i) If ϕ = Eη (H)ϕ for some η < Σ then the limit a'+ (h)ϕ = lim eiHt a' (ht )e−iHt ϕ t→∞
exists. Here ht = e−iωt h.
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
123
ii) The canonical commutation relations [a+ (g), a∗+ (h)] = (g, h)
[a'+ (h), a'+ (g)] = 0,
and
hold true, in form–sense, on χ(H < η)H for all η < Σ. iii) Let m = inf{ω(k) : h(k) = 0} and M = sup{ω(k) : h(k) = 0}. Then a∗+ (h) Ran χ(H ≤ E) ⊂ Ran χ(H ≤ E + M ) a+ (h) Ran χ(H ≤ E) ⊂ Ran χ(H ≤ E − m). iv) Suppose ϕ = Eη (H)ϕ, hi ∈ L2 (Rd ; C) for i = 1, . . . , n and let Mi = sup{ω(k) : hi (k) = 0}. Then a∗+ (h1 ) . . . a∗+ (hn )ϕ = lim eiHt a∗ (h1,t ) . . . a∗ (hn,t )e−iHt ϕ provided that η +
t→∞
n i=1
Mi < Σ.
v) Suppose η < Σ and ϕ ∈ Eη (H)Hpp (H). Then a+ (h)ϕ = 0
for all
h ∈ L2 (Rd ).
√ Remark. For relativistic massive bosons, that is for ω(k) = k 2 + m2 with m > 0, as well as in the case of relativistic electrons and massless bosons, the asymptotic field operators actually exist on a dense subspace of H irrespective of Σ (see [FGS00]). Proof. i) Assume first that h ∈ C0∞ (Rd \{0}). By Cook’s argument it suffices to show that ∞
(Gx , ht )ϕt dt < ∞. (31) 1
This follows from the assumptions (H2) and (H4). For the proofs of ii), iii) and iv) we refer to [FGS00]. v) It suffices to show that a+ (h)ϕ = 0 if Hϕ = Eϕ. Statement v) then follows from the boundedness of a+ (h)Eη (H). Since ht 8 0 weakly as t → ∞ we have s − limt→∞ a(ht )(H + i)−1/2 = 0. Hence a+ (h)ϕ = lim eiHt e−iEt (E + i)1/2 a(ht )(H + i)−1/2 ϕ = 0. t→∞
Next we prove existence of the extended wave operator ˜ ˜ + := s − lim eiHt Ie−iHt Ω t→∞
˜ Since Ω ˜ + agrees with Ω+ on vectors in on a suitable spectral subspace of H. Hpp (H) ⊗ F, this will immediately imply existence of Ω+ .
124
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
Lemma 6 Assume the hypotheses of the theorem above are satisfied. ∗ ∗ a) Suppose n ψ = ϕ ⊗ a (h1 ) . . . a (hn )Ω where ϕ = Eλ (H)ϕ and λ + i=1 Mi < Σ, where Mi is defined as in Theorem 5, iv). Then ˜ ˜ + ψ := limt→∞ eiHt Ie−iHt ψ exists and Ω
˜ + ψ = a∗ (h1 ) . . . a∗ (hn )ϕ. Ω + + ˜ H ˜ for all µ < Σ. ˜ + exists on Eµ (H) b) Ω Proof. Statement a) follows from ˜
e−iHt ϕ ⊗ a∗ (h1 ) . . . a∗ (hn )Ω = e−iHt ϕ ⊗ a∗ (h1,t ) . . . a∗ (hn,t )Ω, the definition of I and Theorem 5, iv). ˜ ˜ is bounded uniformly in t, it suffices to prove (b) Since eiHt Ie−iHt Eµ (H) ˜ H. ˜ By Lemma 30 finite linear com˜ + on a dense subspace of Eµ (H) existence of Ω binations of vectors of the form described in part a) span such a subspace and hence b) follows from a). The following theorem is the main result of this section. Theorem 7 Assume (H1), (H2) and (H4). Then ˜
Ω+ = s − lim eiHt Ie−iHt PB (H) ⊗ 1 t→∞
˜ H, ˜ Ω+ = 1 and hence Ω+ has a unique extension, also exists on ∪µ<Σ Eµ (H) ˜ H. ˜ Ω+ is isometric on (PB (H) ⊗ 1)EΣ (H) ˜ H ˜ and hence denoted by Ω+ , to EΣ (H) Ran Ω+ is closed. Furthermore ˜
eiHt Ω+ = Ω+ eiHt . ˜ H ˜ follows from Lemma 6 b). By Proof. The existence of Ω+ on ∪µ<Σ Eµ (H) Lemma 6 a), the CCR in Theorem 5 ii), and by part v) of that theorem, Ω+ is ˜ H. ˜ So Ω+ is a partial isometry and hence isometric on (PB (H) ⊗ 1) ∪µ<Σ Eµ (H)
Ω+ = 1. All these properties carry over to the closure of Ω+ .
6 Propagation Estimates This section establishes the propagation estimates needed later on to construct the asymptotic observable W (see Eq. (6)) and the Deift Simon wave operator W+ . To begin with, we define a smooth convex function S(y, t) which modifies y 2 /2t for y in a neighborhood of size tδ , δ ∈ (0, 1), of the origin. The operator W will be constructed using S(y, t) in place of y 2 /2t. This will not affect W but allows us to prove its existence.
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
125
Here and henceforth we use the following notation for the various Heisenberg derivatives. Suppose A is an operator in H. Then we define ∂A , ∂t ∂A . D0 A := i[H0 , A] + ∂t Similarly, the Heisenberg derivative d of an operator a on the one boson sector h is defined by ∂a da = i[ω, a] + . ∂t Furthermore, if A(R) is a family of operators on H, and R ∈ R+ , we write DA := i[H, A] +
A(R) = O(Rm )N p
6.1
if
A(R)(N + 1)−p = O(Rm ).
Construction of S(y, t)
Pick m ∈ C0∞ (R+ ) with m ≥ 0, supp(m) ⊂ [1, 2] and dσ m(σ) = 1. Set S0 (y) = dσ m(σ)χ(y 2 /2 > σ) (y 2 /2 − σ). Then S0 is smooth and
S0 (y) =
where b = −
0 y 2 /2 + b
(32)
y 2 /2 ≤ 1 y 2 /2 ≥ 2
dσ m(σ)σ. It follows that S0 (y) = y 2 /2 + a(y) + b
(33)
C0∞ (R3 ).
This formula will allow us to apply Lemma 27 in Appendix A where a ∈ to [iω, S0 ]. It is important that S0 is convex, which is easy to see from the definition. Next we define a scaled version of S0 by S(y, t) = t−1+2δ S0 (y/tδ )
(34)
where 0 < δ < 1. Lemma 8 For all integers n ≥ 0 and all α y2 n α D ∂ S− = O t−(n+1)+δ(2−|α|) . 2t Here D = ∂/∂t or D = ∂/∂t + (y/t) · ∇ and ∂ α is any spacial derivative of order |α|. In particular ∇S =
y + O(t−1+δ ), t
These estimates hold uniformly in y ∈ Rd .
y2 ∂S = − 2 + O(t−2+2δ ). ∂t 2t
126
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
Proof. By definition of S(y, t) and (33), S(y, t) − y 2 /2t = t−1+2δ a(yt−δ ) + t−1+2δ b. The second term clearly enjoys the desired property. For the first term we have ∂ α [t−1+2δ a(yt−δ )] = t−1+δ(2−|α|) (∂ α a)(yt−δ ) where the right hand side is of the form tc h(yt−δ ) with h ∈ C0∞ (R3 ). Both ∂/∂t(tc h) and D(tc h) are again of this form with c replaced by c − 1. This proves the lemma. Lemma 9 Suppose ω ∈ C ∞ (Rd ) and ∂ α ω is bounded for all α = 0. Then dS =
1 ∂S (∇ω · ∇S + ∇S · ∇ω) + + O(t−1 ), 2 ∂t
and for any smooth vector field v(y, t) and Dv = ∂/∂t + v · ∇, d(dS) =(∇ω − v) · S
(∇ω − v) + (Dv ∇S) · (∇ω − v) + (∇ω − v) · (Dv ∇S)
(35) (36)
+ Dv2 S − (Dv v) · ∇S + O(t−1−δ ).
(37)
Proof. The first part follows from (33) and Lemma 27. By definition of the Heisenberg derivative, d(dS) is given by d(dS) = [iω, [iω, S]] + 2[iω, ∂S/∂t] +
∂2S . ∂t2
By Lemma 8, (33), and Lemma 27 ∂S ∂S +∇ · ∇ω + O(t−2 ) ∂t ∂t [iω, [iω, S]] = ∇ωS
∇ω + O(t−1−δ ).
2[iω, ∂S/∂t] = ∇ω · ∇
To prove the second equation an explicit formula for [iω, S] − ∇ω · ∇S is also needed (see the proof of Lemma 27). Since for every smooth vector field v(y, t) ∂S ∂2S ∂S +∇ · ∇ω + 2 ∂t ∂t ∂t =(∇ω − v) · S
(∇ω − v) + (Dv ∇S) · (∇ω − v) + (∇ω − v) · (Dv ∇S)
∇ω · S
∇ω + ∇ω · ∇
+ Dv2 S − (Dv v) · ∇S
(38)
the lemma follows.
6.2
Propagation Estimates
We are now ready to state and prove our propagation estimates. Note that these are basic propagation estimates which are well known in other contexts (see [GS97]).
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
127
Proposition 10 Assume (H1) and (H3), let χ = χ ¯ ∈ C0∞ (R), and suppose α|x| e χ(H) is a bounded operator for some α > 0. Let χ = χ(H) and v = y/t. Then, if λ > 0 is large enough, there exists a constant C such that ∞ 1 dt χψt , χ(λ2 ≤ dΓ(v 2 ) ≤ 2λ2 )χψt ≤ C ψ 2 t 1 for all ψ ∈ H. Here ψt = e−iHt ψ. ˜ and with H and dΓ(v 2 ) Remark. This propagation estimate equally holds on H 2 2 ˜ replaced by H and dΓ(v ) ⊗ 1 + 1 ⊗ dΓ(v ). This is needed for the proof of the remark to Proposition 11. ˜ Proof. h ∈ C0∞ (1/2, 3) with h(r) = 1 on [1, 2], 0 ≤ h ≤ 1 and set h(s) = s 2Pick
˜ ˜ ˜ ˜ ds h (s ). Note that h(s) = h(3) for s ≥ 3. Hence g(s) = h(s) − h(3) for 0 s ≥ 0 and g(−s) = g(s) define a C0∞ -function g on R. The operator B = 2 ˜ −χh(dΓ(v /λ2 ))χ is our propagation observable. Since B is bounded the theorem will follow if we show that DB := [iH, B] +
C ∂B ≥ χ(H) χ(1 ≤ dΓ(v 2 /λ2 ) ≤ 2) χ(H) + integrable terms ∂t t
for some C > 0 if λ is large enough. Henceforth we use the abbreviations h, 2 ˜ ˜ and g to denote the operators h(dΓ(v 2 /λ2 )), h(dΓ(v /λ2 )) and g(dΓ(v 2 /λ2 )), h respectively. Clearly 1 ∂B 2 = χh2 dΓ(v 2 /λ2 ) χ ≥ χh2 χ. ∂t t t
(39)
Next ˜ = χ[idΓ(ω), ˜h]χ + χ[iφ(G), h]χ. ˜ −[iH, B] = χ[iH, h]χ ˜ = [iφ(G), g], Consider first the second term on the right hand side. By [iφ(G), h] the Helffer–Sj¨ ostrand functional calculus (see Appendix A.2), and by (H3) 1 1 ˜
e−α|x|[iφ(G), h](N + 1)− 2 ≤ C e−α|x|φ(−iv 2 G)(N + 1)− 2
C 1 ≤ 2 2 sup e−α|x| φ(iy 2 Gx )(N + 1)− 2 = O(t−2 ). λ t x (40)
˜ is integrable. By Lemma 28 Hence χ[iφ(G), h]χ ˜ = [idΓ(ω), g] [idΓ(ω), h] 1 = g dΓ(∇ω · v/λ + v/λ · ∇ω) + O(t−2 )N λt 1 = hdΓ(∇ω · v/λ + v/λ · ∇ω)h + O(t−2 )N λt C ≤ h(N + 1)h + O(t−2 )N . λt
128
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
Since [χ, h] = O(t−1 ) it follows that ˜ ≤ χ[idΓ(ω), h]χ
C 2 h + O(t−2 ). λt
For C/λ < 1 this in conjunction with (39) proves the proposition. Proposition 11 Assume (H1) and (H5), let χ = χ ¯ ∈ C0∞ (R), and suppose α|x| e χ(H) is a bounded operator for some α > 0. Let f ∈ C0∞ (R), with 0 ≤ f ≤ 1, f (x) = 0 for x ≥ 2, and f (x) = 1 for 0 ≤ x ≤ 1. Denote χ = χ(H) and f = f (dΓ(v 2 /λ2 )), where v = y/t and λ ∈ R. Then, for λ large enough ∞ dt ψt , χf dΓ ((∇ω − v) · S
(∇ω − v)) f χψt ≤ C ψ 2 . 1
˜ and with H, dΓ(v 2 ) and dΓ(P ) Remark. This proposition equally holds on H
˜ dΓ(v 2 ) ⊗ 1 + 1 ⊗ dΓ(v 2 ) and (here P = (∇ω − v) · S (∇ω − v)), replaced by H, dΓ(P ) ⊗ 1 + 1 ⊗ dΓ(P ). This is needed for the proof of Theorem 15. Proof. Let γ(t) = ψt , χf dΓ(dS)f χψt . From ±dS ≤ const(v 2 + 1) it follows that dΓ(dS)f (N +1)−1 is a bounded operator and thus that sup|t|≥1 |γ(t)| < ∞ because of the cut-off χ. Next we show that γ (t) ≥ψt , χf dΓ((∇ω − v) · S
(∇ω − v))f χψt (41) + ψt , B(t)ψt , where B(t) is a bounded operator such that dt|ψt , B(t)ψt | ≤ C ψ 2 . By the Leibnitz rule γ (t) = ψt , χ(Df )dΓ(dS)f χψt + h.c. + ψt , χf (DdΓ(dS))f χψt .
(42)
Only the last term will contribute to (41). DdΓ(dS) = D0 dΓ(dS) + [iφ(G), dΓ(dS)] = dΓ(d(dS)) + φ(−idS G) where d(dS) is given by Lemma 9. Since v = y/t the terms in (37) are of order O(t−1−ε ) where ε = min(δ, 1 − δ). For the terms of (36) we have ± [(Dv ∇S) · (∇ω − v) + (∇ω − v) · (Dv ∇S)] ≤ t2−δ (Dv ∇S)2 + t−2+δ (∇ω − v)2 = O(t−2+δ )(1 + v 2 ) which, thanks to the cutoffs f and χ, gives an integrable contribution to γ (t). This shows that ψt , χf dΓ(d2 S)f χψt ≥ψt , χf dΓ((∇ω − v) · S
(∇ω − v))f χψt + O(t−1−ε ) ψ 2
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
129
To estimate the contribution due to φ(−idS G) use
χf φ(−idSG)f χ ≤ const e−α|x| φ(−idSG)(N + 1)−1/2 . and dS = ∇S · ∇ω − i
2 2 (∂rs S)(∂rs ω) +
r,s
(43)
∂S + O(t−1−δ ). ∂t
This shows that dS = χ(|y| ≥ t )dS + O(t−δ−1 ) and in conjunction with Lemma 8 and (H5) it follows that (43) is integrable. To estimate the contribution in (42) due to Df note that Df = D0 f + [iφ(G), f ]. δ
The second term gives a contribution of order O(t−2 ). This is seen in the same way as (40). Next choose g ∈ C0∞ (R) with supp(g) ⊂ (1, 2) and gf = f and denote g = g(dΓ(v 2 /λ2 )). Then χ(D0 f )dΓ(dS)f χ = χg(D0 f )dΓ(dS)gf χ + O(t−2 ) and hence c |ψt , χ(D0 f )dΓ(dS)f χψt | ≤ g(N + 1)χψt 2 + O(t−2 ) t which is integrable by Proposition 10.
7 The Asymptotic Observable In this section existence of the asymptotic observable W is proved. An auxiliary version Wλ of W will involve a space cutoff at |y| = λt in the bosonic configuration space. W is then obtained in the limit λ → ∞. To define the space cutoff we pick, once and for all, a function f ∈ C0∞ (R) with 0 ≤ f ≤ 1, f (x) = 0 for x ≥ 2 and f (x) = 1 for 0 ≤ x ≤ 1. The space cutoff is the operator f [dΓ(v 2 /λ2 )] or 1 ⊗ f [dΓ(v 2 /λ2 )] on F or H respectively. Here v = y/t and λ ∈ R. For brevity these operators will also be denoted by f if there is no danger of confusion. Since (dS)2 ≤ const(v 4 + 1), which follows from Lemma 9 and Lemma 27, and since f ∈ C0∞ (R), the operator dΓ(dS)f (N + 1)−1 is bounded. Theorem 12 Assume Hypotheses (H1), (H2), and (H5). If χ ∈ C0∞ (R) with supp χ ⊂ (−∞, Σ), then Wλ = s − lim eiHt χf dΓ(dS)f χe−iHt t→∞
exists, is self-adjoint, and commutes with H.
130
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
Proof. To prove existence of Wλ ϕ we use Cook’s argument, i.e. we show that ∞ d dt ψt , χf dΓ(dS)f χϕt ≤ C ψ . dt 1 As in the proof of Proposition 11 one shows that d ψt , χf dΓ(dS)f χϕt = ψt , χf dΓ[(∇ω − v) · S
(∇ω − v)]f χϕt dt + ψt , B(t)ϕt ,
(44)
∞ where B(t) is a bounded t-dependent observable such that 1 dt|ψt , B(t)ϕt | ≤ C ψ
ϕ . Since S
≥ 0 the first term on the right side defines a non-negative sesquilinear form in ψ and ϕ and hence we can apply Schwarz |ψt , χf dΓ[(∇ω − v) · S
(∇ω − v)]f χϕt | ≤ |ψt , χf dΓ[(∇ω − v) · S
(∇ω − v)]f χψt |
1/2
× |ϕt , χf dΓ[(∇ω − v) · S
(∇ω − v)]f χϕt |
1/2
.
This, together with Proposition 11 after an application of H¨ older’s inequality, shows that also the first term in (44) is integrable. This proves existence of Wλ ϕ. Clearly Wλ is bounded and symmetric. To prove that Wλ commutes with H it suffices to show that e−iHt Wλ eiHt = Wλ for all t ∈ R. This follows from −iHt e−iHs Wλ eiHs ϕ − Wλ ϕ = lim eiHt χ[f dΓ(dS)f ]t+s ϕ t χe t→∞
−1 because χ[f dΓ(dS)f ]t+s ). Indeed t χ = O(t
∂ ∂f ∂f ∂ (f dΓ(dS)f ) = f dΓ(dS)f + dΓ(dS)f + f dΓ(dS) ∂t ∂t ∂t ∂t where ∂f /∂t = O(t−1 ) and χf dΓ(d(∂S/∂t))f χ = O(t−1 ). The latter follows from (33) and Lemma 27. In the next step we remove the space cutoff f . This will allow us to prove positivity of W = limλ→∞ Wλ away from thresholds and eigenvalues. Proposition 13 Under the assumptions of Theorem 12, the limit W = limλ→∞ Wλ exists in operator norm sense, and W is given by ϕ, W ψ = lim ϕ, eiHt χdΓ(dS)χe−iHt ψ t→∞
for all ϕ, ψ ∈ D(dΓ(y 2 )) ∩ D(N ). W commutes with H.
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
131
Proof. We pick ϕ, ψ ∈ D(dΓ(y 2 )) ∩ D(N ) and we consider the difference |ϕt , χdΓ(dS)χψt −ϕt , χf dΓ(dS)f χψt | ≤ |ϕt , χ(1 − f )dΓ(dS)f χψt | + |ϕt , χdΓ(dS)(1 − f )χψt | ≤ (1 − f )χϕt
dΓ(dS)f χψt + (1 − f )χψt
dΓ(dS)χϕt . (45) Since (1 − f (s))2 ≤ s2 , we have
(1 − f )χϕt ≤
1 C
dΓ(y 2 /t2 )χϕt ≤ 2 (1/t2 dΓ(y 2 + 1)ϕ + ϕ ), 2 λ λ
(46)
and analogously for ϕ replaced by ψ. The second inequality in (46) follows from Lemma 36 in Appendix F. To handle the factor dΓ(dS)χϕt on the r.h.s. of (45) use that, by Lemma 8 and Lemma 9, dS =
1 y2 (∇ω · y + y · ∇ω) − 2 + O(t−1+δ ). 2t 2t
(47)
Part (iv) of Lemma 37 (with y replaced by y/t) and Lemma 36 in Appendix F lead then to
dΓ(dS)χϕt ≤ C (1/t2 dΓ(y 2 + 1)ϕ + ϕ ). (48) Insering (46) and (48) into (45) and using dΓ(dS)f χψt ≤ C ψ we find |ϕt , χdΓ(dS)χψt − ϕt , χf dΓ(dS)f χψt | ≤
C (1/t2 dΓ(y 2 + 1)ϕ + ϕ ) λ2 × (1/t2 dΓ(y 2 + 1)ψ + ψ ),
for arbitrary ϕ, ψ ∈ D(dΓ(y 2 ))∩D(N ), and for all t ≥ 1. This shows that W exists as a weak limit and that |ϕ, (W − Wλ )ψ| ≤
C
ϕ ψ , λ
which proves that W also exists as a norm limit. Finally, that W commutes with H follows from the fact that Wλ commutes with H for each λ. Using the Mourre inequality from Theorem 4, we next prove positivity of W away from thresholds and eigenvalues. Note that this is the only place where the Mourre inequality is used. Proposition 14 Assume hypotheses (H1), (H2) and (H5) are satisfied. Assume, moreover, that the energy cutoff χ in the definition of W satisfies supp χ ⊂ (−∞, Σ)\S, where S = σpp (H) + m · (N ∪ {0}) is the set of all eigenvalues and thresholds of H. Then W ≥ d χ2 , for some d > 0. In particular, if ∆ ⊂ (−∞, Σ)\S and χ|`∆ = 1, then W ≥ d on RanE∆ (H).
132
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
Proof. By the compactness of σ(H) ∩ (−∞, Σ)\S, and since W commutes with H, it is enough if we prove that, for each x ∈ (−∞, Σ)\S, W |`RanEUx (H) = EUx (H)W EUx (H) ≥ dx EUx (H)χ2 ,
(49)
where Ux is an arbitrarily small neighborhood of x, and dx > 0. If we choose Ux sufficiently small then, by the Mourre estimate (Theorem 4), we can find a function χ ˜ ∈ C0∞ (R) such that χ ˜ = 1 on Ux and ˜ χ(H)[iH, ˜ A]χ(H) ˜ ≥ 2dx χ(H),
(50)
for some dx > 0 (χ ˜ is a smoothed characteristic function of Ux ). Here A = dΓ(a) = ˜ we have EUx (H)χ(H) ˜ = EUx (H), and dΓ([iω, y 2 /2]). Because of the choice of χ thus Eq. (49) follows if we prove that χW ˜ χ ˜ ≥ dx χχ ˜ 2,
(51)
where χ ˜ = χ(H). ˜ To this end we choose ψ ∈ D(dΓ(y 2 )) ∩ D(N ) and let ψt = −iHt e ψ. Then, by Lemma 35, χψ ˜ ∈ D(dΓ(y 2 )) ∩ D(N ) and thus, by Proposition 13, we have ˜ χ ˜ χψt ψ, χW ˜ χψ ˜ = lim ψt , χ χdΓ(dS) t→∞
d ψt , χ χdΓ(S) ˜ χ ˜ χψt , t→∞ dt
(52)
= lim
where the second equality holds because, by Hypothesis (H5), χ[iφ(G), ˜ dΓ(S)]χ ˜= −χφ(iSG) ˜ χ ˜ = o(t−1+δ ). From (52) it now follows that 1 ψt , χ χdΓ(S) ˜ χ ˜ χψt t 1 2 = lim 2 ψt , χ χdΓ(y ˜ /2)χ ˜ χψt , t→∞ t
ψ, χW ˜ χψ ˜ = lim
t→∞
(53)
where, in the second equality, we used the definition of the function S(y). Now we have t 2 2 ˜ /2)χe ˜ −iHt = χdΓ(y ˜ /2)χ ˜+ ds eiHs χ[iH, ˜ dΓ(y 2 /2)]χe ˜ −iHs eiHt χdΓ(y 0 t y2 2 G)χe ˜ −iHs = χdΓ(y ˜ /2)χ ˜− ds eiHs χφ(i ˜ 2 0 t iHs (54) + ds e χA ˜ χe ˜ −iHs 0
t y2 2 G)χe ˜ −iHs = χdΓ(y ˜ /2)χ ˜− ds eiHs χφ(i ˜ 2 0 t s + tχA ˜ χ ˜+ ds dr eiHr χ[iH, ˜ A]χe ˜ −iHr . 0
0
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
133
Note that the operator φ(iy 2 G)χ ˜ is bounded. Moreover the expectation values of 2 χdΓ(y ˜ )χ ˜ and of χA ˜ χ ˜ in the state χψ are finite, because ψ ∈ D(dΓ(y 2 )) ∩ D(N ), ±A ≤ C dΓ(y 2 + 1) and because of Lemma 35 (see Appendix F). Thus, after division by t2 , only the last term in (54) gives a non-vanishing contribution to (53) in the limit t → ∞. By (50) s 1 t ds dr ψr , χ χ[iH, ˜ A]χ ˜ χψr ≥ dx ψ, χ2 χψ, ˜ ψ, χW ˜ χψ ˜ = lim 2 t→∞ t 0 0 which proves Eq. (51) because D(dΓ(y 2 )) ∩ D(N ) is dense in H.
8 Inverting the Wave Operator The Deift-Simon wave operator W+ , to be constructed in this section, inverts the ˜ + with respect to Wλ in the sense that extended wave operator Ω ˜
˜ + W+ = lim eiHt Ie−iHt W+ . Wλ = Ω t→∞
(55)
On spectral subspaces where Wλ is positive and thus invertible, W+ Wλ−1 is then ˜ + . Formally, and when space and energy cutoffs are in fact a right inverse of Ω ignored, then W+ is given by −iHt ˜ ˘ ˘ (56) W+ = s − lim eiHt dΓ(j t , djt )dΓ(S) + Γ(jt )dΓ(dS) e t→∞
˘ t ) = 1 and where jt = (j0,t , j∞,t ) and j0,t + j∞,t = 1. By the last identity, I Γ(j ˘ ˘ IdΓ(jt , djt ) = D0 [I Γ(jt )] = 0. Hence (55) is obvious at least on this formal level. The functions j0,t and j∞,t are constructed as follows. Let j0 , j∞ ∈ C ∞ (Rd ) where j0 (y) = 1 for |y| < 1, j0 (y) = 0 for |y| > 2, and let j∞ = 1 − j0 . Next set j',t (y) = j' (y/ut) where u > 0 is a fixed parameter. By construction of jt and W+ , Eq. (56), photons with velocity u or larger are mapped to the second Fock space where their interaction with the electrons is turned off. First we prove existence of W+ in Theorem 15 and then we prove (55). Theorem 15 together with the Mourre estimate, Theorem 4, is the heart of our proof of AC. Recall from Section 7 that DA, D0 A and da denote Heisenberg derivatives of operators A and a on H and h respectively. If B is an operator on the extended ˜ and if C maps H to H, ˜ then we set Hilbert space H ˜ B] + DB := i[H,
∂B ∂t
˜ := iHC ˜ − CiH + ∂C . DC ∂t ˜ 0 are defined in a similar way using H0 and H ˜ 0 rather The derivatives D0 , and D ˜ than H and H. Finally the Heisenberg derivative db of an operator b mapping the
134
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
one-boson sector h to h ⊕ h (that is, bh = (b0 h, b∞ h) with b0,∞ being operators on h) is defined by ∂b db0 ω 0 = db = i b − b iω + . db∞ 0 ω ∂t The function f ∈ C0∞ (R) and the space cutoff denoted by the same letter are defined as in the previous section, and f˜ = f [dΓ(v 2 /λ2 ) ⊗ 1 + 1 ⊗ dΓ(v 2 /λ2 )] is ˜ the space cutoff on H. The next theorem is the main result of this section. Theorem 15 Assume Hypotheses (H1), (H2), and (H5). If χ ∈ C0∞ (R) with supp χ ⊂ (−∞, Σ), then ˜ ˘ t )dΓ(S) f χe−iHt ˜ 0 Γ(j W+ = s − lim eiHt χ ˜f˜D t→∞
˜ Furthermore exists. Here χ = χ(H) and χ ˜ = χ(H). ˜
e−iHt W+ = W+ e−iHt . Proof. By Cook’s argument we need to show that, for all ϕ ∈ H, there exists a constant C such that ∞ d dt ψ, W+ (t)ϕ ≤ C ψ
dt 1 for all ψ ∈ H, where ˜
W+ (t) = eiHt χ ˜f˜Qf χe−iHt ˘ t )dΓ(S) = dΓ(j ˜ 0 Γ(j ˘ t , djt )dΓ(S) + Γ(j ˘ t )dΓ(dS). Q=D In form sense d ˜ ˜ f˜Qf ]χe−iHt W+ (t) = eiHt χ ˜D[ dt ˜ f˜Qf ] = (Df˜)Qf + f˜(DQ)f ˜ D[ + f˜Q(Df ). The contributions due to Df˜ and Df are dealt with as in Proposition 11 and are ˜ is the sum integrable due to Proposition 10. The operator DQ ˜ =D ˜ 0 Q + i(φ(G) ⊗ 1)Q − Qiφ(G) DQ where the last two terms give a contribution of order t−µ because of Hypothesis (H5). To show this write ˘ t )dΓ(S)idΓ(ω) ˘ t )dΓ(S) − Γ(j Q =i [dΓ(ω) ⊗ 1 + 1 ⊗ dΓ(ω)] Γ(j ∂ ˘ − Γ(j t )dΓ(S) ∂t
(57)
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
135
and commute iφ(G) from the right through the terms in (57) using (14) and (20) (before differentiating with respect to t). This is a lengthy computation which leads to ˘ t )dΓ(S) ˜ 0 Γ(j i(φ(G) ⊗ 1)Q − Qiφ(G) =i [φ(j∞ G) ⊗ 1 − 1 ⊗ φ(j∞ G)] D ˘ t) − i [φ(d(j0 S)iG) ⊗ 1 + 1 ⊗ φ(d(j∞ S)iG)] Γ(j ˘ t) ˜ 0 Γ(j − i [φ(j0 SiG) ⊗ 1 + 1 ⊗ φ(j∞ SiG)] D ˘ t )dΓ(S). − [φ(d(j0 )G) ⊗ 1 + 1 ⊗ φ(d(j∞ )G)] Γ(j Using Hypothesis (H5), Eq. (33) and Lemma 27 one shows that each of these terms is of order t−µ . We demonstrate this for the last term. By Lemma 27 t(dj∞ )G = t ∇ω · ∇j∞ − i (∂rs ω)(∂rs j∞ ) + ∂t j∞ G + O(t−2 )G. (58) r,s
Since derivatives of ω are bounded, and since derivatives of j∞ are of order t−1 and live on {|y| ≥ ut}, Equation (58) in conjunction with Hypothesis (H5) implies −α|x| φ(t(dj∞ )G)(N + 1)−1 = O(t−µ ). e On the other hand dΓ(S/t)f (N + 1)−1 is bounded. Hence ˘ t )dΓ(S)f χ χ ˜f˜ 1 ⊗ φ((dj∞ )G)Γ(j =χe ˜ α|x| f˜ 1 ⊗ e−α|x| φ(t(dj∞ )G)(N∞ + 1)−1 (N∞ + 1)(N0 + N∞ + 1)−1 ˘ t )dΓ(S/t)f (N + 1)−1 (N + 1)2 χ × Γ(j is of order t−µ . ˜ 0 Q we work in the representation F ⊗F = To handle the contribution due to D ˜ 0 Q maps ⊕n≥0 ⊕nk=0 Fn−k ⊗ Fk where Fk is the k-boson subspace of F . Since D Fn to ⊕nk=0 Fn−k ⊗ Fk one has ˜ 0 Qf χϕt = ˜f˜D ψt , χ
∞ n
˜ 0 Qβt,n αt,nk , Pn−k ⊗ Pk D
(59)
n=0 k=0
where αt,nk = Pn−k ⊗ Pk f˜χψ ˜ t , βt,n = Pn f χϕt and Pn is the orthogonal projection F → Fn . Note that χ ˜ = χ ˜ χ(N∞ ≤ n∞ ) for some n∞ large enough, and hence ˜ 0 Qβt,n | for each αt,nk = 0 for k > n∞ . Next we estimate |αt,nk , Pn−k ⊗ Pk D given n and k separately. To this end we identify Fn−k ⊗ Fk with a subspace of L(Rdn ). Then, on Fn , the operator ˘ t )dΓ(S) = Jnk Sn =: Snk Pn−k ⊗ Pk Γ(j
136
J. Fr¨ ohlich, M. Griesemer and B. Schlein
acts by multiplication with a function Jnk Sn where Sn (y, t) = by (22) 1/2 n j0 ⊗ . . . ⊗ j0 ⊗ j∞ ⊗ . . . ⊗ j∞ . Jnk =
k n−k factors
Ann. Henri Poincar´e
n i=1
S(yi , t) and (60)
k factors
˜ 0 Q is given by In terms of Snk the operator Pn−k ⊗ Pk D ˜ 0 Q = D02 Pn−k ⊗ Pk Γ(j ˘ t )dΓ(S) = D02 Snk Pn−k ⊗ Pk D
(61)
where D0 Snk = [iΩ, Snk ] + ∂Snk /∂t and Ω(k1 , . . . , kn ) = ni=1 ω(ki ). Let V = (y1 , . . . , yn )/t ∈ Rdn and let DV = ∂/∂t + V · ∇ denote the material derivative w.r. to V . In the appendix we show that
D02 Snk =(∇Ω − V ) · Snk (∇Ω − V )
+ (DV ∇Snk ) · (∇Ω − V ) + (∇Ω − V ) · (DV ∇Snk ) 1/2 2 2 n O(t−1−δ ) + DV Snk + n k for |y|/t ≤ 2λ (Lemma 40) and that 1/2 n n m α 2 yi /2t = nm+1 O(t−(1+m)+2(δ−|α|) ) Snk − Jnk DV ∂ k i=1
(62)
(63)
(Lemma 39). These are analogs of Lemma 9 and Lemma 8. The binomial factor in (62) and (63) stems from (60) and will be estimated by nk . From Eq. (63) we get for the last term of (62) n yi2 2 2 DV (Jnk Sn ) = Jnk DV Sn − = nk/2+3 O(t−3+2δ ) , (64) 2t i=1 where we used that DV f = 0 for any function f (y, t) which only depends on V and that DV2 (y 2 /2t) = 0. For the second and third term in Eq. (62), Eq. (63) shows that ± (DV ∇Snk ) · (∇Ω − V ) + (∇Ω − V ) · (DV ∇Snk ) ≤ t2−δ (DV ∇Snk )2 + t−2+δ (∇Ω − V )2 = O(t−2+δ )(n + V 2 )nk+4
(65)
Combining (59) and (61) with (62), (64) and (65) we get ∞ n
˜ 0 Qf χϕt ≤ ˜f˜D |αt,nk , (∇Ω − V ) · Snk (∇Ω − V )βt,n | ψt , χ n=0 k=0 ∞ n
+
n=0 k=0
(66)
αt,nk βt,n n
k+4
O(t
−1−ε
)
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
137
where ε = min(1 − δ, δ) > 0. Here nk+1 ≤ nn∞ +1 because αt,nk = 0 for k > n∞ . H¨older’s inequality in the form ∞ n
Ank Bn ≤
n=0 k=0
∞ n
1/2 A2nk
n=0 k=0
∞
1/2 (n + 1)Bn2
(67)
n=0
where Ank , Bn ≥ 0, will be used frequently in the following. It shows that the second term in (66) is bounded from above by f˜χψ ˜ t (N +1)n∞ +5/2 f χϕt O(t−1−ε ) which is integrable. To deal with the first term in (66) we use that
±Snk (y, t)
1/2 n ≤ const × n Sn
(y, t), k 2
for |y| ≤ 2λt and ut1−δ ≥ 2 by Lemma 38. This allows us to estimate the contribu
tion due to Snk χ(|V | ≤ 2λ). The contribution due to Snk χ(|V | > 2λ) is bounded −2 4 by O(t )(n + 1) αt,nk βt,n thanks to the space cutoff f . Together with (66), the Schwarz inequality, and αt,nk = 0 for k > n∞ , this implies that
|αt,nk , (∇Ω − V )·Snk (∇Ω − V )βt,n |
≤ |αt,nk , (∇Ω − V ) · Sn
(∇Ω − V )αt,nk |
1/2
× |βt,n , (∇Ω − V ) · Sn
(∇Ω − V )βt,n |
1/2
nn∞ /2+2
+ O(t−2 )(n + 1)4 αt,nk βt,n
1−δ ≥ 2. Now insert this bound into (66), use that (∇Ω − V ) · Sn
(∇Ω − V ) = for nut
i=1 (∇ω(ki ) − yi /t) · S (∇ω(ki ) − yi /t), sum over k and n, and apply (67) to see that ∞ n
|αt,nk , (∇Ω − V ) · Snk (∇Ω − V )βt,n |
n=0 k=0
12 ≤ f˜χψ ˜ t , dΓ (∇ω − v)S
(∇ω − v) ⊗ 1 + 1 ⊗dΓ (∇ω − v)S
(∇ω − v) f˜χψ ˜ t 1 × f χϕt , N (n∞ +5)/2 dΓ (∇ω − v) · S
(∇ω − v) N (n∞ +5)/2 f χϕt 2 + O(t−2 ) f˜χψ ˜ t (N + 1)9/2 f χϕt . This is integrable w. r. to t by Proposition 11. ˜ To prove the last statement it suffices to show that e−iHs W+ eiHs ϕ = W+ ϕ for all s ∈ R. This follows from ˜ ˜ −iHt e−iHs W+ eiHs ϕ − W+ ϕ = lim eiHt χ[ ˜ f˜Qf ]t+s ϕ t χe t→∞
138
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
if we prove that χ∂/∂t[ ˜ f˜Qf ]χ = O(t−1 ). The contributions due to ∂t f˜ and ∂t f are easily seen to be of order t−1 . As for ∂t Q note that, by Lemma 39, 1/2 ∂ 2 Snk 2 n Pn−k ⊗ Pk ∂t Q = D0 ∂t Snk = ∇Ω · ∇(∂t Snk ) + +n O(t−2 ) k ∂t2 1/2 n = n2 O(t−1 ) k ˜ for |y/t| ≤ 2λ. Use this, (67), and that k ≤ n∞ thanks to the energy cutoff χ. ˜ + W+ as we show in the next lemma. Some By construction of W+ , Wλ = Ω minor technical difficulties in its proof are due to the presence of the cutoffs and due to the unboundedness of I. Lemma 16 Suppose Wλ and W+ are defined as in Theorem 12 and Theorem 15. Then, under the assumptions of these theorems, ˜ ˜ + W+ . Wλ = s − lim eiHt Ie−iHt W+ = Ω t→∞
Proof. By definition of W+ we have, for all ϕ ∈ H, ˜
˜ 0 [Γ(j ˘ t )dΓ(S)]f χϕt + o(1), Ie−iHt W+ ϕ = I χ ˜f˜D
for t → ∞,
(68)
˜
where ϕt = e−iHt ϕ. Note here that e−iHt W+ ϕ is in the domain of I, for all t ∈ R, ˜ + ϕ, for any χ ∈ C ∞ (R), with χ = 1 on supp χ. From because W+ ϕ = χ (H)W 0 (68) it follows now, if we expand the free Heisenberg derivative, that ˜ ˘ t , djt )dΓ(S)f χϕt + I χ ˘ t )dΓ(dS)f χϕt + o(1) Ie−iHt W+ ϕ = I χ ˜f˜dΓ(j ˜f˜Γ(j ˘ t )f dΓ(dS)f χϕt + o(1). ˘ t , djt )dΓ(S)f χϕt + I χ ˜Γ(j = f I χd ˜ Γ(j
(69)
To prove the second equality commute f˜ to the left in the first term, using that, by Lemma 28, [χ, ˜ f˜] = O(t−1 ), and to the right in the second term, using that ˘ t ) = Γ(j ˘ t )f . Choose now χ1 ∈ C0∞ (R), with χ1 = 1 on supp χ, and supp χ1 ⊂ f˜Γ(j (−∞, Σ), and set χ1 = χ1 (H). Then χ1 χ = χ, and thus, by (69), ˜
˘ t , djt )dΓ(S)f χ1 χϕt + I χ ˘ t )f dΓ(dS)f χ1 χϕt + o(1) ˜ dΓ(j ˜ Γ(j Ie−iHt W+ ϕ = f I χ ˘ t , djt )χ1 dΓ(S)f χϕt + I χ ˘ t )χ1 f dΓ(dS)f χϕt + o(1), = f Iχ ˜ dΓ(j ˜ Γ(j (70) where, in the second equality, we used that [f dΓ(S)f, χ1 ]χ = O(t−1 ), for t → ∞, which easily follows, expanding χ1 in a Helffer-Sj¨ ostrand integral, by Hypothesis (H3), by Lemma 28 and because, by assumption, eα|x| χ is a bounded operator, for some α > 0. Below we will show that ˘ t ) χ1 = χ + o (1) and Iχ ˜ Γ(j ˘ t , djt ) χ1 = o (t−1 ). Iχ ˜ dΓ(j
(71) (72)
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
139
Inserting these two equations in (70) it follows, since dΓ(S/t)f χ is uniformly bounded in t, ˜
Ie−iHt W+ ϕ = χf dΓ(dS)f χe−iHt ϕ + o(1) = e−iHt Wλ ϕ + o(1), which proves the lemma. It only remains to prove (71) and (72). We begin proving ˘ t ) = 1H (because, by construction of jt , j0,t + (71). To this end we note that I Γ(j j∞,t = 1) and thus, for any n ∈ N, ˘ t )χ + IE(n,∞) (N∞ )Γ(j ˘ t )χ, ˘ t )χ = IE[0,n] (N∞ )Γ(j χ = I Γ(j
(73)
where E∆ denotes the characteristic function of the set ∆. Now we claim that the norm of the second term on the r.h.s. of the last equation can be made arbitrarily small by choosing n sufficiently large. This follows because ˘ t )χ = IE(n,∞) (N∞ )Γ(j ˘ t )E(n,∞) (N )χ, IE(n,∞) (N∞ )Γ(j ˘ t ) ≤ 1 for all n ∈ N, that which implies, since IE(n,∞) (N∞ )Γ(j ˘ t )χ ≤ IE(n,∞) (N∞ )Γ(j ˘ t ) E(n,∞) (N )(N + 1)−1 (N + 1)χ
IE(n,∞) (N∞ )Γ(j C . ≤ n+2 (74) On the other hand the first term on the r.h.s. of (73) can be written as ˘ t )χχ1 ˘ t )χ = I(N0 + 1)−n E[0,n] (N∞ )(N0 + 1)n Γ(j IE[0,n] (N∞ )Γ(j ˘ t )χ1 + o(1), = I(N0 + 1)−n E[0,n] (N∞ )(N0 + 1)n χ ˜Γ(j
(75)
where the second equality follows because of Lemma 32, and because I(N0 + 1)−n E[0,n] (N∞ ) is a bounded operator (see Lemma 1). Now, for n ∈ N sufficiently large, E[0,n] (N∞ )χ ˜ = χ. ˜ This remark, together with (73), (74) and (75) shows ˘ t )χ1 < ε, for t sufficiently large. This proves (71). that, for any ε > 0, χ − I χ ˜Γ(j ˘ t , djt ) = D0 (I Γ(j ˘ t )) = D0 (1H ) = 0, Eq.(72) follows in a very similar way by IdΓ(j −1 ˘ using that IE(n,∞) (N∞ )dΓ(jt , djt )(N + 1) ≤ const and applying Lemma 33 (see Appendix D). So far the positive parameter u in the construction of W+ was arbitrary. The next lemma now shows that, by choosing u small, we can neglect the possibility to find zero “escaping photons” in states in the range of W+ . This will be important in the proof of asymptotic completeness, Theorem 19. Lemma 17 Assume Hypotheses (H1), (H2) and (H5) hold, and let the Deift-Simon wave operator W+ be defined as in Theorem 15. Then
(1 ⊗ E{0} (N ))W+ ≤ 2 u2 (N + 1)1/2 χ(H) 2 .
140
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e ˜
Proof. For ϕ ∈ H and ψ ∈ H⊗F we define ϕt = e−iHt ϕ respectively ψt = e−iHt ψ. Then by definition of W+ , we have ˜ ˘ t )dΓ(S) f e−iHt χϕ ˜ 0 Γ(j ˜f˜D ψ, W+ ϕ = lim ψ, eiHt χ t→∞ (76) ˜ Γ(j ˘ t )dΓ(S) f χϕt . = lim ψt , χ ˜f˜D t→∞
The last equality follows because, by assumption, the operator eα|x| χ is bounded for some α > 0 and because the norm of the operator ˘ t )dΓ(S) − Γ(j ˘ t )dΓ(S)iφ(G) f (N + 1)−2 (77) e−α|x| (iφ(G) ⊗ 1)Γ(j tends to 0, as t → ∞. To see this write the operator in (77) as ˘ t )iφ(G) dΓ(S)f (N + 1)−2 ˘ t ) − Γ(j e−α|x| (iφ(G) ⊗ 1)Γ(j ˘ t )e−α|x| [iφ(G), dΓ(S)]f (N + 1)−2 . + Γ(j
(78)
Now the operator in the first line equals, by (20), − ie−α|x| {φ((j0,t − 1)G) ⊗ 1 + 1 ⊗ φ(j∞,t G)} (N0 + N∞ + 1)−1 ˘ t ) dΓ(S)f (N + 1)−1 , × Γ(j and tends to 0 as t → ∞. This follows because dΓ(S)f (N + 1)−1 = O(t), while, by ˘ t ) is of order O(t−µ ), for some µ > 1. Hypothesis (H5), the factor on the left of Γ(j To handle the operator in the second line of (78) use that [iφ(G), dΓ(S)] = φ(−iSG), and that, by (H5), sup e−α|x| φ(−iSGx )(N + 1)−1 = sup e−α|x| φ(χ(|y| ≥ tδ )iSGx )(N + 1)−1
x
x
= O(t−1+δ ). This implies that also the term in the second line of (78) tends to zero as t → ∞. From (76) it follows now, because terms involving the Heisenberg derivative of f , or f˜ give, by Lemma 28, a vanishing contribution in the limit t → ∞, that d ˘ t )dΓ(S)f χϕt = lim 1 ψt , χ ˘ t )dΓ(S)f χϕt ψt , χ ˜f˜Γ(j ˜f˜Γ(j t→∞ t dt 1 ˘ ∗ = lim Γ(j ˜ t , f dΓ(y 2 /2t)f χϕt , t ) χψ t→∞ t (79)
ψ, W+ ϕ = lim
t→∞
where, in the last equality, we used Lemma 8 to replace S by y 2 /2t. Consider in particular the case ψ = (1 ⊗ E{0} (N ))ψ; then there exists some α ∈ H with ψ = α ⊗ Ω, and ˘ t )∗ χψ ˜ t = I (Γ(j0,t ) ⊗ Γ(j∞,t )) ((χαt ) ⊗ Ω) = Γ(j0,t )χαt . Γ(j
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
141
For such ψ = α ⊗ Ω, (79) reads ψ, W+ ϕ = lim
t→∞
1 αt , χf Γ(j0,t )dΓ(y 2 )f χϕt . 2t2
(80)
Now we note that |αt , χf Γ(j0,t )dΓ(y 2 )f χϕt | = |αt , χf dΓ(j0,t , j0,t y 2 )f χϕt | ≤ (N + 1)1/2 χαt (N + 1)1/2 χϕt j0,t y 2
≤ 4u t α ϕ (N + 1) 2 2
1/2
(81)
2
χ ,
where, in the last step, we used that j0,t y 2 ≤ sup|y|≤2u t y 2 ≤ 4t2 u2 . Since
α = ψ , it follows from (81) and (80) that |ψ, (1 ⊗ E{0} (N ))W+ ϕ| ≤ 2u2 (N + 1)1/2 χ(H) 2 ψ ϕ .
9 Asymptotic Completeness As explained in the introduction we prove asymptotic completeness by induction in the energy measured in units of m. The first step is the following, essentially trivial lemma. The idea is that AC on Ran Eη (H) as explained in the introduction ˜ in Eq. (2), implies the same property for Ie−iHt on Ran Eη (H) ⊗ F, the photons from F merely contributing to the asymptotically free radiation. Lemma 18 Assume that hypotheses (H1), (H2) and (H4) are satisfied, and let the ˜ + and Ω+ be defined as in Lemma 6 and in Theorem 7, respecwave operators Ω ˜ tively. Suppose Ran Ω+ ⊃ Eη (H)H and µ < Σ. Then for every ϕ ∈ Ran Eµ (H) ˜ there exists a vector ψ ∈ Ran Eµ (H) such that ˜ + (Eη (H) ⊗ 1)ϕ = Ω+ ψ. Ω ˜ H ˜ then ψ ∈ E∆ (H) ˜ H. ˜ If ϕ ∈ E∆ (H) ˜ H ˜ can be approximated by a sequence Proof. By Lemma 30 every given ϕ ∈ Eµ (H) ˜ ˜ of vectors ϕn ∈ Eµ (H)H which are finite linear combinations of vectors of the form (82) γ = α ⊗ a∗ (h1 ) . . . a∗ (hn )Ω, n where α = Eλ (H)α and λ + i=1 Mi < µ, with Mi = sup{|k| : hi (k) = 0}. Let ˜ be of the form (82). Then γ∈H ˜
˜
Ie−iHt (Eη (H) ⊗ 1) γ = Ie−iHt Eη (H)α ⊗ a∗ (h1 ) . . . a∗ (hn )Ω = a∗ (h1,t ) . . . a∗ (hn,t ) e−iHt Eη (H)α.
(83)
142
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
˜ and we may assume β = Eη (H)β ˜ By assumption Eη (H)α = Ω+ β for some β ∈ H by the intertwining relation for Ω+ . From (83) it follows that ˜
Ie−iHt (Eη (H) ⊗ 1) γ = a∗ (h1,t ) . . . a∗ (hn,t )e−iHt Ω+ β ˜
= a∗ (h1,t ) . . . a∗ (hn,t )Ie−iHt (PB ⊗ 1) β ˜ + a∗ (h1,t ) . . . a∗ (hn,t ) e−iHt Ω+ β − Ie−iHt (PB ⊗ 1) β , (84) where PB denotes the orthogonal projector onto Hpp (H). After inserting a factor (N + 1)−n/2 (N + 1)n/2 in the second factor on the r.h.s. of last equation, just in front of the braces, we get ˜
˜
Ie−iHt (Eη (H) ⊗ 1) γ−Ie−iHt (PB ⊗ 1) (1 ⊗ a∗ (h1 ) . . . a∗ (hn ))β
≤ a∗ (h1,t ) . . . a∗ (hn,t )(N + 1)−n/2
˜ × (N + 1)n/2 e−iHt Ω+ β − Ie−iHt (PB ⊗ 1)β . The first factor on the r.h.s. is bounded by a finite constant, uniformly in t. The second factor converges to zero as t → ∞. To see this use that it stays bounded for ˜ all integers n, a fact which follows from the boundedness of (H + i)n/2 Ω+ Eµ (H) n/2 ∗ ∗ ˜ ˜ ˜ and of (N + 1) I Eµ (H). Since (1 ⊗ a (h1 ) . . . a (hn ))β ∈ Eµ (H)H it follows that ˜ + (Eη (H) ⊗ 1)γ = Ω+ (1 ⊗ a∗ (h1 ) . . . a∗ (hn ))β. Ω Hence to each ϕn , as defined at the very beginning, there exists a vector ψn ∈ ˜ such that Ω ˜ + (Eη (H) ⊗ 1)ϕn = Ω+ ψn . The left hand side converges to Eµ (H)H ˜ Ω+ (Eη (H) ⊗ 1)ϕ as n → ∞, and hence the right side converges as well. Since Ω+ is isometric on Hpp (H) ⊗ F it follows that (PB ⊗ 1)ψn is Cauchy and hence has ˜ Thus Ω ˜ + (Eη (H) ⊗ 1)ϕ = Ω+ ψ which proves the first part a limit ψ ∈ Eµ (H)H. ˜ + and of the lemma. The second part follows from the intertwining relations for Ω Ω+ . Theorem 19 (Asymptotic Completeness) Assume hypotheses (H1) through (H5) are satisfied. Then Ran Ω+ ⊃ E(−∞,Σ) (H)H. Proof. The proof is by induction. We show that Ran(Ω+ ) ⊃ E(−∞,km) (H)E(−∞,Σ) (H)H
(85)
holds for all integers k with (k − 1)m ≤ Σ. For k < 0 and |k| large enough (85) is trivially correct because H is bounded from below. Hence we may assume (85) holds for k = n − 1. To prove it for k = n it suffices to show that Ran Ω+ ⊃ E∆ (H)H
(86)
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
143
for compact intervals ∆ ⊂ (−∞, nm) ∩ (−∞, Σ)\S because, by Theorem 4, the union of such subspaces is dense in E(−∞,nm) (H)E(−∞,Σ) (H)Hpp (H)⊥ , because, clearly, RanΩ+ ⊃ E(−∞,Σ) (H)Hpp (H) and because RanΩ+ is closed. Now choose χ ∈ C0∞ (R) such that χ = 1 on ∆ and supp χ ⊂ (−∞, Σ)\S. Let λ, u > 0 and define Wλ and W+ in terms of χ, λ and u as in Theorem 12 and 15. Define moreover the asymptotic observable W = limλ→∞ Wλ as in Proposition 13. By Proposition 14, the operator W : E∆ (H)H → E∆ (H)H is onto and hence for every given ϕ = E∆ (H)ϕ there exists a ψ ∈ Ran E∆ (H) such that ϕ = W ψ. Given ε > 0 we pick u, λ small, respectively large enough so that ˜ + χ(N∞ = 0)W+ ψ < ε,
Ω
(W − Wλ )ψ < ε,
(87)
˜ Then by ˜ + E∆ (H). by using Lemma 17, Proposition 13, and the boundedness of Ω Lemma 16 ˜ + W+ ψ = Ω ˜ + χ(N∞ > 0)W+ ψ + Ω ˜ + χ(N∞ = 0)W+ ψ. Wλ ψ = Ω
(88)
The vector χ(N∞ > 0)W+ ψ has at least one boson in the outer Fock space and thus an energy of at most (n − 1)m in the inner one. More precisely χ(N∞ > 0)W+ ψ = χ(H < (n − 1)m) ⊗ χ(N > 0)W+ ψ. ˜ H ˜ such that Hence by induction hypothesis and Lemma 18 there exist γ ∈ E∆ (H) ˜ + χ(N∞ > 0)W+ ψ = Ω+ γ. Ω This equality together with (87) and (88) shows that ˜ + χ(N∞ > 0)W+ ϕ < 2ε
ϕ − Ω+ γ = W ψ − Ω which proves the theorem.
10 Massless Photons This section is devoted to the case where the bosons are massless photons, but the soft modes do not interact with the particles (electrons). That is H = K ⊗ 1 + 1 ⊗ dΓ(|k|) + φ(G) = H0 + φ(G)
(89)
and (IR)
Gx (k) = 0
if
|k| < m
for some m > 0. As before, we assume that Gx ∈ h for each x ∈ X and that supx Gx < ∞. Then φ(G)(H0 + 1)−1/2 is again bounded and hence H is selfadjoint on D(H0 ). The key idea is to compare H with the modified Hamiltonian Hmod = K ⊗ 1 + 1 ⊗ dΓ(ω) + φ(G) where K and G are as above but the dispersion ω(k) = |k| is modified for |k| < m. We choose ω in such a way that (H1) is satisfied (with m/2 instead of m) and ω(k) = |k| for |k| ≥ m.
144
10.1
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
Asymptotic Completeness
˜ = Asymptotic completeness for H is essentially a corollary of Theorem 19. Let H H ⊗ 1 + 1 ⊗ dΓ(|k|) and let Ω+ and Ω+,mod be defined in terms of H, Hmod , and ˜ mod = Hmod ⊗ 1 + 1 ⊗ dΓ(ω). H Theorem 20 (Asymptotic Completeness) Assume (IR), (H2), and (H5) for the ˜ H ˜ and system defined by (89). Then the wave operator Ω+ exists on E(−∞,Σ) (H) Ran(Ω+ ) ⊃ E(−∞,Σ) (H)H. Proof. We split the Fock space into two Fock spaces, one with interacting photons the other one with non-interacting photons. Henceforth the subindices i and s refer to interacting and soft respectively. Let Ki = {k ∈ Rd : |k| ≥ m}
hi = L2 (Ki )
(90)
Ks = {k ∈ R : |k| < m}
hs = L (Ks ).
(91)
d
2
Then h = hi ⊕ hs and correspondingly F (h) is isomorphic to F (hi ) ⊗ F(hs ) with an isomorphism U as given in Section 2.5. By assumption on G and ω and by (17) U Hmod U ∗ = Hi ⊗ 1 + 1 ⊗ dΓ(ωs ) U HU ∗ = Hi ⊗ 1 + 1 ⊗ dΓ(|k|s ) with respect to the factorization H = (L2 (X) ⊗ F(hi )) ⊗ F(hs ). Here Hi = K ⊗ 1 + 1 ⊗ dΓ(|k|) + φ(G) on Hi := L2 (X) ⊗ F(hi ). It follows that Hmod and H have the same eigenvectors, they are of the form U ∗ (ϕi ⊗ Ωs ) where ϕi is an eigenvector of Hi . Hence PB (H) = PB (Hmod ). Furthermore ˜
˜
eiHt Ie−iHt = eiHmod t e−idΓ(ω−|k|)t Iei[dΓ(ω−|k|)⊗1+1⊗dΓ(ω−|k|)]t e−iHmod t ˜
= eiHmod t Ie−iHmod t because Hmod and dΓ(ω − |k|) commute. This shows that Ω+ = Ω+,mod and hence ˜ mod ≤ µ)H ˜ and that Ran Ω+ ⊃ χ(Hmod ≤ µ)H for all that Ω+ exists on χ(H µ < Σ, by Theorem 19. To reformulate these results in terms of Hi we consider U ⊗ U as a map from ˜ mod ≤ F ⊗ F to Fi ⊗ Fi ⊗ Fs ⊗ Fs . Then U Ω+ (U ∗ ⊗ U ∗ ) exists on (U ⊗ U )χ(H ˜ ˜ ˜ ˜ ˜ µ)H ⊃ χ(Hi ≤ µ)Hi ⊗ Ωs ⊗ Ωs , where Hi = Hi ⊗ 1 + 1 ⊗ dΓ(|k|) on Hi = Hi ⊗ Fi , and Ran U Ω+ (U ∗ ⊗ U ∗ ) ⊃ U χ(Hmod ≤ µ)H ⊃ χ(Hi ≤ µ)Hi ⊗ Ωs . Furthermore U Ω+ (U ∗ ⊗ U ∗ ) = Ωi ⊗ (Is (χ(Ns = 0) ⊗ 1s )) ˜
where Ωi = s − limt→∞ eiHi t Ii e−iHi t (PB (Hi ) ⊗ 1). In fact U I(U ∗ ⊗ U ∗ ) = Ii ⊗ Is and U PB (H)U ∗ = PB (Hi ) ⊗ χ(Ns = 0). It follows that U Ω+ (U ∗ ⊗ U ∗ ) exists ˜ i ≤ µ)H ˜ i ⊗ Fs ⊗ Fs ⊃ (U ⊗ U )χ(H ˜ ≤ µ)H ˜ and that its range contains on χ(H ˜ ≤ µ)H ˜ and by the χ(Hi ≤ µ)Hi ⊗ Fs ⊃ U χ(H ≤ µ)H. Hence Ω+ exists on χ(H ˜ ˜ intertwining relation for Ω+ , the range of Ω+ |`χ(H ≤ µ)H contains χ(H ≤ µ)H. The theorem now follows because µ < Σ was arbitrary and Ω+ = 1.
Vol. 3, 2002
10.2
Asymptotic Completeness for Rayleigh Scattering
145
Relaxation to the Ground State
With the help of AC established in the last section we next show that states below the ionization threshold relax to the ground state in the sense (5) under the dynamics generated by the Hamilton operator (89). We begin by summarizing results due to Bach et al. [BFSS99], [BFS98] on the point spectrum of H that are needed in this section. No infrared cutoff is assumed in the following discussion. Consider the Hamilton operator Hg = K ⊗ 1 + 1 ⊗ dΓ(|k|) + gφ(G) where K = −∆ + V on L2 (Rn ) and V is operator-bounded w.r.t. −∆ with relative bound zero. This assumption allows for typical N -body Schr¨ odinger operators [HS00]. We assume that 1 2 + 1 dk < ∞ (92) sup |Gx (k)| |k| x to ensure self adjointness of Hg on D(Hg=0 ). Following [BFSS99] we furthermore assume that |(k · ∇k )2 Gx (k)|2 <∞ (H6) sup(1 + |x|2 )−M/2 dk |k| x for some M ≥ 0. All excited bound states of Hg will be unstable if their life time as given by Fermi’s Golden Rule is finite. To state this condition we have to introduce some notation. Suppose E0 < E1 < · · · < inf σess (K) are the isolated eigenvalues of K with finite multiplicity. Let mj be the multiplicity of the eigenvalue Ej , and let ϕj,l ∈ L2 (Rn ), for l = 1, . . . mj be an orthonormal base of the eigenspace of K to the eigenvalue Ej . Then for each 0 ≤ i < j and for each k ∈ Rd we define the mi × mj matrix Aij (k) := ϕi,r , Gx (k)ϕj,s . Now, for each j ≥ 0 we define the mj × mj matrix Γj = (93) A∗ij (k)Aij (k)δ(ω(k) − Ej + Ei )dk . i:i<j
The eigenvalues of this matrix are then the resonance widths in second order perturbation theory corresponding to the eigenvalues Ej . To show that Hg has no eigenvalues in neighborhoods of the eigenvalues of the unperturbed Hamiltonian H0 , we therefore need the following assumption: (H7) Fermi Golden Rule. For each j ≥ 1 we have Γj > 0. The following theorem summarizes results from [BFSS99] and [BFS98].
146
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
Theorem 21 i) (Exponential decay, [BFS98]) Suppose µ < inf σess (K) and ε > 0. Then there exists a constant M = M (ε) such that
eα|x|χ(Hg < µ) < M for all α, g with inf σess (K) − µ − α2 − g 2 supx dk |Gx (k)|2 /|k| > ε. ii) (Existence and uniqueness of the ground state, [BFS98, GLL00]) If inf σ(K) < inf σess (K) − g 2 supx dk |Gx (k)|2 /|k| and Gx (−k) = Gx (k) then Eg := inf σ(Hg ) is a non-degenerate eigenvalue of Hg . Moreover if ψg is an eigenvector of Hg corresponding to the eigenvalue Eg , then
Pψg − Pϕ0 ⊗Ω → 0,
as
g → 0,
(94)
where Pψg and Pϕ0 ⊗Ω denote the orthogonal projections onto the spaces spanned by the ground states ψg and ϕ0 ⊗ Ω, respectively. iii) (Absence of excited eigenstates, [BFSS99]) Assume (H6) and (H7). Set ∆ = [E0 + ε, µ], for fixed ε > 0 and µ < inf σess (K). Then σpp (Hg ) ∩ ∆ = ∅ for g > 0 sufficiently small. If the infrared cutoff (IR) is imposed, then assumption (92) simplifies to supx Gx < ∞ and all results of the above theorem then hold for g sufficiently small. Note that m must be small in order for (H7) to hold, because transitions between energy levels Ej with separation less than m are suppressed by the infrared cutoff. Next we prove absence of eigenvalues above and close to Eg for g small, and assuming (IR). As in [BFSS99] we argue by contradiction and prove a virial theorem as well as the positivity of [iH, A] on a spectral interval (Eg , E1 − ε] and for a suitable conjugate operator A. Lemma 22 (Virial Theorem) Assume (IR) and (H3). If Hϕ = Eϕ and E < inf σess (K) − g 2 supx dk |Gx (k)|2 /|k| then ϕ ∈ D(N ) and ϕ, (N − gφ(iaG)) ϕ = 0. Proof. With the notation of the proof of Theorem 20 the eigenvector ϕ is of the form U ∗ (ϕi ⊗ Ωs ). It follows that ϕ is an eigenvector of Hmod and thus in D(Hmod ) ⊂ D(N ). By Theorem 21, part i), eα|x| ϕ ∈ H for some α > 0 and hence Lemma 3 applies to ϕ and Hmod . This shows that ϕ, [dΓ(|∇ω|2 ) − gφ(iaG)]ϕ = 0 which proves the theorem because, by the form of ϕ, dΓ(|∇ω|2 )ϕ = N ϕ.
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
147
Theorem 23 (Positive commutator) Assume (IR), (H0), and (H3). Set ∆ = (Eg , E1 − ε], for some fixed ε > 0 (here E1 is the first point in the spectrum of K above inf σ(K)). Then there is a constant C > 0 such that E∆ (Hg ) (N − g φ(iaG)) E∆ (Hg ) ≥ CE∆ (Hg ), for all g > 0 sufficiently small. In particular, by Lemma 22, it follows that Hg has no eigenvalues in ∆, if g > 0 is small enough. Proof. Using N ≥ 1 − 1 ⊗ PΩ we get E∆ (Hg ) (N − g φ(iaG)) E∆ (Hg ) ≥ E∆ (Hg ) (1 − 1 ⊗ PΩ − g φ(iaG)) E∆ (Hg ) = E∆ (Hg ) − E∆ (Hg ) (1 ⊗ PΩ ) E∆ (Hg ) − gE∆ (Hg )φ(iaG)E∆ (Hg ) −α|x| α|x| 1/2 ≥ E∆ (Hg ) 1 − g sup e
iaGx E∆ (Hg )e
(Ni + 1) E∆ (Hg )
x
− E∆ (Hg ) (1 ⊗ PΩ ) E∆ (Hg ) (95)
where Ni = |k|>m dk a∗ (k)a(k) is the operator counting the number of interacting bosons (which is bounded w.r.t. Hg ), and where PΩ is the orthogonal projector onto Ω. By Hypothesis (H3), and because aGx ≤ const ( yGx + Gx ), the number in the parenthesis in the first term on the r.h.s. of the last equation is larger than C, for any C < 1, if g > 0 is small enough. It remains to show that the last term in (95) converges to 0 as g → 0. To do this we split it into two parts, according to E∆ (Hg ) (1 ⊗ PΩ ) E∆ (Hg ) = E∆ (Hg ) E{E0 } (K) ⊗ PΩ E∆ (Hg ) (96) + E∆ (Hg ) E[E1 ,∞) (K) ⊗ PΩ E∆ (Hg ), where K is the particle Hamiltonian, and E0 and E1 are its ground state energy and its first exited eigenvalue. Since Pψg E∆ (Hg ) = 0 the first term in Eq. (96) can be written as E∆ (Hg ) E{E0 } (K) ⊗ PΩ E∆ (Hg ) = E∆ (Hg ) Pϕ0 ⊗Ω − Pψg E∆ (Hg ), (97) which converges to zero by Theorem 21, part ii). Consider now the second term on the r.h.s. of (96). Choose χ ∈ C∞ 0 (R), such that 0 ≤ χ ≤ 1, χ(s) = 0 if s > E1 − ε/2, and χ(s) = 1 if s ∈ ∆ (this is a smooth version of the characteristic function E∆ ). Then we have on the one hand χ(Hg )E∆ (Hg ) = E∆ (Hg ) and on the other hand E[E1 ,∞) (K) ⊗ PΩ χ(H0 ) = 0. Thus we get E[E1 ,∞) (K) ⊗ PΩ E∆ (Hg ) = E[E1 ,∞) (K) ⊗ PΩ (χ(Hg ) − χ(H0 )) E∆ (Hg ). (98)
148
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
Now if χ ˜ is an almost analytic extension of χ, in the sense of the Helffer–Sj¨ostrand functional calculus (see Appendix A.2), then we have g χ(Hg ) − χ(H0 ) = − ˜ − Hg )−1 φ(G)(z − H0 )−1 . dxdy ∂z¯χ(z π This implies, since χ has a compact support, that
χ(Hg ) − χ(H 0 ) ≤ Cg, for some constant C > 0, and thus, by (98), that E[E1 ,∞) (K) ⊗ PΩ E∆ (Hg ) → 0 as g → 0. This completes the proof of the theorem. The last theorem, together with Theorem 21, proves the following corollary. Corollary 24 Assume Hypotheses (IR), (H3), (H6), and (H7). If µ < inf σess (K), then σpp (Hg ) ∩ (Eg , µ) = ∅, for all g > 0 sufficiently small. With the help of this corollary and Theorem 20 we next prove relaxation to the ground state in the sense of the following theorem. To define the algebra of observables let A denote the C ∗ algebra generated by all Weyl operators W (h) = exp(iφ(h)), with h ∈ L2 (Rd , dk). By taking tensor products of operators in A with bounded operators acting on the Hilbert space Hel = L2 (Rn , dx) of the electrons ˜ one obtains a C ∗ algebra, which we denote by A. Theorem 25 (Relaxation to the ground state) Assume Hypotheses (IR), (H0), and (H3) through (H7). Choose µ < inf σess (K). Then, for sufficiently small values of the coupling constant g > 0, the Hamiltonian Hg exhibits the property of relaxation to the ground state for states with energy less than µ. This means that, if g > 0 is sufficiently small, then, for all A ∈ A˜ and for all ψ ∈ Ran(H ≤ µ), we have lim ψt , Aψt = ψg , Aψg ψ, ψ,
t→∞
(99)
where ψt = e−iHt ψ and ψg denotes the ground state of Hg . Proof. Since the C ∗ algebra A is generated by the Weyl-operators W (h) = eiφ(h) , and because products of Weyl-operators are again Weyl-operators (up to some unimportant phase) it is enough if we prove (99) for A = B ⊗ W (h), where B is a bounded operator on Hel and h ∈ S(Rd ). By Theorem 20 we know that the system we are considering is asymptotically complete. On the other hand we know, from Corollary 24, that the ground state ψg is the only eigenstate of Hg , which lies in the range of the spectral projection χ(Hg ≤ µ). These two results imply that each ψ ∈ Ranχ(Hg ≤ µ) can be written as limit of a sequence of finite linear combinations of states like a∗+ (f1 ) . . . a∗+ (fm )ψg , with fi ∈ S(Rd ). Since we are dealing only with bounded operators, it follows that it is enough to prove (99) for A = B ⊗ W (h) and for ψ=
N i=1
ci a∗+ (f1i )a∗+ (f2i ) . . . a∗+ (fni i )ψg .
(100)
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
149
In this case we have lim e−iHt ψ, Ae−iHt ψ
t→∞
N
=
c¯i cj lim e−iHt
i,j=1 N
=
ni l=1
t→∞
c¯i cj lim t→∞
i,j=1
ni l=1
a∗+ (fli )ψg , Ae−iHt nj m=1
i a∗ (fl,t )ψg , A
nj m=1
j a∗+ (fm )ψg
(101)
j a∗ (fm,t )ψg ,
where we used the definition of the asymptotic fields a∗+ (h). Notice now that the ground state ψg of Hg is in the domain of arbitrary powers of the field-Hamiltonian dΓ(|k|). Moreover we know that the Weyl operators leave D(dΓ(|k|)n ) invariant. This follows by the commutation relations [W (h), dΓ(|k|)] = −φ(i|k|h)W (h) + 1/2 Re (|k|h , h) W (h). These remarks imply that we can rewrite the limit in the r.h.s. of the last equation as lim
t→∞
ni l=1
i a∗ (fl,t )ψg , A
= lim ψg , t→∞
nj m=1 ni l=1
= lim A∗ ψg ,
j a∗ (fm,t )ψg
i a(fl,t ) (B ⊗ W (h))
nj m=1
j a∗ (fm,t )ψg
nj m=1
j a∗ (fm,t )ψg (102) nj ni i ∗ j + lim ψg , B ⊗ m=1 a (fm,t )ψg . l=1 a(fl,t ), W (h)
t→∞
ni l=1
i a(fl,t )
t→∞
If we expand the commutator in the last equation, we get a sum of ni terms. Each i , h)L2 . Now, since we have assumed that of these terms contains a contraction (fl,t i d fl , hr ∈ S(R ), we have i (fl,t , hr ) = dkf i (k)hr (k)ei|k|t → 0 as t → ∞, l and thus the second term on the r.h.s. of (102) vanishes. To handle the first term n on the r.h.s. of (102) we use that, by Lemma 26, limt→∞ l=1 a(fl,t )ψg = 0. Assuming ni ≥ nj , this implies that the first term on the r.h.s. of (102) vanishes if ni > nj , and that lim
t→∞
ni
l=1
i a(fl,t )
nj
m=1
j a∗ (fm,t )ψg = lim ψg ψg , t→∞
i a(fl,t )
ni i l=1 a+ (fl ) ni ∗ i l=1 a+ (fl )ψg ,
= ψg ψg , = ψg
ni l=1
nj m=1
j a∗ (fm,t )ψg
nj ∗ j m=1 a+ (fm )ψg nj ∗ j m=1 a+ (fm )ψg ,
for all ni ≥ nj . Using the antisymmetry of the inner product it is analogously proven that the r.h.s. of (102) vanishes if ni < nj .
150
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
Thus, for arbitrary ni , nj , nj ni
j i )ψg ,A a∗ (fm,t )ψg lim a∗ (fl,t
t→∞
m=1
l=1
= A∗ ψg , ψg
ni l=1
a∗+ (fli )ψg ,
nj m=1
j a∗+ (fm )ψg .
The theorem now follows if we insert this result into (101) and compute the sum over i, j. Lemma 26 Suppose N n > 0 and ϕ ∈ D((Hg + i)n/2 ). Then, if h1 , . . . , hn ∈ L2 (Rd , dk) with hi 2ω = dk |hi (k)|2 (1 + 1/|k|) < ∞, we have lim
t→∞ n i=1
Proof. Notice that
that
n
a(hj,t )ϕ ≤ (
j=1
n
a∗ (hi )(dΓ(|k|) + 1)−n/2 ≤ C
n
a(hj,t ) −
j=1
+
a(hj,t )ϕ = 0.
(103)
j=1
n
n i=1
hi ω . This implies
˜ j,t ))(H + i)−n/2
(H + i)n/2 ϕ
a(h
j=1
n
˜ j,t )ϕ
a(h
j=1
≤
n ˜ j,t ) . . . a(h ˜ n,t ) (H + i)−n/2
a(h1,t ) . . . a(hj,t − h j=1
× (H + i)
n/2
ϕ +
n
˜ j,t )ϕ
a(h
j=1
≤C
n
˜ j ω . . . hn ω +
h1 ω . . . hj − h
j=1
n
˜ j,t )ϕ , a(h
j=1
where we used that (dΓ(|k|) + 1)n/2 (Hg + i)−n/2 < ∞ (see [FGS00]). Because of the last equation it is enough to prove (103) when hj ∈ C0∞ (Rd \{0}). In this case we have M = minj inf{|k| : hj (k) = 0} > 0, and there exists f ∈ C ∞ (Rd ) with f (k) = 0 if |k| < M/2, and f (k) = 1 if |k| ≥ M . Then, on the one hand, dΓ(f ) is bounded w.r.t. Hg (and higher powers of dΓ(f ) are bounded w.r.t. corresponding powers of Hg ). This implies that ϕ ∈ D((dΓ(f ) + 1)n/2 ). On the other hand n −n/2 . Thus, with ψ = (dΓ(f ) + 1)+n/2 ϕ, j=1 a(hj,t ) is bounded w.r.t. (dΓ(f ) + 1) we have n n
a(hj,t )ϕ = a(hj,t )(dΓ(f ) + 1)−n/2 ψ. j=1
j=1
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
151
Since nj=1 a(hj,t )(dΓ(f ) + 1)−n/2 is uniformly bounded in t, it is enough if we show that the r.h.s. of the last equation converges to 0, as t → ∞ for ψ = α ⊗ a∗ (f1 ) . . . a∗ (fm )Ω. To this end we write n
j=1
a(hj,t )(dΓ(f ) + 1)− 2
n
m
a∗ (fi )Ω = (dΓ(f ) + 1 + n)− 2
n
i=1
= (dΓ(f ) + 1 + n)
a(hj,t )
j=1
−n 2
n
n
a(hj,t ) ,
j=1
m
m
a∗ (fi )Ω
i=1
a∗ (fi ) Ω.
i=1
Expanding the commutator, we find a sum of terms containing n contractions (hj,t , fi )L2 . The lemma now follows, because all these contractions converge to 0 as t → ∞.
10.3
QED in Dipole Approximation
As mentioned in the introduction, our methods can be extended to prove AC, as well as our further main results, for atoms described by “non-relativistic QED” in the dipole approximation. In this section we briefly explain how this is accomplished. For an introduction to the standard model of non-relativistic QED and for the justification of the dipole approximation we refer to [BFS98]. Here we merely show how this model fits into our general framework. We consider a non-relativistic electron interacting with the quantized radiation field in Coulomb gauge. (The generalization to N electrons is straightforward.) States of this system are described by vectors in the Hilbert space HT = Hat ⊗ FT , where Hat = L2 (R3 , dx), and FT is the bosonic Fock space over hT = {h ∈ L2 (R3 ; C3 )| k · h(k) = 0 a.e.}, the space of transversal photons. The Hamilton operator is HT = K ⊗ 1 + 1 ⊗ Hf + φ(G)
on HT ,
where K = −∆ + V is assumed to satisfy Hypothesis (H0) and Hf = dΓ(|k|). To describe QED in the dipole approximation, we set Gx (k) = κ(k)g(x)P (k)x
(104)
where x ∈ R3 is the position of the electron, and P (k), for k = 0, is the orthogonal projection onto the plane orthogonal to k ∈ R3 . That is, Pij (k) = δij − ki kj /k 2 . As before we assume κ ∈ C0∞ (R3 ), and κ(k) = 0 if |k| < m, for some m > 0. The factor g ∈ C0∞ (R3 ) is a space cutoff necessary to make Gx bounded as a function of x. From a physical point of view this simplification is legitimate, because the electrons are exponentially localized near the origin (see also [BFS98]). Note that Gx ∈ hT and hence φ(Gx ) is one operator in FT = F (hT ), not a triple of operators as Gx (k) ∈ C3 might suggest.
152
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
The above description of QED in Coulomb gauge is equivalent to the more traditional one in which a choice of polarization vectors ελ (k) ⊥ k, λ = 1, 2, is made and photons are described by functions in L2 (R3 ; C2 ). In fact, using 2 3 2 the polarization vectors an isomorphism L (R ; C ) → hT can be defined by h(k, λ) → λ=1,2 ελ (k)h(k, λ). Under this isomorphism the coupling function Gx (λ, k) = κ(k)g(x)ε∗λ (k) · x, used in the dipol approximation, is mapped to (104). The advantage of our description is that k → Gx (k), unlike Gx (λ, k), is (twice) differentiable as required by Hypothesis (H5). The problem with Gx (λ, k) is that the polarization vectors ελ (k) cannot even be chosen so that they are continuous as functions of k on the unite sphere (an easy application of a famous theorem due to H. Hopf). However the Hilbert space HT is not invariant under the action of many operators which are used in the proof of AC. To circumvent this problem we enlarge the one-boson space by adding hL = {h ∈ L2 (R3 ; C3 )|P (k)h(k) = 0 a.e.}, the space of longitudinal photons. These longitudinal photons do not interact with the electrons and will drop out again at the end. Let H = Hat ⊗ F where F is the bosonic Fock space over h = L2 (R3 ; C3 ) = hT ⊕ hL and let H = K ⊗ 1 + 1 ⊗ Hf + φ(G) on H, where K and G are defined as above. This Hamilton operator is of the form (89) and hypotheses (IR), (H5), and (H2) are satisfied by assumption on κ and K (a proof of (H2) is contained in [GLL00] and [BFS98], see also Theorem 21). Theorem 20 thus tells us that the wave operator ˜
Ω+ = s − lim eiHt Ie−iHt PB (H) ⊗ 1 t→∞
˜ < Σ)H ˜ and that AC holds in the form exists on χ(H ˜ < Σ)H ˜ = χ(H < Σ)H. Ω+ χ(H Using this and the fact that longitudinal photons do not interact with the electron it is now easy to establish AC for HT . To this end recall form Section 2.6 that F is isomorphic to FT ⊗ FL where FL = F (hL ). Hence we may identify F with FT ⊗ FL and F ⊗ F with FT ⊗ FT ⊗ FL ⊗ FL . Then H = HT ⊗ 1L + 1T ⊗ Hf,L on H = HT ⊗ HL , and ˜ = H ⊗ 1 + 1 ⊗ Hf H ˜ T ⊗ (1L ⊗ 1L ) + (1T ⊗ 1T ) ⊗ (Hf,L ⊗ 1L + 1L ⊗ Hf,L ) =H ˜ = H⊗F = H ˜ T ⊗ FL ⊗ FL, where H ˜ T = HT ⊗ FT . Furthermore I = IT ⊗ IL on H as an operator from F ⊗ F = FT ⊗ FT ⊗ FL ⊗ FL to F = FT ⊗ FL , where IT
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
153
and IL are defined in a way analogous to I. Let PL denote the projection onto the vacuum ΩL of FL . Then PB (H) = PB (HT ) ⊗ PL and from eiHf,L t IL e−i(Hf,L ⊗1L +1L ⊗Hf,L )t = IL we get Ω+ = ΩT+ ⊗ ΩL +
˜ < Σ)H ˜ on χ(H
where is the wave operator defined in terms of HT , and ΩL + = IL (PL ⊗ 1L ). It follows that ˜ < Σ) H ˜ T ⊗ ΩL ⊗ ΩL = ΩT χ(H ˜ T < Σ)H ˜ T ⊗ ΩL , (105) Ω+ χ(H + ΩT+
˜ T < Σ)H ˜ T . On the other hand, which show in particular that ΩT+ exists on χ(H using AC for H, the left side of (105) can be written as ˜ < Σ) H ˜ T ⊗ ΩL ⊗ ΩL = (1 ⊗ PL )Ω+ χ(H ˜ < Σ)H ˜ Ω+ χ(H (106) = (1 ⊗ PL )χ(H < Σ)H = χ(HT < Σ)HT ⊗ ΩL which in conjunction with (105) shows that ˜ T < Σ)H ˜ T = χ(HT < Σ)HT . ΩT+ χ(H That is, AC in the sense of Theorem 20 holds for HT .
A
Pseudo Differential Calculus and Functional Calculus
This appendix collects our main tools for computing commutators.
A.1 Pseudo Differential Calculus on Fock Space Lemma 27 Suppose f ∈ S(Rd ), g ∈ C n (Rd ) and sup|α|=n ∂ α g ∞ < ∞. Let p = −i∇. Then i[g(p), f (x)] = i
1≤|α|≤n−1
= (−i)
(−i)|α| α (∂ f )(x)(∂ α g)(p) + R1,n α!
1≤|α|≤n−1
i|α| α (∂ g)(p)(∂ α f )(x) + R2,n α!
where
Rj,n ≤ Cn sup ∂ α g ∞ |α|=n
dk |k|n |fˆ(k)|.
In particular, and most importantly, if n = 2 in the limit ε → 0 i[g(p), f (εx)] = ε∇g(p) · ∇f (εx) + O(ε2 ) = ε∇f (εx) · ∇g(p) + O(ε2 ).
154
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
dk eikx fˆ(k). The first equation follows from g(p)eikx − eikx g(p) = eikx e−ikx g(p)eikx − g(p)
Proof. Let f (x) =
= eikx [g(p + k) − g(p)]
(107)
and Taylor’s formula g(p + k) − g(p) =
(∂ α g)(p)
1≤|α|≤n−1
1
+n
kα α!
dt (1 − t)n−1
0
(∂ α g)(p + tk)k α /α!.
|α|=n
To obtain the second equation write g(p)eikx − eikx g(p) = −[g(p − k) − g(p)]eikx instead of (107). Lemma 28 Suppose ω is in C ∞ (Rd ) and has bounded derivatives. If f ∈ C0∞ (R), then 1 [idΓ(ω), f ] = f dΓ(∇ω · y/t + y/t · ∇ω) + O(t−2 )N t 2 where f = f (dΓ(v )), v = y/t and f = f (dΓ(v 2 )). Proof. The operators dΓ(ω) and f commute with N , hence it suffices to prove the equation on ⊗ns h. On this subspace n n 2 i[dΓ(ω), f ] = vi ) . ω(kj ), f ( j=1
i=1
To evaluate [ω(kj ), f ( ni=1 vi2 )] for given fixed j ∈ {1, . . . , n} consider f ( vi2 ) as a function of yj only, and apply Lemma 27. It follows that n 1 2 ω(kj ), f ( vi ) = 2 f 2yj · ∇ω(kj ) + Rj,t t i=1 where
Rj,t ≤ t−2 sup ∂ α ω ∞ |α|=2
and fˆj (k) = (2π)−d
|k|2 |fˆj (k)|dk
n
e−ik·yj f yj2 +
yl2 dyj .
l=1,l=j
It is easy to see that |fˆj (k)| ≤
Cp (1 + |k|)p
for all
p∈N
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
where Cp only depends on f but not on j or Ct−2 and hence that on ⊗ns h
n l=1,l=j
155
yl2 . It follows that Rj,t ≤
n Cn 1 yj /t · ∇ω(kj ) ≤ 2 .
[idΓ(ω), f ] − f 2 t t j=1
(108)
The same equation holds with ∇ω(kj ) · yj /t instead of yj /t · ∇ω(kj ), as follows from [∇ω(kj ), yj /t] = ∆ω(kj )/t = O(t−1 ). In conjunction with (108) this proves the lemma.
A.2 Helffer-Sj¨ ostrand Functional Calculus Suppose f ∈ C0∞ (R; C) and A is a self-adjoint operator. A convenient representation for f (A), which is often used in this paper, is then given by f (A) = −
1 π
dxdy
∂ f˜ (z) (z − A)−1 , ∂ z¯
z = x + iy,
which holds for any extension f˜ ∈ C0∞ (R2 ; C) of f with |∂z¯f˜| ≤ C|y|, f˜(z) = f (z)
and
1 ∂ f˜ (z) = ∂ z¯ 2
∂f ∂f +i ∂x ∂y
(z) = 0
for all
z ∈ R.
(109) Such a function f˜ is called an almost analytic extension of f . A simple example is given by f˜(z) = (f (x) + iyf (x)) χ(z) where χ ∈ C0∞ (R2 ) and χ = 1 on some complex neighborhood of supp f . Sometimes we need faster decay of |∂z¯f˜| as |y| → 0 in the form |∂z¯f˜| ≤ C|y|n . In that case we work with the almost analytic extension n k (iy) χ(z) f˜(z) = f (k) (x) k! k=0
where χ is as above. We call this an almost analytic extension of order n. For more details and extensions of this functional calculus the reader is referred to [HS00] or [Dav95].
˜ < c)H ˜ B Representation of States in χ(H ˜ < c) proved in this section is used in The representation of states in Ran χ(H Lemma 18 and Theorem 20. Lemma 29 Suppose ω(k) ≥ 0 is continuous an let c > 0. Then the space of lin∗ ∗ 2 d ear n combinations of vectors of the form a (h1 ) . . . a (hn )Ω with hi ∈ L (R ) and i=1 sup{ω(k) : k ∈ supp(hi )} < c is dense in χ(dΓ(ω) < c)F .
156
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
Proof. It suffices to prove the assertion in each n-boson sector Fn separately and without Bose-Einstein symmetrization. That is, given n ∈ N, it suffices to show that the space Dc of finite linear combinations of vectors h1 ⊗ . . . ⊗ hn with n n i=1 sup{ω(k) : k ∈ supp(hi )} < c is dense in χS(c) (⊗ h) where χS(c) is the characteristic function of the set * ) n sup{ω(k) : k ∈ supp(hi )} < c . S(c) = (k1 , . . . , kn ) ∈ Rnd : i=1
Let ψ ∈ χS(c) (⊗ h). By standard arguments one constructs functions ψε ∈ C0∞ (Rnd ), ε > 0, with ψε = χS(c−ε) ψε and ψε − ψ → 0 as ε → 0. Now pick f ∈ C0∞ (Rd ) with supp(f ) ⊂ B(0, 1) and Rd f dk = 1, and let fδ (k) = δ −d f (k/δ), where δ > 0. Let Fδ (k1 , . . . , kn ) = ni=1 fδ (ki ). Then Fδ ∗ψε −ψ → 0 as ε, δ → 0, and, by construction of Fδ and ψε , the function k → Fδ (k − k )ψε (k ) belongs to Dc for every given fixed k ∈ supp(ψε ) if δ ≤ δ(ε). Hence the Riemann sums of Fδ(ε) ∗ ψε (k) = Fδ(ε) (k − k )ψε (k )dk
n
Rnd
belong to Dc as well. Since these Riemann sums converge to Fδ(ε) ∗ ψε uniformly in k, by the uniform continuity of Fδ(ε) and ψε , it follows that Fδ(ε) ∗ ψε ∈ D c and hence that ψ = limε→0 Fδ(ε) ∗ ψε ∈ Dc . Lemma 30 Suppose ω(k) = |k| or ω satisfies (H1), and let c > 0. Then the set of all linear combinations of vectors of the form ϕ ⊗ a∗ (h1 ) . . . a∗ (hn )Ω,
λ+
N
Mi < c
(110)
i=1
where ϕ = χ(H < λ)ϕ for some λ < c, n ∈ N and Mi = sup{ω(k) : hi (k) = 0}, is ˜ < c)H. ˜ dense in χ(H ˜ < c)ϕ ⊗ a∗ (h1 ) . . . a∗ (hn )Ω with ϕ ∈ H, Proof. Obviously the set of vectors χ(H ˜ ˜ Thus it suffices to approximate hi ∈ h and n arbitrary is dense in χ(H < c)H. such vectors by vectors of the form (110). In the sense of a strong Stieltjes integral ˜ χ(H < c) = dEλ (H) ⊗ χ(dΓ(ω) < c − λ). This means that ˜ < c)ϕ ⊗ a∗ (h1 ) . . . a∗ (hn )Ω χ(H E∆i (H)ϕ ⊗ χ(dΓ(ω) < c − λi )a∗ (h1 ) . . . a∗ (hn )Ω = lim sup |∆i |→0
i
where ∆i = (λi−1 , λi ]. The lemma now follows from Lemma 29 applied to χ(dΓ(ω) < c − λi )a∗ (h1 ) . . . a∗ (hn )Ω.
Vol. 3, 2002
C
Asymptotic Completeness for Rayleigh Scattering
157
Number–Energy Estimates
Thanks to the positivity of the boson mass, assumption (H1), one has the operator inequality N ≤ aH + b, (111) for some constants a and b. The purpose of this Section is to prove that also higher powers of N are bounded with respect to the same powers of H. This easily follows from (111) if the commutator [N, H] is zero, that is, for vanishing interaction. Otherwise it follow, as we will see, from the boundedness of adkN (H)(H + i)−1 for all k. Lemma 31 Assume the hypotheses (H1) and (H2) and suppose m ∈ N ∪ {0}. i) Then uniformly in z, for z in a compact subset of C,
(N + 1)−m (z − H)−1 (N + 1)m+1 = O(| Im z|−m ). ii) (N + 1)m (H + i)−m is a bounded operator. iii) If χ ∈ C ∞ (R) with supp χ ⊂ (−∞, Σ), then N m eα|x| χ(H) is a bounded operator, provided α > 0 is small enough. Proof. i) This is proved by induction in m. For m = 0 it suffices to show that N (H + i)−1 is bounded, because (H + i)(H − z)−1 = O(| Im(z)|−1 ) for z in a compact subset of C. The operator N (H0 + i)−1 is bounded because of the positivity of the boson mass, assumption (H1). Since φ(G) is infinitesimally small with respect to H0 , it follows that H0 (H + i)−1 is bounded and hence so is N (H + i)−1 . Next assume that the assertion in i) holds for any non–negative integer less than a given m ∈ N. To prove it for m we commute (z−H)−1 with (N +1)m+1 and get (N +1)−m (z − H)−1 (N + 1)m+1 = (N + 1)(z − H)−1 + (N + 1)−m (z − H)−1 [H, (N + 1)m+1 ](z − H)−1 = (N + 1)(z − H)−1 + (N + 1)−m (z − H)−1 m+1 m + 1 × (−1)l (N + 1)m+1−l adlN (H)(z − H)−1 l l=1 m+1 m + 1 −1 −1 = (N + 1)(z − H) + (N + 1) l l=1
× (N + 1)−(m−1) (z − H)−1 (N + 1)m+1−l il φ(il G)(z − H)−1 .
O(| Im z|−m+1 )
By induction hypothesis and because φ(il G)(z−H)−1 is of order O(| Im z|−1 ), for z in a compact set in C, it follows that the r.h.s. of the last equation is of order O(| Im z|−m ).
158
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
ii) Follows directly from i) if we put z = −i and write (N + 1)m (H + i)−m = (N + 1)m (H + i)−1 (N + 1)−m+1 × (N + 1)m−1 (H + i)−1 (N + 1)−m+2 . . . . . . (N + 1)(H + 1)−1 . iii) We begin to write N m eα|x| χ(H) = N m eα|x| (H + i)−m e−α|x| e+α|x| χ(H)(H + i)m = N m (i + e+α|x| He−α|x| )−m e+α|x| χ(H)(H + i)m = N m (i + H + δHα )−m e+α|x| χ(H)(H + i)m , where δHα = 2iαp·x/|x|− 2α/|x|− α2 , and α > 0 is so small that δHα (H + i)−1 < 1 (this ensures that (i + H + δHα ) = (1 + δHα (H + i)−1 )(H + i) is invertible, with a bounded inverse given by (H + i)−1 (1 + δHα (H + i)−1 )−1 ). The last equation implies iii), since N m (i + H + δHα )−m is a bounded operator. This can be shown in the same way we showed that N m (H + i)−m is a bounded operator, using that N commutes with δHα .
D Commutator Estimates 2 Let j0 , j∞ ∈ C ∞ (Rd ) be real-valued with j02 + j∞ ≤ 1, j0 (y) = 1 for |y| ≤ 1 and j0 (y) = 0 for |y| ≥ 2. Given R > 0 set j#,R (y) = j# (y/R) and let jR : h → h ⊕ h denote the operator defined by jR h = (j0,R h, j∞,R h).
Lemma 32 Assume hypotheses (H1), (H2), (H5). Suppose α > 0, m ∈ N∪{0}, and that jR is as above. Suppose also that χ, χ ∈ C0∞ (R), with supp χ ⊂ (−∞, Σ). Then, for R → ∞, ˘ R )H − H ˜ Γ(j ˘ R ) (N + 1)−m−1 = O(R−1 ), i) e−α|x| (N0 + N∞ + 1)m Γ(j ˜ Γ(j ˘ R ) − Γ(j ˘ R )χ(H) χ (H) = O(R−1 ). ii) (N0 + N∞ + 1)m χ(H) Proof. i) From the intertwining relations (20) and (21) we have that ˘ R )H − H ˜ Γ(j ˘ R ) = − dΓ(j ˘ R , adω (jR )) Γ(j ˘ R ). + [φ((j0,R − 1)G) ⊗ 1 + 1 ⊗ φ(j∞,R G)]Γ(j By Lemma 27 adω (jR ) = O(R−1 ). Hence ˘ R , adω (jR ))(N + 1)−m−1 = O(R−1 ). (N0 + N∞ + 1)m dΓ(j
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
159
To see that the other two terms lead to contributions of order O(R−1 ) write (N0 +N∞ + 1)m [φ((j0,R − 1)G) ⊗ 1 + 1 ⊗ φ(j∞,R G)] = [φ((j0,R − 1)G) ⊗ 1 + 1 ⊗ φ(j∞,R G)](N0 + N∞ + 1)m m m + (−i)l [φ(il (j0,R − 1)G) ⊗ 1 + 1 ⊗ φ(il j∞,R G)](N0 + N∞ + 1)m−l , l l=1
˘ R ) = Γ(j ˘ R )(N + 1)j and use Hypothesis (H5). then use (N0 + N∞ + 1)j Γ(j ii) Let χ ˜ be an almost analytic extension of χ of order m, as defined in Sect. A.2. Then we have ˜ Γ(j ˘ R ) − Γ(j ˘ R )χ(H))χ (H) = − 1 dxdy ∂z¯χ ˜ (N0 + N∞ + 1)m (χ(H) π ˜ −1 e−α|x| (H ˜ Γ(j ˘ R ) − Γ(j ˘ R )H) × (N0 + N∞ + 1)m (z − H) × eα|x| χ (H)(z − H)−1 . Then the statement follows by i) because ˜ −1 (N0 + N∞ + 1)−m+1 = O(| Im z|−m ), (N0 + N∞ + 1)m (z − H)
(112)
m α|x|
and because (N + 1) e ciently small.
χ (H) is a bounded operator, provided α > 0 is suffi-
Lemma 33 Assume hypotheses (H1), (H2), (H5). Suppose α > 0, m ∈ N ∪ {0}. Let jR be as above and set djR = [iω, jR ] + ∂jR /∂R. Suppose also that χ, χ ∈ C0∞ (R), with supp χ ⊂ (−∞, Σ). Then ˘ R , djR )H − Hd ˜ Γ(j ˘ R , djR ) (N + 1)−m−2 = i) e−α|x| (N0 + N∞ + 1)m dΓ(j O(R−2 ),
˜ Γ(j ˘ R , djR ) χ (H) = O(R−2 ). ˘ R , djR )χ(H) − χ(H)d ii) (N0 + N∞ + 1)m dΓ(j
Proof. We begin proving part i). To this end note that ˘ R ,djR )H − Hd ˜ Γ(j ˘ R , djR ) dΓ(j ˘ R , djR )dΓ(ω) − (dΓ(ω) ⊗ 1 + 1 ⊗ dΓ(ω))dΓ(j ˘ R , djR ) = dΓ(j
(113)
˘ R , djR ). ˘ R , djR )φ(G) − (φ(G) ⊗ 1)dΓ(j + dΓ(j Consider the term on the first line on the r.h.s. of the last equation. We have ˘ R , djR )|` ⊗n h ˘ R , djR )dΓ(ω)−(dΓ(ω) ⊗ 1 + 1 ⊗ dΓ(ω))dΓ(j dΓ(j s =U (jR ⊗ · · · ⊗ djR ⊗ · · · ⊗ [jR , ω] ⊗ · · · ⊗ jR ) i=j
+
n i=1
(jR ⊗ · · · ⊗ [djR , ω] ⊗ · · · ⊗ jR ) .
160
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
Each one of the n2 terms on the r.h.s. of the last equation is of order O(R−2 ), by Lemma 27. Hence ˘ R , djR )dΓ(ω) (N0 + N∞ +1)m dΓ(j ˘ R , djR ) (N + 1)−m−2 = O(R−2 ). −(dΓ(ω) ⊗ 1 + 1 ⊗ dΓ(ω))dΓ(j Consider now the term on the second line on the r.h.s. of (113). This is given by ˘ R , djR )φ(G)−(φ(G) ⊗ 1)dΓ(j ˘ R , djR ) = (φ((j0,R − 1)G) ⊗ 1 + 1 ⊗ φ(j∞,R G)) dΓ(j ˘ R ). ˘ R , djR ) + (φ(dj0,R G) ⊗ 1 + 1 ⊗ φ(dj∞,R G)) Γ(j × dΓ(j By Hypothesis (H5) it follows, in the same way as in the proof of Lemma 32, that this term gives a contribution of order O(R−µ−1 ). This completes the proof of part i). Part ii) of the Lemma follows in the same way as in Lemma 32.
E
Mourre Estimate
This section contains the proofs of Lemma 3 (the Virial Theorem) and of Theorem 4 (the Mourre Estimate). Recall that A stands for dΓ(a) with a = 1/2(∇ω · y + y · ∇ω). Proof of Lemma 3 (Virial Theorem). First we prove the virial theorem for a regularized variant, Aε , of A = Aε=0 defined on D(H), and then we let ε → 0. Let aε = 1/2(∇ω · yε + yε · ∇ω) where yε = y(1 + εy 2 )−1 and let Aε = dΓ(aε ). Then iHϕ, Aε ϕ − iAε ϕ, Hϕ = ϕ, dΓ(i[ω, aε ]) − φ(iaε G) ϕ (114) for all ϕ ∈ D(K) ⊗ F0 , which is a core for H. Since all operators in (114) are H-bounded this equation extends to D(H). If ϕ is an eigenvector H then the left side of (114) vanishes because Aε ⊂ A∗ε . Thus it suffices to prove that the right side converges to ϕ, i[H, A]ϕ, as ε → 0, for ϕ ∈ Eµ (H)H, µ < Σ. We show that and
φ(iaε G)ϕ → φ(iaG)ϕ,
(115)
dΓ(i[ω, aε ])ϕ → dΓ(|∇ω| )ϕ
(116)
2
as ε → 0. Eq. (115) follows from aε G−aG → 0 and supx e−α|x| φ((a−aε )G)(N + 1)−1/2 < ∞ by Lebesgue’s dominated convergence theorem if α > 0 is chosen in such a way that eα|x| (N + 1)1/2 ϕ belongs to H. Here we used (H2) and (H3). To prove (116) note that i[ω, yε ] → ∇ω strongly and hence that i[ω, aε ] → |∇ω|2 strongly. Since moreover supε i[ω, aε ] < ∞ this proves (116) for all ϕ ∈ D(N ). The property proved in the following lemma has a well known analog, called local compactness, in the theory of Schr¨ odinger operators. We used it in the proof of Theorem 2 and will need it again in the subsequent proof of the Mourre estimate, Thm. 4.
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
161
Lemma 34 Assume (H1) and suppose f ∈ L∞ (X) and g ∈ L∞ (Rd , dy) with
g ∞ ≤ 1. If f (x) → 0, as |x| → ∞, and g(y) → 0, as |y| → ∞ in Rd , then f (x) ⊗ Γ(g)(H + i)−1/2
is compact.
This lemma follows from the fact that p2 and dΓ(ω) are form bounded w.r. to H and from the positivity of the boson mass. Proof. From m > 0 it follows that χ(N > n)Γ(g)(dΓ(ω) + 1)−1/4 → 0 as n → ∞. Furthermore Γ(g)(dΓ(ω) + 1)−1/4 |` ⊗ns h = g(y1 ) . . . g(yn )[ω(k1 ) + . . . + ω(kn ) + 1]−1/4 which is compact. It follows that Γ(g)(dΓ(ω) + 1)−1/4 is a compact operator on F . Since f (x)(p2 + 1)−1/4 is compact on L2 (X) and (p2 + 1)1/4 ⊗ (dΓ(ω) + 1)1/4 (H + i)−1/2 is a bounded operator, the operator f ⊗ Γ(g)(H + i)−1/2 is compact. The proof of Theorem 4 is by induction in the number of energy steps of size m, the minimal energy of a free boson. Assuming that the theorem holds for λ < min(Σ, (n − 1)m) we prove it for λ < min(Σ, nm). To this end we need a partition of unity in the bosonic configuration space. Let j0 , j∞ ∈ C ∞ (Rd ), with 2 j02 + j∞ = 1, and with 0 ≤ j0 ≤ 1, j0 (y) = 1 if |y| ≤ 1 and j0 (y) = 0 if |y| ≥ 2. Given R > 0 we set j#,R (y) = j# (y/R). Each boson h ∈ h will be split into the two parts j0,R h and j∞,R localized near the origin and near infinity respectively. Recall ˘ R ) does the corresponding localization in Fock space. In from Section 2.6 that Γ(j ˘ R ) through H or any bounded function of H Appendix D we saw how to move Γ(j (see Lemma 32). Since [iH, A] has a structure similar to H, with ω and G replaced by |∇ω|2 and by iaG, respectively, it is easy to show that an analogue of Lemma 32 also holds for [iH, A]. In particular we use that ˘ R )[iH, A] − ([iH, A] ⊗ 1 + 1 ⊗ dΓ(|∇ω|2 ))Γ(j ˘ R ) (N + 1)−1 = o(R0 ), e−α|x| Γ(j (117) as R → ∞. Proof of Theorem 4 (Mourre Estimate). First, we introduce the Mourre constants d(λ) :=
Ω⊥
inf
dΓ(|∇ω(k)|2 )
σpp (H)+dΓ(ω(k))=λ
˜ := d(λ)
inf
σpp (H)+dΓ(ω(k))=λ
dΓ(|∇ω(k)|2 ),
where the superscript Ω⊥ in the definition of d means that we exclude the vacuum sector to compute the infimum. Note that d vanishes only on thresholds, while d˜ vanishes on thresholds and on eigenvalues of H. We introduce, moreover, the ˜ For κ > 0, we set ∆κ = [λ − κ, λ + κ], smeared out versions of the functions d, d. λ
162
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
˜ and then we define dκ (λ) = inf µ∈∆κλ d(µ) and d˜κ (λ) = inf µ∈∆κλ d(µ). By definition of these functions we have the inequality Ω⊥ inf d˜κ (λ − dΓ(ω(k))) + dΓ(|∇ω(k)|2 ) ≥ dκ (λ).
(118)
For all n ∈ N we will show the following statements. H1 (n): Let ε > 0 and λ ∈ [E0 , E0 + nm) ∩ (−∞, Σ). Then there exists an open interval ∆ λ and a compact operator E such that E∆ (H)[iH, A]E∆ (H) ≥ (d(λ)− ε)E∆ (H) + E. H2 (n): Let ε > 0 and λ ∈ [E0 , E0 + nm) ∩ (−∞, Σ). Then there exists an open ˜ − ε)E∆ (H). interval ∆ λ such that E∆ (H)[iH, A]E∆ (H) ≥ (d(λ) H3 (n): Let κ, ε0 , ε > 0. Then there exists δ > 0 such that for all λ ∈ [E0 , E0 + nm − ε0 ] ∩ (−∞, Σ), one has E∆κλ (H)[iH, A]E∆κλ (H) ≥ (d˜κ (λ) − ε)E∆κλ (H). S1 (n): τ is a closed and countable set in [E0 , E0 + nm) ∩ (−∞, Σ). S2 (n): For any closed intervall I ⊂ [E0 , E0 + nm) ∩ (−∞, Σ), with I ∩ τ = ∅, one has dim Ran EI∩σpp (H) < ∞. Note here that all the claims of the theorem follow from these statements, if we prove them for any n ∈ N. Actually, the statements give no new information if n is so large that E0 + nm > Σ. We will prove these statements by induction in n. Since H1 (1) and S1 (1) are obvious, all the statements, for any n ∈ N, follow if we prove the implications: H1 (n) ⇒ H2 (n), H2 (n) ⇒ H3 (n), H1 (n) ⇒ S2 (n), S2 (n − 1) ⇒ S1 (n), and S1 (n) ∧ H3 (n − 1) ⇒ H1 (n). Now the implication S2 (n − 1) ⇒ S1 (n) is trivial. Moreover the implications H1 (n) ⇒ H2 (n), H2 (n) ⇒ H3 (n) and H1 (n) ⇒ S2 (n) are proved by standard abstract arguments, which are typical in the proof of any Mourre inequality. It only remains to prove that S1 (n) together with H3 (n − 1) imply the statement H1 (n). To this end we fix λ ∈ [E0 , E0 + nm) ∩ (−∞, Σ) and ε > 0. Choose now χ ∈ C0∞ (R) with supp χ ⊂ [λ−δ, λ+δ]∩(−∞, Σ), where δ is some positive constant which will be fixed later on. Then we have χ(H)[iH, A]χ(H) ˘ R )∗ E{0} (N∞ )Γ(j ˘ R )χ(H)[iH, A]χ(H) = Γ(j ˘ R )χ(H)[iH, A]χ(H) ˘ R )∗ E[1;∞) (N∞ )Γ(j + Γ(j (119) + , ∗ 2 ˘ ˜ = Γ(qR )χ(H)[iH, A]χ(H) + Γ(jR ) χ(H) [iH, A] ⊗ 1 + 1 ⊗ dΓ(|∇ω| ) ˘ R ) + o(R0 ), ˜ [1;∞) (N∞ )Γ(j × χ(H)E ˘ R ) to the right. The first where we used Lemma 32 and Eq. (117) to commute Γ(j term on the r.h.s. of (119) is compact, by Lemma 34. To handle the second term we want to diagonalize 1 ⊗ dΓ(ω) and 1 ⊗ dΓ(|∇ω|2 ) on the range of E[1;∞) (N∞ ). Using the induction hypothesis S1 (n) we find κ > 0 such that dκ (λ) ≥ d(λ) − ε/3.
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
163
Then, by H3 (n − 1), we know that if δ > 0 is small enough we have + , E∆δλ (H + dΓ(ω(k))) [iH, A] ⊗ 1 + 1 ⊗ dΓ(|∇ω(k)|2 ) E∆δλ (H + dΓ(ω(k))) ε ×E[1;∞) (N∞ ) ≥ E∆δλ (H + dΓ(ω(k))) d˜κ (λ − dΓ(ω(k))) + dΓ(|∇ω(k)|2 ) − 3 ε ×E[1;∞) (N∞ ) ≥ (dκ (λ) − )E∆δλ (H + dΓ(ω(k)))E[1;∞) (N∞ ) 3 2J ≥ (d(λ) − )E∆δλ (H + dΓ(ω(k)))E[1;∞) (N∞ ), 3 where, in the second inequality we used (118). It follows, since supp χ ⊂ ∆δλ , that , + ˜ [1;∞) ≥ (d(λ) − 2J )χ2 (H)E ˜ [1;∞) (N∞ ). ˜ [iH, A] ⊗ 1 + 1 ⊗ dΓ(|∇ω|2 ) χ(H)E χ(H) 3 Now we insert this in the second term on the r.h.s. of (119) and then we com˘ R ) back to the left. Using that mute, using again Lemma 32 and Eq. (117), Γ(j ˘ R )∗ E[1,∞) (N∞ )Γ(j ˘ R ) = 1 − Γ(j 2 ) and that, by Lemma 34, Γ(j 2 )χ(H) is a Γ(j 0,R 0,R compact operator, we find, from (119), χ(H)[iH, A]χ(H) ≥ (d(λ) −
2ε 2 )χ (H) + E + o(R0 ). 3 δ/2
The statement H1 (n) then follows if we choose χ such that χ = 1 on ∆λ , if we multiply (119) from the right and from the left with E∆δ/2 (H) and if we choose R λ sufficiently large.
F Invariance of Domains In this section the invariance of the domain of dΓ(y 2 ) with respect to the action of χ(H) for smooth functions χ is proven. Moreover we prove in Lemma 36 that the norm of dΓ(y 2 /t2 )χ(H)e−iHt ϕ remains uniformly bounded for all t ≥ 1, if ϕ ∈ D(dΓ(y 2 + 1)). These results are used in Proposition 13 to prove the existence of the asymptotic observable W . Lemma 35 Assume Hypotheses (H1), (H2), (H3) are satisfied. Suppose moreover that ϕ ∈ D(dΓ(y 2 )) ∩ D(N ) and that χ ∈ C ∞ (R) with supp χ ⊂ (−∞, Σ). Then we have
dΓ(y 2 )χ(H)ϕ ≤ C( dΓ(y 2 + 1)ϕ + ϕ ). Proof. Find χ1 ∈ C0∞ (R) with χχ1 = χ, and supp χ1 ⊂ (−∞, Σ). Put χ = χ(H) and χ1 = χ1 (H). Then we have dΓ(y 2 )χ = dΓ(y 2 )χχ1 = χdΓ(y 2 )χ1 + [dΓ(y 2 ), χ]χ1 .
(120)
164
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
Now expand the χ in the commutator in an integral, according to the Helffer– Sj¨ ostrand functional calculus (see Appendix A.2). We get 2i 2 [dΓ(y ), χ]χ1 = − ˜ (z − H)−1 dΓ(a)(z − H)−1 χ1 dxdy ∂z¯χ π (121) i dxdy ∂z¯χ + ˜ (z − H)−1 φ(iy 2 G)χ1 (z − H)−1 , π where χ ˜ is an almost analytic extension of χ, in the sense of the Helffer-Sj¨ostrand functional calculus, and a = [iω(k), y 2 /2] = 1/2(∇ω · y + y · ∇ω). The second term on the r.h.s. of the last equation is bounded, because, by hypotheses (H2) and (H3), φ(y 2 G)χ1 < ∞. To handle the first term on the r.h.s. of (121) we commute the factor dΓ(a) to the right of the resolvent (z − H)−1 . Using [dΓ(a), H] = idΓ(|∇ω|2 ) − iφ(iaG), we see that the contributions arising from this commutator are bounded, because |∇ω| < const and because, using again (H2) and (H3), φ(iaG)χ < ∞. Thus we have, from (120) dΓ(y 2 )χ = χdΓ(y 2 )χ1 + C χ (H)dΓ(a)χ1 + bounded,
(122)
where χ is the first derivative of χ. Now we commute the two operators dΓ(y 2 ) and dΓ(a) in the two terms on the r.h.s. of the last equation to the right of χ1 . For example , for the term χdΓ(y 2 )χ1 , we find χdΓ(y 2 )χ1 = χχ1 dΓ(y 2 ) + χ[dΓ(y 2 ), χ1 ] 1 = χdΓ(y 2 ) − ˜1 χ(z − H)−1 [dΓ(y 2 ), H](z − H)−1 dxdy ∂z¯χ π (123) 2i = χdΓ(y 2 ) − 2 ˜1 χ(z − H)−1 dΓ(a)(z − H)−1 dxdy ∂z¯χ πt i + ˜1 (z − H)−1 χφ(iy 2 G)(z − H)−1 . dxdy ∂z¯χ π The third term on the r.h.s. of the last equation is bounded by (H2) and (H3). To handle the second term we commute dΓ(a) to the right. We get 2i ˜1 χ(z − H)−1 dΓ(a)(z − H)−1 − dxdy ∂z¯χ π 2i = − ˜1 χ(z − H)−2 dΓ(a) dxdy ∂z¯χ π 2 + ˜1 (z − H)−2 χdΓ(|∇ω|2 )(z − H)−1 dxdy ∂z¯χ π 2 − ˜1 (z − H)−2 χφ(iaG)(z − H)−1 . dxdy ∂z¯χ π The first term on the r.h.s. of the last equation is proportional to χχ 1 dΓ(a), where χ 1 is the first derivative of χ1 , and thus vanishes, since χ1 is constant on supp χ.
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
165
The other two terms on the r.h.s. of the last equation are bounded. It follows, by (123), that χdΓ(y 2 )χ1 = χdΓ(y 2 ) + bounded. Similarly we find, that χ dΓ(a)χ1 = χ dΓ(a) + bounded. These two results imply, by (122), that + ,
dΓ(y 2 )χϕ ≤ C dΓ(y 2 )ϕ + dΓ(a)ϕ + ϕ . (124) The Lemma now follows using Lemma 37, part (iv), to estimate the second term on the r.h.s. of the last equation. Lemma 36 Assume hypotheses (H1), (H2) and (H3) are satisfied. Suppose moreover that ϕ ∈ D(dΓ(y 2 )) ∩ D(N ) and that χ ∈ C ∞ (R) with supp χ ⊂ (−∞, Σ). Then we have
dΓ(y 2 )e−iHt χ(H)ϕ ≤ C( dΓ(y 2 + 1)ϕ + t2 ϕ ), for all t ≥ 1. Proof. We begin by noting, that eiHt dΓ(y 2 )e−iHt χ(H) − dΓ(y 2 )χ(H) = =2
t
ds eiHs [iH, dΓ(y 2 )] χ(H)e−iHs
0 t
ds e
iHs
dΓ(a) χ(H)e
−iHs
0
−
t
ds eiHs φ(iy 2 G) χ(H)e−iHs ,
0
(125) where a = [iω, y 2 /2] = 1/2(∇ω · y + y · ∇ω). The second integral on the r.h.s. of the last equation is bounded, with norm of order t, because, by (H2) and (H3),
φ(iy 2 G)χ < ∞. To handle the first integral on the r.h.s. of the last equation use the expansion t 2 ds eiHs dΓ(a)χ(H)e−iHs 0 t s = 2t dΓ(a)χ(H) + 2 ds dr eiHr [iH, dΓ(a)] χ(H)e−iHr 0
0
t
s
= 2t dΓ(a)χ(H) + 2 ds dr eiHr dΓ(|∇ω|2 ) χ(H)e−iHr 0 0 t s −2 ds dr eiHr φ(iaG) χ(H)e−iHr . 0
0
Since |∇ω| is bounded and φ(iaG)χ(H) < ∞, both the integrals on the r.h.s. of the last equation are bounded and of order t2 . This implies, by (125), that
dΓ(y 2 )e−iHt χ(H)ϕ ≤ dΓ(y 2 )χ(H)ϕ + 2t dΓ(a)χ(H)ϕ + Ct2 ϕ
= dΓ(y 2 )χ(H)ϕ + 2t2 dΓ(a/t)χ(H)ϕ + Ct2 ϕ . (126)
166
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
By Lemma 37, part (iv), with y replaced by y/t, we have
dΓ(a/t)χϕ ≤ dΓ(y 2 /t2 + 1)χϕ ≤ dΓ(y 2 /t2 )χϕ + C ϕ . The Lemma now follows inserting the last equation into (126) and applying Lemma 35. Lemma 37
(i) For any operator A A2 + (A∗ )2 ≤ AA∗ + A∗ A.
(ii) If a = 1/2(y · ∇ω + ∇ω · y) and ω satisfies (H1) then a2 ≤ const(y 2 + 1). (iii) If a is a symmetric operator in h then dΓ(a)2 ≤ N dΓ(a2 ) . (iv) With a as in (ii) we have dΓ(a)2 ≤ const dΓ(y 2 + 1)2 . Proof.
(i) This follows from 0 ≤ (iA − iA∗ )2 = −A2 − (A∗ )2 + AA∗ + A∗ A.
(ii) By (i), (∇ω · y + y · ∇ω)2 ≤ 2[(∇ω · y)(y · ∇ω) + (y · ∇ω)(∇ω · y)] where (y · ∇ω)(∇ω · y) ≤ ∇ω 2∞ y 2 (∇ω · y)(y · ∇ω) ≤ 2( ∇ω 2∞ + ∆ω 2∞ )(y 2 + 1)
(iii) It suffices to prove this for states ϕnin Ran χ(N = n). such vecFor n n tors ϕn , dΓ(a)2 ϕn = dΓ(a)ϕn 2 ≤ ( i=1 ai ϕn )2 ≤ n i=1 ai ϕn 2 = ϕn , N dΓ(a2 )ϕn , where ai denotes the operator a acting on the i-th boson. (iv) This follows from (ii), (iii), and N ≤ dΓ(y 2 + 1).
G
Some Technical Parts of Theorem 15
Lemma 38 There exists a constant C = C(λ, u) such that 1/2 n ±(Jnk Sn ) (y, t) ≤ Cn Sn
(y, t) k
2
for |y| ≤ 2λt and ut1−δ ≥ 2. Here y = (y1 , . . . , yn ), yi ∈ Rd .
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
167
Proof. We drop n the combinatorial factor in the definition of Jnk for convenience. Let S0 (y) = i=1 S0 (yi ). From (34) and the scaling factor 1/ut in the arguments of Jnk we get tSn
(y, t) = S0
(y/tδ ) and
t(Jnk Sn )
(y, t) = Jnk (z) · S0
(w) + α−2 Jnk (z)S0 (w)
+ α−1 ∇Jnk (z) ⊗ ∇S0 (w) + α−1 ∇S0 (w) ⊗ ∇Jnk (z)
(127)
where w = yt−δ , z = y/ut, α = ut1−δ and hence w = αz, α ≥ 2 and |z| ≤ 2λ/u. Proving the lemma thus amounts to showing that the r.h.s. of (127) (and its negative) are bounded by const × n2 S0
(w). The first term in (127) enjoys the bound. To estimate the other three terms we need some “N -body geometry”. For each a ∈ {0, ∞}n , let 1a : Rnd → Rnd denotes the orthogonal projection onto the subspace {w ∈ Rnd |wi = 0 if ai = 0} and let wa = 1a w. Set
χa (σ, w) =
χ(ws2 /2 > σ).
s:as =∞
r:ar =0
Then
χ(wr2 /2 ≤ σ)
χa ≡ 1. Hence, by (32) and since χa (σ, w) χ(wi2 /2 > σ) = χa (σ, w)δai ,∞ , (128) S0 (w) = dσ m(σ)S0 (σ, w) a
S0 (σ, w) =
n
χ(wi2 /2 > σ)(wi2 /2 − σ) =
χa (σ, w)(wa2 /2 − σa )
(129)
a
i=1
where σa = σ · #{i : ai = ∞}. Furthermore ∇S0 (σ, w) =
χa (σ, w)wa
(130)
χ(σ, w)1a
(131)
a
S0
(σ, w) ≥
a
in the sense that ∇S0 (w) =
dσ m(σ)
S0
(w) ≥
χa (σ, w)wa and
dσ m(σ)
χa (σ, w)1a .
By (128) it suffices to estimate the last three terms of (127) for S0 (w) replaced by S0 (σ, w) uniformly in σ. If wi2 /2 ≤ 1 and α ≥ 2 then |zi | ≤ 1 and hence ∂ β jε (zi ) = 0 for all derivatives of non-zero order β and ε ∈ {0, ∞}. Hence for w such that χa (σ, w) = 0
Jnk (z) = 1a Jnk (z)1a
(132)
∇Jnk (z) = 1a ∇Jnk (z).
(133)
168
J. Fr¨ ohlich, M. Griesemer and B. Schlein
From (129), (132) and (131) we get
(z)S0 (σ, w) ≤ a−2 w2 /2 Jnk (z)
±a−2 Jnk
Ann. Henri Poincar´e
χa (w)1a
a
≤ Cn2 S0
(σ, w). The last two terms in (127) are estimated similarly using (130), (133) and (131). Lemma 39
1/2 y2 m+1 n D ∂ O(t−(1+m)+δ(2−|α|) ) Snk (y, t) − Jnk (y, t) =n k t m α
Proof. The difference Snk (y, t) − Jnk (y, t)y /2t = Jnk (y, t)t 2
−1+2δ
n
[a(yi t−δ ) + b]
(134)
i=1
is of the form t−c f (y/t)h(yt−δ ) where y = (y1 , . . . , yn ) and f, h ∈ C ∞ ∩ L∞ (Rnd ). Under differentiation (134) becomes a sum of terms of the same form where in each term the exponent of t is decreased by at least δ or 1 for differentiation w.r. to y or t (or D) respectively. This accounts for the power in t. From the explicit form the right hand side of (134) and (60) we see that every derivative w.r. to yi (at most) doubles the number of terms, while every derivative w.r. to t multiplies it by n + 1. Lemma 40 For |y|/t ≤ 2λ we have
(∇Ω − V ) D02 Snk =(∇Ω − V ) · Snk + (D∇Snk ) · (∇Ω − V ) + (∇Ω − V ) · (D∇Snk ) 1/2 n + DV2 Snk + n2 O(t−1−δ ) k
where V = y/t ∈ Rnd .
1/2 since it appears in all terms on both sides of the Proof. We drop the factor nk equation. By definition of D0 , ∂ 2 Snk . (135) ∂t2 = Jnk Sn and the Leibnitz rule
D02 Snk = [iΩ, [iΩ, Snk ]] + 2[iΩ, ∂Snk /∂t] + To evaluate the double commutator we use Snk and get for |y/t| ≤ 2λ
[iΩ, [iΩ, Snk ]] =[iΩ, [iΩ, Jnk ]]Sn + 2[iΩ, Jnk ][iΩ, Sn ] + Jnk [iΩ, [iΩ, Sn ]]
Sn + ∇Jnk ⊗ ∇Sn + ∇Sn ⊗ ∇Jnk + Jnk Sn
] ∇Ω (136) =∇Ω · [Jnk + n2 O(t−1−δ )
=∇Ω · Snk ∇Ω + n2 O(t−1−δ )
(137)
Vol. 3, 2002
Asymptotic Completeness for Rayleigh Scattering
169
where we used [iΩ, Jnk ] = ∇Ω·∇Jnk +nO(t−2 ), [iΩ, Sn ] = ∇Ω·∇Sn +nO(t−1 ) (and the same with the order in the dot products reversed) as well as [iΩ, [iΩ, Sn ]] =
∇Ω · Sn
∇Ω + nO(t−1−δ ) (see Lemma9), [iΩ, [iΩ, Jnk ]] = ∇Ω · Jnk ∇Ω + n2 O(t−3 ) α 1−|α| ) for |y/t| ≤ 2λ and |α| ≤ 2. In a similar way one shows and |∂ Sn | = O(t that [iΩ, ∂Snk /∂t] = ∇Ω · ∇ =∇
∂Snk + n2 O(t−2 ) ∂t
∂Snk · ∇Ω + n2 O(t−2 ). ∂t
(138)
The lemma now follows from (135), (137) and (138) as Lemma 9 did from analogous equations.
References [Ara83]
A. Arai, Rigorous theory of spectra and radiation for a model in quantum electrodynamics, J. Math. Phys. 24(7), 1896–1910 (1983).
[BFS98]
V. Bach, J. Fr¨ ohlich and I. M. Sigal, Quantum electrodynamics of confined nonrelativistic particles, Adv. Math. 137(2), 299–395 (1998).
[BFSS99] V. Bach, J. Fr¨ ohlich, I. M. Sigal and A. Soffer, Positive commutators and spectrum of Pauli–Fierz Hamiltonian of atoms and molecules, Comm. Math. Phys. 207(3), 557–587 (1999). [Dav95]
E. B. Davies, The functional calculus, J. London Math. Soc. (2), 52(1), 166–176 (1995).
[DG99]
J. Derezi´ nski and C. G´erard, Asymptotic completeness in quantum field theory Massive Pauli-Fierz Hamiltonians, Rev. Math. Phys. 11(4), 383– 450 (1999).
[DG00]
J. Derezi´ nski and C. G´erard, Spectral scattering theory of spatially cutoff P (φ)2 Hamiltonians, Comm. Math. Phys. 213(1), 39–125 (2000).
[FGS00] J. Fr¨ ohlich, M. Griesemer and B. Schlein, Asymptotic electromagnetic fields in models of quantum-mechanical matter interacting with the quantized radiation field, Adv. Math., to appear. Los Alamos arXiv math-ph/0009033. [GLL00] M. Griesemer, E. H. Lieb and M. Loss, states in non–relativistic QED, Invent. math. 145(3), 557–595 (2001). [GS97]
G. M. Graf and D. Schenker, Classical action and quantum N -body asymptotic completeness, In Multiparticle quantum scattering with applications to nuclear, atomic and molecular physics (Minneapolis, MN, 1995), pages 103–119. Springer, New York (1997).
170
J. Fr¨ ohlich, M. Griesemer and B. Schlein
Ann. Henri Poincar´e
[HS00]
W. Hunziker and I. M. Sigal, The quantum N –body problem, J. Math. Phys. 41(6), 3448–3510 (2000).
[RS79]
M. Reed and B. Simon, Methods of modern mathematical physics: Scattering Theory. Volume 3, Academic Press (1979).
[Spo97]
H. Spohn, Asymptotic completeness for Rayleigh scattering, J. Math. Phys. 38(5), 2281–2296 (1997).
J. Fr¨ ohlich and B. Schlein Theoretical Physics ETH–H¨ onggerberg CH–8093 Z¨ urich Switzerland email:
[email protected] email:
[email protected] M. Griesemer Department of Mathematics University of Alabama at Birmingham Birmingham, AL 35294 USA email:
[email protected] Communicated by Gian Michele Graf submitted 09/04/01, accepted 09/10/01
Ann. Henri Poincar´e 3 (2002) 171 – 201 c Birkh¨ auser Verlag, Basel, 2002 1424-0637/02/010171-31 $ 1.50+0.20/0
Annales Henri Poincar´ e
Self-Adjointness of the Pauli-Fierz Hamiltonian for Arbitrary Values of Coupling Constants F. Hiroshima Abstract. The Pauli-Fierz Hamiltonian describes a system of N electrons minimally coupled to a quantized radiation field. The electrons have spin and an ultraviolet cutoff is imposed on the quantized radiation field. For arbitrary values of coupling constants, self-adjointness and essential self-adjointness of the Pauli-Fierz Hamiltonian are proven by means of a functional integral.
1 Introduction and main results The Pauli-Fierz Hamiltonian [33] governs a system of N -electrons minimally coupled with a quantized radiation field. The N -electrons are assumed to have spin and the quantized radiation field is smeared by an ultraviolet cutoff. The purpose of this paper is to establish self-adjointness and essential self-adjointness of the Pauli-Fierz Hamiltonian for arbitrary values of coupling constants. For sufficiently small values of coupling constants self-adjointness of the PauliFierz Hamiltonian on an explicit domain, say D, has previously been established. In [21] we have shown essential self-adjointness on D for arbitrary values of coupling constants, for N = 1 under certain integrable conditions on the ultraviolet cutoff. In this paper we show self-adjointness on the same domain D for all N ≥ 1 under some weaker conditions on the ultraviolet cutoff.
1.1
The Pauli-Fierz Hamiltonian
Suppose that the N -electrons move in the d-dimensional space, where d ≥ 3. Let L2 := L2 (Rd ). The Hilbert space of state vectors is defined by N [d/2] ⊗ L2 ⊗ F, Htotal := C2
(1.1)
where [k] denotes the integer part of k. Here F is the symmetric Fock space over Cd−1 ⊗ L2 , i.e., F :=
∞ n=0
⊗ns Cd−1 ⊗ L2 ,
172
F. Hiroshima
Ann. Henri Poincar´e
where ⊗ns denotes the symmetric tensor product with ⊗0s [Cd−1 ⊗ L2 ] := C. Under the identification Cd−1 ⊗ L2 ∼ = L2 (Rd × {1, 2, ..., d − 1}) we have F∼ =
∞
⊗ns L2 (Rd × {1, 2, ..., d − 1}).
n=0
In what follows we identify (1.1) as [d/2] [d/2] Htotal ∼ ) ⊗ L2 (RdN ) ⊗ F ∼ ) ⊗ L2 (RdN ; F ). = (⊗N C2 = (⊗N C2
(1.2)
The vacuum vector in F is defined by Ω := {1, 0, 0, 0, ...}. The smeared annihilation and creation operators are denoted by ar (f ) and a†r (f ), f ∈ L2 , r = 1, ..., d − 1, respectively. They are linear in f and satisfy the canonical commutation relations: [as (f ), a†r (g)] = (f¯, g)L2 δrs ,
[ar (f ), as (g)] = 0,
[a†r (f ), a†s (g)] = 0
on the finite particle subspace (0) Ffin := ∪∞ , Ψ(1) , Ψ(2) , ...} ∈ F Ψ(n) = 0, n ≥ m , m=0 {Ψ where (f, g)K denotes the scalar product on K. Note that (Ψ, a†r (f )Φ)F = (ar (f¯)Ψ, Φ)F , and we formally write as ar (f ) = ar (k)f (k)dk,
a†r (f ) =
Ψ, Φ ∈ Ffin ,
a†r (k)f (k)dk.
We write the norm on K as · K . Unless confusion may arise we write simply (f, g) and f for (f, g)K and f K, respectively. Note that in our notation (f, g) is linear in g and anti-linear in f . Under the identification (1.2) the Pauli-Fierz Hamiltonian is given by 2 1 j j σµ ⊗ pµ + ajexµ (xj ) ⊗ 1 − eAµ (xj ) +1⊗V ⊗1+1⊗1⊗Hf , (1.3) HPF := 2 where the summation over repeated indices is automatically understood unless otherwise stated. Here we adopt the unit: c = = 1, e denotes the electron charge regarding as the coupling constant, and pj := (pj1 , · · · , pjd ) and xj := (xj1 , · · · , xjd ) are the-jth electron momentum and its canonical position operators in L2 , respectively. Let σ1 , ..., σd denote 2[d/2] × 2[d/2] matrices satisfying the anti-commutation relations: {σµ , σν } = 2δµν 1, µ, ν = 1, ..., d. Then σ j := (σ1j , · · · , σdj ) is given by σµj := 1 ⊗ · · · ⊗
the jth σµ
N
⊗ · · · ⊗ 1,
µ = 1, ..., d,
j = 1, ..., N,
Vol. 3, 2002
Self-adjointness of the Pauli-Fierz Hamiltonian
173
where σµ appears as the jth factor. V denotes an external potential. We assume that there exists a constant 0 ≤ aV < 1, and bV ≥ 0 such that
V f ≤ aV ∆f + bV f ,
f ∈ D(∆),
(1.4)
where D(T ) denotes the domain of T . ajex : Rd → R denotes an external vector potential assumed to be (1) ajex ∈ C 1 (Rd ),
(2) ajex ∞ < ∞,
(3) ∂µj ajex ∞ < ∞,
(1.5)
where · ∞ denotes the sup norm. The free Hamiltonian in F is defined by Hrad := ω(k)a†r (k)ar (k)dk with the dispersion relation ω(k) := |k|. ∼ L2 (RdN ; F ), the transverse vector poUnder the identification, L (R ) ⊗ F = tential quantized by the Coulomb gauge is defined by 1 ˆ ˆ + ar (k)eikx λ(k) dk, Aµ (x) := √ erµ (k) a†r (k)e−ikx λ(−k) 2 2
dN
where er = (er1 , ..., erd ) denotes polarization vectors satisfying er (k) · es (k) = δrs ˆ ∈ L2 is an ultraviolet cutoff and er (k) · k = 0 for almost everywhere k ∈ Rd . λ function, and physically reasonable choice is ρˆ(k) ˆ λ(k) = , (2π)d ω(k) where eρ denotes the electron charge density; ρ(x)dx = 1, Rd
ˆ ∈ L2 , µ = 1, ..., d, and fˆ is the Fourier transform of f . Note that in the case of kµ λ divA(x) = 0 on some dense domain. We assume that ˆ ˆ λ(−k) = λ(k) which ensures that HPF is a symmetric operator in Htotal . We abbreviate 1 ⊗ 1 ⊗ X, 1 ⊗ X ⊗ 1, X ⊗ 1 ⊗ 1 by X unless confusions may arise. Thus we write (1.3) as HPF =
N 2 1 j p + ajex (xj ) − eA(xj ) + V + Hf + Hspin . 2 j=1
174
F. Hiroshima
Ann. Henri Poincar´e
The spin term is given by Hspin := − where
bµν ex j
1 2
j σµj σνj eBµν (xj ) − bµν (x ) , ex j
1≤µ<ν≤d
denotes the external field, bexj := i∇j ∧ ajex ,
and Bµν the quantized magnetic field, j j 1 ˆ ˆ k ∧ er (k) e−ikx λ(−k) − eikx λ(k) dk. B(xj ) := i∇j ∧ A(xj ) = √ 2 A typical example is as follows: d = 3, (σ1 , σ2 , σ3 ) is the 2 × 2 Pauli matrices: 0 1 0 −i 1 0 σ1 := , σ2 := , σ3 := , 1 0 i 0 0 −1 and V the Coulomb potential smeared by the electron charge eρ, e2 ρ(xi − k)ρ(xj − k ) Vsmear (x1 , · · · , xN ) := − dkdk , | 4π |k − k 3 3 R ×R 1≤i<j≤N
or the Coulomb potential in M static nuclei: VCoulomb (x1 , · · · , xN , y 1 , · · · , y M ) := −
e2 4π
1≤i<j≤N
|e|Zj 1 + , i j |x − x | j=1 i=1 |xj − y i | N
M
where y 1 , ..., y M are constants in Rd , Z1 , ..., ZM positive constants. Over the past few decades a considerable number of studies have been made on the Pauli-Fierz type models (including spin-boson models and the Nelson model [31]) by many authors. In particular spectral analysis of the Pauli-Fierz type models has been investigated in e.g., [2]-[13], [15, 16] [19]-[29], [32, 36, 37]. Ground states are defined as the vectors associated with the bottom of the spectrum of a self-adjoint operator. Existence of ground states of the Pauli-Fierz type Hamiltonian was proven for weak couplings in [2, 4, 19, 20]. In particular, Bach-Fr¨ ohlich-Sigal [5] proved it without infrared cutoffs. For arbitrary values of coupling constants, existence of ground states has been shown by Spohn [37] for the Nelson Hamiltonian, by G´erard [15] for rather general models, and by GriesemerLieb-Loss [16] for the Pauli-Fierz Hamiltonian. The multiplicity of ground states has been established by Bach-Fr¨ohlich-Sigal [4] for the Nelson Hamiltonian, by Hiroshima [20] for a spinless Pauli-Fierz Hamiltonian, and by Hiroshima-Spohn [23] for the Pauli-Fierz Hamiltonian including spin. Resonances have been investigated by H¨ ubner-Spohn [26], Bach-Fr¨ ohlich-Sigal [3, 4], Bach-Fr¨ ohlich-Sigal-Soffer [6], and Jakˇsi´c-Pillet [27], moreover scattering theory by H. Krohn [24], H¨ ubner-Spohn [25], Derezi´ nski-G´erard [8], and Fr¨ ohlich-Griesemer-Schlein [13].
Vol. 3, 2002
1.2
Self-adjointness of the Pauli-Fierz Hamiltonian
175
The Friedrichs extension
To our best knowledge a general way to define a self-adjoint extension of HPF for all values of coupling constants is to take the Friedrichs extension since HPF consists of the quadratic part and a potential V . Let
Kµj := σµj pjµ + ajexµ − eAµ (xj ) . We define the form QPF (Φ, Ψ) by QPF (Φ, Ψ) =
1 j 1/2 1/2 (K Φ, Kµj Ψ) + (Hrad Φ, Hrad Ψ) + (|V |1/2 Φ, |V |1/2 Ψ) 2 µ
with the form domain 1/2
d j 1/2 Q := ∩N ). j=1 ∩µ=1 D(Kµ ) ∩ D(Hrad ) ∩ D(|V |
ˆ ∈ L2 . Since Q is dense, the Friedrichs extension of HPF exists for λ ˆ ∈ L2 . Then the Friedrichs extension of HPF exists. Proposition 1.1 Let λ ˆ is needed to define the Friedrichs It must be noted that no extra conditions on λ extension. As is well known, this extension, however, do not give information of the domain, and there still remains the possibility that several distinct self-adjoint extensions exist. In general, the dynamics of a system is determined by the one-parameter unitary time-evolution generated by the Hamiltonian. A densely defined Hamiltonian, however, may have several distinct self-adjoint extensions. Then we need to choose a particular self-adjoint extension to define a dynamics. In other words, distinct self-adjoint extensions determine distinct time-evolutions. It will be useful to keep this aspect in mind as we examine spectral properties of the Pauli-Fierz Hamiltonian derived by the Friedrichs extension.
1.3
Main results
In [21] we have shown that for arbitrary values of coupling constants HPF is essentially self-adjoint on D(∆) ∩ D(Hrad ) in the case when (1) N = 1 and (2) ˆ ∈ L2 . In this paper we prove self-adjointness on the same domain ˆ √ω, ω 2 λ λ/ ˆ √ω, ω λ ˆ ∈ L2 . Before D(∆) ∩ D(Hrad ) under the assumption (1) N ≥ 1 and (2) λ/ moving on to the main theorems it is desirable to describe why we pay attention ˆ to conditions on λ. 1.3.1 Domain and cutoffs ˆ We shall mention the relaWe shall now look carefully into the assumption on λ. ˆ tionship between D(HPF ) and the ultraviolet cutoff λ.
176
F. Hiroshima
Ann. Henri Poincar´e
ˆ kµ λ ˆ ∈ L2 , µ = 1, ..., d. Then, for Ψ ∈ C ∞ (RdN )⊗[F fin ∩ D(Hrad )] (⊗ Let λ, 0 denotes algebraic tensor product), it holds that Ψ ∈ D(pj · A(xj )) ∩ D(A(xj ) · pj ). By the Coulomb gauge, we have [pjµ , Aµ (xj )]Ψ =
N
divA(xj )Ψ = 0.
j=1
ˆ ωλ ˆ ∈ L2 , then Hence, in particular, if λ, N 2 1 j 1 e2 p − eA(xj ) = p2 − eA(xj ) · pj + A2 (xj ) 2 j=1 2 2
(1.6)
fin ∩D(Hrad )]. Put the right-hand side of (1.6) as K. Clearly holds on C0∞ (RdN )⊗[F j 2 1 N both 2 j=1 p − eA(xj ) and K are closable. Let us denote their closed exten2 N sions by the same symbols 12 j=1 pj − eA(xj ) and K, respectively, and let N be the number operator given by N := a†r (k)ar (k)dk. From the inequalities:
a†r (f )Ψ ≤ f ( N 1/2Ψ + Ψ ),
ar (f )Ψ ≤ f
N 1/2Ψ ,
ar (f )as (g)Ψ ≤ 8 f
g ( N Ψ + Ψ ), ˆ ωλ ˆ ∈ L2 , then where ar = ar or a†r , it follows that, if λ, ˆ 2 ( ∆Ψ + N Ψ + Ψ )
KΨ ≤ C λ
(1.7)
2 N with some constant C. Hence 12 j=1 pj − eA(xj ) is well defined on D(∆) ∩ D(N ). In addition to this, from the inequalities: √ 1/2
a†r (f )Ψ ≤ f / ω
Hrad Ψ + f
Ψ ,
(1.8)
√ 1/2 (1.9)
ar (f )Ψ ≤ f / ω
Hrad Ψ , √ √ √
ar (f )as (g)Ψ ≤ ( f / ω + f )( g/ ω + g + ωg + ωg ) ×( Hrad Ψ + Ψ ),
(1.10)
Vol. 3, 2002
Self-adjointness of the Pauli-Fierz Hamiltonian
177
ˆ √ω, ω λ ˆ ∈ L2 , then it follows that, if λ/
KΨ ≤ C ( ∆Ψ + Hrad Ψ + Ψ ),
(1.11)
ˆ √ω and ω λ . ˆ Thus we see that where C is a constant depending only on λ/ N 1 j j 2 is well defined on D(∆) ∩ D(Hrad ). j=1 p − eA(x ) 2 Let us summarize the main points that have been mentioned above. ˆ ωλ ˆ ∈ L2 , HPF is well defined on D(∆) ∩ Proposition 1.2 (1) In the case of λ, √ ˆ ˆ ∈ L2 , HPF is well defined on D(Hrad ) ∩ D(N ). (2) In the case of λ/ ω, ω λ D(∆) ∩ D(Hrad ). The Pauli-Fierz Hamiltonian can be split out as HPF = −∆/2 + Hrad + eHint + V + H(a), where Hint :=
1 j −p · A(xj ) − A(xj ) · pj + eA(xj )2 + Hspin , 2
and H(a) :=
1 j j p · aex + ajex · pj + (ajex )2 − eajex · A(xj ). 2
ˆ ∈ L2 . From (1.5), (1.8), (1.9), and (1.10), it follows that, ˆ √ω, ω λ Let λ/
Hint Ψ ≤ a (−∆/2 + Hrad )Ψ + b Ψ with some constant a and b, and for arbitrary , > 0
H(a)Ψ ≤ , (−∆/2 + Hrad )1/2 Ψ + b Ψ
(1.12)
with some constant b for Ψ ∈ D(−∆/2 + Hrad ) = D(∆) ∩ D(Hrad ). Then, by (1.4), we obtain
(eHint + V + H(a))Ψ ≤ (|e|a + aV + ,) (−∆/2 + Hrad )Ψ + (b + bV + b ) Ψ . ˆ √ω, ω λ ˆ ∈ L2 , for e such that Then in the case of λ/ |e|a + aV + , < 1, HPF is self-adjoint on D(∆) ∩ D(Hrad ) by the Kato-Rellich theorem. This was established in e.g., [4, 5, 6, 19, 20, 32].
178
F. Hiroshima
Ann. Henri Poincar´e
1.3.2 Main results Several articles also have been devoted to the study of self-adjointness of the Pauli-Fierz type Hamiltonians. It is noteworthy that Arai [1] proves that the PauliFierz Hamiltonian in the dipole approximation is self-adjoint for arbitrary values of coupling constants by means of the Nelson commutator theorem [34, Theorem X.36]. In contrast, self-adjointness of the full Pauli-Fierz Hamiltonian for all values of coupling constants is not obvious. From the argument of Section 1.3.1 we may, therefore, reasonably conclude that it is worth while to prove √ self-adjointness for arbitrary values of coupling ˆ ∈ L2 instead of λ/ ˆ √ω, ω 2 λ ˆ ω, ω λ ˆ ∈ L2 imposed constants under the condition λ/ in [21]. Furthemore by Proposition 1.2 (1) it is interesting to consider essential ˆ ωλ ˆ ∈ L2 but λ/ ˆ √ω ∈ L2 . Our main results in this self-adjointness of HPF for λ, paper are as follows: ˆ √ω, ω λ ˆ ∈ L2 . Then HPF is self-adjoint on D(∆) ∩ Theorem 1.3 Suppose that λ/ D(Hrad ) and bounded from below. Moreover HPF is essentially self-adjoint on any core of −∆/2 + Hrad . ˆ √ω, ω λ ˆ ∈ L2 . Then there exists a constant C such that Corollary 1.4 Let λ/
∆Ψ + Hrad Ψ ≤ C( HPF Ψ + Ψ ),
Ψ ∈ D(∆) ∩ D(Hrad ).
Proof. From Theorem 1.3 and the closed graph theorem,
(−∆/2 + Hrad )Ψ ≤ C ( HPF Ψ + Ψ ),
Ψ ∈ D(∆) ∩ D(Hrad ),
follows with some constant C . Thus the corollary holds true.
Next we consider a more √ subtle case. As we have mentioned in the previous ˆ ω ∈ L2 , HPF is a priori not defined on D(∆) ∩ subsection, for the case of λ/ n D(Hrad ) but well defined on D(∆) ∩ D(Hrad ) ∩ D(N ). Let C ∞ (T ) := ∩∞ n=1 D(T ). ˆ ωλ ˆ ∈ L2 . Then HPF is essentially self-adjoint on Theorem 1.5 Suppose that λ, ∞ ∞ C (∆) ∩ D(Hrad ) ∩ C (N ) and bounded from below. Finally also it is of mathematical interest to consider essential self-adjointness of N 2 1 j p − eA(xj ) + V. 2 j=1
(1.13)
Our method is also available to (1.13). ˆ ωλ ˆ ∈ L2 . Then 1 N pj − eA(xj ) 2 + V is essentially selfTheorem 1.6 Let λ, j=1 2 adjoint on C ∞ (∆) ∩ C ∞ (N ) and bounded from below.
Vol. 3, 2002
Self-adjointness of the Pauli-Fierz Hamiltonian
179
The outline of our strategy is from Step 1 to Step 4. Step 1: Let N 2 1 j H := p − eA(xj ) + Hrad , 2 j=1 then HPF = H + V + Hspin + H(a). For sufficiently small coupling constants, the matrix elements of the heat semigroup e−tH can be expressed by a functional integral. This type of functional integrals were first studied in [30], and there have been numerous related works e.g., [14, 17, 10]. We see that for arbitrary values of coupling constants the functional integral itself is well defined and yields a heat semigroup generated by a self-adjoint operator H. Step 2: we prove that ⊃ HD(∆)∩D(H ) . H rad ˆ √ω, ω 2 λ ˆ ∈ L2 . In This statement was checked in [18] only for the case when λ/ √ 2 ˆ ˆ this paper we relax this assumption to λ/ ω, ω λ ∈ L . b Step 3: we show that D(∆) ∩ D(Hrad ) is invariant under e−tH , i.e., b
e−tH {D(∆) ∩ D(Hrad )} ⊂ D(∆) ∩ D(Hrad ), is essentially self-adjoint on D(∆)∩D(Hrad ) by [34, Theorem which implies that H X.49]. Step 4: We prove that + E)−1 (−∆ + Hrad )(H is closed on D(∆)∩D(Hrad ). is bounded for every E > 0. Thus we conclude that H Hence H is self-adjoint on D(∆) ∩ D(Hrad ) and bounded from below, and H is also self-adjoint on D(∆) ∩ D(Hrad ) and bounded from below. For arbitrary , > 0 there exists c > 0 such that, for Ψ ∈ D(∆) ∩ D(Hrad ),
Hspin Ψ ≤ , (−∆/2 + Hrad )Ψ + c Ψ , and (1.4) derives
V Ψ ≤ aV HΨ + bV Ψ with some 0 ≤ aV < 1 and 0 ≤ bV . This, together with (1.12) and Corollary 1.4, implies that HPF = H + V + Hspin + H(a) is self-adjoint on D(∆) ∩ D(Hrad ). The paper is organized as follows: In Section 2 we construct a self-adjoint extension in terms of a functional integral. In Section 3 we show some relative bounds by using the functional integral representation. In Section 4 we prove the main theorems. Section 5 is devoted to an appendix.
180
F. Hiroshima
Ann. Henri Poincar´e
2 A self-adjoint extension of HPF 2.1
A short review of a probabilistic description
For a moment let us look at H=
N 2 1 j p − eA(xj ) + Hrad 2 j=1
acting in H = L2 (RdN ) ⊗ F. It is useful to take a Schr¨odinger representation of F to investigate the semigroup generated by H. Let W := ⊕d L2 and W0 := ⊕d L2 (Rd+1 ). q and q0 denote bilinear forms defined by 1 dµν (k)fˆµ (k)ˆ gν (k)dk, f, g ∈ W, q(f, g) := 2 Rd q0 (f, g) := dµν (k)fˆµ (k, k0 )ˆ gν (k, k0 )dkdk0 , f, g ∈ W0 , Rd+1
where dµν (k) :=
erµ (k)erν (k)
= δµν
kµ kν − = |k|2
k k ⊗ 1− , |k| |k| µν
k ∈ Rd ,
is the transverse projection. Let (Q, µ) and (Q0 , µ0 ) be probability spaces associated with the Gaussian random variables of zero mean (φ(f ), f ∈ W ) and (φ0 (g), g ∈ W0 ), respectively. Their covariances are given by 1 φ(f )φ(g)µ(dφ) = q(f, g), f, g ∈ W, 2 Q and
φ0 (f )φ0 (g)µ(dφ0 ) = Q0
1 q0 (f, g), 2
f, g ∈ W0 ,
respectively. The finite particle subspace L2fin(Q) is defined by L2fin (Q) := L.H.{:φ(f1 ) · · · φ(fn ):, 1|fj ∈ W, j = 1, ..., n, n ≥ 1}, where L.H.{· · · } is the linear hull of the vectors in {· · · } and :X: denotes the Wick product of X. Let ω : W → W be a self-adjoint operator defined by ω := ⊕d ω(−i∇).
Vol. 3, 2002
Self-adjointness of the Pauli-Fierz Hamiltonian
181
Define Hf and Nf by Hf :φ(f1 ) · · · φ(fn ):=
n
:φ(f1 ) · · · φ(ˆ ω fj ) · · · φ(fn ):,
j=1
Hf 1 = 0, and Nf :φ(f1 ) · · · φ(fn ):= n :φ(f1 ) · · · φ(fn ):, Nf 1 = 0. Let Aµ (λ) := φ(⊕dν=1 δµν λ),
µ = 1, ..., d.
The Hilbert space HS is given by HS := L2 (RdN ) ⊗ L2 (Q), and the following identification HS ∼ = L2 (RdN ; L2 (Q)) is used without notice. We define a symmetric operator HS on HS by HS := HA + Hf , where
1 j (p − eA(λ(· − xj )))2 . 2 j=1 N
HA :=
We define a linear operator T of F to L2 (Q) by ˆ 1 ) · · · Aµ (λ ˆ n ): Ω =:Aµ (λ1 ) · · · Aµ (λn ):, • T :Aµ1 (λ n 1 n • T Ω = 1. It is seen that T can be extended to a unitary operator of F to L2 (Q). We denote the extension as the same symbol T . In particular T : D(N l ) −→ D(Nfl ), l ) −→ D(Hfl ), T : D(Hrad
for arbitrary l ≥ 0. Moreover T implements the equivalences Aµ (x) ∼ = Aµ (λ(· − x)) for each x ∈ Rd , N ∼ = Nf , and Hrad ∼ = Hf . Thus we proved the following proposition.
182
F. Hiroshima
Ann. Henri Poincar´e
ˆ √ω, ω λ ˆ ∈ L2 . Then Proposition 2.1 (1)Let λ/ T HT −1D(∆)∩D(Hf ) = HS D(∆)∩D(Hf ) . ˆ ωλ ˆ ∈ L2 . Then (2) Let λ, T HT −1D(∆)∩D(Hf )∩D(Nf ) = HS D(∆)∩D(Hf )∩D(Nf ) . ˆ ωλ ˆ ∈ L2 . Then (3)Let λ, T
N 2 1 j p − eA(xj ) T −1 D(∆)∩D(Nf ) = HA D(∆)∩D(Nf ) . 2 j=1
Then, instead of H, we shall study self-adjointness of HS for the time being.
2.2
A self-adjoint extension
Let {b(T )}T ≥0 = {bjµ (T )}µ=1,...,d,j=1,...,N,T ≥0 denote the dN -dimensional Brownian motion starting at the origin on a probability space (P, db). We set XT := x + b(T ), XTjµ = xjµ + bjµ (T ), M := RdN × P , and dX := dx ⊗ db. We see that the map Λ : t → λ(· − t) ˆ ˆ ∈ L2 . is an L2 -valued function on Rd . Note that (λ(· − t))(k) = e−ik·t λ(k). Let ω n λ n d 2 Then Λ ∈ Cb (R ; L ), the set of n-times continuously differentiable bounded (up ˆ ωλ ˆ ∈ L2 . Then the L2 to n-times derivatives) L2 -valued functions on Rd . Let λ, valued stochastic integral, t λ(· − Xs )dbjµ (s) ∈ L2 (P ) ⊗ L2 , 0
ˆ ω2λ ˆ ∈ L2 , it is proven that is well defined. Note that, in the case λ, 2n
1 j j jµ jµ λ(· − Xtk/2 Xtk/2 s − lim n ) + λ(· − Xt(k−1)/2n ) n − Xt(k−1)/2n n→∞ 2 k=1
t
λ(· − Xs )dbjµ (s)
= 0
in L2 (P )⊗L2 . Let Ξt : L2 (Q) → L2 (Q0 ) be the second quantization of the isometry ⊕d ξt : W → W0 , which is defined by −itk0 ω(k) e fˆ(k), f ∈ L2 . (ξ t f )(k, k0 ) = √ π ω(k)2 + |k0 |2
Vol. 3, 2002
Self-adjointness of the Pauli-Fierz Hamiltonian
183
Then Ξt also is an isometry and satisfies Ξ∗s Ξt = e−|s−t|Hf .
(2.1)
Since ξs λ is strongly continuous in s in L2 (Rd+1 ), the L2 (Rd+1 )-valued stochastic integral, t
ξs λ(· − Xs )dbjµ (s) ∈ L2 (P ) ⊗ L2 (Rd+1 ), 0
is well defined. Let
Jt0 := Jt0 (X) = e−ieφ(L(t)) ,
where L(t) := L(X, t) :=
N
λ(· − Xs )dbjµ (s). 0
j=1
Moreover let
t
⊕dµ=1
Jt := Jt (X) = Ξ∗0 e−ieφ0 (K(t)) Ξt ,
where K(t) := K(X, t) :=
N
⊕dµ=1
ξs λ(· − Xs )dbjµ (s). 0
j=1
We define, for F, G ∈ H,
t
0 IF,G
(F0 , Jt0 Gt ) dX ,
:= M
and
(F0 , Jt Gt ) dX ,
IF,G := M
where F0 := F (X0 ), Gt = G(Xt ). The key lemma of this paper is as follows. ˆ ωλ ˆ ∈ L2 . Then there exists a nonnegative self-adjoint operLemma 2.2 (1) Let λ, ator HA such that b 0 IF,G = (F, e−tHA G), F, G ∈ HS , A ) ⊃ D(∆) ∩ D(Nf ) with and D(H A ⊃ HA D(∆)∩D(N ) . H f ˆ √ω ∈ L2 . Then D(H A ) ⊃ D(∆) ∩ D(Hf ) with In addition, assume λ/ A ⊃ HA D(∆)∩D(H ) . H f ˆ ωλ ˆ ∈ L2 . Then there exists a nonnegative self-adjoint operator H S such (2) Let λ, that b IF,G = (F, e−tHS G), F, G ∈ HS ,
184
F. Hiroshima
Ann. Henri Poincar´e
S ) ⊃ D(∆) ∩ D(Hf ) ∩ D(Nf ) with and D(H S ⊃ HS D(∆)∩D(H )∩D(N ) . H f f ˆ √ω ∈ L2 . Then D(H S ) ⊃ D(∆) ∩ D(Hf ) with In addition, assume λ/ S ⊃ HS D(∆)∩D(H ) . H f Proof. See Appendix.
3 Relative bounds and invariant domains We first state a fundamental fact. Lemma 3.1 Let K be a Hilbert space. (1) Let T be a densely defined closed operator in K and g ∈ K. Let D be a core of T . Moreover suppose (3.1) |(T f, g)| ≤ Rg f for all f ∈ D with some constant Rg . Then g ∈ D(T ∗ ) with T ∗ g ≤ Rg . (2) Let S be a densely defined closed operator and U a densely defined linear operator in K. Let DS be a core of S and DU the domain of U . Suppose that |(U f, Sg)| ≤ R f
g
(3.2)
for all f ∈ DU and g ∈ DS with some constant R. Then U ∗ S has unique bounded operator extension U ∗ S with U ∗ S ≤ R. In particular, if S is a bounded operator, then U ∗ S is a bounded operator.
3.1
Relative bounds and invariant domains for ∆
Let Hd := −∆/2 + Hf . ˆ ωλ ˆ ∈ L2 , E > 0 and t > 0. Then both ∆k e−tHbS , k ≥ 0, and Lemma 3.2 Let λ, S + E)−1 are bounded operators. In particular, e−tHbS leaves C ∞ (∆) invariant. ∆(H Proof. We fix k ≥ 0. From the functional integral representation (5.19) it follows that b (3.3) |(F, e−tHS G)| ≤ (|F |, e−tHd |G|). 2fin (Q) ∩ C ∞ (Hf )]. Let FR denote the real part of F Take F ∈ D := C0∞ (RdN )⊗[L and FI its imaginary part, i.e., F = FR + iFI . Define subsets of RdN × Q by dN Q+ × Q|(∆k FR )(x, φ) > 0}, R := {(x, φ) ∈ R dN Q− × Q|(∆k FR )(x, φ) < 0}, R := {(x, φ) ∈ R dN × Q|(∆k FI )(x, φ) > 0}, Q+ I := {(x, φ) ∈ R dN Q− × Q|(∆k FI )(x, φ) < 0}. I := {(x, φ) ∈ R
Vol. 3, 2002
Self-adjointness of the Pauli-Fierz Hamiltonian
Define also F1 :=
FR (x, φ), 0,
−FR (x, φ), 0, FI (x, φ), F3 := 0, −FI (x, φ), F4 := 0,
185
(x, φ) ∈ Q+ R, (x, φ) ∈ Q+ R, (x, φ) ∈ Q− R, (x, φ) ∈ Q− R,
F2 :=
(x, φ) ∈ Q+ I , (x, φ) ∈ Q+ I , (x, φ) ∈ Q− I , (x, φ) ∈ Q− I .
Note that Fi ∈ D(∆k ) and ∆k Fi > 0, i = 1, 2, 3, 4. Since |∆k FR | = ∆k F1 + ∆k F2 and |∆k FI | = ∆k F3 + ∆k F4 , we have b
b
b
|(∆k F, e−tHS G)| ≤ |(∆k FR , e−tHS G)| + |(∆k FI , e−tHS G)| ≤ |(|∆k FR |, e−tHd |G|)| + |(|∆k FI |, e−tHd |G|)| =
4
(∆k Fi , e−tHd |G|) ≤
i=1
4
Fi
G .
i=1
Since F1 2 + F2 2 = FR 2 , F3 2 + F4 2 = FI 2 , and FR 2 + FI 2 = F 2 , 4 we have i=1 Fi ≤ 2 F . Hence b
|(∆k F, e−tHS G)| ≤ 2 F
G .
(3.4)
b
Since D is a core of ∆k , the boundedness of ∆k e−tHS follows from Lemma 3.1. By means of (3.3) we have S + E)−1 G)| ≤ (|F |, (Hd + E)−1 |G|). |(F, (H S + E)−1 HS ⊂ D(∆) and Note that ∆(Hd + E)−1 is bounded. Thus, (H S + E)−1 G)| ≤ 2 F
G |(∆F, (H S + E)−1 is bounded by Lemma 3.1. follows in a similar way as (3.4). Hence ∆(H Thus we get the desired results. ˆ ωλ ˆ ∈ L2 . Then the same statements as Lemma 3.2 hold with Lemma 3.3 Let λ, S replaced by H A . In particular, e−tHbA leaves C ∞ (∆) invariant. H Proof. We have
b
|(F, e−tHA G)| ≤ (|F |, e−t(−∆/2) |G|). Hence, the same proof as of Lemma 3.2 yields the desired results.
186
F. Hiroshima
3.2
Ann. Henri Poincar´e
Relative bounds and invariant domains for Hf and Nf
We identify W0 = ⊕d L2 (Rd+1 ) ∼ = L2 (R) ⊗ W . Let ω 0 := 1 ⊗ ω : W0 → W0 . Let E refer to the expectation value with respect to db. The following is useful for b estimating matrix elements of e−tHS . k/2 ˆ ∈ L2 . Then K(t) ∈ D( ˆ ωλ ˆ ∈ L2 and ω k/2 λ ω0 ) and Proposition 3.4 Let λ,
(2m)!
k/2 ˆ 2m tm dm N 2m ω k/2 λ E ω0 K(t) 2m W0 ≤ L2 . 2m
Proof. See [21, Theorem 4. 6]. We define Hf0 : L2 (Q0 ) → L2 (Q0 ) by Hf0 :φ0 (f1 ) · · · φ0 (fn ):=
n
:φ0 (f1 ) · · · φ0 ( ω0 fj ) · · · φ0 (fn ):,
j=1
Hf0 1 = 0. Then
Ξ∗t Hf0 = Hf Ξ∗t .
(3.5)
Let π0 (f ) := i[Hf0 , φ0 (f )]. Thus we have [φ0 (f ), π0 (g)] = iq0 ( ω0 f¯, g).
(3.6)
ˆ ωλ ˆ ∈ L2 . Then it is easily seen that Jt maps D(Hf ) onto itself. From (3.5) Let λ, and (3.6) it follows that e2 ∗ −ieφ0 (K(t)) [Hf , Jt ] = Ξ0 e ω0 K(t), K(t)) Ξt (3.7) −eπ0 (K(t)) + q0 ( 2 holds on D(Hf ). ˆ ωλ ˆ ∈ L2 . Then for Ψ ∈ D(Hf ) we have e−tHbS Ψ ∈ D(Hf ), and Lemma 3.5 Let λ, there exists a constant C such that √ b
[e−tHS , Hf ]Ψ ≤ C(t + t) (Hf + 1)1/2 Ψ . (3.8) Proof. Let F, G ∈ D(Hf ). Note that (Hf F )(x) = Hf (F (x)) for each x ∈ RdN . Then bS −tH (Hf F, e G) = (F0 , Hf Jt Gt ) dX M
Vol. 3, 2002
Self-adjointness of the Pauli-Fierz Hamiltonian
187
(F0 , [Hf , Jt ]Gt ) dX +
=
(F0 , Jt Hf Gt ) dX
M
M
b
(F0 , [Hf , Jt ]Gt ) dX + (F, e−tHS Hf G).
= M
We have by (3.7)
(F0 , [Hf , Jt ]Gt ) dX M
e2 ∗ −ieφ0 (K(t)) = ω0 K(t), K(t)) Ξt Gt dX . F0 , Ξ0 e −eπ0 (K(t)) + q0 ( 2 M
Note that
π0 (K(t))Ξt Ψ ≤ θ (Hf + I)1/2 Ψ with
d−1 1/2 ω0 K(t) ). ω0 K(t) + θ := √ (2 2
Moreover 1/2
ω0 K(t), K(t))| ≤ ω0 K(t) 2 := η. |q0 ( Using Proposition 3.4, we have 1/2
ω0 K(t) 2 + E ωK(t) 2 ) E(θ2 ) ≤ (d − 1)2 (4E ˆ 2) ˆ 2 + tN 2 d ω λ ≤ (d − 1)2 (4tN 2 d ω 1/2 λ ˆ 2 ), ˆ 2 + ω λ ≤ 4tN 2 d(d − 1)2 ( ω 1/2 λ and ˆ 4. E(η 4 ) ≤ 6t2 N 4 d2 ω 1/2 λ By the Schwartz inequality we have (F0 , [Hf , Jt ]Gt ) dX M
1/2
≤ |e|
1/2
F0 2 θ2 dX
(Hf + I)1/2 Gt 2 dX
M
M 2
e + 2
1/2
F0 η dX M
Rd
≤
2
M
It then follows that 1/2 2 2
F0 θ dX ≤ M
1/2
Gt dX
2 4
1/2 (Eθ
2
) F (x) 2L2 (Q) dx
ˆ + ω λ ) F ˆ 4tN 2 d(d − 1)2 ( ω 1/2 λ
,
.
188
F. Hiroshima
and
1/2
F0 η dX 2 4
M
≤
≤
Ann. Henri Poincar´e
1/2 (Eη
Rd
4
) F (x) 2L2 (Q) dx
√ ˆ 2 F . 6t2 N 4 d2 ω 1/2 λ
√ (F0 , [Hf , Jt ]Gt ) dX ≤ C( t + t) F
(Hf + I)1/2 G
Thus
(3.9)
M
with C := max
1√ 4 2 ˆ 2 ˆ + ω λ ), ˆ 4N 2 d(d − 1)2 ( ω 1/2 λ 6N d ω λ . 2
Hence we proved that √ b |(Hf F, e−tHS G)| ≤ C( t + t) + 1 (Hf + 1)G
F . b
Then e−tHS G ∈ D(Hf ) by Lemma 3.1. Moreover, since bS −tH , Hf ]G) = (F0 , [Hf , Jt ]Gt ) dX , (F, [e M
(3.8) follows from (3.9).
ˆ ωλ ˆ ∈ L2 . Let Ψ ∈ D(Hf ) and E > 0. Then, for all k > 0, Lemma 3.6 Let λ, S + E)−k Ψ ∈ D(Hf ) and (H ∞ b S + E)−k Ψ = 1 Hf (H t−1+k Hf e−tHS e−tE Ψdt, (3.10) Γ(k) 0 ∞ where Γ(k) := 0 e−x xk−1 dx is the usual Gamma function. Proof. Note that S + E)−k Ψ = (H
1 Γ(k)
∞
b
e−tHS e−tE t−1+k Ψdt
0
in the strong sense. We have b
b
b
Hf e−tHS Ψ ≤ [e−tHS , Hf ]Ψ + e−tHS Hf Ψ √ ≤ C(1 + t + t) (Hf + 1)Ψ with some constant C. Let Φ ∈ D(Hf ). It is obtained that ∞ 1 b −k (Hf Φ, (HS + E) Ψ) = (Hf Φ, e−tHS e−tE t−1+k Ψ)dt Γ(k) 0
(3.11)
Vol. 3, 2002
Self-adjointness of the Pauli-Fierz Hamiltonian
=
1 Γ(k)
∞
189
b
(Φ, Hf e−tHS e−tE t−1+k Ψ)dt.
0
By virtue of (3.11), S + E)−k Ψ)| ≤ |(Hf Φ, (H ≤
1 Γ(k)
∞
C(1 +
1 Γ(k)
∞
b
Φ
Hf e−tHS e−tE t−1+k Ψ dt
0
√ t + t)e−tE t−1+k dt Φ
(Hf + 1)Ψ .
0
S + E)−k Ψ ∈ D(Hf ) by Lemma 3.1 and (3.10) holds by the strong Hence (H ∞ b integrability of 0 Hf e−tHS e−tE t−1+k Ψdt. Then the lemma follows. ˆ ωλ ˆ ∈ L2 . Then H 1/2 (H S + E)−1/2 is a bounded operator. Lemma 3.7 Let λ, f S G) for G ∈ D(∆) ∩ D(Hf ) ∩ D(N ), Proof. Since (G, Hf G) ≤ (G, HS G) = (G, H we have 1/2
Hf
S + E)−1/2 G ≤ G , (H
G ∈ D(∆) ∩ D(Hf ) ∩ D(N ).
Thus the lemma follows, since D(∆) ∩ D(Hf ) ∩ D(N ) is dense.
ˆ ωλ ˆ ∈ L2 . Then Hf (H S + E)−1 is a bounded operator. Lemma 3.8 Let λ, S + E)−1/2 Ψ ∈ D(Hf ) Proof. Let Ψ ∈ D(Hf ). From Lemma 3.6, it follows that (H −1 and (HS + E) Ψ ∈ D(Hf ). Then we have S + E)−1 Ψ = Hf (H S + E)−1/2 Ψ + [(H S + E)−1/2 , Hf ](H S + E)−1/2 Ψ, S + E)−1/2 Hf (H (H and
S + E)−1/2 , Hf ](H S + E)−1/2 Ψ = [(H ∞ 1 1 b S + E)−1/2 Ψdt √ [e−tHS , Hf ]e−tE (H Γ(1/2) 0 t
holds. We have
S + E)−1/2 , Hf ](H S + E)−1/2 Ψ
[(H ∞√ t+t C S + E)−1/2 Ψ e−tE dt √ (Hf + I)1/2 (H ≤ Γ(1/2) 0 t ∞ √ C S + E)−1/2
Ψ . ≤ (1 + t)e−tE dt (Hf + I)1/2 (H Γ(1/2) 0
(3.12)
190
F. Hiroshima
Ann. Henri Poincar´e
S + E)−1/2 Hf (H S + E)−1/2 is a bounded operator. Hence, by By Lemma 3.7, (H virtue of (3.12), we have S + E)−1 Ψ ≤ C Ψ
Hf (H with some constant C . Since D(Hf ) is dense, the lemma follows.
Next we show a relative bound of Nf . The procedure is similar to that of Hf . We define Nf0 : L2 (Q0 ) → L2 (Q0 ) by Nf0 :φ0 (f1 ) · · · φ0 (fn ):= n :φ0 (f1 ) · · · φ0 (fn ): . Nf0 1 = 0. Then
Ξ∗t Nf0 = Nf Ξ∗t .
Let
(3.13)
π0 (f ) := i[Nf0 , φ0 (f )].
Thus we have
[φ0 (f ), π0 (g)] = iq0 (f¯, g).
For arbitrary l ≥ 0, Jt maps Nfl Jt
=
Ξ∗0 e−ieφ0 (K(t))
D(Nfl )
(3.14)
onto itself and
l e2 0 Ξt −eπ0 (K(t)) + q0 (K(t), K(t)) + Nf 2
(3.15)
holds on D(Nfl ). ˆ ωλ ˆ ∈ L2 . Then both of e−tHbS and e−tHbA leave C ∞ (Nf ) invariLemma 3.9 Let λ, ant. Proof. Let F, G ∈ C ∞ (Nf ). Then b
(Nfl F, e−tHS G) =
M
(F0 , Nfl Jt Gt ) dX .
By (3.15), we easily obtain
Nfl Jt Gt ≤ C1 (X) (Nf + 1)l Gt . Here C1 (X) = C1 (X, t) is a constant depending on K(t) n , n = 0, 1, 2, ..., l, and satisfies by (3.4), E(C1 (X)2 ) ≤ P (t), where P (t) = al tl + · · · + a1 t + a0 with some aj ≥ 0. Hence b
|(Nfl F, e−tHS G)| ≤ P (t) F
(Nf + 1)l G .
Vol. 3, 2002
Self-adjointness of the Pauli-Fierz Hamiltonian
191
b
Thus e−tHS G ∈ D(Nfl ). Since l is arbitrary, b
e−tHS C ∞ (Nf ) ⊂ C ∞ (Nf ). Similarly we have b
(Nfl G, e−tHA G) =
M
(Nfl F0 , e−ieφ(L(t)) Gt ) dX =
M
(F0 , Nfl e−ieφ(L(t)) Gt ) dX .
We also see that
Nfl e−ieφ(L(t)) Gt ≤ C2 (X) (Nf + 1)l Gt (X) with E(C2 (X)2 ) ≤ Q(t), where Q(t) = bl tl + · · · + b1 t + b0 with some bj ≥ 0. Hence b
|(Nfl F, e−tHA G)| ≤ Q(t) F
(Nf + 1)l G . b
Thus the lemma follows in the same way as with e−tHS .
4 Proofs of the main theorems ˆ √ω, ω λ ˆ ∈ L2 . Then HS is self-adjoint on D(∆) ∩ D(Hf ). Lemma 4.1 Let λ/ S ) ⊃ D(∆) ∩ D(Hf ). Since e−tHb S leaves D(∆) ∩ D(Hf ) Proof. Note that D(H S . By Lemmas 3.2 and 3.8, we have invariant, D(∆) ∩ D(Hf ) is a core of H S Ψ + Ψ ),
∆Ψ + Hf Ψ ≤ C( H
S ), Ψ ∈ D(H
(4.1)
with some constant C, and by (1.8)-(1.10) S Ψ = HS Ψ ≤ C ( ∆Ψ + Hf Ψ ),
H
Ψ ∈ D(∆) ∩ D(Hf ),
S is closed on D(∆) ∩ D(Hf ). Thus H S is with some positive constant C . Hence H self-adjoint on D(∆) ∩ D(Hf ). By Lemma 2.2, we have HS D(∆)∩D(Hf ) ⊂ HS . Thus HS is self-adjoint on D(∆) ∩ D(Hf ). ˆ √ω, ω λ ˆ ∈ L2 . From the proof of Lemma 4.1 we conclude that Remark 4.2 Let λ/ S . HS = H
(4.2)
ˆ ωλ ˆ ∈ L2 . Then HS is essentially self-adjoint on C ∞ (∆) ∩ Lemma 4.3 Let λ, ∞ D(Hf ) ∩ C (Nf ).
192
F. Hiroshima
Ann. Henri Poincar´e
b
S ) ⊃ D(∆) ∩ D(Nf ) ∩ D(Hf ). Since e−tHS leaves C ∞ (∆) ∩ Proof. Note that D(H ∞ S is essentially self-adjoint on C ∞ (∆) ∩ D(Hf ) ∩ D(Hf ) ∩ C (Nf ) invariant, H S ⊃ HS D(∆)∩D(N )∩D(H ) . Thus HS is essentially selfC ∞ (Nf ). By Lemma 2.2, H f f ∞ adjoint on C (∆) ∩ D(Hf ) ∩ C ∞ (Nf ). ˆ ωλ ˆ ∈ L2 . Then HA is essentially self-adjoint on C ∞ (∆) ∩ Lemma 4.4 Let λ, ∞ C (Nf ). A ) ⊃ D(∆) ∩ D(N ). Since e−tHb A leaves C ∞ (∆) ∩ C ∞ (Nf ) Proof. Note that D(H A ⊃ A is essentially self-adjoint on C ∞ (∆)∩C ∞ (Nf ). By Lemma 2.2, H invariant, H HA D(∆)∩D(Nf ) . Thus HA is essentially self-adjoint on C ∞ (∆) ∩ C ∞ (Nf ). ˆ ωλ ˆ ∈ L2 . Then Hspin + H(a) in the Schr¨ Lemma 4.5 (1) Let λ, odinger represenˆ ˆ tation is infinitesimally small with respect to HS . (2) Let λ, ω λ ∈ L2 . Then V is relatively bounded with respect to both HA and HS with the relative bound aV < 1. Proof. Since for arbitrary , > 0 we have
(Hspin + H(a))Ψ ≤ , (−∆/2 + Hf )Ψ + b Ψ with some constant b , (4.1) and (4.2) imply (1). (2) follows from [21, Lemma 5.10]. A proof of Theorem 1.3 By Lemmas 4.1 and 4.5, HS +V +Hspin +H(a) is self-adjoint on D(∆)∩D(Hf ) and bounded from below. By Proposition 2.1, HPF is self-adjoint on D(∆) ∩ D(Hrad ). The essential self-adjointness of HPF follows from Corollary 1.4. A proof of Theorem 1.5 By Lemmas 4.3 and 4.5, HS + V + Hspin + H(a) is essentially self-adjoint on C ∞ (∆) ∩ D(Hf ) ∩ C ∞ (N ). Since the unitary operator T in Proposition 2.1 maps T −1 : C ∞ (∆)∩D(Hf )∩C ∞ (Nf ) → C ∞ (∆)∩D(Hrad )∩C ∞ (N ), HPF is essentially self-adjoint on C ∞ (∆) ∩ D(Hrad ) ∩ C ∞ (N ). A proof of Theorem 1.6 By Lemmas 4.4 and 4.5, HA + V is essentially self-adjoint on C ∞ (∆) ∩ C ∞ (Nf ). Since the unitary operator T in Proposition 2.1 maps T −1 : C ∞ (∆) ∩ C ∞ (Nf ) → j j 2 ∞ C ∞ (∆)∩C ∞ (N ), 12 N j=1 p − eA(x ) +V is essentially self-adjoint on C (∆)∩ ∞ C (N ).
5 Appendix ˆ √ω, ω λ ˆ ∈ L2 and In this Appendix we prove Lemma 2.2. For the case when λ/ 2ˆ 2 ω λ ∈ L , Lemma 2.2 has been proven in [21]. We want to remove the assumption ˆ ∈ L2 . We have ω2λ IF,G = (F, Rt G).
Vol. 3, 2002
Self-adjointness of the Pauli-Fierz Hamiltonian
193
Here Rt G := E (Jt G(Xt )) .
(5.1)
Using Markov properties of Ξt [35] and the Brownian motion b(t), we can directly prove that Rt has the semigroup property, i.e., Rt+s G = Rt Rs G. Thus Rt defines the semigroup in HS . It is not easy, however, to see directly from (5.1) that its self-adjoint generator is a certain extension of HS . So let us devote a little more space to proving Lemma 2.2. We define a contractive symmetric operator Qs as follows. For F ∈ HS , let (Q0 F )(x) := F (x), and j j j j (Qs F )(x) := ps (|x − y|)e(−ie/2)Aµ (λ(·−x )+λ(·−y ))(xµ −yµ ) F (y)dy, s > 0, RdN
(5.2) 2 ˆ ω2λ ˆ∈ for almost every x ∈ Rd , where ps (T ) := (2πs)−dN/2 e−T /(2s) , T ∈ R. Let λ, 2fin (Q) and G ∈ HS , a direct calculation shows that L2 . For F ∈ C0∞ (RdN )⊗L g(s) := (Qs F, G) is differentiable in s > 0 and lims→0+ g (s) = −(HA F, G). Hence lim
n→∞
(Qt/n F, G) − (F, G) = −(HA F, G), t/n
t > 0.
ˆ ∈ L2 and F, G ∈ HS . Put fn (s) := (Qs/2n )2n . Then Lemma 5.1 Let λ (F, fn (s)fm (t)G)HS = (F (X0 ), e−ieφ(Lmn (s,t)) G(Xs+t ))L2 (Q) dX .
(5.3)
M
Here Lmn (s, t) :=
N
jµ ⊕dµ=1 lmn (s, t),
j=1 jµ lmn (s, t)
1 2
:=
2m
j j jµ jµ λ(· − Xsk/2 Xsk/2 m +t ) + λ(· − Xs(k−1)/2m +t ) m +t − Xs(k−1)/2m +t
k=1 2n
1 j j jµ jµ + λ(· − Xtk/2 Xtk/2 n ) + λ(· − Xt(k−1)/2n ) n − Xt(k−1)/2n 2 k=1
ˆ ω2λ ˆ ∈ L2 , then In particular, if λ, (F (X0 ), e−ieφ(L(s+t)) G(Xs+t )) dX . lim (F, fn (s)fm (t)G) = m,n→∞
(5.4)
M
Proof. See [18, the proof of Lemma 4.6].
194
F. Hiroshima
Ann. Henri Poincar´e
Lemma 5.2 Let fn ∈ L2 (P ) ⊗ W . Assume that lim E fn − f 2W = 0. n→∞
Then s − limn→∞ eiφ(fn ) = eiφ(f ) on L2 (P ) ⊗ L2 (Q). Proof. It is enough to prove the weak convergence of eiφ(fn ) G for G in a dense domain in L2 (P ) ⊗ L2 (Q). Note that
1/2
eiφ(fn ) − 1 Φ L2 (Q) ≤ C fn W ( Nf Φ L2 (Q) + Φ L2 (Q) ) 1/2
for Φ ∈ D(Nf
) with some constant C. Thus it follows that, for
F ∈ L2 (P ) ⊗ L2 (Q)
G ∈ L∞ (P )⊗D(N f
1/2
and
),
F, (eiφ(fn ) − eiφ(f ) )G 2 2 L (P )⊗L (Q)
iφ(fn ) iφ(f )) = db F (b), e G(b) −e 2 L (Q) P
1/2 db F (b)
fn (b) − f (b) W Nf G(b) + G(b) ≤ P
1/2 1/2 sup Nf G(b) L2 (Q) + G(b) L2 (Q) . ≤ F K E fn − f 2W b∈Rd
Hence the lemma follows.
ˆ ωλ ˆ ∈ L2 . Then there exists a nonnegative self-adjoint operator Lemma 5.3 Let λ, A in HS such that, for F, G ∈ HS , H b
0 IF,G = (F, e−tHA G)HS .
(5.5)
ˆ ∈ L2 . Define ˆ ω2λ Proof. Let λ, St G := E Jt0 G(Xt ) . Then St ≤ 1. From (5.4) it follows that 0 . lim (F, fn (t)G) = (F, St G) = IF,G
n→∞
Thus it is seen that (F, St Ss G) =
lim (F, fn (t)fm (s)G)
m,n→∞
(5.6)
Vol. 3, 2002
Self-adjointness of the Pauli-Fierz Hamiltonian
195
(F (X0 ), e−ieφ(L(t+s)) G(Xt+s )) dX = (F, St+s G).
= M
Hence t, s ≥ 0.
St Ss = St+s ,
(5.7)
Since fn (t), n ≥ 0, is symmetric, we have St∗ = St ,
t ≥ 0.
(5.8)
ˆ ωλ ˆ ∈ L2 and ω 2 λ ˆn be such that λ ˆn , ω λ ˆn, ω2λ ˆ ∈ L2 . Let λ ˆ n ∈ L2 Next we let λ, 2 ˆ ˆ ˆ ˆ and that λn → λ and ω λn → ω λ strongly in L as n → ∞. Let St (n) and Ln (t) ˆ replaced by λ ˆ n , respectively. We have defined by St and L(t) with λ ˆ−λ ˆn 2 = 0. lim E L(t) − Ln (t) 2W = lim tdN λ (5.9) n→∞
n→∞
By Lemma 5.2 and the Lebesgue dominated convergence theorem, we obtain s − lim St (n) = St .
(5.10)
St Sr = s − lim St (n)Sr (n) = s − lim St+r (n) = St+r
(5.11)
s − lim St = 1
(5.12)
n→∞
We easily check that n→∞
n→∞
by (5.7), t→∞
by the definition of St , and finally St∗ = s − lim St (n)∗ = s − lim St (n) = St n→∞
n→∞
(5.13)
by (5.8). Then, (5.11), (5.12) and (5.13) imply that {St }t≥0 is a symmetric strongly continuous one-parameter semigroup. Thus there exists a nonnegative self-adjoint A such that operator H b e−tHA = St , t ≥ 0. Thus we get the desired results.
A . Next we investigate the domain of H ˆ ωλ ˆ ∈ L2 . Then H A ⊃ HA D(∆)∩D(N ) . (1) Let λ, f √ ˆ ∈ L2 . Then H ˆ ω, ω λ A ⊃ HA D(∆)∩D(H ) . (2) Let λ/ f
Lemma 5.4
ˆ ω2λ ˆ ∈ L2 . We have, by (5.2), 2fin (Q), F ∈ HS , and λ, Proof. Let G ∈ C0∞ (RdN )⊗L n
−1 2 1 −tHbA 1 Qt/2n − 1 j n e − 1 G, F = lim G, (Q ) F t/2 n→∞ t 2n t/2n j=0
196
F. Hiroshima
1
= lim
n→∞
0
Qt/2n − 1 G, (Qt/2n )[ns] F t/2n
ds =
1
Ann. Henri Poincar´e
b
−(HA G, e−tsHA F )ds.
(5.14)
0
Since we show that, by (5.14) and ˆ
HA G ≤ C λ ( ∆G + Nf G + G ),
(5.15)
ˆ ∈ L2 . equation (5.14) extends to G ∈ D(∆) ∩ D(Nf ) and F ∈ HS . Suppose ω 2 λ 2ˆ 2 ˆ ˆ ˆ ˆ ˆ Let λn be such that λn , ω λn , ω λn ∈ L and that λn → λ strongly as n → ∞. ˆ replaced by λ ˆ n . Thus (5.18) A (n) and HA (n) are defined by H A and HA with λ H A and HA , respectively. Then, by the A (n) and HA (n) instead of H holds with H definition of HA , b
b
s − lim e−tHA (n) = s − lim St (n) = St = e−tHA .
(5.16)
ˆ n − λ ( ∆F ˆ
(HA (n) − HA )F ≤ C λ
+ Nf F + F ),
(5.17)
n→∞
n→∞
Since it is seen that HA (n)F converges to HA F strongly as n → ∞, Hence, by a limiting argument 1
1 −tHbA b e − 1 G, F = −(HA G, e−tsHA F )ds, (5.18) t 0 ˆ with λ, ˆ ωλ ˆ ∈ L2 and for G ∈ D(∆) ∩ D(Nf ), F ∈ HS . Thus taking holds for λ ˆ ωλ ˆ ∈ L2 , then t → ∞ on both sides of (5.18), we see that, if λ, A ⊃ HA D(∆)∩D(N ) . H f ˆ ∈ L2 and ω 2 λ ˆ ∈ L2 . We have ˆ √ω, ω λ Next we let λ/
HA G ≤ C1 ( ∆G + Hf G + G ), ˆ √ω , λ , ˆ and ω λ . ˆ Hence (5.18) extends to G ∈ where C1 depends only on λ/ 2ˆ 2 ˆ ˆ n /√ω, λ ˆn , ω λ ˆn, ω2λ ˆ n ∈ L2 D(∆)∩D(Hf ). Suppose ω λ ∈ L . Let λn be such that λ √ √ ˆ n / ω → λ/ ˆ ω, λ ˆn → λ ˆ and ω λ ˆn → ωλ ˆ strongly as n → ∞. H A (n) and that λ and HA (n) are defined in the same way as above. It holds that
(HA (n) − HA )G ≤ C2 (n)( ∆G + Hf G + G ), ˆ n − λ)/ ˆ √ω , λ ˆ n − λ ˆ and ω(λ ˆ n − λ) ˆ with where C2 (n) depends on only (λ limn→∞ C2 (n) = 0. Then (5.18) holds for G ∈ D(∆) ∩ D(Hf ) and F ∈ HS . The ˆ √ω, ω λ ˆ ∈ L2 , then limiting argument similar to (5.16) leads to that, if λ/ A ⊃ HA D(∆)∩D(H ) . H f Thus the lemma follows. A ) ∩ D(Hf ) is dense, H S := H A + ˙ Hf is well defined. Since D(H
Vol. 3, 2002
Self-adjointness of the Pauli-Fierz Hamiltonian
197
ˆ ωλ ˆ ∈ L2 . Then H S ⊃ HS D(∆)∩D(N )∩D(H ) . (1) Let λ, f f √ 2 ˆ ∈ L . Then H ˆ ω, ω λ S ⊃ HS D(∆)∩D(H ) . (2) Let λ/ f
Lemma 5.5
Proof. It follows from Lemma 5.4. Next we construct a functional integral representation of e ˆ ω2λ ˆ ∈ L2 . Then Lemma 5.6 Let λ, b
(F, e−tHS G)HS = IF,G ,
bS −tH
.
F, G ∈ HS .
(5.19)
Proof. We show an outline of the proof. See [18] for details. By the Trotter-Kato product formula and (2.1) we have b
b
(F, e−tHS G) = lim (F, (e−tHSA /2 e−tHf /2 )2 G) n
n
n
n→∞
b
b
b
= lim (F, Ξ∗0 (Ξ0 e−tHSA /2 Ξ0 )(Ξ∗t/2n e−tHSA /2 Ξt/2n ) · · · (Ξt e−tHSA /2 Ξ∗t )Ξt G). n
n
n
n→∞
Using (5.2) and the Markov property [35] of Ξt , the right hand side above is = lim (Ξ0 F0 , e−ieφ0 (Kn (t)) Ξt Gt ), n→∞
where Kn (t) =
N
⊕dµ=1
j=1
n −1 (i+1)t/2n 2
it/2n
i=0
ξit/2n λ(· − Xs )dbjµ (s).
Hence we get (5.19). ˆ ωλ ˆ ∈ L2 but ω 2 λ ˆ ∈ L2 . We want to extend (5.19) for the case when λ, ˆ ωλ ˆ ∈ L2 . Then (5.19) holds true. Lemma 5.7 Let λ, Proof. We put
n n n 2 b Pn := e−tHA /2 e−tHf /2
and
n n n 2 b Pn (m) := e−tHA (m)/2 e−tHf /2 .
ˆ replaced by λ ˆm . A with λ A (m) is defined by H Here H bA (m) bA −tH −tH From s − limm→∞ e =e it follows that s − limm→∞ Pn (m) = Pn . It is easily proven that (Ξ0 F0 , e−ieφ0 (Knm (t)) Ξt Gt ) dX . (F, Pn (m)G) = M
Here
N 2 −1
n
Knm (t) :=
⊕dµ=1
j=1 i=0
Let
N 2 −1 j=1 i=0
ti/2n
n
Kn (t) :=
t(i+1)/2n
⊕dµ=1
ξti/2n λm (· − Xs )dbjµ (s).
t(i+1)/2n ti/2n
ξti/2n λ(· − Xs )dbjµ (s).
198
F. Hiroshima
Ann. Henri Poincar´e
Since ξs is an isometry, we have lim E Knm (t) − Kn (t) 2W0
m→∞ −1 N d 2
n
= lim
m→∞
E
−1 N d 2
n
m→∞
E
n 2 −1
m→∞
i=0
ξti/2n (λm (· − Xs ) − λ(· − Xs )) dbjµ (s) W0
2
t(i+1)/2n
(λm (· − Xs ) − λ(· − Xs )) dbjµ (s)
ti/2n
j=1 µ=1 i=0
= lim dN
2
t(i+1)/2n
ti/2n
j=1 µ=1 i=0
= lim
W
t ˆ ˆ 2 = lim tdN λ ˆ m − λ ˆ 2 = 0.
λm − λ m→∞ 2n
Hence, by Lemma 5.2, we have lim (F, Pn (m)G) = (F, Pn G) =
m→∞
(Ξ0 F0 , e−ieφ0 (Kn (t)) Ξt Gt ) dX .
(5.20)
M
Moreover by the triangle inequality, E Kn (t) − K(t) 2W0 ≤
N d
n 2 −1
#
N d
$
t(i+1)/2n
E
2
ξti/2n λ(· − Xs )dbjµ (s) − ξti/2n λ(· − Xti/2n )δbjµ (i)
ti/2n
i=0
+
E
j=1 µ=1
W0 n 2 −1
j=1 µ=1
ξti/2n λ(· − Xti/2n )δbjµ (i) −
i=0
2
t
,
ξs λ(· − Xs )dbjν (s) 0
W0
(5.21) where δbjµ (i) := bjµ (t(i + 1)/2n ) − bjµ (ti/2n ). The first term on the right-hand side of (5.21) is N d lim E n→∞
j=1 µ=1
n 2 −1
i=0
#
t(i+1)/2n ti/2n
$
2
= 0.
λ(· − Xs )dbjµ (s) − λ(· − Xti/2n )δbjµ (t, n)
Noting that ˆ (1 − e−|t−s|ω )λ)| ˆ ≤ |t − s| λ
ω ˆ ˆ ˆ − ξt λ ˆ 2 2 = |(λ, λ ,
ξs λ L
W
Vol. 3, 2002
Self-adjointness of the Pauli-Fierz Hamiltonian
199
ˆ ωλ ˆ ∈ L2 , we see that the second term also converges to zero as n → ∞. Thus, if λ, then lim E Kn (t) − K(t) 2W0 = 0. n→∞
Taking n → ∞ on the both sides of (5.20), we obtain (5.19).
A proof of Lemma 2.2 (1) follows from Lemmas 5.3 and 5.4. (2) follows from Lemmas 5.5 and 5.7.
Acknowledgments. I thank T. Ichinose for his encouragement and H. Spohn for his hospitality of Technische Universit¨ at M¨ unchen. I am grateful to J. L˝ orinczi for a careful reading of my manuscript. I also thank Graduiertenkolleg “Mathematik in ihrer Wechselbeziehung zur Physik” for financial support.
References [1] A. Arai, Rigorous theory of spectra and radiation for a model in quantum electrodynamics, J. Math. Phys. 24, 1896–1910 (1983). [2] A. Arai and M. Hirokawa, On the existence and uniqueness of ground states of a generalized spin-boson model, J. Funct. Anal. 151, 455–503 (1997). [3] V. Bach, J. Fr¨ ohlich and I. M. Sigal, Renormalization group analysis of spectral problems in quantum field theory, Adv. Math. 137, 20-5-298 (1998). [4] V. Bach, J. Fr¨ ohlich and I. M. Sigal, Quantum electrodynamics of confined nonrelativistic particles, Adv. Math. 137, 299–395 (1998). [5] V. Bach, J. Fr¨ ohlich and I. M. Sigal, Spectral Analysis for systems of atoms and molecules coupled to the quantized radiation field, Commun. Math. Phys. 207, 249–290 (1999). [6] V. Bach, J. Fr¨ ohlich, I. M. Sigal and A. Soffer, Positive commutators and the spectrum of Pauli-Fierz Hamiltonian of atoms and molecules, Commun. Math. Phys. 207, 557–587 (1999). [7] V. Betz, F. Hiroshima, J. L˝ orinczi, R. Minlos and H. Spohn, Gibbs measure associated with particle-field system, to be published in Rev. Math. Phys. . [8] J. Derezi´ nski and C. G´erard, Asymptotic completeness in quantum field theory. Massive Pauli-Fierz Hamiltonian, Rev. Math. Phys. 11, 383–450 (1999). [9] J. Derezi´ nski and V. Jakˇsi´c, Spectral theory of Pauli-Fierz operators, J. Funct. Anal. 180, 243–327 (2001). [10] C. Fefferman, J. Fr¨ ohlich and G. M. Graf, Stability of ultraviolet-cutoff quantum electrodynamics with non-relativistic matter, Commun. Math. Phys. 190, 309–330 (1997). [11] J. Fr¨ ohlich, On the infrared problem in a model of scalar electrons and massless, scalar bosons, Ann. Inst. H. Poincar´e 19, 1–103 (1973).
200
F. Hiroshima
Ann. Henri Poincar´e
[12] J. Fr¨ ohlich, Existence of dressed one electron states in a class of persistent models, Fortschritte der Physik 22, 159–198 (1974). [13] J. Fr¨ ohlich, M. Griesemer and B. Schlein, Asymptotic electromagnetic fields in a mode of quantum-mechanical matter interacting with the quantum radiation field, mp-arc 00–374, preprint(2000). [14] J. Fr¨ ohlich and Y. M. Park, Correlation inequalities and thermodynamic limit for classical and quantum continuous systems II. Bose-Einstein and FermiDirac statistics, J. Stat. Phys. 23, 701–753 (1980). [15] C. G´erard, On the existence of ground states for massless Pauli-Fierz Hamiltonians, Ann. H. Poincar´e 1, 443–459 (2000). [16] M. Griesemer, E. Lieb and M. Loss, Ground states in non-relativistic quantum electrodynamics, preprint (2000). [17] L. Gross, The free euclidean Proca and electromagnetic fields, Functional integral and its application, Proceedings of the international conference, the Cumberland Lodge, Windsor Great Park, London in 1974. A. M. Arthurs (editor), Clarendon Press, Oxford (1975). [18] F. Hiroshima, Functional integral representation of a model in quantum electrodynamics, Rev. Math. Phys. 9, 489–530 (1997). [19] F. Hiroshima, Ground states and spectrum of non-relativistic quantum electrodynamics, Trans. of Amer. Math. Soc. 353, 4497–4528 (2001). [20] F. Hiroshima, Ground states of a model in nonrelativistic quantum electrodynamics I, J. Math. Phys. 40, 6209-6222 (1999), II, J. Math. Phys. 41, 661–674 (2000). [21] F. Hiroshima, Essential self-adjointness of translation-invariant quantum field models for arbitrary coupling constants, Commun. Math. Phys. 211, 585–613 (2000). [22] F. Hiroshima and H. Spohn, Enhanced binding through coupling to a quantum field, Ann. H. Poincar´e 2, 1159–1187 (2001). [23] F. Hiroshima and H. Spohn, Ground state degeneracy of the Pauli-Fierz model including spin. mp-arc 01–407, preprint (2001). [24] R. Høegh-Krohn, Asymptotic fields in some models of quantum field theory I, J. Math. Phys. 9, 2075–2080 (1968), II J. Math. Phys. 10, 639–643 (1969), III J. Math. Phys. 11, 185–189 (1969). [25] M. H¨ ubner and H. Spohn, Radiative decay: nonperturbative approaches, Rev. Math. Phys. 7, 363–387 (1995). [26] M. H¨ ubner and H. Spohn, Spectral properties of the spin-boson Hamiltonian, Ann. Inst. Henri Poincar´e 62, 289–323 (1995).
Vol. 3, 2002
Self-adjointness of the Pauli-Fierz Hamiltonian
201
[27] V. Jakˇsi´c and C. A. Pillet, On a model for quantum friction I. Fermi’s golden rule and dynamics at zero temperature, Ann. Inst. H. Poincar´e, 62, 47–68 (1995). [28] E. Lieb and M. Loss, Self-energy of electrons in non-perturbative QED, preprint (1999). [29] J. L˝ orinczi, R. Minlos and H. Spohn, The infrared behaviour in Nelson’s model of a quantum particle coupled to a massless scalar field, preprint (2001). [30] E. Nelson, Schr¨ odinger particles interacting with a quantized scalar field, Analysis in Function Space, Proceedings U. S. A. (1963). [31] E. Nelson, Interaction of nonrelativistic particles with a quantized scalar field, J. Math. Phys. 5 (1964), 1190–1197. [32] T. Okamoto and K. Yajima, Complex scaling technique in non-relativistic massive QED, Ann. Inst. H. Poincar´e 42, 311–327 (1985). [33] W. Pauli and M. Fierz, Zur Theorie der Emission langwelliger Lichtquanten, Nuovo Cimento 15, 167–188 (1938). [34] M. Reed and B. Simon, Methods of Modern Mathematical Physics II, Academic Press (1975). [35] B. Simon, The P (φ)2 Euclidean (Quantum) Field Theory, Princeton Univ. Press (1974). [36] H. Spohn, Ground state(s) of the spin-boson Hamiltonian, Commun. Math. Phys. 123, 277–304 (1989). [37] H. Spohn, Ground state of quantum particle coupled to a scalar boson field, Lett. Math. Phys. 44, 9–16 (1998). Fumio Hiroshima Zentrum der Mathematik Technische Universit¨ at M¨ unchen Gabelsberger straße 49 M¨ unchen Germany Permanent address: Department of Mathematics and Physics Setsunan University Ikeda-naka-machi 17-8 572-8508, Osaka Japan email:
[email protected] Communicated by Bernard Helffer submitted 17/07/01, accepted 19/11/01
Ann. Henri Poincar´e 3 (2002) 203 – 267 c Birkh¨ auser Verlag, Basel, 2002 1424-0637/02/020203-65 $ 1.50+0.20/0
Annales Henri Poincar´ e
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models P. Holick´ y, R. Koteck´ y and M. Zahradn´ık Abstract. We study interfaces between two coexisting stable phases for a general class of lattice models. In particular, we are dealing with the situation where several different interface configurations may enter the competition for the ideal interface between two fixed stable phases. A general method for constructing the phase diagram is presented. Namely, we give a prescription determining which of the phases and which of the interfaces are stable at a given temperature and for given values of parameters in the Hamiltonian. The stability here means that typical configurations of the limiting Gibbs state constructed with the corresponding interface boundary conditions differ only on a set consisting of finite components (“islands”) from the corresponding ideal interface.
0 Introduction Before stating our main result in its full generality in the next section, we shall explain the main idea for a particular model. Namely, we shall consider interfaces between two stable phases for the three-dimensional Blume-Capel model. To every site i ∈ Z3 a spin x(i) is attached attaining the values x(i) ∈ {−1, 0, +1}. The Hamiltonian in a finite volume Λ ⊂ Z3 with boundary conditions z is given as (x(i) − x(j))2 + J (x(i) − z(j))2 − λ x(i)2 − h x(i). HΛ (x|z) = J i,j i,j∈Λ
i,j i∈Λ,j ∈Λ /
i∈Λ
i∈Λ
First two sums are over pairs of nearest neighbours, J > 0 is fixed. It is easy to see that the phase diagram in the (λ, h)-plane and at vanishing temperature consists of three regions of ground states of constant spins, x ≡ +1, x ≡ −1, and x ≡ 0, separated by half-lines h = 0, λ ≥ 0; h = λ, λ ≤ 0; and h = −λ, λ ≤ 0. With the help of Pirogov-Sinai theory [PS] one can show that the phase diagram at small temperatures is a smooth deformation of this zero temperature phase diagram [BS] as indicated in Fig. 1. Notice that the region of the phase 0 is expanding as the temperature grows. This is easy to understand by observing that while the phases + and − have only one type of lowest energy excitation, namely, flipping a single spin to the value 0, the phase 0 allows two excitations of this order, flipping to +1 or −1, which gives the phase 0 an advantage at non-vanishing temperatures when excitations contribute to the free energy. The coexistence of all three phases occurs at the line (λ = λ0 (T ), h = 0), where λ0 (0) = 0 and the function λ0 (T ) is growing in T .
204
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
T
λt (T )
λt (0) = 2J
λ
h Fig. 1. Phase diagram of Blume-Capel model. The line λt (T ) corresponds to coexistence of two different interface patterns on the boundary between stable plus and minus phases. Let us consider now, for h = 0, λ > 0, an interface between plus phase in the upper half-space and minus phase in the lower half-space. Depending on the value of λ, different arrangements of spins on the interface yield the minimal energy. Namely, if λ ≥ 2J, the most convenient way is to switch directly from plus to minus spins — the configuration y (I) , y (I) (i) = 1 whenever i3 ≥ 0 and y (I) (i) = −1 whenever i3 < 0, is a ground state. Indeed, it is not difficult to see that the difference of the energy of any configuration x that differs from y (I) on finite number of sites and that of the configuration y (I) itself, is nonnegative, HΛ (x|y (I) ) ≥ HΛ (y (I) |y (I) ). On the other side, if λ ≤ 2J, it is favourable to separate pluses and minuses by a layer of spins 0; the ground state is the configuration y (II) , y (II) (i) = 1 whenever i3 > 0, y (II) (i) = 0 for i3 = 0, and y (II) (i) = −1 whenever i3 < 0. We will see that this behaviour subsists at small temperatures. Namely, there exists a smooth transition function λt (T ) emanating from the point λt (0) = 2J at T = 0 such that, if λ ≥ λt (T ), the boundary conditions y (I) yield (at temperature T and in the thermodynamic limit) the Gibbs state whose typical configurations differ from y (I) only on small islands — the state corresponding to y (I) is stable. Similarly, for λ0 (T ) + δ < λ ≤ λt (T ) the state corresponding to y (II) is stable. At the coexistence line λ = λt (T ) both states are stable. Notice that the transition is
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
205
of the first order type — the variable that exhibits a discontinuity while passing through the transition line is, for example, the density of spin 0 at the interface. We are actually concerned here with the phenomenon of “prewetting” of the microscopic ± interface by zero spins. Notice that we excluded a small neighbourhood of the coexistence line λ0 (T ). It is expected that for λ very close to λ0 (λ − λ0 ∼ O(e−β ) as β → ∞) the layer of zero spins spreads over several lattice sites, with its thickness growing due to “entropic repulsion” as λ λ0 (it was this type of wetting that was discussed for λ = λ0 in [BL]). Our aim in this paper is to study this type of surface phenomena in a general case. Namely, we are interested in situations where, as in the example above, several different interface configurations may enter the competition for the ideal interface between two fixed stable phases. The simpler case with a single ground state interface in the considered region of parameters (as is the case, for example, for the standard Ising model) is well understood [HKZ] as a straightforward, though rather technically involved, generalization of the standard Dobrushin treatment [D 72]. Simplifying slightly, the main idea is to rewrite the finite volume Gibbs state with interface boundary conditions in terms of the probability distribution of the interface contour separating the regions of two coexisting stable phases. To prove the existence of the interface Gibbs state in the thermodynamic limit, one shows that typical interface contours differ only locally from the ideal ground state interface configuration. To this end one splits the interface contour into regions of ideal interface at different heights — the ceilings — separated by walls. The crucial observation that there is a one-to-one correspondence between interface contours and collections of compatible walls, and that the walls are distributed in an independent way, allows one to use a generalization of the Peierls argument, with walls playing the role of contours, to prove that the probability of walls is dumped and to conclude the existence of interface Gibbs state. The most important fact that was skipped out in the simplified description above, is that the interface is actually surrounded, from above and below, by two different — possibly asymmetric — coexisting phases. Their influence is taken into account by rewriting the corresponding partition functions, with the help of the Pirogov-Sinai theory, in terms of cluster expansions. After separating a suitable normalizing factor, we expand the cluster terms intersecting the interface. As a final result, one has to deal with an interface decorated by clusters. The problem in the general case is that presence of various competing interfaces leads to existence of different types of ceilings. To treat the probability distribution of collection of walls one has to cope with the matching conditions on families of walls. Namely, each wall is “labeled” on any connected component of its boundary, where it is attached to surrounding ceilings, by the type of the corresponding ceiling. For the collection of walls stemming from an interface, these labels on boundaries of different walls attached to the same ceiling must coincide. But this is the starting point of the standard Pirogov-Sinai analysis of probabilities of collections of labeled contours. One only has to be slightly more careful when applying these ideas.
206
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
One point to consider is the nonlinear dependence on the original Hamiltonian (and its parameters) of the decorating cluster terms originated from the coexisting phases surrounding the interface. These have to be attached to the considered walls and one thus has to be prepared to deal with a nontrivial dependence of wall weights on parameters of the original Hamiltonian. Second point is that the decorations may actually stick out of the considered finite volume and as a result one is dealing with infinitely many wall-cluster aggregates intersecting the given finite volume. One thus has to be careful when applying standard cluster expansions to treat correctly these situations having also in mind that the weight factors of terms that are not entirely contained in a given volume may actually depend on it yielding a fixed limiting weight factor only when volume expands to infinity. Last but not least, the weight factors of the wall-cluster aggregates might be actually negative, due to presence of cluster terms whose sign is not determined. There are different strategies how to deal with these problems. Here, we have chosen the most conservative one. Namely, we rewrite the interface partition function (and the corresponding Gibbs state) in such a way that we can use directly (essentially) standard Pirogov-Sinai theory with the role of contours played by “shadows” of wall-cluster aggregates projected to the plane of ideal interface. To get positive weights allowing in the final account, for example, to estimate the probability of external shadows in the standard manner, we add a suitable cluster sum into the exponent contributions to the interface partition function, absorbing it in the same time into a small change of weight of interfaces. The cluster contributions of this added sum can be easily chosen in such a way that the positivity of combined cluster terms is assured. There are at least two alternative approaches. First, one may base all the discussion on an extension of Pirogov-Sinai theory to complex parameters [BI]. The first steps in this direction (dealing only with one type of interface and generalizing thus [HKZ]) are done in [BCF 1]. The study in [BCF 1] is motivated by the investigation of interfaces in quantum lattice models [BCF 2]. Other approach, proposed recently by two of us [HZ], is to develop a new alternative to PirogovSinai theory based on the idea of “expanding away”, one by one, all contours and walls without ever passing through intermediate contour models with their cluster expansions as it is the case in the standard Pirogov-Sinai theory. This method is conceptually promising and we expect that it will allow a treatment of a great variety of models with different types of interface, wetting, and other “stratified” states (i.e. consisting of several interfaces). The paper [HZ] develops the theory in a very general situation and it has to be supplied by a detailed study of the dependence of stability of resulting interfaces on parameters of the original Hamiltonian to draw the phase diagram at nonvanishing temperatures. A great deal of the present paper treats this particular problem in the situation of an interface with different ceilings and has thus its importance even though it is based on a rather standard Pirogov-Sinai approach and does not evoke the approach from [HZ].
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
207
The paper is organized as follows. In Section 1, we set the problem, introduce the class of models to be treated with assumptions (Peierls conditions) that assure a dominance by a particular class of ideal interfaces (ceilings). The main result (Theorem 1) is presented in a general form stressing the smooth dependence of the resulting full phase diagram on the temperature (including the description of regions where particular interfaces become stable — cf. Fig. 1). It follows from Basic Lemma whose proof is postponed, after various preparatory steps in Sections 3 and 4, to Section 4.4. Section 1 contains, in addition to Theorem 1 with its proof following from Basic Lemma, also the characterization of the ground state phase diagram including interface ground states (Proposition 1.1.3) as well as its completeness under the condition of removing of degeneracy (Corollary 1.1.4) with their proofs in Section 1.3. Section 2 is devoted to a brief reformulation and a slight extension of the Pirogov-Sinai theory. First, we summarize the results concerning contour models in a form needed for our purposes. This part includes contour models with boundary dependence (models whose contour weights depend slightly on the boundary of the considered volume and are translation invariant only for contours far from the boundary). The corresponding results are of an independent interest, given the fact that one often obtains a reformulation in terms of such a contour model. Next, we introduce labeled contour models and summarize the Pirogov-Sinai theory in Theorem 2 (characterization of stable phases) proved as a consequence of Proposition 2.2.1 (properties of stable phases) that is proved in Section 2.3. Again, also the situation of labeled contour models with boundary dependence is considered (Corollary 2.2.2). A standard application is the description of periodic Gibbs states in terms of contour models. This is needed in the further treatment and it is presented in Section 3 to set the notation. Finally, in Section 4 we reformulate the Gibbs states with interfaces in terms of labeled contour models with role of contours played by the shadows of walls decorated by the clusters of the stable phases above and below the interface. This is done in a series of steps that yield an expression for the weights of shadows that are sufficiently dumped to allow once more the application of Theorem 2.
1 Setting and the main result 1.1
Setting; ground state phase diagram
We shall consider classical lattice models on a ν-dimensional lattice Zν (ν ≥ 3) with a finite set S of spin values attached to each lattice site i ∈ Zν . The configuration ν space will be denoted by X(= S Z ), the space of restrictions xΛ = (x(i) : i ∈ Λ) of configurations x = (x(i) : i ∈ Zν ) ∈ X to Λ ⊂ Zν by XΛ (= S Λ ). We endow the lattice Zν with the ∞ -metric (ρ(i, j) = maxk=1,...,ν |ik − jk |). Connected (Rconnected) set A ⊂ Zν is then defined as a set whose any two sites i, j ∈ A can be
208
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
joined by a sequence of sites i = i1 , i2 , . . . , j = ik from A such that ρ(i , i+1 ) ≤ 1 ( ρ(i , i+1 ) ≤ R) for all = 1, 2, . . . , k − 1. Every translation invariant Hamiltonian H = (UA ; A ⊂ Zν , diamA < R) of range R with interactions UA : XA → R can be identified with H = (U[A] ), where [A] runs over all (finitely many) equivalence classes of subsets of Zν defined by shifts of an A ⊂ Zν with diamA < R. So the translation invariant Hamiltonians of range R form a finite-dimensional vector space of Hamiltonians denoted by H(R). More precisely, it can be identified with the space of all vectors (UA (xA ); i(A) = 0, diamA < R, xA ∈ XA ), where xA are ordered in a fixed way. Here i(A) ∈ A is a site chosen in A in a fixed canonical way (say, the first site in A in a fixed lexicographic order). Throughout we use · to denote the euclidean norm on H(R), · ∞ the maximum norm on it, and dim H(R) to denote its dimension. Using HΛ (x|z) to denote the Hamiltonian in Λ with boundary conditions z ∈ X, UA (xA ) HΛ (x|z) = A∩Λ=∅ c
with x = z in Λ , we introduce the partition function Z(Λ|z; H) = exp{−HΛ (x|z)}. x=zinΛc
It will be useful to define also Z(Y, Λ|z; H) =
exp{−HΛ (x|z)}
xΛ ∈Y x=zinΛc
for any Y ⊂ XΛ . The Gibbs state in a finite volume Λ ⊂ Zν under a boundary condition x ∈ X with Hamiltonian H is the probability µ(·, Λ|z; H) on X defined by µ({x}, Λ|z; H) =
Z({xΛ }, Λ|z; H) Z(Λ|z; H)
whenever x = z in Λc . The set of all Gibbs states in (possibly infinite) V ⊂ Zν under a boundary condition x ∈ XZν \V with Hamiltonian H introduced by means of the DLR equations (see [HKZ], Section 2.1, for discussion of the situation with V Zν ) will be denoted by G(V |x; H). For V = Zν the boundary condition x is necessarily empty and we get the standard definition of Gibbs states (G(Zν | ∅, H) = G(H)). The inverse temperature β does not appear here as an independent parameter; it is incorporated into the constants of the Hamiltonian H. (See also Remark 4 in Section 1.2.) It is well known that the set of all Gibbs states G(H) is the closed convex hull of all possible weak limits limΛn Zν µ(·, Λn |z; H) of finite volume Gibbs states.
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
209
A configuration x ∈ X is called a ground configuration of a Hamiltonian H = (UA ) if “the energy of every finite perturbation x ˜ of x (˜ x differs from x at a finite number of lattice sites) is not smaller than the energy of x”: (UA (˜ xA ) − UA (xA )) ≥ 0. (1.1) H(˜ x; x) = A
If the above difference is positive for each finite perturbation x˜ = x, we refer to x as to a ground state (isolated ground configuration). We use g(H) (resp. g per (H)) to denote the set of all ground configurations corresponding to H (resp. periodic ground configurations). Our aim will be to discuss phase diagrams including a class of Gibbs states with an (“horizontal”) interface. Before taking into account any excitations, we shall describe the phase diagram at vanishing temperature—the ground state phase diagram. We thus suppose that a set G of configurations is given (to play the role of possible ground states). We shall restrict ourselves to horizontally periodic configurations, i.e. we suppose that every x ∈ G is periodic with respect to translations in the first (ν − 1) coordinates. We use Gper to denote the subset of all x ∈ G that are also vertically periodic, i.e. periodic with respect to the ν-th coordinate, and we put Ghor = G \ Gper . Further we assume that each x ∈ Ghor is identical to two configurations y1 , y2 ∈ Gper above and below certain heights, respectively. Namely, for each x ∈ Ghor there exist two states y1 , y2 ∈ Gper and a pair of constants t1 (x), t2 (x) such that x(i) = y1 (i) once iν ≥ t1 (x) and x(i) = y2 (i) once iν ≤ t2 (x). We may assume that t1 (x) and t2 (x) are chosen as the minimal, resp. maximal, constant with this property. We suppose also that the set G is finite up to vertical translations, i.e. there exists a finite subset of G such that any configuration x ∈ G is a vertical translation of a configuration from the considered finite set. In particular, Gper is finite, Gper = {x1 , . . . , xr }, and there exist a finite constant t (maximal thickness of interfaces) such that t ≥ t1 (x) − t2 (x) for all pairs t1 (x), t2 (x) above1 . To control the suppression of excitations with respect to configurations from G, we rely on an extended Peierls condition. To introduce it, let us first define “the specific energy at the site i” by UA (xA ) (H) (1.2) Ei (x) = |A| Ai
for each configuration x ∈ X. Here |A| refers to the number of sites in A. The no (H) (H) tation EΛ (x) = i∈Λ Ei (x) will be also used. For every periodic configuration x, x ∈ X per , we also define the specific energy ex (H) of x by (H) 1 ex (H) = lim Ei (x) (1.3) ν n→∞ |Vn | ν i∈Vn
1 This
fact implies that the interface introduced below is necessarily connected.
210
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
(with Vnν denoting a cube consisting of nν lattice sites). For any cube Vnν whose side (H) is a multiple of the periodicity of x ∈ X per, one clearly has EVnν (x) = |Vnν |ex (H), (H) i.e. ex (H) = |V1ν | i∈V ν Ei (x). n n It is useful to introduce an averaged specific energy2 at site i for any configuration x by the right hand side of the last equation. Namely, let p be a common multiple of periods of all configurations from Gper as well as horizontal periods of all configurations from Ghor , for any i ∈ Zν let Vp (i) be the cube Vp (i) = j ∈ Zν ; j = i + k, k = (k1 , . . . , kν ) ∈ {0, 1, . . . , p − 1}ν , and for any configuration x ∈ X let (H)
Ei
(x) =
1 (H) Ej (x). |Vp (i)| j∈Vp (i)
It is easy to verify that for any Λ, the sum a local boundary term, i∈Λ
(H)
Ei
(H)
(x) = EΛ (x) +
j∈Λc
(H)
Ej
(x)
i∈Λ
(H)
Ei
(H)
(x) differs from EΛ (x) by
|Vp∗ (j) ∩ Λ| (H) |Vp∗ (j) ∩ Λc | − . E (x) j |Vp∗ (j)| |Vp∗ (j)| j∈Λ
Here, Vp∗ (j) = {i ∈ Zν ; j ∈ Vp (i)} = i ∈ Zν ; i = j − k, k ∈ {0, 1, . . . , p − 1}ν . The explicit form of the boundary term is not very relevant; an important fact is, however, that whenever x and x ˜ differ only on a finite set, then (H) (H) (H) (H) x) − EΛ (x) = E i (˜ x) − E i (x) (1.4) EΛ (˜ i∈Λ
i∈Λ
once Λ is sufficiently large. More exactly, the equality holds if x(i) = x ˜(i) implies (H)
that d(i, Λc ) > R + p. Clearly, E i (x) = ex (H) for any x ∈ Gper . Let, now, an integer d ≥ R be chosen so that it is surpassing both periodicity p as well as interface maximal thickness t, d > max{R − 1, p, t}. Consider the set of all elementary cubes consisting of dν lattice sites. A bad cube of a configuration x ∈ X is an elementary cube D for which xD differs from yD for every y ∈ Gper . Notice that the choice of d ensures that the only configurations with no bad cube are those from Gper . The boundary B(x) of x is the union of all bad cubes of x. ˜ is its finite perturbation (differing from x on a finite set of If x ∈ Gper and x 2 The reader who is ready to sacrifice the subtlety of the case of periodic but not necessarily translation invariant configurations in Gper can skip the present paragraph, suppose that all (H) (H) x ∈ Gper are translation invariant, and replace everywhere E i (x) by Ei (x). Actually, as explained after the formulation of Basic Lemma in Section 1.3 below, by introducing “block spins” one can rewrite the model (possibly lowering the upper bound on allowed temperatures) in such a way that all configurations in Gper become translation invariant as well as all configurations in Ghor become horizontally translation invariant.
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
211
lattice sites), then, necessarily, B(˜ x) is finite. Any connected component Γ of B(˜ x) is calleda contour (of x ˜) and we use ∂(˜ x) to denote the set of all contours of x˜, B(˜ x) = Γ∈∂(˜x) Γ. Notice that the configuration x ˜ coincides with one of the states x). The only x ∈ Gper that coincides with x ∈ Gper on every component of Z \ B(˜ x ˜ on d-boxes that intersect both Γ and the only infinite component of Zν \ Γ, is called the external boundary condition of the contour Γ of x. Finally, we use e0 (H) for minx∈Gper ex (H) and denote g0per (H) = {x ∈ Gper ; ex (H) = e0 (H)}. The standard Peierls condition3 , with ρ > 0 and with respect to Gper , that is used in the Pirogov-Sinai theory, can be formulated as the bound (H) (per) E i (xΓ ) − ex (H) ≥ ρ|Γ| (P ) i∈Γ
for any x ∈ Gper , any contour Γ, and any configuration xΓ such that Γ is its only contour with external boundary condition x. Notice that if all configurations in Gper are actually translation invariant, E i can be simply replaced by Ei , yielding Peierls condition in the form (H)
EΓ (xΓ ) − ex (H)|Γ| ≥ ρ|Γ|.
(P(per) )
Let us remark that one can normalize the condition with respect to the minimum e0 (H) instead of ex (H) [BK]. This is a useful trick that enables, with some additional care in relevant estimates, to get a uniform validity of the theory far away from lines of coexistence. However, we will restrict our considerations to a small neighbourhood of a fixed Hamiltonian anyway and will thus abstain from this extension. Turning now to the configurations containing interfaces, we first introduce the notion of a wall. Namely, consider a configuration x ∈ Ghor and its excitation x ˜ differing from x on a finite set of lattice sites. Let I(˜ x) denote the infinite connected component of B(˜ x) (notice that B(˜ x) has only one infinite component (cf. footnote 1)) and let I(˜ x), an interface, be the pair (I(˜ x), x ˜I(˜x) ). Notice that I(˜ x) is splitting Zν \ I(˜ x) into two infinite components. Denoting, for any i ∈ Zν , by C(i) the column {(i1 , . . . , iν−1 , n); n ∈ Z} and by Cd (i) its d-neighbourhood, we use C(˜ x) to denote the set of those sites i of I(˜ x) for which there exists a configuration y ∈ G such that I(˜ x) ∩ Cd (i) = I(y) ∩ Cd (i) and x ˜ = y on it. The set C(i) ∩ I(y), for such i ∈ C(˜ x), is called a y-ceiling column. x) \ Further, a pair w = (W, x˜W ), where W is a connected component of I(˜ C(˜ x), is a wall of I(˜ x). We denote by W(I(˜ x)) the collection of all walls of I(˜ x). Whenever w = (W, x˜W ) is a wall, there exists a configuration yw ∈ G and its perturbation xw such that w is the only wall of xw , W(I(xw )) = {w}. For any wall 3 The
word standard here corresponds to the fact that here we are dealing only with “contours immersed in periodic configurations” in contrast with “walls of an interface” as will be the case below. On the other hand, we are dealing here with a Peierls condition valid uniformly on a neighbourhood of Hamiltonians — this type of the Peierls condition is sometimes [EFS] called “extended Peierls condition”.
212
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
w, we use IW = I(yw )∩( i∈W C(i)). Notice that this definition is consistent—the right hand side above is well defined only from geometry of W . Indeed, any wall has an “outside rim” from which the set I(yw ) can be uniquely read of. A Hamiltonian H = (UA ) is said to fulfill the Peierls condition (P) with respect to the set G (and with constant ρ > 0) if it satisfies (P
(per)
) as well as
(H) (H) E i (xw ) − e0 (H) − E i (yw ) − e0 (H) ≥ ρ|W | i∈W
(P
(hor)
)
i∈IW
for any wall w and the corresponding configuration yw ∈ Ghor and its excitation xw . Again, if all configurations in Gper are translation invariant, the condition takes simpler form (H) (H) EW (xw ) − e0 (H)|W | − EIW (yw ) − e0 (H)|IW | ≥ ρ|W |.
(P(hor) )
The symbol (P) denotes that both (P(per) ) and (P(hor) ) are satisfied. The following lemma is related to Lemma 2.1 from [S]. Lemma 1.1.1 Let H satisfy the Peierls condition with respect to a non-empty finite set Gper ⊂ X per. Then all periodic ground configurations of H are ground states, their set g per (H) coincides with the set g0per (H), and g0per (H) minimizes specific energy over all X per , g0per (H) = {x ∈ X per; ex (H) = e0 (H)}. Proof. 1. We first show that each element x of Gper with ex (H) = e0 (H), is a ground state of H. This, in particular, implies g0per(H) ⊂ g per (H). Let y differ from x on a non-empty finite subset of Zν . Let Λ ⊂ Zν be finite and such that x(i) = y(i) implies that dist(i, Λc ) > p + R. (per)
By (1.4), (P (H)
), and the choice of x and Λ, we get (H)
H(y; x) = EΛ (y) − EΛ (x) = (H) (H) (H) (y) − E i (x) = i∈Λ E i (y) − ex (H) = i∈Λ E i (H) (H) = i∈B(y) E i (y) − ex (H) + i∈Λ\B(y) (E i (y) − ex (H)) > ρ|B(y)| > 0. (H)
Notice that for i ∈ / B(y) the values E i (y) equal to some of ez (H) with z ∈ Gper and so they are greater or equal to ex (H) = e0 (H). Hence the claim of the first step is verified. 2. Let ey (H) > e0 (H) for some periodic configuration y. Then we may consider a configuration y˜ that is equal to y outside of some large cube Λ ⊂ Zν and coincides with x in Λ, where x ∈ g per (H). It is obvious that H(˜ y; y) < 0 if Λ is large enough due to the inequality ey (H) > e0 (H) = ex (H). So y is not a ground configuration. 3. Finally, let y be a periodic configuration and y ∈ / Gper . Let us consider x x configurations yΛ such that yΛ equals to y on a sufficiently large cube Λ ⊂ Zν and
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
213
to x ∈ g0per (H) on Λc . We may notice that by the finite range of the potential ey (H) − ex (H) = limν Λ→Z
x H(yΛ ; x) , |Λ|
where the limit is e.g. over any sequence of cubes Λ with diameters tending to infinity. We get, similarly as in step 1, for sufficiently large neighbourhood V of Λ that (H) x (H) x x H(yΛ ; x) = EV (yΛ ) − EV (x) > ρ|B(yΛ )|. x There is c > 0 independent of Λ such that |B(yΛ )| > c|Λ| as y is periodic. Summing up what we observed, we get that ey (H) > ex (H) implying the inclusion {x ∈ X per; ex (H) = e0 (H)} ⊂ Gper and finishing thus the proof.
Our next aim is to study all horizontally periodic ground configurations of H. hor In analogy with ex (H) for x ∈ X per, we consider a configuration x ∈ G∩Xy,z per hor with y, z ∈ g0 (H), y = z, where Xy,z is defined as the set of all horizontally hor if periodic configurations asymptotically coinciding with y and z (i.e. x ∈ Xy,z x(i) = y(i) whenever iν ≥ t1 and x(i) = z(i) whenever iν ≤ t2 for some t1 , t2 ∈ Z), and define (H) 1 (H) ex (H) = lim Ei (x) − Ei (y) . (1.5) ν−1 n→∞ |Vn | ν−1 i:(i1 ,...,iν−1 )∈Vn
Notice that only finite number of terms in the sum does not vanish and that the configurations y and z can be read off from x and we do not introduce them explicitly into the notation. Notice also that (H) ex (H) = E i (x) − e0 (H) , i∈C(0)
where the sum is again only formally infinite. Notice, that even though we keep the same notation as in (1.3), no confusion can arise since these two definitions concern disjoint classes of configurations x; periodic ones in (1.3) and horizontally periodic ones above. Similar notion to e0 (H), that according to Lemma 1.1.1 equals min{ex (H); x ∈ X per}, is hor ey,z 0 (H) = min{ex (H); x ∈ G ∩ Xy,z }. We also use the notation hor g0y,z (H) = {x ∈ Ghor ∩ Xy,z ; ex (H) = ey,z 0 (H)}.
Finally, we say that a set G is admissible if it is non-empty and, up to vertical translations, finite set of horizontally periodic configurations such that for each hor hor for some y, z ∈ Gper , y = z, and that G ∩ Xy,z = ∅ for x ∈ Ghor it is x ∈ Xy,z per each y, z ∈ G , y = z.
214
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
Lemma 1.1.2 Let G be admissible and let H satisfy the Peierls condition (P) with respect to G and some ρ > 0. Then a horizontally periodic configuration x is a ground state, equivalently a ground configuration, of H if and only if either x ∈ g0per (H) hor or x ∈ Xy,z for some y, z ∈ g0per (H), y = z, and ex (H) = ey,z 0 (H). Moreover, such configuration x automatically belongs to G. Proof. According to Lemma 1.1.1, it suffices to consider only non-periodic configurations. 1. We first show that each x ∈ g0y,z (H) for some y, z ∈ g per (H), y = z, is a ground state. Let x ˜ differ from x on a non-empty finite subset of Zν and let Λ ⊂ Zν be a finite cube such that x ˜(i) = x(i) implies that dist(i, Λc ) > R + p. Let xI be the unique configuration with B(xI ) consisting only of the infinite component of B(˜ x). Then H(˜ x; x) = H(˜ x; xI ) + H(xI ; x). (per)
Using (P ), we get that H(˜ x; xI ) ≥ 0, or even that H(˜ x; xI ) > 0 if xI = x (and I I thus x ˜ = x ) when also H(x ; x) = 0. (H) (H) If xI = x, we rewrite H(xI ; x) = i∈Λ E i (xI ) − E i (x) , by adding and subtracting the term e0 (H)|Λ|, as the sum over disjoint columns C of the terms of the form (H) (H) E i (xI ) − e0 (H) − E i (x) − e0 (H) . i∈C∩Λ
i∈C∩Λ
For any x ¯-ceiling column C this term equals to ex¯ (H)−ey,z 0 (H) that is nonnegative by the definition of ey,z 0 (H). Noticing also that for a column C that intersects a (H)
wall w of xI every contribution E i (xI ) − e0 (H) with i ∈ C ∩ W c is nonnegative, and using the bound eyw (H) ≥ ex (H) = ey,z 0 (H), and the transcription of the definition of ex (H) after (1.5) above, we are left with (H) (H) H(xI ; x) ≥ E i (xI ) − e0 (H) − E i (yw ) − e0 (H) > 0 w
i∈W
i∈IW hor
by the Peierls condition (P ). 2a. Let x be a horizontally periodic ground configuration that is not periodic. If, in the upper half space, there are only finitely many horizontal layers of width d with x equal to some element of g per(H) on them, then we easily obtain a per contradiction with (P ) by considering cubes Λ that are large enough and take x on Λ and some x0 ∈ g per (H) on Λc . Notice that the volume of bad cubes in Λ
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
215
is of the order of the volume of Λ. So there must be infinitely many such layers of a configuration y ∈ g per(H) in the upper half-space and infinitely many layers of z ∈ g per(H) (possibly equal to y) in the lower half-space . Further, whenever there are two disjoint layers on which x equals to the same element y of g per (H), it necessarily equals y also in-between of these two layers. Indeed, let us suppose that x is not equal to y in-between. Considering parallelepipeds Λ that have the upper mentioned layer just above its top and the lower mentioned layer just below its bottom and changing x on the d-neighbourhood of the boundary of Λ to y may increase the energy at most by a contribution of the order of the size of the side-wise boundary of Λ (excluding its top and bottom), whereas changing additionally x to y everywhere inside Λ causes a decrease of the per energy by a contribution of the order of the size of the top of Λ (applying (P and the fact that between those two layers there must be a layer containing, due to horizontal periodicity of x, a periodic horizontal grid of bad cubes). Since the size of the top, or bottom, of Λ are asymptotically larger than that of the rest of the boundary of Λ, the energy for Λ large enough decreases and so x is not a ground hor for some y, z ∈ g per(H), configuration. This contradiction shows that x ∈ Xy,z y = z. per hor (H), y = z, x ∈ Xy,z , then x is not 2b. If ex (H) > ey,z 0 (H) for y, z ∈ g hor a ground configuration because replacing it by some element x ¯ of Xy,z for which y,z ex¯ (H) = e0 on sufficiently large cube, with its top in the upper region for both configurations and its bottom in the lower region for both configurations x and x¯, we increase the energy. hor hor , we get that ex (H) > ey,z ) investi3. If x ∈ / G and x ∈ Xy,z 0 (H) from (P 1 w gating the limit of ex (H) = |V ν−1 H(x ; w) for some sequence of cubes V tending n Vn | n
y,z c hor to Zν . Here xw Vn is equal to w on Vn for some w ∈ G ∩ Xy,z with ew (H) = e0 (H) and to x on Vn .
Our aim now is to describe “the phase diagram at zero temperature” — the ground state phase diagram. Consider a Hamiltonian H0 with the set G0 of all horizontally periodic ground states of H0 , G0 = g(H0 ). To describe the ground state phase diagram in a (sufficiently small) neighbourhood of H0 in H(R) means to specify, for every subset G ⊂ G0 , the set of all Hamiltonians H for which G is the set of all horizontally periodic ground configurations, G = g(H). In fact we are describing the phase diagram for H from some cone containing a neighbourhood of H0 . Proposition 1.1.3 (ground state phase diagram) Suppose that H0 satisfies the Peierls condition (P) with a constant ρ0 > 0 and with respect to an admissible set G0 of horizontally periodic ground states, G0 = g(H0 ). ¯ : H ¯ − H0 < Then there exists ε > 0 such that for each H ∈ Kε (H0 ) = {β H εH0 , β > 1} the set g(H) of all horizontally periodic ground configurations is contained in G0 , each x ∈ g(H) is a ground state and either •
x ∈ Gper ¯ ∈ Gper ¯ (H) : x 0 and ex (H) = min{ex 0 }
(1.6)
216
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
•
Ann. Henri Poincar´e
hor x ∈ G0 ∩ Xy,z for some y, z ∈ g per(H), y = z, and hor ex (H) = min{ex¯ (H) : x¯ ∈ G0 ∩ Xy,z }
(1.7)
Moreover, for each 0 < ε¯ < ε there exists a constant ρε¯ such that limε¯→0+ ρε¯ = ρ0 H and each H ∈ Kε¯(H0 ) satisfies the Peierls condition (P) with the constant ρε¯ H . 0 Proof. To show the validity of Peierls condition for the Hamiltonian H with respect 0 to G0 and a suitable ρε¯, it suffices to compare it with the condition for H H H close to H0 . Hence the characterization from Lemmas 1.1.1 and 1.1.2 is valid implying the claims. Remark. Notice that only non-empty admissible sets g(H) ⊂ g(H0 ) will appear. To enable the realization of all admissible G ⊂ G0 in a neighbourhood of H0 (“full ground state phase diagram”) it is necessary to assume that a condition of removing of degeneracy is fulfilled. To formulate it and to enable a global study of the phase diagram at a non-vanishing temperature it is convenient to extend the functional H → ex (H) defined by (1.5) only for those Hamiltonians H for which y, z ∈ g per (H). This can be done for example by defining 1 ex (H) = lim [Ei (x) − min ey (H)] (1.8) ν−1 n→∞ |Vn y∈Gper | 0 i∈I(x) (i1 ,...,iν−1 )∈Vnν−1
hor hor for every horizontally periodic x ∈ Xy,z with y, z ∈ g per (H). Using |G ∩ Xy,z |∼ver hor to denote the number of elements of G ∩ Xy,z taken up to vertical translations, we formulate the condition of removing of degeneracy as the following assumption.
(RD) An affine subspace H0 ⊂ H(R) removes degeneracy of G0 = g(H0 ), an admissible set of horizontally periodic configurations, if the set of hor N0 = |g0per (H0 )| − 1 + (|g0 (H0 ) ∩ Xy,z |∼ver − 1) per y,z∈g0 (H0 ) y=z
linear functionals {ex (H) − ex0 (H), x ∈ g0per(H0 ), x = x0 } ∪ hor , x = xy,z , y, z ∈ g0per(H0 ), y = z} ∪{ex (H) − exy,z (H), x ∈ g0 (H0 ) ∩ Xy,z hor with arbitrarily chosen x0 ∈ Gper and xy,z ∈ G0 ∩ Xy,z is a set of linearly independent functionals on H0 . Notice that choosing H0 = H(R) with R large enough, degeneracy is always removed. (It is sufficient to choose R larger than the smallest periods of configurations in G0 as well as “thickness” of I(x) for all x ∈ G0 ∩ X hor .) If we considered G consisting of translation invariant configurations, already single site potentials (“external fields”) would be enough to remove the degeneracy. Let us also remark that one might consider a more general “manifold of parameters” H0 ⊂ H(R).
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
217
Corollary 1.1.4 (completeness of the ground state phase diagram) Supposing the validity of (RD) (with H0 = H(R)) in addition to the assumptions of Proposition 1.1.3 and denoting Hgr (G) = {H : G = g(H)}, one has Hgr (G) ∩ Kε (H0 ) = ∅ for each admissible G ⊂ g0 (H0 ) and Hgr (G1 ) ∩ Kε (H0 )⊂ = Hgr (G2 ) ∩ Kε (H0 ) whenever G1 ⊃ G . Actually the set H (G) is the intersection of the corresponding gr = 2 number of hyperplanes and open half-spaces in H(R) whose boundaries are yielded by equalities contained in (1.6) and (1.7).
1.2
The main result
Our aim is to show that the phase diagram including all horizontally periodic states is a small distortion of the ground state phase diagram described in Proposition 1.1.3 and Corollary 1.1.4 above. A Gibbs state µ of a Hamiltonian H ∈ H(R) is said to be a perturbation of a (ground) configuration x ∈ G0 if for µ-almost all configurations x ˜ there exists a connected subset M ⊂ Zν such that x ˜ differs from x only outside M , x ˜M = xM , the R-components of its complement Zν \ M are finite and, if x ∈ G0 ∩ X hor , also the set M0 = {i ∈ I(x)|M ⊃ I(x) ∩ C(i)} is connected. Theorem 1 Let H0 be a translation invariant Hamiltonian, H0 ∈ H(R), fulfilling the Peierls condition (P) with respect to an admissible set of ground states G0 = g(H0 ) and a sufficiently large ρ0 . Let further an affine subspace H0 ⊂ H(R) containing H0 remove degeneracy of G0 (condition (RD)). Then there exist constants ε > 0, c > 0, and a one-to-one mapping T from Kε (H0 ) onto a subset of H(R) so that (i) T (H0 ∩ Kε (H0 )) ⊂ H0 ; (ii) for any x ∈ G, with G ⊂ g(H0 ) admissible, there exists a Gibbs state that is a perturbation of x whenever H ∈ T (Hgr (G) ∩ Kε (H0 ))(=: H(G)); (iii) T (H) − H ≤ e
H −(ρ0 −c) H 0
for each H ∈ Kε (H0 );
(iv) (T (H1 ) − H1 ) − (T (H2 ) − H2 ) ≤ e for all H1 , H2 ∈ Kε (H0 ).
H H
−(ρ0 −c) min( H1 , H2 ) 0
0
H1 − H2
Remarks. 1. We expect that the phase diagram is complete in the sense that the only horizontally periodic Gibbs states for H ∈ H(G) are those corresponding to x ∈ G. Even though we have not a proof of this fact in a general situation, it should follow from the completeness of the ground state phase diagram by a method similar to that of [Z]. 2. The Gibbs states from (ii) satisfy an exponential decay of correlations. Also, explicit integral formulas describing them can be written using the integration with respect to measures on families of external contours and on families of walls which describe the interface of the states corresponding to x ∈ Ghor . We shall only describe the probabilities of external contours in Section 3 and of “shadows” in
218
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
Section 4 and we abstain from a discussion how to reconstruct all the probability using them. It could be done similarly like in [HKZ] Section 6. 3. The affine space H0 may be given in the form H0 = {Hα = H0 + N k=1 αk Hk }. It is customary to choose Hamiltonians Hk ∈ H(R) in a suitable way so that their number N may be taken as the minimal possible (N = N0 from (RD)). One could replace H0 by an N -dimensional smooth manifold defined by a mapping H(α), α ∈ RN (with N ≤ N0 ), from a neighbourhood V of a point α(0) ∈ RN into H(R) with H0 = H(α(0) ). The condition of removing of degeneracy then would be stated as the condition of maximality of the rank of the N0 × N matrix with the entries
∂ ex (H(α)) − ex0 (H(α)) , x ∈ g per(H0 ), x = x0 , and ∂αk α=α(0)
∂ , ex (H(α)) − exy,z (H(α)) ∂αk α=α(0) hor , x = xy,z , y, z ∈ g per(H0 ), y = z. x ∈ g(H0 ) ∩ Xy,z We shall not pursue the case of a general manifold H0 any further. 4. Usually one considers pairs, a Hamiltonian H and a temperature T , and assigns them the Gibbs states defined by T1 H (we put Boltzmann constant k = 1). In Theorem 1 we included the temperature into the Hamiltonian to avoid an overparametrization (multiplying both, the Hamiltonian and the temperature, by the same factor, we get exactly the same set of Gibbs states as originally). One can reformulate Theorem 1 exposing explicitly the temperature. Namely, one may introduce the temperature-depending mapping TT (H) = T · T (
H ). T
It describes the phase diagram at the slice of constant temperature T : the Gibbs states yielded by T1 TT (H) correspond (by (ii)) to the ground states of H T (the same as the ground states of H). Thus, for a fixed temperature T , one actually considers H0 as a space of parameters and the mapping TT shows how one should deform the ground state phase diagram to get the phase diagram at the temperature T . The condition (iii) shows that limT →0 TT = id, the identical mapping, and that the limit is attained exponentially fast. Notice that iv) yields a Lipschitz condition for the mapping (T, H) → (T, TT (H)). We derive Theorem 1 from the following lemma on the existence of a function characterizing the presence of stable states whose proof is the main content of the present paper and is presented in Sections 3, 4 and 5. We use here and in what follows ∂α+ ¯ to denote the directional one-sided ¯ (α) + derivative in the direction α ¯ , i.e. ∂α¯ f (α) = limt→0+ f (α+tα)−f . Let us recall t that the hamiltonian H = (U[A] ; A ⊂ Zν , diamA < R) is an element of the finitedimensional space H(R).
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
219
Basic Lemma. Let the assumptions of Theorem 1 be fulfilled (with ρ0 sufficiently large). Then there exist ε > 0 and mappings hx : Kε (H0 ) → R (H → hx (H)) for each x ∈ G0 such that (i) a) for H ∈ Kε (H0 ) and x ∈ Gper such that hx (H) = minx˜∈Gper hx˜ (H) (such x 0 is then called stable), there exists an extremal Gibbs state µ ∈ G(H) that is a perturbation of x; hor b) for H ∈ Kε (H0 ) and x ∈ G0 ∩ Xy,z , where y, z ∈ Gper 0 , y = z, are stable per elements of G0 and hx (H) = minx˜∈G0 ∩Xy,z hor hx (H) (x is stable), there exists ˜ an extremal Gibbs state µ ∈ G(H) that is a perturbation of x;
(ii) there exists c > 0 (independent of ρ0 ) such that a)
|hx (H) − ex (H)| ≤ e
−(ρ0 −c)
H H0
(1.9)
for each H ∈ Kε (H0 ) and b)
+ + ∂H ¯ hx (H) − ∂H ¯ ex (H) ≤ e
−(ρ0 −c)
H H0
¯ H
(1.10)
¯ ∈ H(R). for any H ∈ Kε (H0 ), H Remark (passing to horizontally translation invariant setting). Since we suppose that G0 is finite up to vertical translations, we may and shall suppose that all are actually translation invariant and all elements of Ghor are elements of Gper 0 0 horizontally translation invariant by considering a modified model. Namely, we can choose some partition of Zd into a grid of cubes with edges of a length that is some (e.g. the smallest possible) common multiple of the periods of all concerned periods. The set of “spins” attained at such a cube B consist then of all configurations on B with values in S. This changes the number of “spins” in dependence on the periods only. In the proofs that follow in Sections 3, 4 and 5, we use this observation and suppose that G0 consists of translation invariant and horizontally translation invariant configurations. Proof of Theorem 1. Denoting the parameter ε of the cone from Basic Lemma by ε˜, the assertion (i) implies that it is enough to find a suitable mapping T (H) solving the equations hx (T (H)) − hx0 (T (H)) = ex (H) − ex0 (H) for x, x0 ∈ Gper 0 , x = x0 , and hx (T (H)) − hxy,z (T (H)) = ex (H) − exy,z (H) for y, z ∈ Gper 0 , y = z and (1.11) hor , x = xy,z x, xy,z ∈ G0 ∩ Xy,z for all H ∈ Kε (H0 ) for some 0 < ε < ε˜. We see from Basic Lemma (i) immediately that, if x is a horizontally invariant ground configuration of H, then x is stable with respect to T (H). The number of different equations in (1.11) is N0 (see (RD) above). To get a unique solution T (H) satisfying (i) of the theorem we add the following
220
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
equations. Namely, consider a decomposition H(R) = H(0) ⊕ H(1) into two orthogonal subspaces such that dim H(0) = N0 , H(0) + H0 ⊂ H0 , and H(0) + H0 removes degeneracy of G0 . Thus every H ∈ H(R) can be decomposed uniquely as H = H0 + H (0) + H (1) , H (0) ∈ H(0) , H (1) ∈ H(1) . Denoting H (1) = Π(1) (H) we consider the additional equations Π(1) (T (H)) = Π(1) (H).
(1.12)
The right hand sides of equations (1.11), (1.12) define an invertible linear mapping L from H(R) onto RN0 × H(1) while the left-hand sides define a mapping F that differs only slightly from L on Kε˜(H0 ) according to Basic Lemma (ii). More precisely, we put L(H) = (ex (H) − ex0 (H); x ∈ Gper 0 \ {x0 }), per (1) (ex (H) − exy,z (H); y = z, y, z ∈ Gper (H) 0 , x ∈ G0 ∩ Xy,z \ {xy,z }), Π and
F (H) = (hx (H) − hx0 (H); x ∈ Gper 0 \ {x0 }),
per (1) (hx (H) − hxy,z (H); y = z, y, z ∈ Gper , x ∈ G ∩ X \ {x }), Π (H) . 0 y,z y,z 0
If we express L in suitable orthogonal coordinates, we get by (RD) that the absolute value of the jacobian of L, |jL (H)|, is strictly positive for all H ∈ H(R). ˜ = T (H) fulfilling (1.11) and (1.12), i.e. fulfilling Our task thus is to find H ˜ = L(H) for all H ∈ Kε (H0 ) for some ε. Equivalently, we want the equation F (H) to solve the equation ˜ = H − L−1 P (H), ˜ H (1.13) ˜ = F (H) ˜ − L(H) ˜ is a mapping fulfilling the bound where P (H) ˜ H
˜ ∈ Kε˜(H0 ) ˜ ≤ e−(ρ0 −c) H0 for H P (H) due to Basic Lemma (ii)a), and ˜ H ˜ H
˜ 1 ) − P (H ˜ 2 ) ≤ e−(ρ0 −c) min( H10 , H20 ) H ˜1 − H ˜ 2 for H ˜ 1, H ˜ 2 ∈ Kε˜(H0 ) P (H due to Basic Lemma (ii)b). ˜ of the mapping Q : H ˜ → H − L−1 P (H) ˜ in (1.13) To find the fixed point H for a fixed H ∈ Kε (H0 ), we use the Banach contraction principle. We have ˜ 2 ) ≤ L−1 e−(ρ0 −c)(1−ε) H ˜1 − H ˜ 2 ˜ 1 ) − Q(H Q(H ˜ 1, H ˜ 2 ∈ Kε (H0 ). for H
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
221
Since the constant L−1 e(ρ0 −c) (1−ε) is smaller than one if ρ0 is sufficiently large (we may and shall suppose that ε ≤ 12 ), it suffices to show that Q maps ˜ | H ˜ − H ≤ ε˜ } into itself for H ∈ K ε˜ (H0 ). This allows us to take U ε˜ (H) = {H 2 2 2 ε = 2ε˜ . ˜ − H ≤ ε˜ . Then Let H 2 ˜ − Q(H) ≤ L−1 e−(ρ0 −c)(1−˜ε) Q(H)
ε˜ ε˜ ≤ . 2 2
Hence, for sufficiently large ρ0 (chosen independently of H ∈ Kε (H0 )), we get that Q is a contraction on Uε (H). The unique solution T (H) in Uε (H) fulfills the assertions (i) and (ii) of Theorem 1 according to (i) of Basic Lemma. Using (ii) of Basic Lemma, we get the asked properties (iii) and (iv) of Theorem 1 easily.
2 Labeled contour models We present here a brief reformulation of the essential part of Pirogov-Sinai theory and its slight extension to a form needed for our application. The task is to grasp some control over a description of “labeled contour models” that arise in the study of (“physical”) Gibbs states. The characteristic feature of the Pirogov-Sinai theory is a reformulation in terms of “contour models”. The following presentation follows essentially [Z], but it brings some improvements (see especially Theorem 2 and Proposition 2.2.1; compare also the paper [BK] using some ideas from a preliminary version of the present paper).
2.1
Contour models
We use here the word contour simply for any finite connected (in the sense of nearest neighbours) non-empty subset of Zν (ν ≥ 2). Given any contour Γ ⊂ Zν , we define its exterior, ExtΓ, to be the only infinite connected component of Zν \ Γ. The interior, IntΓ, of Γ is the union of the other (finite) connected components of Z \ Γ. We denote V (Γ) = Γ ∪ IntΓ. A set ∂ of contours is compatible if any pair of distinct elements of ∂ has a disconnected union. A contour model is given by introducing a contour functional4 (“contour weights”) Ψ which maps the set of contours to [0, ∞). Considering a set L of contours, we define the contour model partition function in L by Ψ(Γ) , (2.1) Z(L; Ψ) = ∂⊂L Γ∈∂
with the sum taken over all compatible families ∂ of contours from L. Notice that this definition makes sense not only for any finite set L, but supposing that the 4 We restrict our attention here to the case of real-valued contour functionals that will arise in our context.
222
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
sum and products converge, also for some infinite L. In dealing with general sets L we follow the abstract setting from [KP]. Given a volume Λ ⊂ Zν , we shall be concerned with two particular cases of sets of contours, namely the set L(Λ) of all contours in Λ, L(Λ) = {Γ : dist(V (Γ), Λc ) > 1}, and the set LΛ of all contours intersecting Λ, LΛ = {Γ : V (Γ) ∩ Λ = ∅}. For a finite L we can also introduce the contour model probability distribution in L by Ψ(Γ) µ({∂}, L; Ψ) = Γ∈∂ (2.2) Z(L; Ψ) for any compatible ∂ in L. As usual, the empty product is put equal to one. For an infinite L we can introduce the compact space of families of contours from L as a closed subspace of {0, 1}L with the product topology, and consider the weak limit of measures (2.2). It is useful to introduce a special symbol, e.g. ρ(∂, L; Ψ), for the “correlations” ¯ L; Ψ). µ({∂¯ : ∂ ⊂ ∂}, We may now summarize the main facts concerning contour models with a contour functional Ψ satisfying the inequality Ψ(Γ) ≤ exp(−τ |Γ|) for any contour Γ,
(2.3)
i.e. with Ψ being a τ -functional. Most of the assertions of the following proposition as well as their proofs can be found, for example, in [S, Se, Br]. In view of an application to “volumes” of the form LΛ we rely on an abstract version of contour models and cluster expansions as presented in [KP, D 96]. In particular, we define the distancedist(∂, L) from a set ∂ of compatible contours to a set L of contours as minC Γ∈C |Γ|, where the minimum is taken over all clusters whose union is sets of contours C (i.e. Γ Γ = ∅. connected) such that C ∩ L = ∅ and Γ∈∂ Γ∈C Let us introduce also the notion of external contours. Namely, a contour Γ ∈ ∂ ¯ for each Γ ¯ ∈ ∂, Γ ¯ = Γ. If ∂ is a family of is an external contour of ∂ if Γ ⊂ Ext Γ ¯ for some external contours such that every Γ ∈ ∂ is either external or Γ ⊂ Int Γ ¯ contour Γ of ∂, then we say that external contours exist. We use ϑ(∂) to denote the set of all external contours. Proposition 2.1.1 (contour models) There exist constants τcl ≡ τcluster (ν) and ccl ≡ ccluster(ν) (both depending only on the dimension ν of the lattice) such that, whenever L is an arbitrary set of contours and Ψ is a τ -functional with τ ≥ τcl , then the weak limit µ(·, L; Ψ) = lim µ(·, K; Ψ) K L
over the system of finite subsets ordered by inclusion exists, and : a) For µ-almost all families ∂ there exists the set ϑ(∂) of external contours. b) Whenever ∂ is a compatible family in L, we have ρ(∂, L; Ψ) ≤ exp(−(τ − 1) |Γ|). Γ∈∂
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
223
c) Whenever ∂ is a compatible family of contours in the intersection of sets L1 , L2 , we have |ρ(∂, L1 ; Ψ) − ρ(∂, L2 ; Ψ)| |Γ|) exp(−(τ − ccl ) dist(∂, L1 L2 )). ≤ exp(−(τ − 1) Γ∈∂
(Here L1 L2 is the symmetric difference L1 L2 = (L1 \ L2 ) ∪ (L2 \ L1 ).) Hence also, for any finite non-empty set L ⊂ L1 ∩ L2 and any mapping ϕ of families of contours to real numbers, living5 on L, with |ϕ(∂)| ≤ ϕ, one has
ϕ(∂)µ(d∂, L1 ; Ψ) − ϕ(∂)µ(d∂, L2 ; Ψ)
≤ ϕ exp(−(τ − ccl ) dist(L, L1 L2 )). d) Let us suppose further that Ψ is translation invariant. Then the limiting “pressure” 1 p(Ψ) = limν log Z(L(Λ); Ψ) Λ Z |Λ| (with the limit in the van Hove sense) exists. The partition function Z(L(Λ); Ψ) satisfies the approximation log Z(L(Λ); Ψ) = |Λ| p(Ψ) + ε |∂Λ| with |ε| ≤ exp(−(τ − ccl )), and |p(Ψ)| ≤ exp(−(τ − ccl )). e) Let us suppose now that a family of translation invariant τ (α) -functionals Ψ(α) , τ (α) ≥ τcl , is given, depending on a parameter α from an open set Ω ⊂ Rn and suppose that the one-sided derivative in the direction α ¯ fulfills for some α ∈ Ω, α ¯ ∈ Rn , the bound
+ (α)
∂α¯ Ψ (Γ) ≤ exp(−τ (α) |Γ|)||¯ α|| (α) for every contour Γ. Then the derivative ∂α+ ) satisfies the bound ¯ p(Ψ
+
∂α¯ p(Ψ(α) ) ≤ exp(−(τ (α) − ccl ))||¯ α||. 5 We say that ϕ lives on L if ϕ(∂ ) = ϕ(∂ ) whenever the set of all contours from ∂ that are 1 2 1 contained in L coincides with the set of all contours from ∂2 contained in L.
224
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
Proof. The assertions a)-c) are standard and their proof may be based on the cluster expansion as presented in the statement a) of Proposition 2.1.3. The statement d) follows from a rather straightforward extension of standard proofs of cluster expansion (Proposition 2.1.3 b)). The constant ccl (ν) can be derived from the constant c# = c# (ν) determining the number of “contours of given length”,
{Γ | Γ ! 0, |Γ| = k} ≤ ec# (ν)k .
(2.4)
Namely, it can be shown (see [KP]) that ccl (ν) = c# (ν) + a +
log(1 + a) ∼ c# (ν) + 1.58 log a
(2.5)
√
with a = 5−1 2 . One often meets a situation where the contour functional ΨΛ (Γ) depends on the volume Λ once the contour Γ crosses the boundary of Λ. Also in this case one has a full control of the limiting contour model. Proposition 2.1.2 (contour models with dependence on the boundary) Let V ⊂ Zν be arbitrary and let a sequence Vn of sets converging to V be given, Vn " V . Let for any Vn a τ -functional ΨVn be given so that ΨVm (Γ) = ΨVn (Γ) whenever Γ∈ / LVm Vn 6 and suppose that τ ≥ τcl with τcl from Proposition 2.1.1. Then the limit ΨV = lim ΨVn n→∞
is uniquely determined and the limiting measure µ(·, LV ; ΨV ) = lim µ(·, LVn ; ΨVn ) n→∞
exists. It coincides with the measure µ(·, LV ; ΨV ) = lim µ(·, K; ΨV ) K LV
from the preceding proposition. Further, for any (possibly infinite) V, V1 , V2 ⊂ Zν , one has: a) Whenever ∂ is a compatible family in LV , we have ρ(∂, LV ; ΨV ) ≤ exp(−(τ − 1)
|Γ|).
Γ∈∂ 6 It
means also that ΨΛ (Γ) coincides with Ψ
Z(Γ) for every contour Γ in Λ. ν
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
225
b) Whenever ∂ is a compatible family of contours in the intersection of sets LV1 , LV2 , we have |ρ(∂, LV1 ; ΨV1 ) − ρ(∂, LV2 ; ΨV2 )| |Γ|) exp(−(τ − ccl ) dist(∂, LV1 LV2 )) ≤ exp(−(τ − 1) Γ∈∂
with ccl = ccl (ν) from Proposition 2.1.1. For any finite non-empty set Λ ⊂ V1 ∩ V2 and any mapping ϕ, living on LΛ , of families of contours to real numbers, with |ϕ(∂)| ≤ ϕ, one has
ϕ(∂)µ(d∂, LV1 ; ΨV1 ) − ϕ(∂)µ(d∂, LV2 ; ΨV2 )
≤ ϕ exp(−(τ − ccl ) dist(Λ, V1 V2 )). Suppose, further, that Ψ ≡ ΨZν is translation invariant. Then c) The limiting “free energy” (or “pressure”) p(Ψ) = limν Λ Z
1 log Z(LΛ ; ΨΛ ) |Λ|
(with the limit in the van Hove sense) exists with p(Ψ) the same as in c) of the preceding proposition7 . The partition function Z(LΛ ; ΨΛ ) satisfies also the approximation log Z(LΛ ; ΨΛ ) = |Λ| p(Ψ) + ε |∂Λ| with |ε| ≤ exp(−(τ − ccl )). Proof. Due to the conditions on the functionals ΨVn , the value ΨVn (Γ) stays constant, for every Γ, once n is sufficiently large. Taking into account the possibility to verify the convergence of measures by proving the convergence of correlations ρ(∂, LVn ; ΨVn ) for finite families ∂ and observing that it may be approximated, from a cluster expansion, by restricting the contour model on a finite number of contours, the statements of Proposition 2.1.2 follow from cluster expansions (Proposition 2.1.3 below) and Proposition 2.1.1 (cf. [HKZ] Proposition B.2). Let us also briefly summarize few standard facts about the cluster expansion in a form suitable for our purposes [KP, D 96]. A proof of b) in Proposition 2.1.3 appears in several papers [BK, DKS]. Propositions 2.1.1 and 2.1.2 follow easily from Proposition 2.1.3 below. The results do not serve for the proofs of the above propositions only; we use the explicit form of the expansion in an essential way in Section 4 when studying the probability of interfaces. 7 Notice that it means in particular that the statement 1 c) is valid also with L replaced by Λ sets L(Λ) , where L(Λ) ⊂ L(Λ) ⊂ LΛ .
226
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
Proposition 2.1.3 (cluster expansion) There exist constants τcl ≡ τcl (ν) and ccl ≡ ccl (ν) such that the following statements hold true. a) Whenever Ψ is a τ -functional with τ ≥ τcl , then there exists a (“cluster”) functional ΨT that maps all non-empty clusters of contours (i.e. sets C of contours with connected support C = Γ∈C Γ) to R such that log Z(L; Ψ) =
ΨT (C)
(2.6)
C⊂L
for every finite set L of contours. The functional ΨT satisfies, for every i ∈ Zν , the bound
ΨT (C) exp (τ − ccl ) |Γ| ≤ 1, (2.7) C:Ci
Γ∈C
with the sum taken over all clusters with support C = Γ∈C Γ containing a fixed site i ∈ Zν . The value ΨT (C) depends only on values of Ψ(Γ) with Γ ∈ C, and it is translation invariant once the functional Ψ is translation invariant. b) Suppose further that τ (α) - functionals Ψ(α) , τ (α) ≥ τcl , depend on a parameter α from an open set Ω ⊂ Rn and satisfy for every contour Γ, α ∈ Ω, and any α ¯ ∈ Rn , the inequality
+ (α)
∂ Ψ (Γ) ≤ exp(−τ (α) |Γ|)||¯ α||. (2.8) α ¯ Then, for every i ∈ Zν and all α ∈ Ω, one has
∂ + Ψ(α)T (C) exp (τ (α) − ccl ) |Γ| ≤ ||¯ α||. α ¯ C:Ci
(2.9)
Γ∈C
Remark. Assuming, instead of (2.8), the existence of the Fr´echet derivative and a bound on its norm, we actually get the existence of the Fr´echet derivative also for Ψ(α)T , with a corresponding bound. Resuming over all clusters with coinciding support and denoting Ψ T (C) = ΨT (C), we will get a formulation that is particularly suitable for our C:
S
Γ∈C
Γ=C
implementations. Corollary 2.1.4 Let τcl and ccl be the constants from Proposition 2.1.3. If Ψ is a τ -functional with τ ≥ τcl , then log Z(L(Λ); Ψ) = Ψ T (C) (2.10) C⊂Λ(0)
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
227
for every finite Λ ⊂ Zν . The sum above runs over all connected subsets C of the set Λ(0) = {i ∈ Λ dist(i, V c ) > 1}. The functional Ψ T (C) satisfies the bound
Ψ T (C) exp (τ − ccl ) |C| ≤ 1. (2.11) Ci
Let, moreover, Ψ be translation invariant. Then the limits p(Ψ) = limν Λ Z
1 1 log Z(L(Λ); Ψ) = limν log Z(LΛ ; Ψ) Λ Z |Λ| |Λ|
(the limit in the van Hove sense) exist and satisfy the bounds
p(Ψ) ≤ exp{−(τ − ccl )} and
+
∂α¯ p(Ψ) ≤ exp{−(τ − ccl )}||¯ α||.
(2.12) (2.13)
The function p(Ψ) is explicitly given by p(Ψ) =
Ψ T (C) . |C|
(2.14)
C0
Remark. Notice, that the assumption of the translation invariance of Ψ in Propositions 2.1.1 c), 2.1.2 c), and the above corollary is actually not necessary. It is enough to introduce pi (Ψ) =
ΨT (C) Ψ T (C) = |C| |C|
C:Ci
and to replace the terms |Λ| p(Ψ) by
Ci
pi (Ψ).
i∈Λ
For periodic Ψ the limit p(Ψ) is obtained as a mean of pi (Ψ) over the cell of periodicity.
2.2
Labeled contour models
We consider a finite set Q = {1, . . . , r} of “labels” and call γ = (Γ, λ) a labeled contour if its support Γ ≡ Γ(γ) ⊂ Zν is a finite non-empty connected set (a contour), and λ = λ(γ) assigns to each connected component of the boundary ∂Γ of Γ some q ∈ Q. A labeled contour γ is called a q-contour if the label assigned to its external boundary (the boundary of Γ∪Int Γ) is q. A family x of labeled contours is said to be compatible and matching if their supports are compatible and their labels match (i.e., considering a connected component, say C, of Zν \ {Γ(γ) : γ ∈ x}, all
228
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
connected components of the boundaries of Γ’s adjacent to C are labeled by the same label q; we say that C is a q-component ). Compatible and matching families x of labeled contours play a role of configurations for labeled contour models 8 . Whenever x is such that for the family ∂ = {Γ(γ) : γ ∈ x} of its supports, there exists the set ϑ(∂) of external contours (in the sense of the definition preceding Proposition 2.1.2), we introduce the set ϑ(x) (ϑq (x)) of external labeled contours of the family x as the subset of those γ = (Γ, λ) ∈ x whose support Γ is an external (q−)contour of ∂. We say that x is included in V ⊂ Zν if dist(V (Γ(γ)), V c ) > 1 for γ ∈ x. Whenever Λ ⊂ Zν , we consider the set X(Λ) of compatible matching families x of labeled contours that are included in Λ. We introduce the set XqΛ as the set of all compatible matching families of labeled contours that intersect Λ and are such that external contours exist and all external contours as well asq all = contours not belonging to X(Λ) are q-contours. Let us also denote X Λ q XΛ . If x ∈ XΛ , the union of m-components of Λ \ {Γ(γ) : γ ∈ x} together with supports of all m-contours of x is denoted by Λm (x). Clearly, the number of labeled contours γ, for which supp γ = Γ, is bounded by |S||Γ| . A labeled contour functional Φ (“an exponential of a contour Hamiltonian”) maps the set of labeled contours into [0, ∞). We also assume that a vector ϕ = (ϕ1 , . . . , ϕr ) ∈ Rr (of “specific energies of some translation invariant configurations”) is given. We consider a labeled contour model with a boundary condition q ∈ Q by introducing, for any Λ ⊂ Zν and any X(Λ) , X(Λ) ⊂ X(Λ) ⊂ XΛ , the (labeled contour model) partition functions exp − ϕm |Λm (x)| Φ(γ). (2.15) Z(X(Λ) |q; Φ, ϕ) = x∈X(Λ)
γ∈x
m∈Q
Similarly we introduce the (labeled contour model) probability distribution ϕm |Λm (x)| Φ(γ). µ({x}, X(Λ) | q; Φ, ϕ) = Z −1 (X(Λ) |q; Φ, ϕ) exp − m∈Q
γ∈x
(2.16) Again, the notation for partition functions and probability distributions above is distinguishing them from those for a lattice spin model (c.f. Section 1.1) only by using different variables. When stressing the fact that we are dealing with partition functions and probability distributions of a labeled contour model, we will use the notation Z cont and µcont . Remarks. 1. For any Λ ⊂ Zν , the sets X(Λ) (resp. XΛ ) may be embedded into ν the product space {1, . . . , r}Z by assigning to every x that configuration from ν {1, . . . , r}Z which attains the value q at all lattice sites in Λq (x). On the sets X(Λ) ν (resp. XΛ ) we consider the topology inherited from the compact space {1, . . . , r}Z . 8 We use the same notation as for spin configurations of classical lattice models in view of the existence of a natural identification of a class of lattice configurations with a given compatible and matching family of labeled contours (see Section 3).
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
229
Let us notice that for an infinite Λ, even if x = lim xn with xn such that ϑ(xn ) consists of q-contours and in the same time ∂(xn ) converges, in the topology of families of contours used in Section 2.1, to a compatible family of contours ∂ such that external contours ϑ(∂) exists, it might be that ϑ(x) = ϑq (x) with q = q. 2. Notice that there are small formal differences in our formulation and that of [S]. First, our definition of contours does not specify configurations on the support. We also found useful to consider directly the “weights of contours”, thus the ˜ in [S, Z]. functional Φ above corresponds to exp(−Φ) In the standard Pirogov-Sinai theory, the functional Φ˜ as well as the r-tuple ϕ = (ϕ1 , . . . , ϕr ) are linear in the Hamiltonian H (they are given in terms of explicit formulae — see e.g. (1.4) and (1.6) from [Z]). However, in the present application, we shall meet a more general situation. 3. Readers accustomed to standard Pirogov-Sinai formulation may also notice that the partition function Z(X(Λ)|q; Φ, ϕ) corresponds, roughly speaking, to the (relative) diluted partition function multiplied by the factor exp(ϕq |Λ|) (since the standard approach is based on relative Hamiltonian with respect to the energy corresponding to the external boundary condition). The crucial part of the Pirogov-Sinai theory is formulated in Theorem 2 below. First, we need a definition and some more notation. Definition. A phase q is called cs -stable in Λ (with respect to Φ and ϕ) if Z(X(Λ)|λ; Φ, ϕ) ≤ exp(cs |∂Λ|), Z(X(Λ)|q; Φ, ϕ) for every labeling λ (i.e. a label λ(b) ∈ Q for every connected component b of the boundary ∂Λ, chosen in a compatible way). A phase q is said to be cs -stable (with respect to Φ and ϕ) if it is cs -stable in Λ for every non-empty finite Λ. The main aim of the Pirogov-Sinai theory is to provide a characterization of stable phases showing, in the same time, that any stable phase gives rise to a distinct Gibbs state that is a perturbation of the corresponding ground configuration. We split these claims into two statements. First, in Theorem 2, we characterize the stability (see (S)) in terms of certain functions hq (Φ, ϕ) that are close to external fields ϕq . Proposition 2.2.1 then describes the properties of the stable phases. Theorem 2 (characterization of stable phases). Let, for every α from an open set Ω ⊂ Rn of parameters and any Λ ⊂ Zν , a nonnegative labeled contour functional (α) ΦΛ and a vector ϕ(α) of “specific energies” be given (with labels from a finite set (α) (α) Q) such that ΦΛ1 (γ) = ΦΛ2 (γ) whenever supp γ ⊂ Λ1 ∩ Λ2 . We denote Φ(α) = (α)
ΦZν . We suppose that a continuous function τ (α) is given so that for every labeled contour γ, and all α ∈ Ω, we have (1)
(α)
ΦΛ (γ) ≤ exp(−τ (α) |Γ(γ)|).
230
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
Then, there exist a constant τ = τlabeled(cs , ν, |S|) and functions hq (Φ(α) , ϕ(α) ) characterizing cs -stability9 whenever τ (α) ≥ τ . Namely, the phase q is cs -stable with respect to Φ(α) and ϕ(α) if and only if hq (Φ(α) , ϕ(α) ) = min(hm (Φ(α) , ϕ(α) )).
(S)
m
The functions hq (Φ(α) , ϕ(α) ) can be chosen in such a way that, denoting h(Φ , ϕ(α) ) = minm (hm (Φ(α) , ϕ(α) )), one has
Z(Λ|q; Φ(α) , ϕ(α) ) ≥ exp −hq (Φ(α) , ϕ(α) ) |Λ| − ε |∂Λ| (α)
and
Z(Λ|q; Φ(α) , ϕ(α) ) ≤ exp −h(Φ(α) , ϕ(α) ) |Λ| + ε |∂Λ|
with ε = e−(τ −ccl ) + e−(τ −ccl −c# −log |S|−1) . Moreover, there exists a constant c = clabeled(cs , ν, |S|) such that
≤ exp −(τ (α) − c − ccl )
hq (Φ(α) , ϕ(α) ) − ϕ(α) q (α)
(α)
for q = 1, . . . , r and α ∈ Ω. Supposing, moreover, for any α ∈ Ω and α ¯ ∈ Rn , the bounds on the (one(α) (α) sided) directional derivatives of ΦΛ and ϕq ,
+ (α)
∂ Φ (γ) ≤ exp −τ (α) |Γ(γ)| ||¯ (2) α|| and α ¯ Λ
+ (α)
∂ ϕ ≤ M ||¯ α|| for some M > 0, (3) α ¯
q
there exists constants τ = τ (cs , M, ν, |S|) and c = c (cs , ν, |S|) (possibly larger than those above) such that for τ (α) > τ we have:
+
≤ exp −(τ (α) − c − ccl ) ||¯
∂α¯ hq (Φ(α) , ϕ(α) ) − ϕ(α) α|| q and, denoting h(Φ(α) , ϕ(α) ) = minm (hm (Φ(α) , ϕ(α) )), also
+
∂α¯ h(Φ(α) , ϕ(α) ) ≤ M + exp −(τ (α) − c − ccl ) ||¯ α|| for α ∈ Ω and α ¯ ∈ Rn . Remark. Notice, in particular, that if Ω is convex, the functions hq (Φ(α) , ϕ(α) ) − (α) ϕq are, as functions of α, Lipschitz with the constant exp −(τ (α) − c − ccl ) . The theorem follows from an explicit construction of functions hq (Φ(α) , ϕ(α) ) ¯ introduced below. The expression in terms in terms of contour functionals Ψ and Ψ of those functionals also yields the properties of stable phases. 9 Stability implies a good control of the corresponding states. These implications, as well as an explicit construction of the characterizing functions hq (Φ(α) , ϕ(α) ), are presented in Proposition 2.2.1 and Corollary 2.2.2.
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
231
Proposition 2.2.1 (properties of stable phases) Under assumption (1) of Theorem 2: a) Existence of contour functionals Ψ For every finite Λ, α ∈ Ω and q = 1, . . . , r, (α) there exist uniquely determined nonnegative contour functionals Ψq,Λ such that (Λ) ; Ψ(α) Z(X(Λ) |q; Φ(α) , ϕ(α) ) = exp −ϕ(α) q |Λ| Z(L q )
(2.17)
for any L(Λ) , L(Λ) ⊂ L(Λ) ⊂ LΛ with X(Λ) = {x ∈ XΛ ; γ ∈ X implies Γ(γ) ∈ L(Λ) }, and (α)
(α)
µ({x : ϑ(x) = ϑ}, X(Λ) | q; ΦΛ , ϕ(α) ) = µ({∂ : ϑ(∂) = ϑ}, L(Λ) ; Ψq,Λ ) (α)
(2.18)
(α)
for any collection ϑ of external q-contours. Moreover, Ψq,Λ1 (Γ) = Ψq,Λ2 (Γ) when(α)
ever10 V (Γ) ⊂ Λ1 ∩ Λ2 . The functional Ψq,Λ satisfies, for any Λ, the bound
(α)
Ψ (Γ) ≤ exp −(τ (α) − c ) |Γ| , q,Λ
whenever Γ is such that q is cs -stable in every component of Int Γ (with respect to Φ(α) and ϕ(α) ) and c ≥ c# (ν) + cs + log |S|. (α) ¯ Taking Ψ(α) = Ψ ν , nonnegative contour b) Existence of contour τ -functionals Ψ q q,Z
(α)
¯ q exist, such that, for every contour Γ, we have functionals Ψ (α) (α) ¯ • Ψq (Γ) ≤ Ψq (Γ), (α) (α) ¯ •Ψ whenever q is cs -stable in Int Γ,
q (Γ) =
Ψq (Γ)
¯ (α) • Ψq (Γ) ≤ exp −(τ (α) − c ) |Γ| with c ≥ c# + cs + log |S|. Supposing that the functional Φ(α) is translation invariant (in an obvious way ¯ (α) with respect to shifts in Zν ), the functional Ψ can be chosen to be translation q invariant. Further on, there exists τ = τ (cs , ν, |S|) such that for τ ≥ τ we have ¯ (α) The functions hq (Φ(α) , ϕ(α) )(≡ c) Description of hq (Φ(α) , ϕ(α) ) in terms of Ψ ¯ (α) , ϕ(α) )) defined by hq (Ψ ¯ (α) ), hq (Φ(α) , ϕ(α) ) = ϕq − p(Ψ q (α)
¯ q ) defined in Corollary 2.1.4, characterize the stability (i.e., satisfy the with p(Ψ equivalence (S)). Further on, whenever α is such that q is cs -stable with respect to Φ(α) , ϕ(α) , one has hq (Φ(α) , ϕ(α) ) = − lim 10 Supposing
1 log Z(X(Λ)|q; Φ(α) , ϕ(α) ) = h(Φ(α) , ϕ(α) ). |Λ| (α)
(α)
equality ΦΛ1 (γ) = ΦΛ2 (γ) for supp γ ∩(Λ1 Λ2 ) = ∅, we get Ψq,Λ1 (Γ) = Ψq,Λ2 (Γ) whenever dist(V (Γ), Λ1 Λ2 ) > 1.
232
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık (α)
Ann. Henri Poincar´e
(α)
Supposing, moreover, the smoothness of ΦΛ and ϕq (the conditions (2) and (3) of Theorem 2), there exist a constant τ = τ (cs , M, ν, |S|) (possibly larger then τ (cs , ν, |S|) above) such that for τ (α) > τ we have : a ) Smoothness of contour functionals Ψ(α) For any Λ,
+ (α)
∂α¯ Ψ (Γ) ≤ exp −(τ (α) − c ) |Γ| ||¯ α|| q,Λ with c ≥ cs + c# + log |S|+ 1 + ln 2, whenever Γ (and α) are such that q is cs -stable in every component of Int Γ (with respect to Φ(α) and ϕ(α) ). ¯ (α) There exist functionals Ψ ¯ (α) satisfying b ) Smoothness of contour functionals Ψ q b), with c ≥ 2cs + c# + log |S| + 1, and
+ (α)
∂α¯ Ψ ¯ q (Γ) ≤ exp −(τ (α) − c ) |Γ| ||¯ α|| with c ≥ 4cs + c# + 2 + 2 log(|S| + 1), for every Γ and every α ∈ Ω. (α) ¯ (α) Observation. Notice that, by b) and c), hq (Φ, ϕ) = minm (hm (Φ, ϕ)) iff Ψ = Ψq . q Further, the notion of cs -stability of a phase q actually does not depend on cs . Hence, in the following, we will just say that a phase q is stable. (However, changes of cs may lead to changes of τ .)
Remarks. 1. Clearly, there exists a choice of sufficiently large c (cs , ν, |S|) such that it can
(α)
(α) ¯ q (Γ) from a) be used simultaneously in the bounds on Ψq (Γ) and Ψ
+ (α)
+ (α)
¯ and b) as well as in the bounds on ∂α¯ Ψq (Γ) and ∂α¯ Ψq (Γ) from a ) and b ) . 2. The limit in c) is over finite volumes in the van Hove sense and defines the quantity corresponding to “pressure” in the physical model based on a Hamiltonian at a particular temperature. (α)
3. Studying the contour model probabilities µ(·, Λ; Ψq ) for stable q and their limits for simply connected Λ tending to infinite simply connected volumes, we can describe labeled contour model probabilities in an infinite volume under boundary condition q by conditioning over external contours (cf. Proposition 2.1.2). 4. It suffices to suppose (1) (with sufficiently large τ (α) ), to get the existence (α) ¯ (α) = (Ψ ¯ (α) of Ψ(α) = (Ψq ), Ψ q ) fulfilling a), b), and c). However, we were not able to follow exactly the method described in [Z] or [BI] to prove the smoothness b ) (and hence also a )). Therefore, we extend the assertion of Theorem 1 from [Z] in Lemma 2.3.1 to get a), b), and c) for a wider class ¯ (α) fulfilling b ) among ¯ (α) . We shall later find a functional Ψ of functionals Ψ them (Lemmas 2.3.1 and 2.3.2 and the final part of the proof of Theorem 2). 5. If we take H ∈ H(R) for the parameter α and Kε (H0 ) for the set Ω, the H (see Proposition 1.1.3). function τ (H) can be chosen as τ (H) := ρε H 0
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
6. It is enough to suppose the estimates from (1) and (2) for the sum is over q-contours γ = (Γ, λ) with Γ fixed.
γ
233
(α)
ΦΛ (γ), where
7. The first point in b) can be weakened to ¯ (α) (Γ) ≤ econst|Γ| Ψ(α) (Γ), Ψ q q and the second point can be weakened by introducing a stronger notion of stability (see [BK]). We postpone the proof of Proposition 2.2.1 to Section 2.3 below. Theorem 2 then easily follows. Proof of Theorem 2. The only claims that are not directly included in Proposition 2.2.1 are the bounds on derivatives of hq (Φ(α) , ϕ(α) ) and h(Φ(α) , ϕ(α) ). Using b ) ¯ (α) of Proposition 2.2.1, the bound (2.13) on p(Ψ q ), and the assumption (3), we get
+
∂ hq (Φ(α) , ϕ(α) ) ≤ M + exp −(τ (α) − c − ccl ) ||¯ α||. α ¯
(α) To estimate ∂α+ , ϕ(α) ) , we first notice that if, for a fixed α, there exists ¯ h(Φ q ∈ Q such that h(Φ(α) , ϕ(α) ) = hq (Φ(α) , ϕ(α) ) and h(Φ(α) , ϕ(α) ) < hm (Φ(α) , ϕ(α) ), m = q, the claim immediately follows from the bound above. If h(Φ(α) , ϕ(α) ) = ¯ ⊂ Q, and h(Φ(α) , ϕ(α) ) < hq (Φ(α) , ϕ(α) ), q ∈ Q \ Q, ¯ we get, hq (Φ(α) , ϕ(α) ), q ∈ Q for a fixed direction α ¯, (α) (α) ∂α+ , ϕ(α) ) = min ∂α+ , ϕ(α) ). ¯ h(Φ ¯ hq (Φ ¯ q∈Q
¯ the directional Indeed, if the right hand side is attained for several q’s in Q, derivative on the left hand side equals any one of them, since they are equal anyway. We will use the results of Theorem 2 and Proposition 2.2.1 together with the results on contour models to describe the limit Gibbs states obtained under special boundary conditions as it is customary in the standard Pirogov-Sinai theory. However, in our application, the partition function of the “physical” model in Λ is described in terms of a labeled contour model with contours that may reach out of the volume Λ and with contour functional that (for such contours) depends on Λ. This is the reason why we formulate the following result on the description of labeled contour models under a stable boundary condition in a slightly more general situation than it is customary. Let thus a translation invariant functional Φ and a vector ϕ be as above, and suppose that sequences of finite sets Λn " V and functionals ΦΛn are given so that • ΦΛm (γ) = ΦΛn (γ) whenever m, n, and γ with support Γ are such that {γ} ∈ XΛn ∩ XΛm and dist(V (Γ), Λm Λn ) > 1, • 0 ≤ ΦΛn (γ) ≤ exp{−τ |Γ(γ)|} for all n and all γ.
234
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
We study partition functions Z(X(Λ) |q; ΦΛ , ϕ) and the probabilities µ(·, X(Λ) | q; ΦΛ , ϕ) for X(Λ) ⊂ X(Λ) ⊂ XΛ . Corollary 2.2.2 For τ sufficiently large and a phase q stable with respect to Φ, ϕ (as in Theorem 2), there exists, for a (possibly infinite) simply connected set V ⊂ Zν and any Y ⊂ XV , a unique probability measure µ(·, Y | q; ΦV , ϕ) defined as a limit of µ(·, Y ∩ XΛ | q; ΦΛ , ϕ) with finite Λ converging to V . For almost all x ∈ XV there are external q-contours — the set ϑ(x) is defined and consists of q-contours. Moreover, for every bounded continuous function f : XV → R, there exist (τ − c − cs )-functionals Ψq,V such that f (x)µ(dx, XV |q; ϑ(x) = ϑ(∂))µ(d∂, LV |Ψq,V ). f (x)µ(dx, XV |q; ΦV , ϕ) = (2.19) Here µ(dx, XV |q; ϑ(x) = ϑ(∂)) is naturally defined as the probability µ(dx, XV |q; ΦV , ϕ) under the condition ϑ(x) = ϑ(∂). The measures µ(d∂, LV |Ψq,V ) fulfill the assertions of Proposition 2.1.2. (a) (a) Let Ei (V ), i ∈ V , a ∈ R, be the set Ei (V ) = {x ∈ XV ; ∃γ ∈ x such that (a) V (Γ) ! i, |Γ| ≥ a}. There exists a constant C such that µ(Ei (V ), XV |q; ΦV , ϕ) ≤ Ce−τ a for any i ∈ V and a ∈ R. Proof. For any finite Λ we define Ψq,Λ by Ψq,Λ (Γ) =
q {γ}∈XΛ γ: supp γ=Γ
ΦΛ (γ)
Z(X(IntΛ Γ)|λ(γ); Φ, ϕ) . Z(X(IntΛ Γ)|q; Φ, ϕ)
(2.20)
By λ(γ) in the numerator we indicate that the partition function is considered with the boundary condition induced by λ(γ). Here IntΛ (Γ) is the union of those components of Int Γ whose distance from Zν \Λ is at least 1. In principle, also those components that are not fully contained in Λ are contributing, in a multiplicative way, to the numerator as well as denominator above. However, since for any γ in the sum above the label of the boundary of any such component is q, the concerned contributions to the numerator and denominator cancel. By induction, one can first verify that (α) (α) (Λ) ; Ψq,V ). Z(X(Λ) |q; ΦV , ϕ(α) ) = exp −ϕ(α) q |Λ| Z(L The functionals Ψq,Λ clearly satisfy the condition (about the independence on Λ for contours sufficiently far from the boundary) from Proposition 2.1.2 and the limiting functional Ψq,V as well as the measure µ(d∂, LV ; Ψq,V ) are well defined. To prove (2.19) we follow the proof of Proposition 3.4 from [HKZ]. The equality (2.19) holds for every finite Λ. Supposing that f is a cylindric function living on Λ, the right hand side of (2.19) can be approximated, up to a set of small measure, say E, by a cylindrical continuous function living on a large finite Λ(E).
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
235
The equality (2.19) for such a continuous bounded function then follows from Proposition 2.1.2 b). (a) The bound on probability of Ei (V ) is a standard implication of Proposition 2.1.2 a) applied to the contour functional Ψq,V .
2.3
Proof of Proposition 2.2.1
Since the assertion of Proposition 2.2.1 is an improvement of the results from [S] and [Z], we present here only the necessary changes or complements to the proofs from [Z]. In the same way as in Section 2.2, we often omit the superscript (α) if it cannot cause any misunderstanding. Sometimes we write Z(Λ|q) instead of Z(X(Λ)|q; Φ(a) , ϕ(α) ) and Z(Λ; Ψ) instead of Z(L(Λ); Ψ). 1. Proof of a) Recall that the equality (2.18) is fulfilled iff Ψ is defined by (2.20) (cf. proof of Corollary 2.2.2). The estimate a) follows easilyfrom the definition of stability of q in Int Γ and the bound exp (c# + log |S|) |Γ| on the number of γ’s in (2.20). ¯ explicitly as in [Z], 2. Implication b) ⇒ c) Our strategy is not to define Ψ ¯ that ensure, in particular, the conclusion c) of but to isolate those properties of Ψ ¯ (α) satisfying the assumpProposition 2.2.1. Then we look for a suitably smooth Ψ tions of the following lemma whose assertion (iv) is essentially identical to c). A ¯ (α) is given in Lemma 2.3.2. To satisfy the assumptions of the particular choice of Ψ following Lemma 2.3.1, one has to assume that τ − cs − c# − log |S| ≥ τ˜0 (cs , ν, |S|) with τ˜0 (cs , ν, |S|) determined in course of the proof of Lemma 2.3.1. Thus τ needed for validity of Proposition 2.2.1 c) can be taken as τ = τ˜0 (cs , ν, |S|) + cs + c# + log |S|. ˜ q are translation-invariant (non-negative) τ -functionals for q = 1, . . . , r, If Ψ ˜ r ), we define the following quantities: ˜ ˜ 1, . . . , Ψ and Ψ = (Ψ ˜ ϕ) = ϕq − p(Ψ ˜ q ), hq (Ψ, ˜ ϕ) = min hq (Ψ, ˜ ϕ), h(Ψ, q
˜ ϕ) = hq (Ψ, ˜ ϕ) − h(Ψ, ˜ ϕ). aq (Ψ, ˜ = {Ψ ˜ q } is a contour Lemma 2.3.1 There exists τ˜0 ≡ τ˜0 (cs , ν, |S|) such that if Ψ functional satisfying ˜ •Ψ ˜ q (Γ) ≤ Ψq (Γ), b) ˜ •Ψ q is cs -stable in Int Γ,
q (Γ) =
Ψq (Γ) whenever
˜
• Ψq (Γ) ≤ exp −˜ τ |Γ| and if τ˜ ≥ τ˜0 , then, denoting ε˜0 ≡ ε˜0 (˜ τ ) := ε(˜ τ ) + ε(˜ τ − c# − log |S| − 1) with ε(τ ) = e−(τ −ccl) , the following holds:
236
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
(i) If q is not cs -stable in Λ, then ˜ ϕ) |Λ| > (cs − 2˜ aq (Ψ, ε0 ) |∂Λ| .
˜ ϕ) |Λ| − ε˜0 |∂Λ| . (ii) Z(Λ|q; Φ, ϕ) ≥ exp −hq (Ψ,
˜ ϕ) |Λ| + ε˜0 |∂Λ| . (iii) Z(Λ|q; Φ, ϕ) ≤ exp −h(Ψ, ˜ ϕ) = 0. Whenever q is cs -stable, (iv) q is cs -stable iff aq (Ψ, ˜ ϕ) = − lim 1 log Z(Λ|q; Φ, ϕ) hq (Ψ, |Λ|
1 = min lim − log Z(Λ|λ; Φ, ϕ) (= h(Φ, ϕ)), λ |Λ| with the minimum taken over all multi-indices λ. Proof. Assuming (i)–(iii) for all proper subsets of Λ, one can prove (ii) and (iii) following the inductive proof of Theorem 1.7 in [Z]. To be more precise we give some commentary to it. The proof of (ii) remains τ ) = e−˜τ +ccl here). In the proof of (iii) we unchanged (it suffices to take ε˜0 ∼ ε(˜ 11 get the formula [Z, (1.44)] and we intend to apply [Z, Main Lemma, (2.13)] to the functional ε0 (˜ τ ) − h + ϕq ) |Γ| ≡ τ¯ |Γ| Ξ(Γ) = (˜ τ − c# − log |S| − 2˜ for a fixed q. Observe now that the equation τ ∗ = τ¯ − ε(τ ∗ ) has, for every τ¯ ≥ τcl , a solution τ ∗ such that ε(τ ∗ ) → 0 for τ¯ → ∞. For the auxiliary functional Ξ∗ (Γ) ≡ τ ∗ |Γ| := Ξ(Γ) − ε(τ ∗ ) |Γ| of [Z, (2.10)] we need τ ∗ ≥ τcl , i.e., ε0 (˜ τ ) − h + ϕq ≥ τcl + ε(τ ∗ ) τ¯ = τ˜ − c# − log |S| − 2˜ to be able to use the results of Section 2.1. Therefore we choose τ˜0 large enough to ensure that for some δ > 0 we have 2ε(˜ τ0 − c# − log |S| − 2δ) < δ, and
ϕq ≥ h(˜ τ0 ) − δ,
τ0 − c − 2δ), ε(τ ∗ ) < δ. τ˜0 ≥ τcl + c# + log |S| + 2δ + ε(˜
11 We only will have a factor 2˜ ε0 instead of 3ccl in (1.44). This can be traced back to the fact that, in the induction hypothesis used for (1.38), the role of 2ccl is played by ε˜0 .
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
237
Then, for τ˜ > τ˜0 , one has τ˜ − c# − log |S| − 2ε(˜ τ ) − h + ϕq ≥ τ˜ − c# − log |S| − 2δ ≥ τcl + ε(˜ τ − c# − log |S| − 2δ) ≥ τcl + ε(τ ∗ ) because τ ∗ ≥ τ˜ − c − 2δ. With a particular choice of δ, say δ = 1/3, the choice of τ˜0 depends on ν and |S| only. Hence, we bound the right hand side of [Z, (1.44)] by ˜ ϕ)) exp{−h |Λ| + ε(˜ τ ) |∂Λ|}Z q (Λ, Ξ, aq (Ψ, ˜ ϕ)) the partition function from [Z] Main Lemma, (2.12) (dewith Z q (Λ, Ξ, aq (Ψ, noted by Z(Λ, H) there) defined with the functional Ξ above and the parameter ˜ ϕ) defined before the present lemma. This yields [Z] (2.13) in the form aq (Ψ, ˜ ϕ)) ≤ eε(τ Z q (Λ, Ξ, aq (Ψ, or, equivalently,
Z(Λ; Ξ∗ ) ≤ ep(Ξ
∗
∗
)|∂Λ|
)|Λ|+ε(τ ∗ )|∂Λ|
,
.
The bound (iii) follows using [Z, Main Lemma] and the inequality ε(τ ∗ ) ≤ ε(˜ τ− c# − log |S| − 1). The only difference between [Z, Theorem 1] and our Lemma 2.3.1 (i) – (iii) concerns now the derivation of (i) and stems from our definition of cs -stable sets. Indeed, if q is not cs -stable in Λ, there exists a multi-index λ such that Z(Λ|λ) > exp cs |∂Λ| . Z(Λ|q) From (ii) and (iii) we have ˜ ϕ)|Λ| − ε˜0 |∂Λ| , and Z(Λ|˜ ˜ ϕ)|Λ| + ε˜0 |∂Λ| Z(Λ|q) ≥ exp −hq (Ψ, q) ≤ exp −h(Ψ, and thus (i) follows. ˜ ϕ) = 0. Then q is stable by (i) once τ˜ To prove (iv), suppose first that aq (Ψ, is so large that cs − 2˜ ε0 ≥ 0. ˜ q (Γ) for every Γ by Let, on the other side, q be cs -stable. Then Ψq (Γ) = Ψ ˜ Thus the condition b). ˜ q ) = exp ϕq |Λ| Z(Λ|q; Φ, ϕ) Z(Λ; Ψq ) = Z(Λ; Ψ and
˜ q ) = hq (Ψ, ˜ ϕ) = − lim log Z(Λ|q) . ϕq − p(Ψ (2.21) |Λ| ˜ ϕ) = 0, we know already that q0 is stable and Taking now any q0 such that aq0 (Ψ, thus Z(Λ|q0 ) ≤ exp cs |∂Λ| . exp(−cs |∂Λ|) ≤ Z(Λ|q) ˜ ϕ) = hq (Ψ, ˜ ϕ) and aq (Ψ, ˜ ϕ) = 0. Hence hq0 (Ψ,
238
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
Further, if q is cs -stable, we have Z(Λ|λ) ≤ exp cs |∂Λ| Z(Λ|q) for any λ. Thus, taking into account (2.21), we get log Z(Λ|λ) log Z(Λ|q) lim − ≥ lim − = h(Φ, ϕ). |Λ| |Λ| Hence
1 log Z(Λ|λ; Φ, ϕ) . h(Φ, ϕ) = min lim − λ |Λ|
˜ of Lemma 2.3.1. ˜ satisfying b) 3. The choice of Ψ Lemma 2.3.2 Let c˜ ≥ cs . The functionals Z(Int Γ|λ(γ)) Z(Int Γ|λ(γ)) exp c˜|∂ Int Γ| ˜ Ψq (Γ) = , Φ(γ) min , ˜ Z(Int Γ|q) maxλ˜ Z(Int Γ|λ) γ
(2.22)
with the sum taken over all labeled q-contours with Γ = Γ(γ) and with labeling ˜ satisfy λ(γ), and with the maximum in the second term taken over all labelings λ, ˜ the assumptions b) of Lemma 2.3.1 above with τ˜ = τ − c˜ − c# − log |S|. Proof. It follows clearly from (2.22) and the definition of stability.
Remark. 1. In particular, combining Lemma 2.3.2 with Lemma 2.3.1 (for τ˜ = τ − c˜ − c# − log |S| ≥ τ˜0 ), we can conclude that • there exists a cs -stable q0 for every α; • for any cs -stable q one has − lim
log Z(Λ|q) log Z(Λ|λ) = min −lim ; λ |Λ| |Λ|
• Z(Λ|q) ≤ exp −h(Φ, ϕ) |Λ| + ε˜0 |∂Λ| for every q; • Z(Λ|q) ≥ exp −h(Φ, ϕ) |Λ| − ε˜0 |∂Λ| for every cs -stable q. ˜ 2. Once the existence of a stable q0 is established, we notice that replacing λ ˜ q (Γ) satisfying assumptions of Lemma 2.3.1 by a constant q˜ in (2.22) leads to a Ψ with τ˜ = τ − 2cs − c# − log |S| (taking c˜ and thus τ˜ sufficiently large, we have E˜0 ≤ cs ). ˜ for Ψ, ¯ we proved b) and c) from Proposition 2.2.1. The task now 3. Taking Ψ ¯ so that it also is, assuming conditions (2) and (3) from Theorem 2, to choose Ψ satisfies b ) (cf. the step 5 below).
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
239
4. Estimates of derivatives of partition functions and proof of a ). Lemma 2.3.3 For sufficiently large τ ≡ τ (cs , M, ν, |S|), τ (α) > τ , and taking ε˜0 ≡ ε˜0 (τ (α) ) as in Lemma 1, one has12
+ (α)
∂ Z (Λ | λ) ≤ M |Λ|(1 + ε˜0 ) exp −h(α) |Λ| + ε˜0 |∂Λ| ||¯ α||, (2.23) α ¯ whenever Λ is a union of simply connected finite sets and λ is an arbitrary labeling. Moreover,
Z (α) (Λ | λ)
+
α||, (2.24)
∂α¯
≤ exp |∂Λ| ||¯ (α) Z (Λ | q) if q is cs -stable in Λ with respect to Φ(α) , ϕ(α) . Proof. The first inequality can be proven in a rather straightforward way, by induction in |Λ| (cf. also [BK] (Lemma 2.2)). Namely, considering the derivative of S (α) e−ϕq |Λ\ Int γ| Φ(α) (γ)Z (α) (Int γ | λ(γ)), Z (α) (Λ | q) = ϑ
γ∈ϑ
with the sum over collections ϑ of external q-contours in Λ (including the empty one), we get (taking, without loss of generality, ||¯ α|| = 1)
+ (α)
∂α¯ Z (Λ | q) ≤ ≤ M |Λ|Z (α) (Λ | q) +
∂ + Φ(α) (γ)Z (α) (Int γ | λ(γ)) e−ϕ(α) q |γ| Z (α) (Ext γ | q) ≤ α ¯ γ
≤ exp{−h
(α)
(α) 1 + M | Int γ|(1 + ε˜0 ) e−τ |γ|+3˜ε0 |γ| ≤ |Λ| + ε˜0 |∂Λ|} M |Λ| + γ
≤ M |Λ|(1 + ε˜0 ) exp{−h(α) |Λ| + ε˜0 |∂Λ|}. (α)
The first term corresponds to the derivative of ϕq using assumption (3) of Theorem 2. In the second term, we summed over contours not affected by the derivative, yielding Z (α) (Ext γ | q) with Ext γ = Λ \ (Int γ ∪ γ). To get the second inequality ˜ from Lemma 2) to bound Z (α) (Λ | q), we use Lemma 1 (iii) (with the help of Ψ (α) (α) Z (Int γ | λ(γ)), and Z (Ext γ | q), the assumption (2) of Theorem 2 to bound (α) (α) (γ)|, induction hypothesis (2.23) to bound |∂α+ (Int γ | λ(γ))|, and the |∂α+ ¯Φ ¯Z (α) (α) fact that ϕq ≥ h − ε˜0 . To get the last inequality, we use the bound
(α) 1 + M | Int γ|(1 + ε˜0 ) e−τ |γ|+3˜ε0 |γ| ≤ ε˜0 ,
γ:γ∪Int γi
where the sum is taken over all γ encircling a fixed site i. 12 We
explicitly indicate here the dependence on α.
240
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
To get (2.24), we use (2.23) and Lemma 1 (ii) and (iii) (with existence of ˜ assured by Lemma 2) and get needed Ψ
Z (α) (Λ | λ)
+
ε0 |∂Λ|
∂α¯ (α)
≤ M |Λ|(1 + ε˜0 ) exp 2˜ Z (Λ | q) ε0 |∂Λ| ≤ exp |∂Λ| +M |Λ|(1 + ε˜0 ) exp 4˜
for τ (α) sufficiently large (and thus ε˜0 sufficiently small).
Proof of a ) of Proposition 2.2.1. We differentiate the right hand side of (2.20). (α)
(α)
(X(IntΛ Γ)|λ(γ);Φ,ϕ) Evaluating the derivative of the product ΦΛ (γ) Z Z (α) , we use the (X(IntΛ Γ)|q;Φ,ϕ) assumption (2) of Theorem 2, the cs -stability of q in IntΛ Γ, assumption (1) of Theorem 2, and the bound (2.24) for the set IntΛ Γ, to get the bound (again, ||¯ α|| = 1)
Z (α) (X(IntΛ Γ)|λ(γ); Φ, ϕ)
+ (α)
∂α¯ ΦΛ (γ)
≤ Z (α) (X(IntΛ Γ)|q; Φ, ϕ)
≤ e−τ
(α)
|Γ| cs |Γ|
e
+ e−τ
(α)
|Γ| |Γ|
e
≤ e−(τ
(α)
−cs −1−ln 2)|Γ|
that yields (a ) with c = cs + 1 + ln 2 + c# + log |S|.
5. Definition of Ψ satisfying b), and simultaneously b’), of Proposition 2.2.1 ˜ q from (2.22) is the use of According to 3. and 4., the only problem with Ψ ˜ non-smooth. Therefore we introduce a the min in its definition which makes Ψ “smooth version of min” first and then we apply it to define Ψq by a natural modification of (2.22). Lemma 2.3.4 For any η > 0, there is a function minη : Rr → R, r ∈ N, such that (i) minη (u) ≤ min(u) for u = (u1 , . . . , ur ) ∈ Rr ; (ii) minη (u) = ui whenever ui ≤ min{uj | j = i} − η; (iii) minη ∈ C ∞ (Rr ) and it is 1-Lipschitz; (iv)
∂ ∂ui
minη (u) = 0 if ui ≥ min(u) + 2η.
Proof. We may define minη as the convolution of min and a nonnegative function ϕη ∈ C ∞ (Rr ) that is symmetric (i.e. of the form ψη (u)), fulfills ϕη (u)du = 1 Rr
and ϕη (u) = 0 if u ≥ η2 . Namely, we put minη = ϕη ∗ min. Then (i) holds because ϕη ∗ min(u) = ϕη (u − v) min(v)dv ≤ ϕη (u − v)vi dv = ϕη (v)(ui − vi )dv = Rr
Rr
Rr
ϕη (v)dv −
= ui Rr
The last integral is zero because of the symmetry of ϕη .
ϕη (v)vi dv = ui . Rr
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
To verify (ii), notice that ϕη (v) min(u − v)dv = ϕη ∗ min(u) = v≤ η2
and
241
ϕη (v)(ui − vi )dv = ui . v≤ η2
The condition (iii) follows immediately from the fact that min is 1-Lipschitz ϕη (v)dv = 1 with ϕη (v) ≥ 0.
Rr
Finally, we verify (iv). Let |˜ ui − ui | < η and u ˜j = uj for j = i. Then minη (˜ u) = ϕη (v) min(˜ u − v)dv = ϕη (v) min(u − v)dv = minη (u). Here v< η2
v< η2
˜i − vi ≥ min(u) + we used that u˜i ≥ min(u) + η and u holds and Lemma 4 is proved.
η 2
≥ min(u − v). So (iv)
Proof of b) and b ) of Proposition 2.2.1. For η > 0 and any contour γ, we introduce the shorthand
Z(Λ|λ(γ)) exp (cs + η)|∂Λ| Z(Λ|λ(γ)) , log ; q˜ = q , M(γ, η) = minη|∂Λ| log Z(Λ|q) Z(Λ|˜ q) where Λ = Int Γ; q, q˜ ∈ {1, . . . , r}. Put now (α) Ψq (Γ) = Φ(α) (γ) exp M(γ, η) ,
(2.25)
γ
where the sum is over q-contours with support Γ. (α) First we show that the functional Ψq fulfills the three assumptions of ˜b) of Lemma 1 (with τ˜ = τ − c , c = c# + log |S| + cs + ε˜0 + η + 1) and thus b) of Proposition 2.2.1 (with c = c# + log |S| + 2cs + 1) . (α)
(α)
Namely, the first point (Ψq ≤ Ψq ) follows from (2.20), (2.25) and the property (i) of minη (cf. Lemma 4). (α)
(α)
The second point (Ψq (Γ) = Ψq (Γ) if q is cs -stable in Λ) needs moreover the definition of cs -stability of q in Λ and the property (ii) of minη (cf. Lemma 4). (α) To show the third point (|Ψq (Γ)| ≤ exp −(τ (α) − c )|Γ| ), we use the fact (α) that, by Lemma 2, there is a cs -stable q0 . So |Ψq (Γ)| ≤ exp −(τ (α) − c# − (α) log |S| − cs − ε˜0 − η))|Γ| by the definition of Ψq , the assumption (1) of Theorem 2, Lemma 4 (i), and the definition of cs -stability of q in Λ. Using the choice η = 1, we get b) with c = c# + log |S| + 2cs + 1. (α)
Now, we shall show that Ψq fulfills the estimate (b ). First, notice that the following auxiliary estimate holds due to Lemma 4 (i),
∂
exp minη (log t1 , · · · , log tr )
∂ti
∂ 1
≤ exp minη (log t1 , · · · , log tr ) · minη (u) · ≤ 1 ∂ui ti
242
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
for t1 , . . . , tr > 0. We differentiate ∂α+ ¯
Φ(α) (γ) exp M(γ, η)
γ
=
γ
(α) (α) ∂α+ (γ) exp M(γ, η) + Φ (γ)∂α+ ¯Φ ¯ exp M(γ, η) . (2.26) γ
(α) We are going to use (2) of Theorem 2 to estimate ∂α+ (γ) and (1) of Theorem ¯Φ (α) 2 to estimate Φ (γ). Using the existence of a cs -stable q0 we get further that
exp M(γ, η) ≤
Z(Λ|λ(γ)) exp (cs + η)|∂Λ| Z(Λ|λ(γ)) , log ; q˜ = q ≤ ≤ exp min log Z(Λ|q) Z(Λ|˜ q) ≤ exp (2cs + η)|∂Λ| . To establish the estimate (b ) it remains to estimate the derivative
+
∂α¯ exp M(γ, η) ≤ exp M(γ, η) ∂α+ ¯ M(γ, η) ≤ ∂ ∂ minη|∂Λ| (log t1 , . . . , log tr )ti =tqi (α) · + (tqj (α)) , ≤ exp M(γ, η) ∂tj ∂α¯ j
Z(Λ|λ(γ)) Z(Λ|q)
Z(Λ|λ(γ)) exp (cs +η)|∂Λ| Z(Λ|˜ q)
and tq˜(α) = where we use the notation tq (α) = for q˜ = q. We consider the product of the two terms of the last bound. The first one is bounded by exp (2cs + η)|Γ| as we already noticed. The otherone may be reduced to the sum over j’s such that tqj (α) < min(t) · exp 2η|∂Λ| by Lemma 4 (iv). Now we may use our auxiliary estimate of the partial derivatives over tj ’s. Since we confined ourselves to those special j’s, we may use Lemma 3 to estimate the partial tq (α) and tq˜(α) derivatives of for those q and q˜ for which tq (α) ≤ exp cs |Γ| and tq˜(α) ≤ exp (2cs + η)|Γ| . Inserting the above bounds into (2.26), we finally get
(α)
+ (α) exp (2cs + η)|Γ|
∂α¯ Ψq (Γ) ≤ ec# |Γ| e−τ +|S| exp (2cs + η)|Γ| exp (2cs + η)|Γ| ||¯ α||. Hence, taking again η = 1, we get (b ) with c = c# + 4cs + 2 + 2 log(|S| + 1) .
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
243
3 Periodic Gibbs states In this short section we recall how the standard Pirogov-Sinai theory leads to a description of Gibbs states of classical lattice models in terms of contour models. Indeed, after reformulating a model with Hamiltonian H in terms of a labeled contour model, we can apply results of Section 2.2. = Recall that we suppose that all periodic configurations from Gper 0 {x1 , . . . , xr } are actually translation invariant. This does not mean a loss of generality as explained in the remark following Basic Lemma 1.2. Considering, in the standard way (in the present context, cf. [[HKZ, Lemma 3.1]), the “diluted” partition functions (H) Z d (Λ|xq ; H) = exp{−EΛ (x)}, x=xq in (Λc )R (H)
with (Λc )R = {i ∈ Zν ; dist(i, Λc ) ≤ R+1}, we introduce ϕ(H) = (ϕq ) and Φ(H) in such a way that we can replace the diluted sums Z d (Λ|xq ; H) by the corresponding labeled contour model partition functions Z cont (X(Λ)|q; Φ(H) , ϕ(H) ) discussed in ˙ Section 2.2 (cf(2.15)); we use here the superscript “cont” to stress that we are concerned with partition functions (resp., probability distributions) of a labeled contour model. Namely, let us introduce the labeled contour model (with labels Q = = {x1 , x2 , . . . , xr }) with {1, . . . , r}, where we denote Gper 0
and Φ(H) (γ) =
ϕ(H) = exq (H) q
(3.1)
(H) (H) exp −EΓ (x) + EΓ (xq ) ,
(3.2)
x∼γ
whenever γ is a q-contour. The sum is over configurations on Γ that can be ex(H) tended to a configuration having γ as its contour (for definitions of EA (x) and ex (H) see (1.2) and (1.3)). Then, one can easily verify that Z d (Λ|xq ; H) = Z cont (X(Λ)|q; Φ(H) , ϕ(H) )
(3.3)
for every finite Λ. Moreover, the state µ({x}, Λ|xq ; H) can be linked with the labeled contour model probability distribution (2.16), µ({x}, Λ|xq ; H) = µcont ({z}, X(Λ) | q; Φ(H) , ϕ(H) ). x∈XΛ x∼z
Here the sum is over all spin configurations x ∈ XΛ such that the set of their labeled contours is just z (it means that one is summing only over configurations on γ∈z supp Γ with boundaries fixed by labeling of contours γ ∈ z).
244
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
Next, we verify that the assumptions of Theorem 2 for the labeled model described by Φ(H) and ϕ(H) are satisfied. Recall that the Hamiltonian H0 satisfies the Peierls condition (P(per) ) with respect to the set G0 and constant ρ0 . We can choose ε so that each H ∈ Kε (H0 ) satisfies, according to Proposition 1.1.3, the Peierls H , where limε→0 ρε = ρ0 , condition (P(per) ) with a sufficiently large constant ρε H 0 if ρ0 is large enough. Now, we consider the open set Kε (H0 ) ⊂ H(R) to play the role of the set Ω from Theorem 2 (thence the Hamiltonians H stand for the parameters α of Theorem 2). The functional Φ(H) (γ) defined by (3.2) satisfies, for any q-contour γ, the condition (1) of Theorem 2. Indeed, using (P(per) ) and the estimate |S||Γ| on the H
−(ρ
−log |S|)|Γ|
cardinality of the set {x; x ∼ γ} we get Φ(H) (γ) ≤ e ε H0 . To verify the condition (2), we consider the derivative
+ (H) (H) + (H) (γ)| = ∂H exp − (Ei (x) − Ei (xq ) . |∂H ¯Φ ¯ x∼γ
(3.4)
i∈Γ
Referring to (1.2), notice that ∂UA ∂(zA ) ( 0
0
i∈A
UA (xA ) |A| )
≤
i∈A,[A]=[A0 ]
1 |A|
≤ 1,
whenever A0 is such that the canonically fixed site i(A0 ) equals 0 (cf. the discussion at the beginning of Section 1.1) and a configuration zA0 ∈ S A0 is fixed. Therefore, taking again into account that the cardinality of {x; x ∼ γ} is at most |S||Γ| and (H) (H) −ρ H |Γ| ¯ = (U ¯A0 (zA0 )|i(A0 ) = 0), the that e−(EΓ (x)−EΓ (xq )) ≤ e ε H0 , we get, for H bound + (H) (γ)| ≤ e|Γ| log |S| e |∂H ¯Φ
H |Γ| 0
−ρε H
(H)
+ (|∂H ¯ Ei
(H)
+ (x)| + |∂H ¯ Ei
(xq )|) =
i∈Γ
=e
U (x ) ∂
A A U A0 (zA0 ) + ∂UA0 (zA0 ) |A| i∈Γ A0 ,zA0 i∈A
∂ UA ((xq )A )
+ U A0 (zA0 ) ≤ ∂UA0 (zA0 ) |A|
H |Γ|(log |S|−ρε H ) 0
×
i∈A
≤e
H +log |S|) 0
|Γ|(−ρε H
≤e
¯ |Γ|2 dim H(R)H
H |Γ|(−ρε H +log |S|+1+log 2+log dim H(R)) 0
¯ H.
Hence, (H)
+ (H) |∂H (γ)| ≤ e−τΦ ¯Φ
with (H)
τΦ
= ρε
|Γ|
¯ H,
H − log |S| − 1 − log 2 − log dim H(R). H0
(3.5) (3.6)
Thus Φ(H) satisfies the conditions (1) and (2) from Theorem 2 and Proposition (H) (H) 2.2.1 with the constant τΦ playing the role of τ (α) . Further, ϕq , in the place
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
245
(α)
of ϕq , satisfies the condition (3) from Theorem 2 with M = 1. Recalling the notation ε˜0 , ccl , and c (cf. Lemma 2.3.1, Proposition 2.1.3, and Theorem 2), and applying Theorem 2, the equality (3.3), and Lemma 2.3.1, we obtain Proposition 3.1 Under assumptions of Theorem 1: a) For every H ∈ Kε (H0 ) as above, defining Φ(H) and ϕ(H) by (3.1) and (3.2), we obtain a labeled contour model that satisfies (1) – (3) with the respective (H) constants τ (H) = τΦ and M = 1. Moreover, Z d (Λ|xq ; H) = Z cont (X(Λ)|q; Φ(H) , ϕ(H) ) for every finite Λ. b) For every x ∈ Gper 0 , there exists a function hx (·) on Kε (H0 ) such that there exists an extremal Gibbs state µ ∈ G(H) that is a perturbation of x whenever hx˜ (H). Moreover, hx (H) = h(H) = minx˜∈Gper 0 (H)
−τh |hx (H) − ϕ(H) , x |≤e (H)
+ (H) −τh |∂H , ¯ (hx (H) − ϕx )| ≤ e
(3.7) (3.8)
and, finally, (H)
+ −τh ¯ )H |∂H ¯ h(H)| ≤ (1 + e
for every x ∈
Gper 0
and any H ∈ Kε (H0 ). Here, and in what follows, (H)
τh Further, for every x ∈
(H)
= τΦ
− ccl − c .
Z d (Λ|x; H) ≥ exp −hx |Λ| − ε˜0 |∂Λ| Gper 0 .
(3.9)
(3.10) (3.11)
If xq is stable, then
Z d (Λ|xq ; H) ≤ exp −h|Λ| + ε˜0 |∂Λ|
(H)
and there exists a (τΦ
(3.12)
− c )-functional Ψq such that
Z d (Λ|xq ; H) = e−exq (H)|Λ| Z(L(Λ); Ψq )
(3.13)
for every finite Λ.
4 Gibbs states with interfaces The aim of the present section is to start from the situation of Theorem 1 and to rewrite the partition functions Z(Y, Λ|y; H), y ∈ Ghor 0 , in terms of labeled contour models that can be treated by the methods of Section 2.2.1. In fact, we are doing this only for a class of sets Y of configurations with the aim to estimate the probability of a particular interface. Throughout this section we suppose that the assumptions of Theorem 1 (i.e. also of Basic Lemma) are satisfied. In particular,
246
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
ρ0 is assumed to be sufficiently large, with the exact bounds specified in the course of the exposition in the present section. We shall proceed in several steps that culminate in Proposition 4.4 below that yields an expression of probabilities of some sets of configurations from X in terms of corresponding interfaces which, in their turn, are related to certain labeled contour model. The probability measure that we have in mind is constructed, following Dobrushin [D 72], by a suitable limit starting from Gibbs states in finite volumes with boundary conditions y ∈ Ghor 0 .
4.1
Interfaces in finite volumes
To consider Gibbs states in Basic Lemma for non-periodic elements of G0 we fix a particular H ∈ Kε (H0 ), two different configurations from Gper 0 , say xp , xq with p, q ∈ {1, . . . , r}, and a configuration y ∈ G0 ∩ Xxhor to play the role of boundary p ,xq conditions. The Hamiltonian H being fixed, we shall often omit a reference to it from the notation. Following [HKZ] we say that a configuration x has a y-interface if its boundary B(x) has a unique infinite component I(x), I(x) \ I(y) has only finite components, and Zν \ I(x) has exactly two infinite R-components. We say that the sites in one of them are lying above and those in the other one below the interface. Further, we use I(x) to denote the pair (I(x), xI(x) ) and say that I is a y-interface if there exists a configuration x that has a y-interface I = I(x). We begin with a study of the partition function Z(Λ|y; H) in a fixed finite ν volume Λ ⊂ Zν . If x is any configuration (i.e. x ∈ S Z , ν ≥ 2) that equals y in Λc , then x has an interface ([HKZ], Lemma 4.1). We use J (y, Λ) to denote the set of interfaces of configurations x considered above. Our aim is to study the probability of interfaces and thus we begin with rewriting, for a fixed I ∈ J (y, Λ), the partition function Z(I, Λ|y; H) defined as Z(I, Λ|y; H) = Z({x : I(x) = I}, Λ|y; H) in five steps. (H)
Step 1 (sum over interfaces) Notice first that the energy EI∩Λ (x) does not depend (H) on xI c and will be denoted by EI∩Λ (I). The volume ΛR (= {i ∈ Zν : d(i, Λ) ≤ R + 1}) is, by means of an interface I ∈ J (y, Λ) with the support I, split up into several parts: I ∩ ΛR , the R-components Λm (I), m ∈ {p, q}, of ΛR \ I containing the sites lying above or below I, respectively, and, finally, the remaining finitely many R-components of ΛR \ I denoted by Intj I, j = 1, 2, . . . . The configuration x equals to some xq(j) ∈ Gper = {x1 , x2 , . . . , xr }, in IR ∩ Intj I if I(x) = I. 0 We recall that, given the Hamiltonian H, the “physical” partition function in Λ under the boundary condition z ∈ X is expressed by exp{−HΛ (x|z)}, Z(Λ|z; H) = where HΛ (x|z) =
A∩Λ=∅
x=z in Λc
UA (xA ) with x = z in Λc .
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
247
Using now directly concerned definitions, we get Z(I, Λ|y; H), Z(Λ|y; H) = I∈J (y,Λ)
where |A ∩ ΛR | = UA (y) Z(I, Λ|y; H) exp − |A| A⊂Λc (H) = exp{−EI∩ΛR (I)} Z d (Λm (I)|xm ; H) Z d (Intj I|xq(j) ; H). (4.1) j
m∈{p,q}
The equalities above correspond to Lemma 4.2 from [HKZ] that differs only by the usage of the relative diluted partition function Θ instead of Z d . Step 2 (cluster expansion) From now on we suppose that xp and xq are stable with respect to the considered Hamiltonian H and so we may apply the PirogovSinai theory as formulated in Section 2. We express the diluted partition functions (H) (H) with boundary conditions xp , xq in terms of contour functionals Ψp and Ψq (H)T and thus, referring to (3.13) and (2.10), in terms of cluster functionals Ψp and (H)T Ψq . Namely, |A ∩ ΛR | d Z(I, Λ|y; H) exp − UA (y) Z (Intj I|xq(j) ; H)× = |A| j A⊂Λc (H) Ψ(H)T (C) − exm (H) |Λm (I)| . (4.2) × exp −EI∩ΛR (I) + m m∈{p,q} C⊂Λm (I) (H)
(H)
Here, the contour functionals Ψp and Ψq satisfy, according to Proposition (H) 2.2.1, the bounds a) and a ) with the sufficiently large constant τΦ − c . Further, (H)T (H)T and Ψq satisfy, in view of Proposition 2.1.3 (if the cluster functionals Ψp (H) (H) τΦ − c ≥ τcl ), the bounds (2.7) and (2.9) with the constant τh defined in Proposition 3.1 above. Step 3 (extraction of bulk terms) We are going to extract a bulk term independently of I using the fact that (see (S) in Theorem 2, Proposition 2.2.1 c), (2.14) of Corollary 2.1.4 and (3.1)) ϕ(H) − p
Ψ(H)T (C) p Ci
|C|
= ϕ(H) − q
Ψ(H)T (C) q Ci
|C|
.
Denoting this expression by h(H) (notice that, since the phases xp and xq are stable, one has h(H) = − limν Λ Z
1 1 log Z d (Λ|xq¯; H) = − limν log Z(Λ|z; H) Λ Z |Λ| |Λ|
248
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
for any q¯ ∈ {1, . . . , r} and z ∈ X (the first equality, for stable q¯, follows by (3.13), (3.1), (2.10) and Proposition 2.2.1 c); the other one is standard), we get |A ∩ ΛR | = Z(I, Λ|y; H) exp h(H) |ΛR |} exp{− UA (y) |A| A⊂Λc
= exp − EI∩ΛR (I) − h(H) |I ∩ ΛR | log Z d (Intj I|xq(j) ; H) + h(H) |Intj I| × exp j
× exp −
Ψ(H)T (C) m
m∈{p,q} CΛm (I)
|C ∩ Λm (I)| . (4.3) |C|
Step 4 (extraction of surface terms) The next step is to extract a surface term that does not depend on I. If C ⊂ Zν and m ∈ {p, q}, we write C ∗ m whenever there exist i, j ∈ C such that |i − j| = 1, i ∈ ∂ΛR ∩ ∂Λm (I(y)), j ∈ (ΛR )c . We put Λ χΛ m (C) = 1 if C ∗ m and χm (C) = 0 otherwise. Using this notation, we have |A ∩ ΛR | |C ∩ ΛR | + Z(I,Λ|y;H)exp h(H)|ΛR |− UA (y) Ψ(H)T (C) m |A| |C| A⊂Λc m∈{p,q} C∗m log Z d (Intj I|xq(j) ; H)+h(H) |Intj I| = exp − EI∩ΛR (I)−h(H) |I ∩ ΛR | + × exp
m∈{p,q} C∩I=∅
j
|C ∩ ΛR | |C ∩ Λm (I)| Ψ(H)T (C) χΛ − . (4.4) m m (C) |C| |C|
The above equality corresponds to Lemma 4.3 in [HKZ]. Notice that the factor on the left-hand side of (4.4) does not depend on I. |C∩ΛR | |C∩Λm (I)| Step 5 (positivity of cluster terms) The terms Ψm (C)(χΛ ) m (C) |C| − |C| are not necessarily positive, a feature that would be useful in further application of our version of the Pirogov-Sinai theory. However, the terms depending on C may be turned into explicitly positive by adding a suitable sum in the exponent and absorbing it into a small change of weight of interfaces in the same time. For a reasonable choice of the added sum (to secure the positivity of cluster terms and to (H) (H) allow the proof of Lemma 4.3 below) we shall use the bounds |Ψm (Γ)| ≤ e−τh |Γ| (H)
(H)T
(H) and Ψm (C) ≤ e−τh |C| with τh from Proposition 3.1 (see also Step 2). Hence (H)T
|A ∩ ΛR | UA (y) Z(I, Λ|y; H) exp h(H) |ΛR | − |A| A⊂Λc (H) |C ∩ ΛR | −τh |C| = + 2 × exp Ψ(H)T (C) e m |C| C∗p m∈{p,q} C∗m
C∗q
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
249
(H) = exp − EI∩ΛR (I)−h(H) |I ∩ ΛR | + log Z d (Intj I|xq(j) ; H)+h(H) |Intj I| j
|C ∩ ΛR | |C ∩ Λm (I)| Ψ(H)T − ) (C)(χΛ m m (C) |C| |C| m∈{p,q} C∩I=∅ (H) Λ × exp 2 e−τh |C| (|C ∩ I ∩ ΛR | + χΛ (C) · χ (C)) p q
× exp
C∩I=∅
(H) e−τh |C| |C ∩ I ∩ ΛR | . (4.5) × exp −2 C∩I=∅
We added the terms 2
e
(H) −τh |C|
and added and subtracted 2
(H)
e−τh
|C|
C∩I=∅
C∗q C∗p
(H) κ(τh ) |I
|C ∩ I ∩ ΛR |. The idea is that the latter equals (H) (H) e−τh |C| κ(τh ) = 2
∩ ΛR | with (4.6)
C:0∈C
and may be absorbed into a controllable change of energy, while the former does not depend on I and may be extracted as a “border of surface term”. In the same |C∩ΛR | m (I)| time, whenever the term χΛ − |C∩Λ is non-vanishing (in any case, m (C) |C| |C| Λ its absolute value is bounded by 1), the term |C ∩ I ∩ ΛR | + χΛ p (C) · χq (C) is at Λ least 1. Indeed, if χm (C) = 0 and in the same time C ∩Λm (I) = ∅, then necessarily C ∩ I ∩ ΛR = ∅ (we took into account that C ∩ I = ∅). If, say, χΛ p (C) = 1 and in Λ the same time χq (C) = 0 and C ∩ ΛR = C ∩ Λp (I), then C ∩ I ∩ ΛR = ∅. Finally, Λ the claim is trivial if χΛ p (C) = χq (C) = 1. As a result, |C ∩ ΛR | |C ∩ Λm (I)| Λ − + (C) χ (C) Ψ(H)T m m |C| |C| (H) Λ + e−τh |C| |C ∩ I ∩ ΛR | + χΛ (C) · χ (C) ≥ 0 (4.7) p q
(H)T
(H) since Ψm (C) ≤ e−τh |C| . ˜ Λ|y, H) to denote the left-hand side of (4.5), we stress that the Using Z(I, ˜ ratio Z/Z does not depend on I and thus the probabilities of interfaces defined by Z˜ or Z do not differ. To evaluate those probabilities, we recall the notion of walls in order to rewrite (4.5) in terms of them. A pair w = (W, xW ), where W is a connected component of I(x) \ C(x), is a wall of I(x). We denote by W(I(x)) the collection of all walls of I(x). For any w ∈ W(I(x)) we put (H)
E (H) (w) = EW (x).
(4.8)
Let us recall also that given a wall w, we use xw ∈ X to denote the configuration for which I(xw ) = B(xw ) and w is the only wall of I(xw ), and yw to denote
250
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
the unique element of G0 which differs from xw in at most finite number of sites. Recall that the set G0 ∩ Xxhor consists of configurations that are vertical shifts of p ,xq finitely many “representatives” xp,q;1 , . . . , xp,q;np,q . Let s(w) be such that xp,q;s(w) is a vertical shift of yw . We use π to denote the projection of Zν onto Zν−1 defined by π(i1 , . . . , in−1 , in ) = (i1 , . . . , in−1 ), and recall that IW = I(yw ) ∩ π −1 (W ). Notice further that each set Intj I is surrounded by the support W of a single wall w of I and use j ◦ W to denote that this is the case. Let Ip,q;s be the support of the interface of xp,q;s , s = 1, 2, . . . , np,q and Bp,q;s (I) be the set {i ∈ BR = π(ΛR ) : π −1 (i) ∩ I is a vertical shift of the “xp,q;s -ceiling column” }∪ {π(W ); s(w) = s} for I ∈ J (y, Λ). Let Tp,q;s be the number of sites in π −1 (0) ∩ Ip,q;s , i.e. the “thickness” of Ip,q;s , and T be the maximum of all Tp,q;s ’s over p, q ∈ {1, . . . , r}, s ∈ (H) (H) {1, . . . , np,q }. Let us recall that 2 e−τh |C| |C ∩ I ∩ ΛR | = κ(τh )|I ∩ ΛR |. C∩I=∅
˜ (H) (w) to denote the “modified energy of the wall w”, Further, we use E ˜ (H) (w) = E (H) (w) − h(H) |W | + κ(τ (H) ) |W | − E h − (log Z d (Intj I|xq(j) ; H) + h(H) |Intj I|), (4.9) j◦W (H)
and ϕp,q;s to denote the “modified specific energy of an s-ceiling between xp above and xq below”, (H) (H) ) − h(H) Tp,q;s . ϕ(H) p,q;s = Eπ −1 (0)∩Ip,q;s (xp,q;s ) + κ(τh
(4.10)
(H)
We also use ΦΛ (w; p, q) to denote the “modified weight corresponding to a wall w of an interface separating xp and xq ”, the wall functional (H) (H) ˜ (w) − ϕ(H) ΦΛ (w; p, q) = exp − E p,q;s(w) |π(W )|
(4.11)
and, finally, we define the cluster functional
|C ∩ ΛR | |C ∩ Λm (I)| (H) ΦΛ,I (C; p, q) = exp − (C) χΛ Ψ(H)T + m m (C) |C| |C| m∈{p,q} (H) Λ (C)χ (C) − 1. (4.12) + e−τh |C| |C ∩ I ∩ ΛR | + χΛ p q
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
251
Using this notation, we rewrite (4.5) as ˜ Λ|y; H) = Z(I,
np,q (H) ΦΛ (w; p, q)
w∈W(I)
s=1
×
exp{−ϕ(H) p,q;s |Bp,q;s (I)|}×
(H) 1 + ΦΛ,I (C; p, q) , (4.13)
C:C∩I=∅,C∩BR =∅
and notice that by (4.7) we have (H)
ΦΛ,I (C; p, q) ≥ 0.
4.2
(4.14)
Interfaces in cylinders with finite base
˜ ˜ |y; H) = Our next aim is to show that Z(V I∈J (y;V ) Z(I, V |y; H) can be defined −1 for an infinite cylinder V = π (B) with a finite base B(= π(V )) ⊂ Zν−1 and to ˜ V |y; H). study Z(I, Step 6 (a wall bound) The main aim of this step is to prove that, for any finite volume Λ ⊂ Zν and any interface I of a configuration which equals y in Λc , ˜ Λ|y; H) ≤ ecI |BR |(H+1) exp −(ρε H − cI ) Z(I, H0
|W | ,
(4.15)
w∈W(I)
for some “interface” constant cI , once ρε is sufficiently large. (H) To estimate the wall functional ΦΛ (w; p, q) defined by (4.11), we bound first (cf. (4.9) and (4.10)) (H)
(H)
E (H) (w) − h(H) |W | + κ(τh ) |W | − ϕp,q;s(w) |π(W )| = (H) (H) = EW (xw ) − e0 (H)|W | − EIW (yw ) − e0 (H)|IW | + (H)
+ (e0 (H) − h(H))(|W | − |IW |) + κ(τh
)(|W | − |IW |).
(4.16)
H First of all we use Peierls condition (Phor ) with the constant ρε H (see 0 Proposition 1.1.3) to the first parenthesised part on the right-hand side and we get
(H) H (H) |W |. EW (xw ) − e0 (H)|W | − EIW (yw ) − e0 (H)|IW | > ρε H0
(4.17)
By (3.7) we get
(H)
(e0 (H) − h(H))(|W | − |IW |) ≤ e−τh |W |(1 + T ).
(4.18)
252
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
(H)
We estimate the artificial term κ(τh ), used to achieve the positivity of using its definition (4.6) and the bound (2.4), by
(H) ΦΛ,I (C; p, q),
(H)
0 ≤ κ(τh
)≤2
∞
(H)
e−(τh
−c# )k
(H)
≤ e−(τh
−c# −log 3)
(4.19)
k=1 (H)
once τh
is sufficiently large. Hence (H)
κ(τh
(H)
)(|W | − |IW |) ≥ −e−(τh
−c# −log 3)
|W |T.
(4.20)
Further, due to (3.12) we have the bound (H) ε˜0 (τh )|∂ Intj I| ≤ log Z d (Intj I|xq(j) ; H) + h(H)| Intj I| ≤ j◦W
j◦W
≤
(H) 2ν ε˜0 (τh )|W |
(H)
≤ e−(τh
−cZ )
|W | (4.21)
for a constant cZ . Here we used that (H)
) ≤ 2e−(τh
(H)
≤ e−(τh −cε ) , for a suitable constant cε , by Lemma 2.3.1. Here j◦W means any sum over a subset of {j; j ◦ W }. Substituting the just derived inequalities (4.17), (4.18), (4.20) into (4.16), we get (H)
ε˜0 (τh
(H)
E (H) (w) − h(H) |W | + κ(τh ρε
−ccl −c# −log |S|−1)
(H)
) |W | − ϕp,q;s(w) |π(W )| ≥
(H) (H) H |W | − e−τh |W |(1 + T ) − e−(τh −c# −log 3) |W |T. (4.22) H0
Using further (4.21) and having in mind (4.11) with (4.9) and (4.10), we conclude that (H)
ΦΛ (w; p, q) ≤ (H) (H) (H) H |W |+e−τh |W |(1+T )+e−(τh −c# −log 3) |W |T +e−(τh −cZ ) |W | ≤ exp −ρε H0 H − cw )|W | , (4.23) ≤ exp −(ρε H0 where cw is a positive constant which can be chosen arbitrarily small if taking (H) τh , i.e. ρε , sufficiently large in the same time. Now we estimate the contribution of the specific ceiling energies to the ˜ Λ|y; H). Due to (4.10), (4.19), and the inequality |h(H)| ≤ logarithm of Z(I, (H) −τh H + e following from (3.7), we get (H)
−(τh |ϕ(H) p,q;s | ≤ t(H + e
−c# −log 3)
(H)
+ H + e−τh ) ≤ cϕ (H + 1),
(4.24)
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
253
for some real constant cϕ . It follows that p,q
ϕ(H) p,q;s |Bp,q;s (I)| ≤
n
s=1
max
s=1,...,np,q
(H)
ϕp,q;s π(ΛR ) ≤ cϕ ( H + 1) BR . (4.25) (H)
Finally, we get bounds for the nonnegative cluster functionals ΦΛ,I (C; p, q). (H)
Using (4.14) from Step 5, (4.12), the fact that |Ψm (C)| ≤ e−τh |C| , and the inequalities 0 ≤ ex − 1 ≤ 2x for x nonnegative and small enough, we get (H)T
(H)
0 ≤ ΦΛ,I (C; p, q) ≤ 4(|C| + 2)e−τh (H)
|C|
(H)
≤ e−(τh
−cC )|C|
,
(4.26)
where cC is a real constant. So we can estimate the third product from (4.13) by
(H) 1 + ΦΛ,I (C; p, q) ≤
C:C∩I=∅,C∩BR =∅
(H) 1 + e−(τh −cC )|C| ≤
i∈I∩ΛR C:dist(C,i)≤|C|
≤ exp cπ (|I ∩ ΛR |) ≤ exp cπ (|BR | +
|W |)
(4.27)
w∈W(I)
for some constant cπ . Here, we used (H) 1 + e−(τh −cC )|C| ≤ exp C:dist(C,i)≤|C|
(H)
e−(τh
−cC )|C|
C:dist(C,i)≤|C|
(H) ≤ exp |C|ν e−(τh −cC )|C| . (4.28) C:i∈C
Applying the estimates (4.23), (4.25), and (4.27) to (4.13), we get ˜ Λ|y; H) ≤ Z(I, H ≤ exp{− − cw )|W |} exp{cϕ (H + 1)|BR |} exp{cπ |BR | (ρε H0 w∈W(I)
+
w∈W(I)
H − cI ) cπ |W |} ≤ ecI |BR |(H+1) exp −(ρε H0
|W |
w∈W(I)
with a suitable constant cI , getting thus (4.15). Step 7 (partition functions in infinite cylinder sets) Let us use χV to denote, for V = π −1 (B), the analogue of χΛ . Further, we introduce the analogs of the wall and (H) (H) cluster functionals ΦΛ (w; p, q) and ΦΛ,I (C; p, q), defined by (4.11) and (4.12). We denote the corresponding functionals, obtained by substituting V for Λ, by (H) (H) ΦV (w; p, q) and ΦV,I (C; p, q).
254
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
To avoid overloading the functionals with an additional label, we slightly abuse the notation observing that the functionals defined above can be easily distinguished by the number of their indices and arguments from general labeled (H) contour functional of Section 2.2 as well as from the functionals ΦB (A; p, q) and (H) ΦB (σ; p, q), introduced in the next section, that yield the labeled aggregate and shadow models. In correspondence to Lemma 4.7 in [HKZ], and due to the estimate (4.26), we may use the fact that (H) (H) ΦV,I (Ck ; p, q) 1 + ΦV,I (C; p, q) = C:C∩I=∅,C∩VR =∅
{Ck } finite Ck ∩I=∅,Ck ∩VR =∅
k
to obtain the following expression for the limit partition function (H) ˜ Λ|y; H) = ˜ V |y; H) = lim Z(I, Z(I, ΦV (w; p, q)× Λ V
{Ck } finite Ck ∩I=∅,Ck ∩VR =∅
w∈W(I)
np,q
×
exp{−ϕ(H) p,q;s |Bp,q;s (I)|}
s=1
(H)
ΦV,I (Ck ; p, q). (4.29)
k
Thus the probabilities of any interface I ∈ J (y, Λ) defined by µ {x ∈ X : I(x) = I}, Λ|y; H = converge to the probability of I ∈ J (y, V ) = µ {x ∈ X : I(x) = I}, V |y; H =
˜ Λ|y; H) Z(I, ˜ ¯ ¯ I∈J (y,Λ) Z(I, Λ|y; H)
Λ⊂V
J (y, Λ) defined by
˜ V |y; H) Z(I, . ˜ ¯ ¯ I∈J (y,V ) Z(I, V |y; H)
(4.30)
The sum in the denominator above converges since, according to (4.15), it H −(ρε H −cI )|W | 0 can be bounded by ecI |BR |(H+1) i∈B , where W stands i∈W e for connected subsets of Zν (possible supports of walls shifted vertically to intersect B). Here we have used the important fact that the interfaces from J (y, V ) are uniquely determined by their walls even if we “forget” their vertical position. This observation goes back to “(y, V )-admissible families of standard walls ([HKZ] Lemma 2.2)” and originates in [D 72]. (H) Now we shall recall or derive some estimates on the functionals ΦV (w; p, q) (H) and ΦV,I (C; p, q) and their derivatives as well as a bound on the derivative of (H)
ϕp,q;s needed later. Lemma 4.3 Let xp , xq ∈ G0 be two stable translation invariant configurations at H ∈ Kε (H0 ). Let s ∈ {1, . . . , np,q } be arbitrary.
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models (H)
255
(H)
Then ΦV (w; p, q) is translation invariant, ΦV,I (C; p, q) is translation invariant “inside V ”, and the following bounds are fulfilled. (H) (H) (a) 0 < ΦV (w; p, q) ≤ e−τI |W | ; (H)
(b) 0 ≤ ΦV,I (C; p, q) ≤ e−τI (H)
|C|
;
(H) + (H) (a ) |∂H ΦV (w; p, q)| ≤ H exp{−τI |W |}; (H) + (H) ΦV,I (C; p, q)| ≤ H exp{−τI |C|}; (b ) |∂H + (H) (c) |∂H ϕp,q;s | ≤ HMs . Ms(≡ Mshadow ) and cI are suitable constants, Here ρε log H0 − cI H.
(H)
and τI
(H)
= τh
−
Proof. (a) Taking the limit Λ " V in (4.23), we get (H) H (H) − cw )|W | ≤ e−(τh −cw )|W | 0 ≤ ΦV (w; p, q) ≤ exp −(ρε H0 by (3.6) and (3.10). (b) Taking the limit Λ " V in (4.26) we get (H) (H) 0 ≤ ΦV,I (C; p, q) ≤ exp −(τh − cC )|C| . (H) (a ) By the definitions (4.9) of E˜ (H) (w) and (4.11) of ΦΛ (w; p, q), with Λ replaced by V , we have (H)
+ ∂H ΦV (w; p, q) =
(H) (H) + exp −E (H) (w) + (h(H) − κ(τh ))|W | + ϕp,q;s(w) |π(W )| × = ∂H
Z d (Intj I|xq(j) ; H) = exp{−h(H)| Intj I|} j◦W (H) (H) (H) + = ΦV (w; p, q)∂H −E (H) (w) + (h(H) − κ(τh ))|W | + ϕp,q;s(w) |π(W )| + (H) (H) + exp −E (H) (w) + (h(H) − κ(τh ))|W | + ϕp,q;s(w) |π(W )| × Z d (Int I|x ; H) Z d (Intj I|xq(j) ; H) j q(j) + ∂H . (4.31) × I|} exp{−h(H)| Int exp{−h(H)| Intj I|} j j =j j◦W
j ◦W
To get the sought estimate (a ), we begin by estimating the first summand in (4.31). Due to (a), it suffices to show that (H) (H) + −E (H) (w) + (h(H) − κ(τh ))|W | + ϕp,q;s(w) |π(W )| | ≤ C (1) |W | H |∂H (H)
with a suitable constant C (1) that can be chosen independent of τh
(H)
for large τh
.
256
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
Indeed, this follows from (3.9), and the following bounds:
+ (H)
(H)
∂ ¯ E (w) ≤ ∂ +¯ Ei (x) ≤ dim H(R) H W H H
(4.32)
i∈W (H)
by linearity of Ei
(x) in H (see (4.8) to understand the first inequality); (H)
+ ∂H ¯ τh
= ρε
¯ H H0
(4.33)
which follows from (3.10) and (3.6); (H)
+ |∂H ¯ κ(τh
)| ≤ 2
(H)
e−τh
|C|
|C|ρε
¯ H H0
0∈C
(H)
¯ −(τh ≤ He
−log
ρε H0
−cκ )
¯ const, ≤ H
(4.34)
with a suitable constant cκ , follows from (4.6) and (4.33); + (H) ¯ |∂H ¯ ϕp,q;s | ≤ Ms H
(4.35)
follows from (4.10) using an analogy with (4.32), (3.9), and (4.34). To estimate the other summand in (4.31), we use (4.22), (4.21) and, moreover,
Z d (Int I|x ; H) j q(j)
+
¯ exp{cZ W }.
∂H¯
≤ H exp{−h(H)| Intj I|}
(4.36)
To show the last bound, we use (2.23) of Lemma 2.3.3, (3.12), (3.9), and the ν bound | Intj I| ≤ |∂ Intj I| ν−1 ≤ e|∂ Intj I| to get
Z d (Int I|x ; H) |(∂ + Z d (Int I|x ; H)) exp{−h(H)| Int I|}| j j q(j) j ¯ q(j)
+
H +
∂H¯
≤ exp{−h(H)| Intj I|} exp{−2h(H)| Intj I|} + |Z d (Intj I|xq(j) ; H)) exp{−h(H)| Intj I|} | Intj I| ∂H ¯ h(H)| ≤ exp{−2h(H)| Intj I|} (H) ¯ Intj I| exp ε˜0 |∂(Intj I)| Ms (1 + ε˜0 ) + (1 + e−τh ) ≤ He ¯ cZ |W | , ≤ H| (4.37)
+
where cZ is a suitable constant. We used the obvious inequality | Intj I| ≤ |∂(Intj I)|ν .
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
257
(b ) From the definition (4.12) we have
|C ∩ VR | |C ∩ Vm (I)|
+ (H) Ψ(H)T − )+ (C)(χVm (C) |∂H ¯ ΦV,I (C; p, q)| ≤ exp m |C| |C| m∈{p,q} (H)
+ e−τh |C| (|C ∩ I ∩ VR | + χVp (C)χVq (C)) × (H) + (H)T + (H) × |∂H (C)| + (|C| + 1) exp{−τh |C|}|∂H ||C| ≤ ¯ Ψm ¯ τh m∈{p,q}
¯ exp{−(τ (H) − log ≤ H h
ρε − cC )|C|} H0
(4.38)
with some constant cC . + (H)T (C)|, Here we used the bound (2.9) from Proposition 2.1.3 to estimate |∂H ¯ Ψm as well as the bound (4.33) above. (c) The needed estimate was already proved in (4.35) above.
4.3
Aggregate and labeled shadow models
We shall now project the walls, as well as clusters decorating the interface, to Zν−1 (⊂ Zν ), by applying the projection π. In that way, using the equalities (4.29) and (4.30), we introduce a shadow model, in terms of which we shall grasp the characteristic features of interfaces. The series of preceding steps will be concluded by Proposition 4.4 that describes the properties of the labeled shadow model as well as the relations to the probabilities of interfaces via an intermediary aggregate model. Given an interface I = (I, xI ) ∈ J (y, V ) (where V = π −1 (B) as in Section 4.2 and y ∈ Xxhor ) and a finite family {Ck } of clusters Ck , Ck ∩I = ∅, Ck ∩VR = ∅ (we p ,xq say that “{Ck } decorates I at V ”), and using W(I) to denote the collection of walls of I, we first consider connectedcomponents, to be called shadows of (I, {Ck }), of the projection π( w∈W(I) W ∪ k Ck ). If (W, C) is a collection of walls of I, W ⊂ W(I), and clusters from {Ck }, C ⊂ {Ck }, such that Σ = π( w∈W W ∪ C∈C C) is a shadow, we introduce an aggregate A of (I, {Ck }) as the “piece of interface above the shadow” together with the corresponding clusters. Namely, A = (I A , C), where I A = (IA , xIA ) with IA = π −1 (Σ) ∩ I. The shadow Σ is called the support of A. We identify aggregates that can be moved one into another by a vertical translation (by shifting both, the corresponding walls and clusters). Notice, that the relative position of concerned clusters with respect to the set IA is, in a given aggregate, always fixed. An aggregate A = (I A , C) with support Σ is naturally labeled since for any i ∈ ∂Σ, the column π −1 (i) is an “s-column” of xI with some s ∈ {1, . . . , np,q }. It is clear that this s is identical for all i’s from each component of ∂Σ. We denote this labeling by λ(A). As a result, the shadow Σ corresponding to an aggregate A may be labeled by λ(A) as described above. A shadow Σ endowed with such a labeling λ = λ(A)
258
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
will be called a labeled shadow (of (I, {Ck })). We write σ = (Σ, λ), λ(σ) = λ, and supp σ = Σ. Notice that the labeled shadows of a pair (I, {Ck }), as above, form compatible matching families in the sense of Section 2.2, whereas, strictly speaking, labeled aggregates do not fit into that scheme, neither as a system of contours in Zν nor in Zν−1 . To include them would need a slight generalization of the setting from Section 2.2. Labeled shadows are “labeled contours” in the sense of Section 2.2. Considwith fixed p, q ∈ Q, p = q, ering now generally a configuration xp,q;s ∈ G0 ∩ Xxhor p ,xq such that xp , xq are stable, and s ∈ (1, . . . , np,q ) (in the role of the triplet y, xp , xq from the preceding subsection), we introduce the weight of an aggregate A = (I A , C) compatible with V and y by (H)
ΦB (A; p, q) =
(H)
ΦV (w; p, q)
w∈W(I A )
(H)
ΦV,I (C; p, q).
(4.39)
C∈C
Summing over all aggregates A corresponding to a fixed labeled shadow σ (the same support and the labeling λ(σ) = λ(A) corresponding to xA as above), we get the shadow weight (H)
ΦB (σ; p, q) =
(H)
ΦB (A; p, q).
(4.40)
A:supp A=supp σ λ(A)=λ(σ) (H)
(H)
n
p,q Finally, we use ϕp,q to denote the vector (ϕp,q;s )s=1 . It is defined as in (4.10) with stable xp , xq . Preparing for a direct application of the Pirogov-Sinai theory from Section 2 to the model defined by labeled shadows, notice first that, even though we took with a volume B ⊂ Zν−1 only interfaces from J (xp,q;s , V ) with the walls inside of V = π −1 (B), the clusters Ck may stick out of V and, correspondingly, the aggregates may not be contained in V and the shadows may not be contained in B. Anticipating this feature, we actually considered, in Section 2, the generalization allowing for volume depending contour weights. Thus, for any finite B ⊂ Zν−1 , we introduce the generalized ensemble aggr XB (p, q) of all compatible and matching families of labeled aggregates with walls in V = π −1 (B) and with clusters intersecting V and the ensemble Xshad B (p, q) of all aggr families of shadows that correspond to families from Xaggr B (p, q). The set XB (p, q) (Xshad B (p, q)) is actually the union, over s ∈ (1, . . . , np,q ), of all sets characterized by the external labels of all external aggregates (shadows) being fixed to equal s — i.e. families of aggregates (shadows) consistent with boundary conditions xp,q;s . The set Xshad B (p, q) will play the role of the abstract set XΛ from Section 2.2. Recall that, in accordance with the notation from the first paragraph of shad Section 2.2, for any A ∈ Xaggr B (p, q) (S ∈ XB (p, q)), we use Bs (A) (Bs (S)) to denote the set of sites in B that are outside of the supports of all aggregates A ∈ A
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
259
(shadows σ ∈ S) and are labeled by the label s of the corresponding ceiling. Now, we define Z aggr Xaggr Z aggr {A}, Xaggr B (p, q)|s0 ; H = B (p, q)|s0 ; H , A∈Xaggr (p,q) B ext. label s0
where Z
aggr
np,q (H)
aggr
{A}, XB (p, q)|s0 ; H = ΦB (A; p, q) exp − ϕ(H) p,q;s Bs (A) A∈A
s=1
(4.41) if A ∈ Xaggr (p, q) with external label s ; otherwise it is not defined (or, rather, it 0 B is put to equal 0). Similarly, Z shad Xshad Z shad {S}, Xshad B (p, q)|s0 ; H = B (p, q)|s0 ; H S∈Xshad (p,q) B ext. label s0
with np,q
(H)
(p, q)|s ; H = Φ (σ; p, q) exp − ϕ(H) Z shad {S}, Xshad 0 B p,q;s Bs (S) , B σ∈S
s=1
(4.42) whenever S ∈ Xshad B (p, q) with external label s0 ; otherwise it is put to equal 0. shad Notice that Xaggr B (p, q) (respectively, XB (p, q)) contains an infinite number of configurations of aggregates (shadows). Nevertheless, the convergence of (4.41) and (4.42) is an easy consequence of the bounds (a) and (b) from Lemma 4.3. The probability of a compatible matching family of labeled aggregates A ∈ Xaggr B (p, q) with external label s0 is, correspondingly, defined by Z aggr {A}, Xaggr aggr B (p, q)|s0 ; H aggr . µ {A}, XB (p, q)|s0 ; H = (4.43) Z aggr Xaggr (p, q)|s ; H 0 B Similarly, for S ∈ Xshad B (p, q) with external label s0 we introduce Z shad {S}, Xshad B (p, q)|s0 ; H shad shad . µ {S}, XB (p, q)|s0 ; H = Z shad Xshad (p, q)|s ; H 0 B
(4.44)
These measures can be used to evaluate the probability of perturbations of the (H) interface of xp,q;s0 , once a sufficient decay of functional ΦB (σ; p, q) in dependence on the size of | supp σ| is guaranteed as we shall see in Section 4.4. Proposition 4.4 (interfaces in terms of aggregates and shadows) Let xp , xq be two translation invariant ground states of H0 that are stable for some H ∈ Kε (H0 ) for ε small enough, ρ0 large enough, and let B ⊂ Zd−1 be finite. Then the following claims hold.
260
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
(i) For any s ∈ {1, . . . , np,q }: aggr aggr Z shad Xshad XB (p, q)|s; H B (p, q)|s; H = Z ˜ π −1 (B)|xp,q;s ; H). Z(I, = I
Further, aggr µshad {S}, Xshad µ {A}, Xaggr B (p, q)|s; H = B (p, q)|s; H , A
where the sum runs through A ∈ Xaggr B (p, q) such that supp A = supp σ for all A ∈ A and the labels corresponding to mathcalA and A coincide, λ(A) = λ(S). Finally, using W(A) to denote the set of all walls corresponding to the collection of aggregates A, we have µ {x ∈ X : W(I(x)) = W}, π −1 (B)|xp,q;s ; H = µaggr {A}, Xaggr B (p, q)|s; H A:W(A)=W
for any admissible collection of walls W. (H)
(H)
(H)
n
p,q (ii) The functional ΦB (σ; p, q) and the vector ϕp,q = (ϕp,q;s )s=1 define a labeled contour model that fulfills all the assumptions of Theorem 2 and Proposition 2.2.1. Namely,
(H)
• ΦB (σ; p, q) is nonnegative, • ΦB (σ; p, q) = ΦB (σ ; p, q) = ΦZν−1 (σ; p, q) whenever σ and σ are such that supp σ, supp σ ⊂ B and σ is a translation of σ (“the functional (H) ΦB (·; p, q) is, inside B, translation invariant and independent of B”). Also, (H) (H) there exists a constant τs differing from τI by a fixed constant depending only on ν and |S| and a constant Ms , such that the following bounds hold: (H)
(H)
(H)
(H)
(H)
(1) ΦB (σ; p, q) ≤ exp{−τs |supp σ|} for any labeled shadow σ consistent with boundary conditions xp,q,s in B; (H) + (H) ¯ (2) |∂H | supp σ|}; ¯ ΦB (σ; p, q) ≤ H exp{−τs + (H) ¯ (3) |∂H ¯ ϕp,q;s | ≤ Ms H.
Proof. (i) It follows from (4.39) – (4.44) with the use of (4.29) and (4.30). (H) (ii) The functional ΦB (A; p, q) is nonnegative by its definition (4.39) and by (4.14).
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
261
The translation invariance inside B follows from the definitions (4.39), (4.11), (4.12), from the translation invariance of the functionals involved and the invariance inside B of χVm (C), and by inspecting cardinalities of the sets involved in (4.12) for sets C fully contained in V . (H)
The corresponding claims for ΦB (σ; p, q) follow from (4.40). (1) We use first Lemma 4.3(a) and (b) to get, by (4.39) and (4.40), that (H)
0 ≤ ΦB (σ; p, q) ≤
(H)
exp{−τI
(H)
|W | − τI
w∈W(I)
(I,C)
|C|},
C∈C
where the sum is over aggregates that correspond each to a pair of an interface I ∈ I(xp,q;s , V ) and of a finite family C such that C ∩ VR = ∅ and C ∩ I = ∅, for all C ∈ C, and π( C∪ W ) = Σ = supp σ as above (i.e. over (I, C) with C∈C
w∈W(I)
the only aggregate A = (I A , C)). |C| + Since, obviously, A = (W(I), C) ≡ C∈C
w∈W(I)
|W | ≥ | supp σ|, we
have (H)
ΦB (σ; p, q) ≤ e−(τI (H)
−ζ)| supp σ|
exp{−ζ(
(I,C)
w∈W(I)
|W | +
|C|)},
C∈C
(H)
for any ζ > 0, ζ < τI . We shall show that, for some sufficiently large ζ, in dependence on ν and |S| only, the last sum is at most ec| supp σ| for suitable c which also depends on |S| and ν only. We begin with an observation. Namely, introducing the notion of an extent of an aggregate A as ext A = IA ∪ C, we notice that it is a connected set of cardinality at most |W |+ C∈C w∈W(I A ) (1 + T ) |C| ≤ (1 + T )A. In the same time we may suppose, referring to the C∈C
identification of aggregates under vertical translations, that each ext A contains an element i of supp σ. As a consequence, there are at most ecs A aggregates A with i ∈ ext A and given A. Namely, every such A can be identified with a connected path over ext A of the length at most 2ν(T + 1)A, each point of which is moreover equipped with a label saying whether this point of the path belongs to a wall, to a cluster, or to the remaining part W∪ C ext A \ w∈W(I A )
C∈C
of the interface I A . The label will also say what spin at this point is attained if it is a point of a wall. The number of needed labels is thus depending on ν and |S|
262
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
only. As a result, we get (H)
ΦB (σ; p, q) ≤ e−(τI (H)
−ζ)| supp σ|
exp{−ζ(
(H)
≤e
−ζ)| supp σ|
(H) −(τI −ζ)| supp σ|
|W | +
w∈W(I)
A=(I,C)
≤ e−(τI
|C|)} ≤
C∈C
e−ζA ≤
i∈supp σ A:i∈ext A
(H)
e−(ζ−cs )k ≤ e−(τI
−ζ−1+log(1−e−cs ))| supp σ|
i∈supp σ k≥| supp σ|
≤ e−τs
(H)
| supp σ|
,
where ζ ≥ 2cs . (2) By differentiating (4.40) and using (4.39) as well as the estimates (a), (b), (a ), (b ) of Lemma 4.3, we obtain (H)
+ |∂H ¯ ΦB (σ; p, q)| ≤
+
A=(I,C) w∈W(I)
(H)
¯ exp{−τ H I
¯ exp{−τ (H) H I
¯ w∈W(I)
¯ − |w|
¯ w∈W(I)
A=(I,C) C∈C (H)
¯ −(τI ≤ He
(H)
¯ − τI |w|
(H) τI
−ζ−1)| supp σ|
¯ |C|}+
¯ C∈C
¯ ≤ |C|}
¯ C∈C
exp{(−ζ + 1)A}.
A=(I,C)
Here the sum
above is taken over all aggregates A with supp A = supp σ
A=(I,C)
and λ(A) = λ(σ). Now, we use the same argument as in (1) with ζ − 1 instead of ζ, so that for ζ − 1 ≥ 2cs we get + −τs |∂H ¯ ΦB (σ; p, q)| ≤ e
(H)
(H)
| supp σ|
.
(3) This is the estimate of Lemma 4.3(c) above.
4.4
Gibbs states with interfaces. Proof of Basic Lemma
To conclude the proof of Basic Lemma, it remains to define, for any y ∈ Ghor 0 , the functions hy : Kε (H0 ) → R such that the statements (i) (b) and (ii) of Basic Lemma hold. As anticipated in Section 4.1, we fix a particular H ∈ Kε (H0 ), two different configurations from Gper 0 , say xp , xq with p, q ∈ {1, . . . , r}, and a configuration y ∈ G0 ∩ Xxhor to be identified with a triplet (p, q, s), s ∈ {1, . . . , np,q }. The p ,xq (H)
(H)
(H)
corresponding functionals ΦB (A; p, q), ΦB (σ; p, q), and vectors ϕp,q describing the aggregate and shadow models in the preceding Section 4.4 were defined (or considered) only if xp and xq were stable and distinct. We may however notice that (H) ϕp,q;s are by (4.10) actually well-defined for general pairs of (distinct) xp and xq
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models (H)
263
(H)
from Gper 0 . To define ΦB (A; p, q) and ΦB (σ; p, q), we then use formulas (4.39) (H) and (4.40) referring to (4.11) and (4.12). While ΦV (w; p, q) again is well-defined (H) even if xp , xq are arbitrary elements of Gper 0 , we need to define ΦV,I (C; p, q) so that the bounds (1), (2), and (3) of Proposition 4.4 (ii) hold true. This can be easily achieved if we simply replace Ψp and Ψq in (4.12), and further on, by Ψp and Ψq , respectively, where Ψp and Ψq are the auxiliary functionals from Proposition 2.2.1 (b ), so that we may apply all the cluster expansion technique to them to get (ii) of Proposition 4.4 for all p, q’s. This is the starting point for a second use of the Pirogov-Sinai strategy. This (and fixed p, q) playing the role of reference time with different y ∈ G0 ∩ Xxhor p ,xq states and thus the label in Theorem 2 and Proposition 2.2.1 taking values s with s ∈ {1, . . . , np,q }. The function hy (H) from Basic Lemma is, for y corresponding (H) to the triplet (p, q, s), defined to be equal to the function hs (Φ(H) (·; p, q), ϕp,q ) whose existence is assured by Theorem 2. The claim (ii) of Basic Lemma is then an immediate consequence of Theorem 2 (approximating hy (H) (and its derivative) (H) (H) in terms of ϕp,q;s ) and the equation (4.10) relating ϕp,q;s to ey (H) (cf. (3.7), (3.9), and (4.34)). (Of course, if either xp or xq is not stable, we lose the equality (4.13) and as a result we cannot claim anything like (i) of Proposition 4.4.) It remains to prove the statement (i) (b) of Basic Lemma. Therefore we assume that both xp and xq are stable, and that y = xp,q;s is stable for the (H) (H) shadow model determined by ΦB (σ; p, q) and ϕp,q . We are going to derive the properties of the Gibbs measure µ(·|y, H) obtained from measures µ(·, Vn = π −1 (Bn )|y, H) as a weak limit with Bn growing up to Zν−1 , which were investigated in Step 7 above. Such measures are Gibbs states as they are weak limits of µ(·, Λn |y, H) for volumes Λn = Bn × [−kn , kn ], with kn growing to infinity sufficiently quickly (cf. [D 72, HKZ (3.2)]). To show that almost every configuration is a perturbation of y, it suffices to verify that for every ε and any Λ ⊂ Zν , there exist constants a, b, d > 0 so that, defining X(Λ, a, b) = x ∈ X; w ∈ W(I(x)), V (W ) ∩ Λ = ∅ ⇒ diamW ≤ a and γ a contour of x, V (Γ) ∩ Λ = ∅ ⇒ diamΓ ≤ b , with V (W ) denoting the union of W and all finite components of Zν \ W , we have µ X(Λ, a, b), π −1 (B)|y; H > 1 − ε
(4.45)
whenever B ⊂ Zν−1 is such that dist(π(Λ), B c ) ≥ d. To this end we first show that it is unprobable that shadows intersecting a given finite volume in Zν−1 are large and similarly for corresponding aggregates in Zν . Finally we estimate the probability of existence of large contours intersecting a fixed finite volume in Zν .
264
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
(c)
Let thus EM be the set of all shadow configurations containing a large shadow intersecting or surrounding a fixed finite set M ⊂ Zν−1 . Namely, for any such M and c > 0 we introduce (c) EM = S ∈ Xshad B (p, q); there exists σ ∈ S such that V (Σ) ∩ M = ∅ and diamΣ > c , with Σ = supp σ and V (Σ) (for any finite set Σ ⊂ Zν−1 ) denoting the union of Σ with all finite components of Zν−1 \ Σ. Using the main results of Section 2 (Theorem 2, Proposition 2.2.1, and Corollary 2.2.2) to the shadow model, we may show, in a standard way, that for a given finite M ⊂ Zν−1 and a positive ε, constants d = d (M, ε) and c = c(M, ε) may be found such that (c) 1 µshad EM , Xshad B (p, q)|s; H ≤ ε 2
(4.46)
for all B ⊂ Zν−1 such that dist(M, B c ) ≥ d . For a fixed configuration S of shadows of diameter at most c and with V (Σ)∩ M = ∅ for every σ ∈ S, consider the event consisting of those configurations of walls and clusters (configurations of aggregates) whose set of shadows intersecting M is fixed and equals S. The conditional probability of each of these events, given such a fixed S, is independent of B for B sufficiently large. Since there are countably many such events, and due to (4.46), we may find a = a(M, ε) such that, supposing dist(M, B c ) > d , we get 1 µaggr (A(M, a), Xaggr B (p, q)|s; H) > 1 − ε, 2 with A(M, a) = {A ∈ Xaggr B (p, q); [w ∈ W(A) and V (W ) ∩ π −1 (M ) = ∅] ⇒ diamW ≤ a}. Comparing µaggr and µ (Proposition 4.4 (i)), we get the bound 1 µ X(M, a), π −1 (B)|y; H > 1 − ε 2
(4.47)
for
X(M, a) = x ∈ X; w ∈ W(I(x)) and V (W ) ∩ π −1 (M ) = ∅ ⇒ diamW ≤ a .
Given a finite Λ ⊂ Zν with projection π(Λ) = M , xΛ is fully determined if we know the contours γ with V (Γ) ∩ Λ = ∅ and walls w with V (W ) ∩ Λ = ∅. Let I be a given interface compatible with the boundary condition y on Λc . If γ is a contour such that V (Γ) ∩ Λ = ∅ and, in the same time, V (Γ) ⊂ Int W = V (W ) \ W
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
265
for some wall, then necessarily V (W ) ∩ Λ = ∅. Thus it suffices to consider contours that are either above or below I, Γ ⊂ ( j Intj I ∪ I)c . Introducing X(I, Λ, b) = {x ∈ X; I(x) = I and [γ is a labeled contour of x, Γ ⊂ ( Intj I ∪ I)c , V (Γ) ∩ Λ = ∅] ⇒ diamΓ ≤ b}, j
we get
1 µ X(I, Λ, b), π −1 (B)|y, I; H > 1 − ε, (4.48) 2 independently of I for B ⊂ Zν−1 with dist(M, B c ) > d for some b = b(M, ε) and d = d(M, ε) > d . Here we consider the conditional probability under the condition that the interface with support I is present. The bounds (4.47) and (4.48) together yield (4.45). Remark. Notice that having (4.45) for Λ’s that are singletons, we get that almost all configurations with respect to the limit Gibbs state are perturbations of y. In fact, we may prove an explicit expression for the limit probability µ(·|y; H) by means of the limit conditional probability “µ(·, π −1 |I)” and the limit probability on families of walls “ limν−1 µI (·, π −1 (Bn )|y; H)” similarly as (2.1) was proved Bn Z
in [HKZ, Section 6.2.2]. However, it is a little bit more complicated in our more general situation. The probabilities µI can be expressed using their projections to families of external walls. Also we don’t have the exponential estimates as (a) and (b) of (i) and (ii) in [HKZ, Theorem 2]. However, one may check that the above estimate (4.45) is sufficient to carry on the proof of the respective identity.
References [BCF 1] C. Borgs, J. Chayes and J. Fr¨ ohlich, Dobrushin states for classical spin systems with complex interactions, Jour. Stat. Phys. 89, 895 (1997). [BCF 2] C. Borgs, J. Chayes and J. Fr¨ ohlich, Dobrushin states in quantum lattice systems Jour. Stat. Phys. 89, 591 (1997). [BI]
C. Borgs and J. Imbrie, A Unified Approach to Phase Diagrams in Field Theory and Statistical Mechanics, Commun. Math. Phys. 123, 305–328 (1989).
[BK]
C. Borgs and R. Koteck´ y, A Rigorous Theory of Finite-Size Scaling at First-Order Phase Transitions, Journ. Stat. Phys. 61, 79–119 (1990).
[BL]
J. Bricmont and J. L. Lebowitz, Wetting in Potts and Blume-Capel Models, Journ. Stat. Phys. 46, 1015–1029 (1987).
[BS]
J. Bricmont and J. Slawny, Phase transitions in systems with a finite number of dominant ground states, Journ. Stat. Phys. 54, 89–161 (1989).
266
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık
Ann. Henri Poincar´e
[Br]
D. C. Brydges, A short course on cluster expansions, Critical phenomena, Random Systems, Gauge Theories, Les Houches 1984, North-Holland, Amsterdam, 129–183 (1986).
[D 96]
R. L. Dobrushin, Estimates of Semi-invariants for the Ising Model at Low Temperatures, Amer. Math. Soc. Transl. (2) 177, 59–81 (1996).
[D 72]
R. L. Dobrushin, Gibbs State Describing Coexistence of Phases for a Three-Dimensional Ising Model, Teor. Ver. Pril. 17, 619–639 (1972), English transl. in Theor. Probability Appl. 17, 582–600 (1972).
[DKS]
R. L. Dobrushin, R. Koteck´ y and S. Shlosman, The Wulff construction: a global shape from local interactions AMS, Translations of Mathematical Monographs 104, Providence, Rhode Island (1992).
[EFS]
A.C.D. van Enter, R. Fern´ andez and A. Sokal, Regularity properties and pathologies of position-space renormalization-group transformations: Scope and limitations of Gibbsian theory, J. Stat. Phys. 72, 879–1167 (1993).
[HKZ]
P. Holick´ y, R. Koteck´ y and M. Zahradn´ık, Rigid interfaces for lattice models at low temperatures, Journ. Stat. Phys. 50, 755–812 (1988).
[HZ]
P. Holick´ y and M. Zahradn´ık, Stratified low-temperature phases of stratified spin models. A general Pirogov-Sinai approach, preprint.
[KP]
R. Koteck´ y and D. Preiss, Cluster Expansion for Abstract Polymer Models, Commun. Math. Phys. 103, 491–498 (1986).
[PS]
S. Pirogov and Ya. G. Sinai, Phase diagrams of classical lattice systems, Theor. Mat. Phys., 25, 1185–1192 (1975), 26, 39–49 (1976).
[Se]
E. Seiler, Gauge theories as a problem of constructive quantum field theory and statistical mechanics Lecture notes in physics, Vol. 159, Springer Verlag Berlin, Heidelberg, New York (1982).
[S]
Ya. G. Sinai, Theory of phase transitions: Rigorous results, Pergamon Press, Oxford-New York-etc (1982).
[Z]
M. Zahradn´ık, An Alternate Version of Pirogov-Sinai Theory, Commun. Math. Phys. 93, 559–581 (1984).
P. Holick´ y Department of Mathematical Analysis Charles University Sokolovsk´a 83, Praha 8 Czech Republic email:
[email protected]
Vol. 3, 2002
Phase Diagram of Horizontally Invariant Gibbs States for Lattice Models
Roman Koteck´ y Center for Theoretical Study Charles University Jilsk´ a 1, 110 00 Praha 1 Czech Republic email:
[email protected] Miloˇs Zahradn´ık Department of Mathematical Analysis Charles University Sokolovsk´a 83, Praha 8 Czech Republic email:
[email protected] Communicated by Jennifer Chayes submitted 12/12/00, accepted 16/11/01
To access this journal online: http://www.birkhauser.ch
267
Ann. Henri Poincar´e 3 (2002) 269 – 295 c Birkh¨ auser Verlag, Basel, 2002 1424-0637/02/020269-27 $ 1.50+0.20/0
Annales Henri Poincar´ e
The Infrared Behaviour in Nelson’s Model of a Quantum Particle Coupled to a Massless Scalar Field J. L˝ orinczi, R. A. Minlos and H. Spohn Abstract. We prove that Nelson’s massless scalar field model is infrared divergent in three dimensions. In particular, the Nelson Hamiltonian has no ground state in Fock space and thus it is not unitarily equivalent with the Hamiltonian obtained from Euclidean quantization. In contrast, for dimensions higher than three the Nelson Hamiltonian has a unique ground state in Fock space and the two Hamiltonians are unitarily equivalent. We also show that the Euclidean Hamiltonian has no spectral gap.
1 Introduction A quantum particle coupled to a free Bose field can be described by the Hamiltonian 1 − ∆ + V (q) ⊗ 1 + 1 ⊗ ω(k)a∗ (k)a(k)dk (1.1) HgN = 2 Rd 1 1 + ˆ(k)eik·q ⊗ a(k) + ˆ(−k)e−ik·q ⊗ a∗ (k) dk. d/2 2ω(k) (2π) Rd In the first term we have the Schr¨ odinger operator Hp = −(1/2)∆ + V for the quantum particle. V is assumed to be a confining potential growing at infinity, hence Hp has a purely discrete spectrum. The second term is the energy of the Bose field with dispersion relation ω(k) ≥ 0. For a shorthand we set Hf = Rd ω(k)a∗ (k)a(k)dk. a∗ (k), a(k) are the Bose creation and annihilation operators satisfying the canonical commutation relations [a(k), a∗ (k )] = δ(k − k ), [a(k), a(k )] = 0 = [a∗ (k), a∗ (k )]. The last term in (1.1) describes a translation invariant interaction between the particle and the field, which is linear in the field operators. is the charge distribution and mollifies the coupling between particle and field. We assume to be smooth, radial and of rapid decay. ˆ denotes the Fourier transform of . We use the normalization Rd (x)dx = e ≥ 0; e then cali2 d ⊗ Fsym with Fsym brates the strength of the interaction. HgN acts on L 2(R , dq) 2 d the symmetric Fock space over L (R , dk). If Rd |ˆ (k)| (ω(k)−1 + ω(k)−2 )dk < ∞, then HgN is bounded from below and the interaction is infinitesimally Hf -bounded. Therefore HgN is a self-adjoint operator with domain D(Hp ⊗ 1) ∩ D(1 ⊗ Hf ). introduces an ultraviolet cut-off, which may be removed by letting (x) → eδ(x) or, equivalently, ˆ(k) → (2π)−d/2 e. In his famous paper [17] Nelson proved that in
270
J. L¨ orinczi, R. A. Minlos and H. Spohn
Ann. Henri Poincar´e
the case ω(k) = k 2 + m2b , mb > 0, at the expense of renormalizing the energy, HgN stays well-defined in the point charge limit. Here we will keep the ultraviolet cut-off and will study the infrared behaviour, which is of interest only for massless bosons, mb = 0, i.e. ω(k) = |k|. (1.2) We call HgN with dispersion relation (1.2) the massless Nelson model and denote the corresponding Hamiltonian by HN . For d = 1, 2 the Nelson Hamiltonian is not bounded from below, thus we require d ≥ 3. In fact, we will mostly concentrate on dimension d = 3 for which, as proved below, the massless Nelson model is infrared divergent. Before stating our project specifically let us explain in general terms the phenomenon of infrared divergence in the context of Hamiltonians of the type (1.1). It is convenient somewhat to generalize and start from HG = K ⊗ 1 + 1 ⊗ Hf + (v(k) ⊗ a∗ (k) + v(k)∗ ⊗ a(k)) dk. (1.3) Rd
K is a self-adjoint operator acting on the particle space K and it is assumed to have a purely discrete spectrum. k → v(k) is a function of operators acting on K such that HG is bounded from below. Clearly, HgN is the special case where K = L2 (Rd , dq), K = Hp , v(k) = (2π)−d/2 e−ik·q ˆ(−k)/(2ω(k))1/2 . The Hamiltonian HG acts on K ⊗ Fsym . We say that HG is infrared divergent if it has no ground state in K ⊗ Fsym . The physical ground state has then an infinite number of very low energy bosons forcing a representation of the canonical commutation relations different from the Fock representation. HG should be thought of only as a formal object and one has first to construct the Hilbert space containing the physical ground state, as well as the corresponding Hamiltonian. Ideally, one would like to decide whether the model is infrared divergent or not directly in terms of K and v in (1.3). At present, this issue is only partially understood. G´erard proved [9] that if, beside some conditions ensuring the selfadjointness of H, Rd
ω(k)−2 v(k)(K + 1)−1 2 dk < ∞
(1.4)
holds, then HG has a ground state in K ⊗ Fsym . In the particular case (1.1), if |ˆ (k)|2 ω(k) + ω(k)−2 + ω(k)−3 dk < ∞, (1.5) Rd
then HgN has a unique ground state in L2 (Rd , dx)⊗Fsym , see [24]. Thus (1.5) must be contravened for HgN to get infrared divergent. We note that for the massless Nelson model in three dimensions indeed |ˆ (k)|2 dk = ∞. (1.6) 3 R3 |k|
Vol. 3, 2002
The Infrared Behaviour in Nelson’s Model
271
On the other hand, by (1.5) the massless Nelson model is infrared convergent for d ≥ 4. The integrals in (1.4), resp. (1.5), becoming infinite is only a necessary condition for infrared divergence to occur. To conjecture sufficient conditions we have to rely more closely on model systems. The best understood model is the spin-boson Hamiltonian where K = C2 , K = σx , v(k) = λ(k)σz , and σx , σz are the Pauli matrices. Some evidence comes also from the generalized spin-boson models studied J in [1], where v(k) = j=1 Bj λj (k) with Bj = Bj∗ acting on K. These examples suggest that, in addition to (1.4) turning infinite, v(k) must have non-vanishing ground state expectation (Ψ, v(k)Ψ)K for small k, with Ψ ∈ K the ground state of K. If this happens, the model is divergent for all couplings e = 0. If, however, v(0) has zero ground state expectation, then cancellations may occur making the model stay infrared convergent in spite of the integral in (1.4) being infinite. This is the case for the spin-boson Hamiltonian with v(k) = αλ(k)(σz + m) and 2 |λ| /ω 2 dk = ∞. If m = 0, the model is infrared divergent for all α = 0. If d R m = 0, cancellations appear and for sufficiently small α the model is infrared convergent while for large α it becomes divergent [24]. The same phenomenon takes place in the dipole approximation of the massless Nelson model in d = 1 where K = L2 (R, dq), K = Hp , v(k) = q · ˆ(k) [19]. If V (q) = V (−q), the onset of infrared divergence depends on how strong the coupling is; otherwise the model is infrared divergent no matter how large the (non-zero) coupling is. For the three dimensional Nelson model’s Hamiltonian we have e (Ψ, v(k)Ψ) = + O(1), 2|k|
k → 0.
(1.7)
Thus we expect, and will actually prove, that HN is infrared divergent for every coupling e = 0. Fr¨ ohlich [8] considered the massless Nelson model for V = 0 at fixed total momentum p in which case the Hamiltonian acting on Fsym is H(p) =
1 2
p−
ka∗ (k)a(k)dk
2
R3
Hf +
1 (2π)3/2
+ R3
1 2|k|
(ˆ (k)a(k) + ˆ(−k)a∗ (k))dk.
(1.8)
With an infrared cut-off and for sufficiently small |p|, H(p) has a unique ground state in Fsym . The removing of the infrared cut-off yields a physical ground state lying outside Fock space, and a non-Fock representation of the canonical commutation relations. In [8] a C ∗ -algebraic approach was developed. In contrast, in our work we follow the route of constructive quantum field theory by introducing a suitable probability measure on the space of Brownian paths. In terms of this path measure we can define both the physical ground state and the Hamiltonian governing the excitations relative to the ground state. While the explicit construction
272
J. L¨ orinczi, R. A. Minlos and H. Spohn
Ann. Henri Poincar´e
will be done in full detail in what follows, we now explain some of the particular features of the path measure associated with HN . In the Euclidean picture we have the trajectories t → qt of the particle, resp. t → ξt (x), x ∈ Rd , of the field appearing as basic ingredients. For assigning them a statistical weight we define the formal Euclidean action for the particle by 1 2 q˙ + V (qt ) dt, Sp ({qt }) = (1.9) 2 t R and by
Sf ({ξt }) =
R
Rd
1 (∂t ξt (x))2 + (∇ξt (x))2 dxdt 2
(1.10)
for the field. From (1.1) we deduce that the particle-field interaction restricted to the time interval [−T, T ] is given by Sint,T ({qt , ξt }) =
T −T
Rd
(x − qt )ξ(t, x)dxdt =
T
−T
(ξt ∗ )(qt )dt
(1.11)
with ∗ denoting convolution. Thus, heuristically, the full Euclidean path measure is 1 dqt e−Sp ({qt }) dξt (x) e−Sf ({ξt }) e−Sint,T ({qt ,ξt }) . (1.12) Z d+1 t∈R
(t,x)∈R
The first factor above provides an a priori weight for the paths qt . The corresponding path measure is denoted by N 0 , see Section 2.1 below. The second factor provides an a priori weight for the paths ξt . Clearly, this is a Gaussian measure, denoted by G in Section 2.2 below. In the spirit of the Feynman-Kac formula we define a semigroup through the quadratic form (F¯ , e−tH G)(q, ξ) = EN 0 ×G [e−
Rt 0
(ξt ∗)(qs )ds
F¯ (q0 , ξ0 )G(qt , ξt )],
(1.13)
t ≥ 0, where F, G are suitable functions. e−tH is a contractive semigroup on L2 (Rd , dN0 ) ⊗ L2 (S , dG) (here dN0 and dG are the t = 0 distributions under N 0 and G, respectively), moreover it is unitarily equivalent with e−tHN , t ≥ 0, on L2 (Rd , dq) ⊗ Fsym . To construct the physical ground state and its associated Hilbert space we have to give sense to the path measure (1.12) in the infinite time limit T → ∞. With respect to the limit measure (whenever it exists), the process t → qt , ξt (x) is jointly a time-reversible stationary Markov process. Therefore this stochastic process has a canonically defined semigroup Tt , t ≥ 0, self-adjoint on the Hilbert space weighted by the t = 0 stationary measure. Thus Tt has a self-adjoint generator so that Tt = e−tHeuc , t ≥ 0, which by definition is the Euclidean Hamiltonian Heuc . Clearly, by construction, Heuc ≥ 0 and Heuc 1 = 0, i.e. Heuc has an eigenvalue at the bottom of its spectrum. If HN is infrared convergent, then Heuc and H are unitarily equivalent and one may work with either of them on an equal
Vol. 3, 2002
The Infrared Behaviour in Nelson’s Model
273
footing. However, in the infrared divergent case they are not unitarily equivalent. Since Heuc accommodates the ground state, it should be regarded as the proper Hamiltonian of the system. Once path measures are introduced, one might wonder how infrared divergence is reflected in their properties. As explained in more detail below, within the context of the Nelson model a sharp criterion can be found: If HN is infrared divergent, then the t = 0 distribution of the path measure for the interacting system is singular with respect to the t = 0 distribution of the path measure for the non-interacting system. A short outline of the paper is the following. In Section 2 we recall the basic processes leading up to Section 3 where the definition of the joint particle-field path measure appearing in (1.12) is given. Based on the cluster expansion in [13], under suitable assumptions on V and for sufficiently small e, we establish the existence of the infinite time path measure. Thereby Heuc will be well defined. In Section 4 we discuss infrared divergence in terms of the path measure for the interacting system and prove that the massless Nelson model in d = 3 is infrared divergent. In Section 5 we prove that for d ≥ 4 it is however infrared convergent. Finally, in Section 6 we show that Heuc has no spectral gap. We remark that the quantized Maxwell field is less infrared divergent since by minimal coupling the Bose field couples to the momentum and not to the position as in (1.1). As proved in [3, 11], the Pauli-Fierz operator has a ground state in Fock space also in three dimensions, which suggests that in the case of the Maxwell field Heuc and H are unitarily equivalent.
2 Path measures for the particle and the field 2.1
Euclidean quantization of the particle
On the Hilbert space L2 (Rd , dq) we define the particle Hamiltonian 1 Hp = − ∆ + V (q). 2
(2.1)
The external potential V : Rd → R is a multiplication operator. We consider two classes of potentials. (P1) V is bounded from below, continuous, and having the asymptotics V (q) = C|q|2α + o(|q|2α )
(2.2)
for large |q|, with some constant C > 0 and exponent α > 1. (P2) V is of Kato-class bounded from below as V (q) ≥ C|q|2α with some constant C > 0 and exponent α > 0.
(2.3)
274
J. L¨ orinczi, R. A. Minlos and H. Spohn
Ann. Henri Poincar´e
For details on the Kato class we refer to [5]. In either case above Hp has discrete spectrum and a unique, strictly positive ground state ψ0 lying at its bottom Ep , i.e. Hp ψ0 = Ep ψ0 . Associated with Hp we consider the stationary P (φ)1 -process qt with time t = 0 distribution (invariant measure) dN0 = ψ02 (q)dq,
(2.4)
as defined through the stochastic differential equation dqt = ∇ ln ψ0 (qt ) dt + dBt
(2.5)
where dBt denotes Brownian motion. The path measure of this process will be denoted by N 0 ; for more details see [5] and references therein. (Here and henceforth we use calligraphic letters for path measures and straight characters for time t = 0 distributions.) Lemma 2.1 N 0 -almost all paths Q = {qt } are continuous and for any T > 0 they satisfy (2.6) |qt − qs | ≤ C|t − s|1/8 ∀s, t ∈ [−T, T ], where C = C(Q, T ) > 0 is independent of s, t. Moreover, N 0 -almost surely |qt | ≤ C1 (ln(|t| + 1))1/(α+1) + C2 , ∀t ∈ R,
(2.7)
where C1 > 0 is a constant dependent only on the dimension of the space, C2 = C2 (Q), and α is the exponent appearing in (P1), resp. (P2).
Proof. See [5]. The process qt is time-reversible. Its stochastic semigroup is generated by 1 Hpeuc f = − ∆f − ∇ ln ψ0 · ∇f. 2
(2.8)
Hpeuc is self-adjoint on the Hilbert space L2 (Rd , dN0 ). The map f → f /ψ0
(2.9)
from L2 (Rd , dx) to L2 (Rd , dN0 ) is unitary and transforms Hp − Ep into Hpeuc .
2.2
Euclidean quantization of the free field
We introduce the massless free field in d + 1 dimensions, d ≥ 3. For any f ∈ S(Rd+1 ), i.e. the Schwartz space over Rd+1 , we denote ξ(f ) = f (t, x)ξt (x)dtdx. (2.10) Rd+1
Vol. 3, 2002
The Infrared Behaviour in Nelson’s Model
275
According to the action (1.10) the quantum field {ξ(f ), f ∈ S(Rd+1 )} is described by the generalized Gaussian measure G on S (Rd+1 ) with zero mean and covariance covG (ξ(f1 ), ξ(f2 ))
= EG [ξ(f1 )ξ(f2 )] = (−∂t2 − ∆)−1 f1 , f2 ˆ f1 (k0 , k)fˆ2∗ (k0 , k) = dkdk0 . k02 + k 2 R Rd
(2.11)
Here fˆ1 , fˆ2 ∈ S(Rd+1 ) denote the Fourier transforms of f1 , f2 ∈ S(Rd+1 ). The scalar product is taken in the space L2 (Rd+1 , dtdx). G is a probability measure on the space S (Rd+1 ) taken together with its associated Borel σ-field [7, 15]. The particle trajectories are a stochastic process qt taking values in Rd . Following this example, it will be convenient to regard the field configurations ξt as a stochastic process taking values in a suitable Hilbert space to be specified below. To carry out this construction first note that the covariance (2.11) can be extended to test functions of the form ft (x , t ) = ϕ(x )δ(t − t) ϕ
(2.12)
with ϕ ∈ S(Rd ). Thus the random process ϕ ξt (ϕ) = ξ(ft ), ϕ ∈ S(Rd ), t ∈ R,
is well-defined. From (2.11) and (2.13) it follows that ˆ ∗2 (k) −|k||s−t| ϕ ˆ 1 (k)ϕ 1 e dk. EG [ξs (ϕ1 )ξt (ϕ2 )] = 4 Rd |k|
(2.13)
(2.14)
Clearly, ξt (ϕ) is a stationary Gaussian process. Its invariant (t = 0) measure G defined on S (Rd ) is again Gaussian with mean 0 and covariance given by (2.14) at s = t. Moreover, the process ξt (ϕ) is time-reversible [10]. Lemma 2.2 {ξt , t ∈ R} is a Markov process.
Proof. See [7, 16].
Next we construct a Hilbert space BD such that ξt ∈ BD and t → ξt is normcontinuous with probability 1. Let D be the positive self-adjoint operator in the space L2 (Rd , dx) given by a jointly continuous symmetric kernel ˆ )dk ˆ D(k, k )ξ(k (2.15) (Dξ)(k) = Rd
with ker D = {0}. We introduce the real Hilbert space BD ⊂ S (Rd ) with norm ˆ 1 )ξˆ∗ (k2 )dk1 dk2 D(k1 , k2 )ξ(k (2.16) ξ2BD = Rd ×Rd
276
J. L¨ orinczi, R. A. Minlos and H. Spohn
Ann. Henri Poincar´e
ˆ and ξˆ∗ (k) = ξ(−k). Clearly,
1 EG ξt 2BD = EG ξ0 2BD = 4 Hence, if
Rd
Rd
D(k, k) dk. |k|
D(k, k) dk < ∞, |k|
(2.17)
(2.18)
the measure G is concentrated on the space BD and the random process ξt takes its values from this set; for more details see [10]. By an easy calculation we obtain that 2 D(k, k) 1 4 −|k||t−s| (1 − e )dk + (2.19) EG (ξt − ξs BD ) = 4 |k| Rd 1 D(k1 , k2 )2 (1 − e−|k1 ||s−t| )(1 − e−|k2 ||s−t| )dk1 dk2 . 2 Rd Rd |k1 ||k2 | Thus if in addition to (2.18) we assume that D(k, k)dk < ∞ and D(k1 , k2 )2 dk1 dk2 < ∞,
(2.20)
then, using that |k|−1 (1 − e−|k||s−t| ) ≤ |s − t|, we obtain
EG ξt − ξs 4 ≤ C |t − s|2 .
(2.21)
Rd
From (2.21) and Kolmogorov’s criterion of continuity of random processes (see [10, 21]) we conclude Lemma 2.3 Under conditions (2.18) and (2.20) almost all paths of the process ξt are continuous in the metric of BD . Moreover, for any finite interval T > 0 these paths have the property that ξs − ξt BD ≤ C |t − s|1/8 , ∀s, t ∈ [−T, T ]
(2.22)
where C = C({ξt }, T ) is independent of s, t. Lemma 2.3 states that with probability 1 the process ξt is realized on C(R, BD ). We denote the corresponding Gaussian path measure again by G. Since the staeuc tionary Markov process ξt is reversible, its stochastic semigroup e−tHf acting on L2 (BD , dG) is self-adjoint and Hfeuc is the generator of this semigroup. Lemma 2.4 There exists a unitary map Fsym → L2 (BD , dG) transforming the free field Fock space Hamiltonian Hf into Hfeuc . This map can be constructed by using the Itˆo-Wick transformation, see [22].
Vol. 3, 2002
The Infrared Behaviour in Nelson’s Model
277
For concreteness we give an explicit form of the kernel D(k1 , k2 ) which satisfies (2.18) and (2.20) above. Consider the operator ¯ = (−∆k + |k|2 )−(d+1) . D
(2.23)
The eigenvalues of this operator are λn =
1 , (n1 + ... + nd + d/2)d+1
(2.24)
¯ is with multi-index n = (n1 , ..., nd ) of integer entries ni ≥ 0, i = 1, ..., d. Hence D ¯ a nuclear and positive operator, and its kernel D(k1 , k2 ) is jointly continuous [4]. Let |k|1 denote the function |k|, if |k| < 1, |k|1 = 1, otherwise. We set ¯ 1 , k2 )|k2 |1 . D(k1 , k2 ) = |k1 |1 D(k
(2.26)
It is then easily seen that this kernel satisfies conditions (2.18) and (2.20). For later use we also note that the level sets KC = {ξ ∈ BD : (D−1/2 ξ, ξ)BD ≤ C}
(2.27)
are compact in BD .
3 Path measure for the interacting system In the previous section we defined the stochastic processes qt and ξt governing the particle, resp. the free field trajectories. The joint field-particle process on the space C(R, Rd × BD ) is then given by the path measure P 0 = N 0 × G. In order to implement the interaction as in (1.12) we take P 0 as reference process and modify it with a density given by the exponential of Sint , compare with (1.11),
T 1 d PT = exp − (ξt ∗ )(qt )dt d P 0 , (3.1) ZT −T where
ZT =
exp −
T
−T
(ξt ∗ )(qt )dt d P 0
(3.2)
is the normalizing partition function. Since G is a Gaussian measure, by an easy computation we obtain
T 1 T |ˆ (k)|2 −|k||t−s| ZT = exp cos(k · (qt − qs )) e dtds dk dN 0 . 8 −T −T |k| d R (3.3) This integral is well-defined for every T > 0.
278
J. L¨ orinczi, R. A. Minlos and H. Spohn
Ann. Henri Poincar´e
We are interested whether the sequence of measures PT has a limit as T → ∞. The following notion of convergence will be used. Let M be a metric space, and C(R, M ) the space of continuous paths {Xt } with values in M . For any interval IT = [−T, T ] ⊂ R let MT ⊂ C be a sub-σ field of the Borel σ-field C generated by the evaluations {Xt : t ∈ IT }, i.e. MT = σ(Xt , t ∈ IT ). We say that a sequence of probability measures {µn } on C(R, M ) converges locally weakly to the probability measure µ if for any 0 ≤ T < ∞ the restrictions µn |MT converge weakly to the measure µ|MT , see [10]. We have then the following result. Theorem 3.1 Let d ≥ 3, 0 < e ≤ e∗ , with sufficiently small e∗ > 0, and V satisfy (P1). Then there exists the local weak limit PT → P as T → ∞. P is a probability measure of a stationary reversible Markov process on the path space C(R, Rd ×BD ). Moreover, the stationary distribution P of P can be obtained as the weak limit P = limT →∞ PT , where PT is the distribution of {q0 , ξ0 } under PT . Proof. For any fixed path Q = {qt : t ∈ R} ∈ C(R, Rd ) consider the conditional distribution Q (3.4) PT = PT ( · |{qt } = Q) viewed as a probability measure on C(R, BD ). Then the distribution PT on C(R, Rd × BD ) can be written as Q (3.5) PT (S × A) = PT (A)dNT (Q), ∀T > 0 S
where A ⊂ C(R, BD ) and S ⊂ C(R, Rd ), and NT is the Q-marginal of PT . Since the process ξt is Gaussian and the coupling is linear in it, this marginal can be computed explicitly and is given by
T T 1 exp − W (qs − qt , s − t)dsdt dN 0 , (3.6) dNT = ZT −T −T where W (q, t) = −
1 8
Rd
|ˆ (k)|2 cos(k · q) e−|k||t| dk. |k|
(3.7)
NT is a Gibbs measure for the finite interval [−T, T ]. The reversible diffusion process N 0 is the reference measure, and W is the (effective) pair interaction potential. We denote the distribution of q0 under NT by NT . We use identity (3.5) to show the existence of the T → ∞ limit of PT . We will do this by examining separately the limits of PTQ and NT , starting with the former measure. PTQ is a Gaussian measure with covariance given by (2.11) and mean ϕ ϕ(x)gTt (x; Q)dx, ϕ ∈ S(Rd ), EP Q [ξt ( )] = (3.8) T
Rd
Vol. 3, 2002
The Infrared Behaviour in Nelson’s Model
279
where gTt (x; Q)
= =
T ˆ(k) 1 cos(k · (x − qτ )) e−|k||t−τ | dkdτ (3.9) − (2π)d/2 −T Rd 4|k| 1 eik·x gˆTt (k; Q)dk (2π)d/2 Rd
and gˆTt (k; Q)
ˆ(k) =− 4|k|
T
e−ik·qτ e−|k||t−τ | dτ.
(3.10)
−T
This function has the properties |ˆ gTt (k; Q)| ≤
|ˆ (k)| 2|k|2
(3.11)
|ˆ gTt (k; Q) − gˆTs (k; Q)| ≤
C|ˆ (k)| |t − s| 2|k|
(3.12)
for any s, t ∈ R, |s − t| < 1, and with some constant C > 0. Furthermore, we have that for any t and Q lim gˆTt (k; Q) = gˆt (k; Q) (3.13) T →∞
exists in the norm topology of BD , uniformly with respect to Q and t in every bounded interval. The T → ∞ limit of (3.9) is given explicitly by ∞ ˆ(k) 1 cos(k · (x − qτ )) e−|k||t−τ | dkdτ. g t (x; Q) = − (3.14) d/2 (2π) −∞ Rd 4|k| From (3.12) it follows that gˆTt (k; Q) as a function of t belongs to C(R, BD ) and sup ˆ gTt ( · ; Q)BD < ∞
(3.15)
t
uniformly in Q and T ≤ ∞. Moreover, as follows from (3.11), for arbitrary Q, T , and t, gˆTt (·, Q) ∈ KC for some suitable C > 0, with KC defined by (2.27). Thus by (3.12) and (3.15) the set of restrictions gTt (·, Q), |t| ≤ R to any interval [−R, R], ∀Q ∈ C(R, Rd ), ∀T ≤ ∞, is a compact set in C([−R, R], BD ). Hence it follows that the restrictions PTQ |MR form a weakly compact set of measures. The characteristic functional of the Gaussian measure PTQ is iξ(f ) χQ Q [e ] T (f ) = EPT
∞
∞
fˆ(s, k)fˆ∗ (t, k) −|k||t−s| e dt ds dk = exp −1/8 |k| −∞ −∞ Rd ∞ dt gˆTt (k; Q)fˆ(t, k)dk . × exp i −∞
Rd
×
280
J. L¨ orinczi, R. A. Minlos and H. Spohn
Ann. Henri Poincar´e
Q Q From (3.13) it follows that χQ T (f ) → χ (f ) as T → ∞, where χ (f ) is the Q characteristic functional of the Gaussian measure P with covariance (2.14) and mean ϕ ϕ(x)g t (x; Q)dx. EP Q [ft ] = (3.16) Rd
Thus from the convergence of characteristic functionals χQ T and the weak comQ pactness of the restrictions PT |MR the weak local convergence limT →∞ PTQ = P Q follows for any Q ∈ C(R, Rd ), see [10]. In order to complete our argument we need to control the NT → N limit. This problem has been investigated in [13] by a cluster expansion technique and some of the results relevant in our context are quoted in Theorem 3.2 below. In particular, the local weak limit NT → N exists, as stated under (1) in Theorem 3.2. This then completes the proof of the first claim. Since PT has the Markov property and is reversible, the limit distribution P has the same properties. Stationarity of P follows by point (2) of Theorem 3.2 and the obvious equality d P θs Q (ξt−s ) = d P Q (ξt )
(3.17)
θs being the time-shift acting on paths. The existence of the weak limit P = limT →∞ PT follows by the argument above by borrowing statement (3) of Theorem 3.2 below. This completes the proof of the theorem. We conclude this section by stating some facts on N and related measures. Results on the existence and uniqueness of Gibbs measures relative to Brownian motion are available for some time [19], and a more general framework has been considered more recently in [13]. In the following theorem we list some results taken from [13] to be used below adapted for the case of Nelson’s model. For details we refer to this paper, in particular for results on the uniqueness of N which we do not discuss here. Theorem 3.2 For d ≥ 3, sufficiently small e > 0, and potentials satisfying (P1) the following properties hold: 1. There exists the local weak limit NT → N as T → ∞. 2. The measure N is invariant with respect to time shifts and time reflections. 3. The t = 0 distributions NT , resp. N, of the measures NT , resp. of N , are mutually absolutely continuous with respect to N0 , i.e. there exists a constant c > 0 such that 1 dNT ≤ (q0 ) ≤ c, (3.18) c dN0 uniformly in T and q0 . Moreover, dN dNT (q0 ) → (q0 ) as T → ∞ dN0 dN0 pointwise.
(3.19)
Vol. 3, 2002
The Infrared Behaviour in Nelson’s Model
281
4. For N0 -almost all q the conditional distribution NT ( · |q0 = q) converges locally weakly to N ( · |q0 = q). 5. N -almost all paths Q = {qt } ∈ C(R, Rd ) have the property that for all t ∈ R |qt | ≤ C1 (ln(|t| + 1))1/(α+1) + C2 (Q)
(3.20)
where C1 > 0 is a constant, and C2 is a function of the path. 6. For any bounded functions F1 , F2 on Rd the following estimate holds on their covariance: EN [F1 (qs )F2 (qt )] − EN [F1 (qs )]EN [F2 (qt )] ≤ C
sup |F1 | sup |F2 | |s − t|γ + 1
(3.21)
with some γ > 0 and a suitable prefactor C > 0.
4 The infrared divergent case : d = 3 After having introduced the path measures for the interacting system, we turn now to compare the Fock quantization and the Euclidean quantization of the interacting system. The first one is constructed through the Feynman-Kac formula, EP 0 [F (q0 , ξ0 )G(qt , ξt )e−
Rt 0
(ξs ∗)(qs )ds
] = (F¯ , e−tH G)H0 , t ≥ 0,
(4.1)
which defines the semigroup exp(−tH) on H0 = L2 (Rd × BD , dP0 ). Its generator, the Hamiltonian H, is unitarily equivalent with HN . For the Euclidean quantization we start with the semigroup Tt associated with the time reversible Markov stochastic process {qt , ξt } distributed according to P as defined through EP [F (q0 , ξ0 )G(qt , ξt )] = (F¯ , Tt G)H ,
t ≥ 0,
(4.2)
on the Hilbert space H = L2 (Rd × BD , dP). Tt is a symmetric contractive semigroup. Hence there exists a self-adjoint semibounded operator Heuc generating it, i.e. Tt = exp(−tHeuc ), which by definition will be viewed as the Hamiltonian of the system obtained in Euclidean quantization. N