This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
0. The lemma 2 is proved. 3. // o diagram with (n~,n+) external lines consists only from pairs of conjugated vertices then n~ = n+ and the diagram has on n+ connected parts. Each connected diagram has (1,1) external lines and is half-planar, i.e. it can be drown in the half-plane without self intersections. LEMMA 0}. Its dual space [S]*a is the space of generalized functions. By identifying (L2) with its dual we get the following continuous inclusion maps: 0 there exists a constant K > 0 such that \F(Z)\ 0. Hence 0. Thus y? 6 [£]a and the converse of the theorem is proved. ■ 3 3.1 (x), it is well known that it appears a stability condition which, after a simple change of variables of the type 00, there exists a critical value TJC (percolation thresh old) for which the overlapping disks form an infinite cluster "spanning the entire domain of volume V". The dimensionless factor T)c has been obtained by numerical simulations 13 ' 14 ?7C w 1.18 (\) | A € Spectrum} are square integrable functions. However, the integrals ( l M ) = y"dA
o € V. and a selfadjoint Hamiltonian H with domain D(H) dense in U, has a solution given by the one parameter group of (strongly continuous) unitary operators W{t) inH:2 4>{t) = U\t)4>0=e-iHt o o
We will not present here the simple proof of this lemmata. Now the theorem follows from the above three lemma.
32
5
Entangled Commutation Relations
5.1
Stochastic limit for N point correlators in one-particle states
The following theorem has been proved in 1 3 . THEOREM
2. The stochastic limit
lim Ax(p,k,t)
A-»0
= B~{p,k,t),
lim A$(p,k,t)
X-yO
=
B+(p,k,t)
exists in the sense of the convergence of the matrix elements (ti = ±) lim < 0\c(q)Ax1{pi,kutl)...Ax-(Pn,kn,tn)c+(q')\0
>
= (
(5.47) +
as distributions and the limiting operators B~ = B and B gled commutation relations B{p, k, t)B+ (p',k',t')
satisfy the entan
= 2n6(t - t')8(p - p')5(k - k') ■ 6{E{p, k))n(p)
[n{p'), B*(p, k, t)) = (±)(S(p' -p)-
S(p' - p + k))B*(p, k, t)
[n{p),n(p')}=0 s
(5.48)
(5.49) (5.50)
+
Here *o * ^ e vacuum in the new Hilbert space, B(p,k,t)YliC (qi)^o = 0. We use the same notations for the creation and annihilation operators of cparticles in the original and in the new Hilbert spaces. n(p) is the operator density of the c-particles, n(p) — c+(p)c(p). If we set B(p, k, t) = b(t) ® B(p, k) where b{t)b+{t') =
2n6(t-t')
then we obtain the relations (1.2) B{p, k)B+(p', k1) = n(p)8(E(p, k))S(p - p')S(k - k') The more general theorem will be proved in the next section. Note that to get non-zero in the RHS of (5.48) we have to choose suitable dispersion relations, so that there are non-trivial solutions of the equation E(p, k) — 0.
33
5.2
Stochastic limit for N point correlators in n particles states
In this section the main result of this paper will be proved. Let us consider the matrix element N
<*;,n^A*m)
(5-51)
t=l
where A\ and A% are given by (4.31) and (4.32) and m
n
c+(f) = Jc+(k)f(k)dk
, c(f') = J'c(k)f'(k)dk
Lemma 2 claims that if diagram corresponding to (5.51) doesn't consist only of pairs of conjugated vertices then its contribution vanishes in the limit A —> 0 (in the sense of distributions). One can see that to select diagrams corresponding to (5.51) such that all their vertices form pairs of conjugated vertices one has to consider diagrams that satisfy the following requirements: a) n = m, b) If n ^ 1 then the diagram must be disconnected and it must include n connected parts. Each connected part corresponds to Ni point correlator of composite operators between one particle stats and
c) Each connected part contains only pairs of conjugated vertices. To describe all diagrams satisfying above requirements it is suitable to do the following. Let us first apply the Wick theorem to represent the product A 's in normal form with respect to only of operators c and c+ N
Y[A£i = Y, i=l
:
KpA£i}(a,a+,c,c+)
:c
(5.52)
P
Here : : c means the normal product with respect to operators c and c + , and summation on p involves all terms that appear in the process of application of the Wick theorem. Applying the Wick theorem to
34
nc(/f)^nc+(/,-) i
j
we have f[c(f'i)f[Ac):f[c+(fi)= t
«=i
(5-53)
j=\
symbol j means that we take into account the diagram with / contractions (c-lines). For example, A+Ax = Klt+- + K2,+(5.54) Ki,+- = I dpdp'dkdk'a+{k)c+{p - k)c+{p')c{p)c(p' #2,+- = f
k')a{k')v(p,k)v{p'k')
dpdkdk'c+{p-k)a+{k)a(k')c(p-k')v{p,k)v(p,k)
(see fig. 5)
_^
<~
Figure 5: Friedrichs diagrams representation of (5.54)
For the product of AA+ we have AxAt = Ki,-++ K3,-+
(5.55)
where Klt-+ = fdpdp'dkdk'v(p,k)v(p',k')c+(p)c+{p'
- k')c(p - k)c(p')a{k)a+ (*')
K2,-+ = J dpdkdk'v(p, k)v(p -k + k', k')c+(p)c(p -k + k')a(k)a(k')
35
Figure 6: Friedrichs diagrams representation of (5.55)
(see fig. 5) To select the non-zero vacuum average of (5.53) that consists from dia grams satisfying requirements a), b) and c) we take the factorized products
I J {
£ j^ji=Ni=i
1
1
1
(5-56)
1
Now (5.56) is a sum of products of Wick monomials containing a and a+ operators. Diagrams corresponding to (5.51) that survive in the limit A -> 0 consist from n continuous c-lines connected one annihilator operator, some number of vertices and one creator operators. These c-lines enter in different connected parts of the same disconnected diagram. Beside c-lines the diagrams contain a-lines, that connect only vertices that are crossed by the same c-line. Each of connected part is represented by a half-planar diagram. This consideration proves the following T H E O R E M 3. The stochastic limit of the matrix element (5.51) can be repre sented as follows
im(oiTTc(/n (oi n cifj) n A? n c+^)i°>=
A-K)
= *«m £p
j=i
£
N
m
<=i
j=i
<°lc(/i)S£- • • • ^
c+(/P(i))|0>-
pe n "partitions"
•<0|c(/J)B e 'i .. . e ^ c + ( / p ( 2 ) ) | 0 ) . . . (0\c(fn)B^
. .. ^
c + (/ p ( n ) )|0>
Here Vn is the group of transpositions, the sum over "partitions" includes the sum overall subsets ( i i , . . . , ^ , ) ( j i , . . . ,jk2)... (lly... ,lkn) of {1,...,N} preserving the order and fci + &2 + ... + kn = N.
36
Conclusion The theorem 3 from the previous section describes the stochastic limit of com posite operators in n-particle states. It gives the second quantized general ization of interacting commutation relations from the previous works 5 , 1 3 . It would be interesting to find an operator expression of the obtained fomula as well as the corresponding path integral representation. Acknowledgments. I.Ya.A. and I.V.V. are grateful to the Centro Vito Volterra Universita di Roma Tor Vergata for the kind hospitality. This work is supported in part by INTAS grant 96-0698, I.Ya.A. is supported also by RFFI-99-01-00166 and I.V.V. by RFFI-99-01-00105 1. L. Streit, A new look at functional integration. In: Capri 1993, Proceed ings, Advances in dynamical systems and quantum physics, 307-325. 2. L. Streit, White noise analysis and functional integrals, in: "Mathemat ical approach to fluctuations", T.Hida, ed., World Sci., 1993 3. I.Ya.Aref'eva and I.V.Volovich. Large N QCD and q-deformed quantum field theories, Nucl.Phys.B 462 (1996) 600 4. M. B. Halpern and C. Schwartz, The Algebras of Large N Matrix Me chanics, hep-th/9809197. 5. L. Accardi, Y.G. Lu and I.V.Volovich, Quantum theory and its stochastic limit, Springer Verlag (in press) 6. L. Accardi, S.V. Kozyrev and I.V. Volovich, Dynamical q-deformation in quantum theory and the stochastic limit, J. Phys. A: Math.Gen.32(1999)3485-3495 7. I.Ya. Aref'eva and I.V. Volovich, Quantum group particles and nonarchimedean geometry, Phys. Lett. B268(1991)179-187 8. I.Ya.Aref'eva and I.V.Volovich. Quantum group gauge theory, Mod.Phys.Lett.A6(1991) 893-907. 9. I.Ya. Aref'eva and I.V. Volovich, Noncommutative Gauge Fields on Poisson Manifolds, hep-th/9907114 10. R. Vilela Mendes, Geometry, stochastic calculus and quantum fields in a non-commutative space-time, math-ph/9907001 11. L. Accardi, Y.G. Lu and I.V. Volovich, Interacting Fock spaces and Hilbert module extensions of the Heisenberg commutation relations, Pub lications of HAS, Kyoto,1997 12. L. Accardi, Y.G. Lu, Comm. Math. Phys. 180 (1996) 605 13. L. Accardi, I.Ya. Aref'eva and I.V. Volovich, Non-Equilibrium Quantum Field Theory and Entangled Commutation Relations, hep-th/9905035 14. V.S.Vladimirov, Equations of Mathematical Physics, Academic Press, 1986
37
D E R H A M COMPLEX OVER P R O D U C T MANIFOLDS: DIRICHLET FORMS A N D STOCHASTIC D Y N A M I C S SERGIO ALBEVERIO Institut fur Angewandte Mathematik, Universitat Bonn, Wegelerstr. 6, D 53115 Bonn; SFB 256 (Bonn); BiBoS Research Centre (Bielefeld); CERFIM (Locarno); Ace. Arch. (USJ) ALEXEI DALETSKII Institut fir Angewandte Mathematik, Universitat Bonn, Wegelerstr. 6, D 53115 Bonn; SFB 256 (Bonn); Institute of Mathematics, Kiev (Ukraine) YURI KONDRATIEV Institut fur Angewandte Mathematik, Universitat Bonn, Wegelerstr. 6, D 53115 Bonn; SFB 256 (Bonn); BiBoS Research Centre (Bielefeld); Institute of Mathematics, Kiev (Ukraine) We define de Rbam complex over a product manifold (infinite product of compact manifolds), and Dirichlet operators on differential forms, associated with differentiable mesures (in particular, Gibbs measures), which generalize the notions of Bochner and de Rham Laplacians. We give probabilistic representations for corresponding semigroups, and study properties of the corresponding stochastic dynamics.
1
Introduction
Various questions of stochastic analysis on product manifolds (that is, on infinite products of compact manifolds), have received a strong interest in resent times. This interest is strongly motivated by applications to models of statistical mechanics. Various aspects of the study of Dirichlet operators associated with Gibbs measures on product manifolds, and corresponding semigroups, were considered by many authors, see ', ! for a detailed review of the literature. Let us remark that the Dirichlet operator of a differentiable measure, which is actually an infinite dimensional generalization of Laplace-Beltrami operator, has a natural supersymmetric extension. Namely, we can consider Dirichlet operators on differential forms over a product manifold M, extending the notions of Bochner and de Rham Laplacians. The study of the latter operators and corresponding semigroups on finite dimensional manifolds was the subject of many works, and leads to deep results on the border of stochastic analysis, differential geometry and topology, and mathematical physics, see
38
e.g. 1 7 , 1 8 . In an infinite dimensional situation, such questions were discussed in the flat case in 10 , n , 12 , 6 . A regularized heat semigroup on differential forms over the infinite dimensional torus was studied in 13 . The study of such questions on product manifolds reflects the non-trivial interplay of measure theory and differential geometry (which is essential in an infinite dimensional setting). The present paper represents a step in this direction. The structure of it is as follows. In Section 2 we introduce main notations and geometrical objects. Section 3 is devoted to the definition of the de Rham complex over M . In Section 4 we introduce Bochner-Dirichlet and de Rham-Dirichlet op erators in the de Rham complex, and prove an analog of the Weitzenbock formula. Then we formulate the main result which characterizes Markov and hypercontractivity properties of the corresponding semigroups. Section 5 is devoted to the development of a probabilistic technique needed for the proof of this theorem. In particular, we consider SDE on tensor bundles over M and give with their aid probabilistic representations of semigroups. Section 6 is devoted to a discussion of the obtained results in the framework of Gibbs measures. We give here only main ideas of proofs. More extended presentation see in6. Dedication. It is a great pleasure for the authors to dedicate this work to Ludwig Streit on the occasion of his 60th birthday. His work on the interface of mathematical physics and stochastic analysis was always a great inspiration for us, and we are very grateful to him. Acknowledgments. We are glad to thank K. D. Elworthy, M. Rockner and A. Thalmaier for useful and stimulating discussions. The financial sup port SFB-256, DFG-research projects and INTAS-project No.378 is gratefully acknowledged. 2
Setting
Let A, B be Banach spaces and % be a Hilbert space. We will use the following general notations: - (•, •) - pairing of .4 and A', A' being the dual space; ( ) ' ) « - t n e scalar product in H; - tne norm m
— II'IU
-4;
- C{A, B) - the space of bounded linear operators A -* B; -C(A) = C(A,A); - Cn(A,B) - the space of bounded n- linear operators A -> B\ - HS{H, B) - the space of Hilbert-Schmidt operators H-+B.
39 Let M be a compact connected smooth JV-dimensional manifold. Let us assume that M is equipped with the Riemannian structure given by the operator field G(x) : TXM -f T*M, (•, -)T.M =< G(x); • >. The distance on M which corresponds to this Riemannian structure will be denoted by p. Let us consider the integer lattice Z d , d > 1, and define the space M, which is an infinite product of manifolds Mk = M: M =Mzi := xk€ZjJkfk 9 x = (xk)k€z<< • M is endowed with the product topology. Given A C Z M9iHi
A
(1)
d
= (zk)kGA € M A
(2)
A
denotes the natural projection of M onto M . We denote by Vk> the derivative of the function <j> w.r.t. the variable x k (which will be always identified with the vector field over M (gradient of
r
Let us remark that for u € TC {M) we have Vu € J C will use the notations < X(x),Y(x)
m_1
(4)
(M ->TM). We
>x= Y, (*k(*),*k(*))T„kM, divX = Yl div^X^, (5) kez* kez"4 dtVk meaning the divergence with respect to x k . The space M has a Banach manifold structure with the Banach space lb(Zd -*■ B.N) of bounded sequences Y = (yk)kez*,*k € KN, equipped with the norm imi„ := sup ||Yk||RW , (6) kez*
40
as the model. However, this norm being not smooth, one gets difficulties in using this manifold structure for the purposes of stochastic analysis. The way of overcoming of this difficulty was proposed in 3 , 4 . In these works, an analog of Riemannian structure on M was introduced. On a heuris tic level, the tangent space to M at the point x can be identified with the space x k 6 Z i T z k M . In order to define a differentiable structure on M , it is natural to consider some Hilbert subspace of x k 6 Z .iT x k M. Let h := h(Zd -* R^.) be the space of summable sequences p = (pk)kez,i of positive numbers. For a f i x e d p e l i let us define the space T P l , = I X 6 x kGZ< ,T xk Jlf : Y, P* H*kHr.kM < ° 4 , { kez<« J
(7)
equipped with the natural scalar product <*• y )p,» = £ **(*><> y " ) r . k M • (8) kez d The scalar product (X,Y) in the spaces T P)X wiU play the role of a Riemannian-like structure for M . The space M equipped with this struc ture will be denoted by M p . The bundle over M p with fibres TPiX will be called the tangent bundle of M p and denoted by TMp. The fibres TPiX will be denoted by T x Mp. The bundle T M P is not the tangent bundle to M p in the proper sense. Nevertheless T M P gives us the possibility to define analogues of various differentiable structures on M p . In 3 , 4 , the spaces C n (Mp -> B) of difierentiable mappings of M into a Banach space B, resp. C"*(Mp -> TMp) of differentiable vector fields over M , were introduced. That is, / € C x (Mp -> TMp), iff the matrix V/(ar) := (V,/ k (x)),
(9)
of partial (covariant) derivatives generates a bounded operator in T X M P , con tinuous in x. We will use the notation T M :=2 1 M 1 , where 1 is the weight sequence with elements Ik = 1. The space M p possesses the metric pp defined by
Pp{x,y)2 = 5Z P(*k,*k)9Jfc, kez d
(10)
which makes it a complete metric space. Let /i be a probability measure on M differentiable in the sense that the following integration by parts formula holds true: for any u G TCX (M -> R 1 ) ,
41
and any vector field X e TC1 (M ->TM) / J2 (Vku(x),X k (a;))T„ k M **(*) = - / / ? > ( * ) **(*), ^ kez* ^
(H)
with some /?£ € L 2 ( M , / J ) . /?£ is called the logarithmic derivative of n in the direction X. We assume that /3£ is given by 0x(«) = E (^W-■Xk(s))T. k M + dfoXkfr)), kez^
(12)
where /3M(x) = (/3k 0*0) e C 1 ( M P -»■ TM P ) for some weight sequence p£ We will call 0^ the (vector) logarithmic derivative of \i. 3
h.
Differential forms and de R h a m complex
The simplest differential forms over M are the forms with both image and domain consisting of cylinder elements. The space of m— times differentiable forms of the order m of such type will be denoted by .Ffi™. We will use the notation 7Tln := U m ^n™. Each v € .Ffi„ has the form v x
() =
v
53 ki
*i
k-(*)i
(13)
knCZi
where Ukx k„(«) € T ^ M A ... A T ^ M, and the sum is finite. Actually, each w G T£ln can be regarded as a differential form on the manifold MA for some finite A C Zd. For such forms, the symbols V* resp. A* will denote the covariant derivative resp. the Bochner Laplacian (Afcw(a:) := TrV^w(x)), w.r.t. X*. We will use the notation Vw = (Vfcw)fc€Zd, Aw = X) AjfcW. For w e TQ.n we have Vw e , H } n . We will use the notation <w{x),v(x)
>x=
^ ki
(win
k„(a;),Uk,
k.Wk^AfA.-.Ar^M.
k„&Zd
(14) Let / x b e a differentiable measure on M with the logarithmic derivative p. We introduce spaces of integrable forms L* iln as the completions of F£ln w.r.t. the norm \\-\\s given by
IMC = /
E («ki,...,k„(a0>t'ki i k,,...,k„6Z'
^ ( x ) ) ^ ^ MA...AT,k, Af
d/x(a;). (15)
42
In order to define de Rham complex over M, we need spaces of smooth forms. Given a Hilbert space )C, we define the space A"AC which is the n—fold antisymmetric tensor power of the Hilbert space AC. We introduce in the stan dard way the exterior multiplication A : A"AC x A m AC -»• A n + m A C ,
(16) 0 A tp:=^±I^.ASn+m((l> n'.ml
<8> VO,
ASn+m : ®n+mIC -* A n + m K. being the antisymmetrization operator, and creation resp. annihilation operators a*(Jfc):AnAC->An+1AC, k G K, (17) am(k)
resp. a(k) : An+1AC -»• AnAC
(18)
(the adjoint operator). Let us define the bundle A T M with fibres AT Z M. Sections of this bundle can be considered as differential forms over M (we identify the space T^M with TmM). Each section w of A T M can be represented as a sum of sections with cylinder images, w(*) =
5Z k,
«ki.....k.0O. <^i,....k„(*)€T^MA...Ar x k t M,
(19)
k„€Z<<
converging at each x in the norm of AT^M. We define the (covariant) deriva tive Vu> as an infinite block-operator matrix, Vw(s) : = (Vjc^ kl> ...,k I> (a:))i )kl) ... )k „ eZ , ,
(20)
where VtWk, k n (i) e C(TX\M,Tx]Si MA...AT lk> M) (if exists). We can now define the space C 1 (M P -+ AnTM) of differentiable forms, requiring that Vw(s) G C(TXMP, A n T x M),
(21)
and is bounded uniformly in x G M. Similarly we define the spaces C m (M p -¥ ATM). We wiU use the notation ft™(M„) := Cm(Mp -¥ A T M ) .
(22)
43
We introduce the exterior differential dn^CMp^n^CMp),
(23)
dnwix) = (n + l)ASn+1 (Vw(z)),
(24)
setting
where ASn+i : (£) n+1 TXM -»• A"+1 TXM is the antisymmetrization opera tor, and Vw(x) € C(TXMP -> A"TXM) is identified with the element of (AnTsM)
k„(aO = ailu...tit,(x)hx\Sl A ... A hx^ , hx^ € T^ Af,
akl
k„€C
m
(Mp-^R1).
We have as usual dnWk1 k.(ar) = da k l
k„(z) A / i ^ A ... A hx\^ , (25)
dnw = '^2
kn,
where daki,...,k„ is the 1— form defined by the derivative of the function "k,
k„-
Let us observe that d„ can be considered as unbounded operator I/£ft„ -*• L^H„ + i. We denote by d* : l£fl„+i -)■ l£ft„ the adjoint operator. The following fact follows in a straightforward manner from the integra tion by parts formula. Proposition 3.1 Let us assume that ^ e C^Mp -*• TMP). Then the oper ator e£ is densely defined, its domain of definition contains the space . H V n , and for u> € .Ffi„+i we have d*w € flJ,(Mp). In what follows, we will also use the spaces L"(M -* A"TM,,/j) of integrable forms with values in A n TM„ with the norm given by (JIHIA-T.M.
*<*))*•
44
4
Dirichlet operators on differential forms
In this section we define Dirichlet forms and Dirichlet operators in spaces of differential forms over M, generalizing the notion of Bochner and de Rham Laplacians. Let n be a differentiable measure on M such that the corresponding log arithmic derivative &P belongs to C1 (M p -t TM P ) for some weight sequence p € h. We introduce two Dirichlet forms on l£ft„. For «,« 6 T£ln we define the pre-Dirichlet form £f? resp. £* associated with /x, setting £«(«,») := f ^ ( V k u ( x ) , Vkt;(x))T.M dn(x), J
(26)
k
resp. £*(u, v) := (du, d«) t jn„ +1 + (<**". d * u )/.jn„_ 1 •
(27)
Analogously to the classical pre-Dirichlet form £M on functions (see e.g. 7 ), the forms £j? resp. £j* have generators Hjf resp. H£ acting in L2fi„ on the domain Tiln as H*u(x) = - E Aku(x) - ^(/3 k (a;), Vku(x))A»T.M®T.kAf, k
(28)
k
resp. H£ = dd* + d*d,
(29)
and are therefore closable. We preserve the same notations for their closures resp. generators of the closures. Let us observe that in the case of a single manifold M and fi a Riemannian volume measure on it, the operator Hjf resp. Hj* coincide with the Bochner resp. de Rham Laplacian, see e.g. , 15. We will call them the BochnerDirichlet resp. the de Rham-Dirichlet operator associated with p. Our next goal is to find a relation between the operators H * and H^. / .\i=i,-.Af
/ ,\i=i
N
Let f e^J be an orthonormal basis in T^M. Then f e£ 1 is obviously the orthonormal basis for T Z M, (for any weight sequence q). We set (4)* := aV k )> 4 :=«(«£)• Let us introduce the operator R-W= E
E
[^i.«(^k)(o k )X(a k )*oL]
k € Z d i,j,M,l
in A n T x M,, where Rij,i is the curvature tensor on M.
(30)
45
Lemma 4.1 Rn(x) is a bounded operator in each space AnTzM,, q arbitrary. Because of the assumptions on /?", we have V/3**(x) G £(T x Mp). Because of the symmetry of Levi-Civita connection, the operator V/J^x) is symmetric w.r.t. the scalar product of TZM, and therefore belongs also to £(Tj.M p -i). Let us define the operator [V/?"(x)]An := V/3"(i)®lO...®l (31) +1 ® V0M(x) ®... ® 1 +... + 1 ®... ® 1 ® V0M(x) in A n T x M. We have [V0"(x)]An G £ ( A T I M p - i ) and [V0*(x)]An € £(A n T z M p ). The following result is an extension of the classical Weitzenbock formula. Theorem 4.1 For u G ?Sln H««(i) = H*u(x) + Rn{x)u(x) - [V/3"(x)]An u(x).
(32)
n
Let us introduce the operator R£(x) € £(A T I M p ), R>(x) = Rn(x)-[V0"(x)]An. (33) Formula (32) implies that R£(x) does not depend on the choice of the basis in (30). Both Hjf resp. H * are non-negative self-adjoint operators in L2ft„ and generate therefore strongly continuous contraction semigroups Tff(t) resp. T*(t) in it. The following result characterizes the properties of these semi groups. Theorem 4.2 Let us assume that 0" 6 C4(MP -► TMP). Then : 1) both Hjf and H* are essentially self-adjoint operators on TCPQA); 2) for any weight sequence q the semigroup T^(t) leaves invariant the space C{M -¥ A"TM,), and for v G C(M -¥ ATM,) we have II T J?(*)«(*)|IA-T.M 1 = T "W ll«HA-TM, (*)!
(34)
3) for the weight sequence q = p or q = p _ 1 the semigroup T*{t) leaves invariant the space C(M -t ATM,), and for v G C(M -> ATM,) tue /w«e ||T«(t) V (x)|| A „ T>Mq < e ^ T ^ t ) HHIA.TM, (*).
(35)
where r is such that r • (h,h)A„TmM < (R£(x)h, /I)A»T,M, for each h e AT*M, and x G M; 4) let us assume that there exists a constant ro is such that ro ■ (h, fe)A,TiM < (RJJ(X)/I,/I) A „ T M M for each h G AT^M, and x G M; then the semigroup
46 T£(*) leaves invariant the space C(M -+ ATM), and for v G C(M -4 A"TM) we have ||T?(*M*)|| A ,. T . M < e-tr" T„(t) ||t,|| A „™ (*).
(36)
The prooffollowsfrom a probabilistic representations of the semigroups Tjf(i) and T*(i), which will be given in Section 5. Let us recall that a semigroup T(t) (acting on functions on M) is called hypercontractive, if for all 1 < s < a < oo T{t) : L'(M,M) -*• L"(M,n),
(37)
and T(t) is a contraction when ^
> e" 2tA
(38)
for all t > 0 and some A > 0. Formula (34) emphasizes a certain Markov property of Tjf (t) and implies its hypercontractivity provided TM is hypercontractive. An analogous state ment for T*(£) requires positivity of R£. The following result is the corollary of Theorem 4.2. Theorem 4.3 Let us assume that in the framework of Theorem 4-2 r > 0 resp. r0 > 0. Then: 1) the semigroup T*(i) is Markov in the sense that for q = p orq = p~l resp. g= l
IW<Mz)| A „ T . M , < T„(«) IMU-™, (*);
(39)
2) the semigroup T*(<) is contractive in each L*(M -> A" TM.g,fi), where q = p - 1 resp. q = 1; 5^ «/TM(t) is Aj/peram£racttve, Men T%(t) is also hypercontractive (with the same X), in the sense that T*(t) : L'(M -► AnTMq,(i) -> L a (M -+ AnTM,,/x), where q=p~l 5
(40)
resp. q = 1, ond T^(t) is a contraction when (38) holds.
Probabilistic representations of semigroups
The aim of this section is to obtain probabilistic representations of the semi groups Tjf(t) and T*(i), and to prove with their aid Theorem 4.2. The construction of the diffusion process in M, which gives the stochastic dynamics associated with the classical Dirichlet form £M, is given in 2 - 4 . This process was constructed as the strong solution to an infinite system of SDE
47
d(k(t) = &«(*))# + P(&(t)) o dufcft), k € Z d ,
(41)
in the Stratonovich form on M, where w^, k € Zd, are independent Wiener processes in a given Euclidean space R n such that M c R " isometrically, and P(m) : R n -¥ TmM is the orthoprojector. We will also write this system in the form of one SDE #(*) = J ^ K W ) * + P(«*)) o dw{t)
(42) d
on M, where w is the cylinder Wiener process in /C := /2(Z -► R"), and P(a;) is the block-diagonal operator K. -> r x M with diagonal blocks Pkkfa) = P(xk). It is easy to see that A(x) £ HS(K.,TXMP) for each p eh. Theorem 5.1 (3, 4) Let &> € Cl(Mp -»> TM P ). TAen fAe SDE (42) has a unique solution £x(t) for any initial data x € M. £x{t) depends continuously (in the mean square sense) on the initial data x. The corresponding generator coincides on TC1 (M) with the Dirichlet operator of measure /*. Our next goal is to construct the parallel translation of differential forms along solutions of (42). For this, we need an analog of Levi-Civita connection onM. Let OM be the orthonormal frame bundle over M. The space OM has obviously the structure of a compact manifolds which fits into the framework of previous sections. We define the product manifold 0M:=x k 6 Z d0M.
(43)
This space will play the role of orthonormal frame bundle over M, with fibres O s M:=x k6Z
(44)
We denote by ir: OM -¥ M the corresponding projection. Given the Levi-Civita connection on OM, we equip OM with the product connection. This means the following. Let us consider the tangent bundle T(OM)p to OM. By the definition, for each z G OM T,(OM) p = © k 6 Z , VPLT^OM).
(45)
We define the horizontal tangent space ffT^(OM)p resp. the vertical tangent space VTz(OM)p as HTz(OM)p:= ©keZ<< JpZHTzk(OM), (46) VTz(OM)p:= © k 6 Z ,
^pZVTzk(OM),
48
where HT^OM) resp. VT^OM) is the horizontal resp. vertical tangent space of OM associated with the given connection. We have obviously the decomposition Tz(OM)p = HT,(OM)p © VTz{OM)p.
(47)
Let us observe that the corresponding covariant derivative of a vector field X G Cl(Mp -+ TM,,) coincides with the derivative dX defined in above. We denote by X the horizontal lift of the vector field X over M to OM. It follows directly from our definition of HTz(OM) that (*(*)) k = £ ( * k ) ,
(48)
where Y* is the horizontal lift of the vectorfieldY* := Xk(n(z\)) over M, A = Zd\ {k}. We have then obviously that X G C n ((OM) ->• T (OM) ) provided X G C n (M p -*• TMP). Let now £(t) be the Brownian motion with the drift a G C 1 (Mp -¥ TMP) described by SDE (42). We consider its horizontal lift ~f(t) G OM defined by the SDE dj(t) = a(j(t))dt + P( 7 (t)) o dw(t),
(49)
where P(z)h := P(z)h, with initial data 7(0) such that TT(7(0)) = £(0), which is by Theorem 5.1 uniquely solvable. As in the case of a single manifold M (see 17) we have n(g(t)) = £(t). Given n > 1 and a weight sequence q of positive numbers (not necessarily decreasing), we consider the bundle V := An TMq over M. Let v G V x and define by p(g)v the natural action of g G O {TXM) := xkeZ*0 (TX\M) on Vx, O (TmM) being the space of orthogonal linear transformations of TmM. We can now define the parallel translation p
W)--Vil0)-+Vm
(50)
along the solution £(t) of (42) as **«*)« = PWMO)- 1 )*, (51) where j(t) solves (49). The transformation P^ is obviously orthogonal. Let J be a continuous operator field on M, J(x) G C(VX). We define the mapping J : OM -> V€(0) by the formula J(*) := p ( « b ~ M - , ^ M * ) ) p ( * 0 . «b = 0(0), and consider the equation
(52)
<*?(*) =
(53)
J(l(t)Ht)dt
49 in Vf( 0 ). This equation is uniquely solvable for any initial data by general theory of SDE in Hilbert spaces (see e.g. 1 6 ), and defines a multiplicative functional of the process f(t) and consequently of £(t). Let us consider the semigroup T*'J(t) acting in the space C(M -* V'), which is defined by the expression < T^J(t)w{x),y
> = E < w(^{t)),PUt)Vv(t)
>,
(54)
where nv(t) is the solution to (53) with the initial condition y, and denote by Hi,J its generator. Proposition 5.1 Forut T(?$A -> V ) we have &<Ju{x) = l&u{x)+
< a ( i ) , V u ( i ) >x +Jm{x)u(x).
(55)
st
Proof. Let us first assume that a = 0 and J = 0. This case reduces obviously to the case of a single manifold, and the corresponding generator is equal to \ A by the finite dimensional theory 1T. In the general case, applying Ito formula to the function $(z,v) =< u(n(z)), P(ZZQX)V > on O M x V x , we obtain the additional first-order term V , * ( z , « ) [3(*)] + V.*(z,t>) [J(z)v] = < Vu(*(*)) [a(ir(x))],p(zz^1)v + < u(-w{z)),J(z)p(zZo1)v
> (56)
>,
which implies the result. The following statement follows in a straightforward way from the Ito formula and Gronwall inequality. Proposition 5.2 The semigroup T^'J(t) satisfies the following estimate: \\T*'J(t)v(x)\\{yl)m
< e tc T<(*) ||«|| v , (x),
(57)
v 6 C ( M -» V ) , where c is such that cl > J(x) for each x, as operators in V,. Remark 5.1 Let us consider the bundle W := A n T M , i , where the weight sequence q1 is such that q£ < ?k for each k. Then we have V x C W x for each x e M . Let us assume that there exists a constant c\ such that Cyl > J(x) for each x, as operators in Wx. Similar arguments show that the process n(t) satisfies also the estimate
lln(tM0||W{(1)=e*Mb?(0)llw.0,
(58)
50
which implies that T €,J (f) leaves invariant the space C(M -► W ) , and for v € C(M -► W ) we have ||T € ' J (*)«(*)|| (W , ) . < e te ' I*(t) || V || W , (x).
(59)
Now we can construct probabilistic representations for Bochner and de Rham semigroups and to obtain with their aid a proof of Theorem 4.2. Let us observe that on TQ,n we have H« = -H«,
(60)
where H € is the generator of the parallel translation along the paths of the "stochastic dynamics" process f associated with /i, defined by SDE (42). If ( - H ^ , ^ n „ ) is essentially self-adjoint, the semigroup T^(i) has a simple probabilistic interpretation. Indeed, in this case Tjf(i) is the unique semigroup with the generator (—H^,^fi„) , and therefore for w € C(M -»• AnTM) we have T«(t)u;(x) = E(P t M& (*))).
(61)
In order to obtain a probabilistic representation for the semigroup T^ associated with H*, let us observe that for w € FSln we have H«'- R -w = -H*w.
(62)
The corresponding semigroup T*' -R S (t) leaves invariant the space C(M -»• An TMp-i), and coincides on this space with T^, provided H* is essentially selfadjoint on ^ n „ . The latter fact holds true if /?" € C 4 (M P -> TM,) (in this case both semigroups T*(f), T* , - R - (t) leave invariant the space C 2 (M -> An TMp-i), which obviously contains in the domain of Hjf and H*). This implies Theorem 4.2. 6
Stochastic dynamics for lattice models associated with Gibbs measures on product manifolds
Let us consider a family of potentials U = {UA)A€{I, U\ 6 (?(AfA). Let ft(k) be the family of all sets A G fi containing the point k G Zd. We will assume the following: (Ul) £
sup |l/ A (x)| < oo
Aen(k)* 6 M
(63)
51
for any k € Zd. Let n be a Gibbs measure on the Borel er-algebra B(M), associated with the family of potentials U. We denote by G{U) the family of all such Gibbs measures. G{U) is non-empty under the condition (63), see e.g. 19, 18. Let us now assume that the family of potentials U satisfies (in addition to (Ul) the following conditions: (U2) Us € C 1 (M A ) for each A, and sup V |||V k tf A |||™ < oo, k
(64)
6zdA6n
where |||Vk[/A|||TM :=sup l6M ||V k £/ A (a;)|| TiikM ; (U3) U\ G C2 (A) for each A, and there exists C < oo such that SU
P E E IIIVjVkt/A|||TM®TM < C, (65) Jez<
SU
P E
k6Z 1
Elll V Ji V J' VkC/A IH® 4 ™ <00 '
' JiJj6Z''A6n
52
(66) 8U
P
£
V
V
£lH J'- J«
Vk:/A
8
HI® ™
where ll|Vj1...VjmVkyA|||®-.+iTi»#:= 8U
P liyi,...Vj m V k C/A(s)||T.S Af®-®T.; Af®T,UM-
This condition implies obviously that there exists a weight sequence p € h such that /?" 6 C 4 (M„->rM p ). We summarize our discussion in the following Theorem 6.2 Let the family of interactions U satisfies (Ul), (U2), (US) and (66), and \i e G{U). Then there exists a weight sequence p€ h such that the statements of Theorems 4-2 and 4-S hold true. Remark 6.2 An application of an approximation technique similar to the one developed in3,7 gives the possibility to prove the essential self -adjointness of operators Hjf and H * on FSln only under the conditions (Ul), (U2), (US), and therefore to avoid the additional condition (66) in Theorem 6.2. References 1. S. Albeverio: Some applications of infinite dimensional analysis in math ematical physics, Helv.Phys.Acta 70 (1997), 479-506. 2. S. Albeverio, A. Daletskii, Yu. Kondratiev: Infinite systems of stochastic differential equations and some lattice models on compact Riemannian manifolds, Ukr. Math. J. 49 (1997), 326-337. 3. S. Albeverio, A. Daletskii, Yu. Kondratiev: Stochastic analysis on prod uct manifolds, in the book: "Stochastic Dynamics", ed. H. Crauel and M. Gundlach, Springer 1999 (Proceedings of the conference "Random Dynamical Systems", Bremen, April 28 - May 2, 1997). 4. S. Albeverio, A. Daletskii, Yu. Kondratiev. Stochastic equations and Dirichlet operators on product manifolds, Univ. Bonn, Preprint No. 591 SFB 256 (1999). 5. S. Albeverio, A. Daletskii, Yu. Kondratiev: Stochastic analysis on prod uct manifolds: Dirichlet operators on differential forms, Univ. Bonn, Preprint No. 598 SFB 256 (1999). 6. S. Albeverio, Yu. Kondratiev: Supersymmetrie Dirichlet operators, Ukr.Math.J. 47 (1995), 583-592.
53
7. S. Albeverio, Yu. Kondratiev, M. Rockner: Uniqueness of the stochastic dynamics for continuous spin systems on a lattice, J.Func.Anal., 133, No.l (1995), 10-20. 8. S. Albeverio, Yu. Kondratiev, M. Rockner: Quantum fields, Markov fields and stochastic quantization, in the book: Stochastic Analysis: Mathematics and Physics, Nato ASI, Academic Press (1995) (A. Car doso et al.,eds). 9. S. Albeverio and M. Rockner: Dirichlet forms on topological vector spaceconstruction of an associated diffusion process, Probab. Th. Rel. Fields 83 (1989), 405-434. 10. A. Arai: Supereymmetric extension of quantum scalar field theories, in Quantum and non-commutative analysis, H.Araki et al. (eds.), 73-90, Kluwer Academic Publishers, Holland (1993). 11. A.Arai: Dirac operators in Boson-Fermion Fock spaces and supereym metric quantum field theory, Journal of Geometry and Physics 11 (1993), 465-490. 12. A.Arai, I.Mitoma: De Rham-Hodge-Kodaira decomposition in oo-dimensions, Math. Ann.,291 (1991), 51-73. 13. A. Bendikov, R. Leandre: Regularized Euler-Poincare number of the in finite dimensional torus, to appear. 14. Yu. M. Beresansky and Yu. G. Kondratiev: "Spectral Methods in Infinite Dimensional Analysis", NaukovaDumka, Kiev, 1988.[English translation: Kluwer Akademic, Dordrecht/Norwell, MA, 1995]. 15. L. Cycon, R. G. Froese, W. Kirsch, B. Simon: "Schrodinger Opera tors with Applications to Quantum Mechanics and Global Geometry", Springer, 1987. 16. Yu. L. Daletcky and S. V. Fomin: "Measures and Differential Equations in Infinite-Dimensional Space", Kluwer Academic, Dordrecht, Boston, London, 1991. 17. K. D. Elworthy: Geometric aspects of diffusions on manifolds, Lecture Notes in Math., 1362, Springer Verlag, Berlin and New York, 276-425 (1988). 18. V. Enter, R. Fernandez, D. Sokal: Regularity properties and Pathologies of Position-Space renormalization-group transformations, J. Stat. Phys. 2, Nos. 5/6 (1993), 879-1168. 19. H. O. Georgii: "Gibbs measures and phase transitions", Studies in Math ematics, Vol. 9, de Gruyter, Berlin, New York (1988). 20. L.Gross: Hypercontractivity and logarithmic Sobolev inequalities for Clifford-Dirichlet forms, Duke Math. J., 43 (1975), 383-386 .
54
REAL TIME R A N D O M WALKS ON P - A D I C N U M B E R S SERGIO ALBEVERIO Institut fur Angewandte Mathematik, Universitdt Bonn, Bonn, Germany WITOLD KARWOWSKI Institute of Theoretical Physics University of Wrocllaw, Wroclaw, Poland Dedicated to Ludwig Streit on the occasion of his 60 birthday The field of p-adic numbers is a complete metric space under a non-Archimedian metrics. For this reason it is a suitable mathematical framework for the description of hierarchical phenomena appearing in many disciplines. Here we summarize some investigation of the real time random walks on p-adic based on the properties of transition functions obtained by solving the Kolmogorov equations.
1
Introduction
The last two decades have seen a steady growing interest in p-adic analysis. A substantial part of mathematical investigations has received momentum from physical motivations. There have been two main physical ideas leading towards application of local fields and in particular the p-adic numbers 1 . One came from particle physics. It was based on the conjecture that the space time at Planck distances may have a structure better described by a field other than the one of the real numbers. This has given motivation to exten sive mathematical work which resulted in elaborating interesting structures going under the names of p-adic quantum mechanics and p-adic quantum field theory 2,3 . In this paper we will relate mainly to another idea coming from statisti cal physics, in particular in connectioin with models describing relaxation in glasses. The non exponential nature of those relaxations has been interpreted as a consequence of a hierarchical structure of the state space, which can in turn be put in connection with p-adic structures. In fact p-adic numbers can be interpreted as having a "tree structure". In the physical literature various models of "diffusion" or random walks on (finite or infinite) trees have been investigated 1,15 . In particular L. Brekke and M. Olson4 constructed a class of random walks on p-adic numbers and discussed the relation with relaxations in glasses. This and pure mathematical interest motivated systematic studies of ran-
55 dom walks indexed by continuous real time with the field of p-adic numbers as state space. This note is intended to present some methods and results ob tained in this direction. The Levy processes on p-adic numbers have been con structed by different methods see Evans 16 and Figa-Talamanca 5 , but we shall be mostly concerned with the approach developed by Albeverio, Karwowski7,8 and later extended by Karwowski, Vilela-Mendes 9 , Albeverio, Karwowski, Zhao 10 and Albeverio, Karwowski, Yasuda 13 . For the relation betrween Figa-Talamanca 5 and Albeverio, Karwowski 7|8 see Husssmann 6 . In Section 2 we give basic properties of p-adic numbers and show how the translation and rotation invariant Levy processes are obtained by solving the corresponding Kolmogorov equations. We also present spectral properties of the generator. In Section 3 we generalize this method to cover corresponding processes taking values in weighted state space. In Section 4 we discuss conditions for a process to be recurrent and also hitting and exit times for p-adic balls. In Section 5 we discuss a p-adic trace formula and its analogy with the Selberg trace formula. 2
Levy Processes on Qp
We begin with basic definitions and facts about p-adic numbers 14 . Let p > 1 be a prime number. A p-adic number a is associated with the formal power series oo
a=J2^iP^
(2-1)
i=N
where AT is an integer and a,- = 0 , 1 , . . . , p - 1. With addition and multiplica tion defined in the natural way for formal power series the set Qp of all p-adic numbers becomes a field. Given a G Qp, set i0 for the smallest value of i in the sum (2.1) for which at ^ 0. Then we put
NI P =p- i 0 .
(2-2)
The map a -+ ||o|| p defines a norm in Qp, which has the non Archimedian triangle property ||a + % ^ m a x { | | a | | p , | | 6 | | p } ,
(2.3)
56 and Qp with this norm is a complete separable locally compact totally dis connected space (with the cardinality of the continuum). The series (2.1) converges with respect to the || || p norm. Let a G Qp, Mel, then the set K(a>PM)
= {x € Q P ; \\a - x\\p ^ pM} M
is called a sphere (or a ball) of radius p consequences of (2.3) i) If x e K{a,pM)
then K{x,pM)
=
(2.4)
centered at a. We note the following K(a,pM)
ii) If at e Qp, 1=1,... ,p, \\a, - ak\\p = pM+\
k # I then ^
K{ahpM)
=
M+1
K(ak,p
)
M
in) Let /C be the family of all disjoint balls of radius pM. Then K.M is countable and writing K.M = {Ki}^ we have U Ki = Qp. iv) K(a,pM)
is open and compact.
Let £ denote the er-algebra generated by the family of all balls in Qp. Then the set function /z defined on the balls by M
(K{a,pM))
=pM,
a € Qp, M <E Z
(2.5)
can be uniquely extended to a measure on £ ) p also denoted by /*. It is a Haar measure for the additive group in Qp. Put G+ for Qp as an additive group and G. = {x € QP, \\x\\p = 1}
(2.6)
for Qp \ {0} as a multiplicative group. Then G. defines a group of automor phisms Qz, z E G* of G+ by Qza = za, a € G+. We put G for the semidirect product of C?» and G+ relative to 0 : G = G. x e G+
(2.7)
i.e. G = {ff€ [z,a]:zeG*,
a€G+}
(2.8)
and 5i52 = [z\z2,ziai
+02].
(2.9)
Defining action of G on Qp by 0Z = [z, a]x = zx + a,
x e Qp, g € G
(2.10)
57
we find that G is a doubly transitive group of isometries for Qp We also have (see 1 3 ): Proposition 2.1 /z is the Haar measure for the group G. There are different starting points possible for the construction of random processes. Before presenting the approach proposed in Albeverio et al7-8 we briefly mention two other a) Levy-Khinchine representation 12 ' 16 . Put {x} for the fractional part of x G Qp. The normalized additive character on Qp (G+) is defined by X{x)= exp(27ri{x}) and then the Fourier transform of a complex valued function <j> e L1 (Qp, fi) is defined by
m = f x&MzMdx) Any Levy process on Qp can be defined by the LeVy-Khinchine formula
*(t,fl = exp< tj\x(xO-lMdx)\ where v is a cr-finite measure on £
(2.11)
satisfying v({x € Qp; \\x\\p > pM}) <
oo, M € Z. v is called a Levy measure and $(t,£) is the Fourier transform of the transition function for a Levy process. b) The generator of a Markovian semigroup 17 ' 18 . Consider the real space L2(Qp\n). Then the notion of Dirichlet form and Markovian semigroup are well defined and their relations with the Hunt processes follow from general theory 19 . Vladimirov, Volovich, Zelenov and Kochubei 2,17 ' 18 have taken as the starting point the generators of the Marko vian semigroups Da, a > 0 defined on the space D(QP) of locally constant functions with compact support by the formula
Da
+w = g lp-C-i / w* - f) - *(*)] Wvw;0"1^) ■
(2-12)
Qy
R e m a r k 2.1 Recently H. Kaneko22 modified formula (2.12) and obtained new random walks which apparently do not belong to any class discussed in this note.
58
The construction proposed in Albeverio et al 7 ' 8 is based on the Kolmogorov equations. We begin by taking KM for the state space. Then the system of forward and backward Kolmogorov equations reads PKiKi (*) = -a{Kj)PKtKi
(*) + £
u(Kf, Ki)PKiK,
(t)
(2.13a)
resp.
PKtKi(t) = -a{Ki)PKiKi{t) + Y,^K^Kf)P^f^t>>
(2-13b)
/#* with t > 0, i,j € N and the initial condition P K ^ . (0) = dy. a(Kj) is interpreted as intensity of the state Kj and u{K^Kj) as the infinitesimal transition probability from the initial state K{ to the target state Kj. Till now K,M is just a countable set. Its non Archimedian structure must be imprinted in the coefficients u(Ki,Kj). We shall do it as follows We say that a sequence of real numbers A = {a(n)}„ 6 z belongs to the class A iff a) b)
a(n) ^ a(n + 1), lima(n)=0.
(2.14)
n—>oo
Let A G A. If for any M e Z and m € N put U(M) = a(M - 1) - a(M) and u{M,m) = (p - l ) - 1 ? - " " - 1 C/(M + m ) . MKif\Kj
= 0 then dist p (J^i, Kj) = pM+m for some m € N. Then we define «(lfi,iir i ) = «(M,m)
(2.15)
a(Kj) = Y,u(Ki,Kj)
(2.16)
Requiring further
we immediately obtain a(Kj) = a(M).
59
The equations (2.13) with the coefficients specified by (2.15) and (2.16) can be explicitly solved. Further by shrinking the initial ball to a point while keeping the target ball fixed one arrives at the formulas which can be extended to the transition function of a random walk with Qp as the state space 7 ' 8 . Namely oo
Pt(x, K(a,pM))
=p-1{p-
l)53p-»exp{-T M .H*}
(2.17a)
t=0
if x eK(a,pM)
and
M\\ Pt(x,K(a,pM))=
— p—m
P'1 (P - 1) 5 Z P~* exp{-TAf+m+< *} «=0
- exp{-T W + m _i t}
(2.17b)
if dist p (x, K(a,pM)) = pM+m. These formulas hold for all x,a € Qp, M G Z. We put here rn = ( p - l ) _ I [ p a ( n ) - a ( n + l)] ,
neZ.
(2.18)
The transition function Pt{x,B) (t > 0 , i £ Qp, B G £ ) determined by the formulas (2.17) can be seen as the integral kernel for a Markovian semigroup (Tt,t>0)mL2(Qp]ti). Set Tt — e~Ht. Then H is a positive self adjoint operator in this space. Its action on the indicator functions for the balls can be given explicitly, which provides sufficient information for the spectral analysis of H. It turns out that the spectrum of If is essential pure point
<*{B) = {M B E z • The corresponding eigenvectors are defined as follows. Given a e Qp and n € Z. Put a.1 = a and choose 0 2 , . . . , o p so that ||a* - a.j\\p = p" i ^ j i,j = l,...,p. Let further Xi(x) stand for the indicator function of K{a,i,pn). Then for p
any real numbers 6 1 , . . . bp such that £ bi = 0 we have _•
1
y
v
t=l
i=l
60
The Dirichlet form £ corresponding to H has only the jump part in the Beurling-Deny representation i.e.
W,9) =
I
(/(*) - /(»)) (g{x) - g(y)) J(dx,dy)
Q„xQp\d
for all f,g € D[£]. J is a symmetric positive Radon measure on Qp x Qp supported off the diagonal. J is uniquely determined by J(K{a,pN),K(b,pM)) if &stp(K{a,pN),K{b,pM))
= ^"-"^[ain =pn
- 1) - a(n)}
(n > maxM,N)
J(K(a,pN),K(a,pN))=pNa(N) Since G (2.7) is a group of isometries it follows that the transition functions (2.17) are G invariant. This results in the following Theorem 2.1 2 0 The class of random walks on Qp constructed above is iden tical with the class of all G« symmetric L6vy processes defined by (2.11). It naturally includes the processes defined by {Da,a > 0} (2.12) with the fol lowing relation between a and {o(n)} n 6 z
3
n(n\ P-1 P ^ a(n) = . p l-p-<*-i Random Walks with Weighted Target States
The procedure presented in Sec. 2 can be generalized to yield a wider class of random walks which are not necessarily LeVy processes. In Karwowski et al 9 a class of processes which are not necessarily LeVy processes has been constructed. Similarly as in Albeverio et al 7 ' 8 one begins by solving the Kolmogorov equations (2.13), but the coefficients u(Ki,Kj) are denned differently. Let p be a nonnegative locally Ll{y) function. For B 6 £ ) p we set PB = fp(x)n(dx). Taking {a(n)}„ 6 Z G A we put B
u(Ki,Kj)=pK.U(M
+ m)
where as before U(N) = a(N - 1) - a(N) and distp(Ki, K5) = pM+m, m G N. Since in our notation Ki is the initial state and Kj the target state we see that in this formulation the target states are weighted. The method of solving equations (2.13) developed in Albeverio et al 7,8 has been modified in
61
Karwowski et al9 to also cover this case. To formulate the results we shall need more notation. For any a G Qp and n e Z put p*=
j
p{x)p.{dx).
(3.1)
K(a,p") n
Note that if p ^ ||a||p then K(a,pn) = K(0,pn) and p£ = p°. For the target state Kj = K(b,pM) we define -WhM = a{Ki) = Y,*(Ki*Ki)
(3-2)
and - W ' = X > ( » + * " 1) " a(" + * + 1)K+* • Jfc=0 10
These quantities are finite
(3-3)
iff
«=1
The formulas for transition functions depend on whether / p{x)p,{dx) is finite or not. In the first case we normalize p so that / p(x)p,(dx) = 1. Then Q,
Pt (x,K(a,pM))
=^
{ l + f ; ( 1 - J - ) exp {Mfc.,}} P I i=M ™ <+ l / J
(3.4a)
if x€ K{a,pM) and Pt (x,K(a,pM))
= p%, \l + f ) ( 1 - - U exp {*W?+1}
-i-exp{
(3.4b)
if distp(x, K(a,pM)) =pn,n> M. If J p(x)n(dx) = oo then OP
P t (*, K(a,pM)) = p% f ; ( 1 - J - ) exp {tW?+1}
(3.5a)
62
if xtK(a,pM)
and
P^tfCa.p^-pdfM i - ^ r - ) « P {***+!} v ,,,„
Pi
«+!/
_JLexp{tVO|
(3.5b)
if dist p (i,i<:(a,p M )) = p", n > M. Remark 3.1 When / p(x)p,{dx) < oo thenPt{x,K{a,pM))
-+ paM as
t-too
i.e. paM, a € Q p , M e Z defines the invariant measure for the process. Remark 3.2 If a{N) = 0 for some N € Z then Pt(x,K(a,pM)) = 0 when ever distp(x,K(a,pM)) > pN i.e. the process is confined to the ball K{x,pN). This is a consequence of the non Archimedian distance property. Indeed if a(N) = 0 then U(n) — a(n - 1) - a(n) — 0 for all n > N which means that the process can not make a jump larger than pN. But the p-adic dis tance between K(x,pN) and any ball disjoint with it is longer than pN and this distance must be taken in one jump which is impossible. Remark 3.3 Ifa{N + 1) < a(N) = a(N - m) for allmGN then the process does not make jumps smaller than pN+1. In this case the process is equivalent to one with K.N as the state space. Remark 3.4 If p{x) = 1 then the resulting process is G invariant. However, given A G A the transition functions (2.17) and (3.5) are different. Never theless both formulations are equivalent. It can be seen by direct computations that (3.5) defined by p(x) = 1 and A £ A is identical with (2.17) defined by A' € A where oo
o'(M)=p"
1
a M
{p-l)- a{M)-J2 (
+ i)P~i ■
»=i
Similar arguments work in the converse direction. R e m a r k 3.5 The processes discussed in this section have been used in a model for turbulent cascades n . In order to consider Pt(x, B) as the integral kernel for a Markovian semigroup (Tt, t > 0) we have to take L2(QP, pp.) for the underlying Hilbert space. Then explicit formula for the action of the generator H on the indicator functions can be given9. It follows further that H has pure point spectrum cr(H) = {Wf}i6Z,o6Qp-
63
The eigenvectors are denned as follows. Given a G Qp and n G Z we define Xi(x)' symmetric case. Then
»=i
* = 1,. . . , p as in the G
»=i p
where c* G R, i = l,...,p
satisfy X) CtPJi' = 0-
Passing to the corresponding Dirichlet form we find that it is p\x symmetric and consists of the jump part only as should be expected. The jump measure J(dx, dy) is defined by J(K(a,pM),K(a,pM))
= -\waM
p%
and J{K(a,pM),K{b,pM)) if distpCK, (a,pM), K(b,pM)) 4
= \p%pbM[a{M + m-\)-a{M = pM+m,
+ m)}
meN.
The Global Properties
Now we shall list some global properties of the Markovian semigroups (Tt,t> 0) constructed in previous sections 10 . First {Tt,t > 0) is conservative i.e. Pt{x,Qp) = 1. This property holds generally provided in case />«> = / p{x)n{dx) < oo the function p is normalQP
ized so that />«> = 1Concerning the transience, resp. recurrence problem results have been obtaind in 10 , we summarize them here. The semigroup (Tut > 0) of the bounded operators in I?{Qv\pp) is irreducible iff a{n) > 0 for all n G Z (see Remark 3.2). Recall that a Markovian semigroup which is irreducible is either transient or recurrent. The simplest case is when p^ = 1 and a(n) > 0, n G Z. Then the constant function 1 G Lx{pp) and Ttl(x)
oo
= Pt(x,Qp)
= 1 so that fTtl{x)dt = oo o and (Tt, t> 0) is not transient. As it is irreducible it must be recurrent. oo
If Poo = oo then (Tt,t > 0) is recurrent iff / Pt{x, B)dt = oo for some B o
64
with / p(x)p,(dx) > 0. This by direct examination is seen to be equivalent to B
-T,e%Awr- = ^
(4-1)
where M is such that p°M > 0. If (r t , t> 0) is determined by (2.17) then it is recurrent12 iff °°
E
■ 1
P _ , - 7 T T = OO.
•„
°(»)
v This condition can also be obtained by and passing from A «=0 ' specifying This condition can also be obtained by specifying p = 1 and passing from A to A' according to Remark 3.3. We add a short comment on the probabilistic interpretation of the constants W^. Let & be the random walk denned by the transition function (3.4) or (3.5). Then TK(a,pM) = inf {t > 0;6 i K(a,pM)} is the exit time for the ball K(a,pM). We have Theorem 4.1 10 If a(M) > 0 and p% ^ poo then TK(a,PM) ** almost surely finite and its expectation is ExTK{p.,vM) — ~WM)~X> x e K(a,pM). If p = 1 or equivalently the transition functions are defined by (2.17) then
ExTK{a,p")
Denote by
=a(M)-1. TK^IPM^ the
hitting time for a ball K(a,pM) i.e.
rK{a>pM) = inf {* > 0;& 6 K(a,pM)}
.
Recall that recurrence is said to be positive or persisting if ExTK(a
Trace Formula
We shall close this review with a result which bears similarity with the trace formulas for the Laplace-Beltrami operator for hyperbolic spaces. We shall make our point using probabilistic interpretation13.
65
Consider the complex upper half-plane C+ with the Poincare metrics. Then SL(2, B)/ / {1, —1} is a group of isometries of the half-plane C+. Given a discrete subgroup T consisting of hyperbolic elements there is a fundamental domain F for T i.e. an open subset of C+. such that JiFn
72.F=0
if
71,72 GT
and
717^72
and
Then identification of the points x G F with 72, 7 G T defines a compact Riemann surface M. The Laplace-Beltrami operator A on C+ (resp. AM on M) generates a random walk Xt, t ^ 0 on C+. (resp. Rt,t ^ 0 on M). Let Pt(x,y) (resp. qt(x,y)) be the transition density for Xt (resp. for Rt). Define Qt(x,y) = ^2Pt{x,jy) -rer
x,y£M.
It follows that qt = qt i.e. qt is the transition density for Rt. Moreover Tre A M t = / pt(x,x)dx + ^ F
/ pt(x,'yx)dx F
-i*'
where e is unit of T, can be expressed in terms of lengths of closed geodesies in M i.e. the distances between x and 7 a;, 7 G T21. When looking for a p-adic analog of this formula we first observe that we do not have a diffusion but rather a family of jump processes. Thus instead of the Laplace-Beltrami operator on C+ we have a class of generators for the random walks on Qp. We already mentioned that G is a group of isometries for Qp. For certain discrete family Y of p-adic translations (r cannot be a group and we assume that 7 G T implies ||7|| p > 1) 2^(0,1) becomes a "fundamental domain" for Qp. Let Pt(x, K(a,pM)) x, a G Qp, M G Z be given by (2.17) i.e. we consider G symmetric processes. We define the transition density pt{x, y), t > 0, and assume it to be bounded on Qp x Qp.(this can be realized by a suitable choice of the sequence a(n),) Then qt(x,y) = pt{x,y) + ^2pt{x,n/ + y) 7er
x,y G K(0,1)
66
is a transition density for the process confined to K(0,1). generator of the process in K(0,1) then T r e - " " ' = ft (0,0) = p t (0,0) + $ > ( 0 , 7 ) •
If — H0 is the
(5.1)
If the process in Qp is denned by a(n) = p~" then the right hand side can be expressed in terms of ||7|| p i.e. the p-adic distances between 0 and 7. The process in K(0,1) can be viewed as obtained from the process on Qp by identification of the points x € A"(0,1) and x + 7, 7 G T. Conversely if — Ho generates a process on K(0,1) such that T r e - " 0 * = ft(0,0) < 00 then there is a family of processes on Qp such that (5.1) holds. References 1. R. Rammal and G. Toulouse, Ultrametricity for Physicists, Rev. Modern Phys. 58, 765-78 (1986) 2. V.S. Vladimirov, I.V. Volovich, E.I. Zelenov, p-adic Numbers in Mathe matical Physics, World Scientific, Singapore (1993) 3. A. Khrennikov, p-adic Valued Distributions in Mathematical Physics. Kluver, Dordrecht (1994) 4. L. Brekke, M. Olson, p-adic Diffusion and Relaxation in Glasses, Preprint EFI Chicago (1989) 5. A. Figa - Talamanca, Diffusion on compact ultvametvic spaces. In: Noncompact Lie Gruops and Some of Their Applications, Ed. E.A. Tanner, R. Wilson, Kluwer, Dordrecht (1994) 6. S. Hussmann, Random walks on p-adic tree. SFB 237 Preprint No. 359 '97 (1997) 7. S. Albeverio, W. Karwowski, Diffusion on p-adic Numbers, pp. 86-99 in •.Gaussian Random Fields, Ed. K. Ito, H. Hida, World Scientific, Singa pore (1991) 8. S. Albeverio, W. Karwowski, A Random Walk onp-adics: The Generator and its Spectrum, Stochastic Processes and their application, 53, 1-22 (1994) 9. W. Karwowski, R. Vilela-Mendes, Hierarchical Structures and Asymmet ric Processes on p-adics and Adeles, J. Math. Phys. 35, 4637-4650 (1994) 10. S. Albeverio, W. Karwowski, X. Zhao, Asymptotic and Spectral Results for Random Walks on p-adics, Stochastic Processes and their Applica tions, 83,39-59 (1999)
67
11. R. Lima, R. Vilela-Mendes, Stochastic Processes for the Turbulent Cas cades, Phys. Rev. E 5 3 , 3536-3540 (1996) 12. K. Yasuda, Additive Processes on Local Fields, J. Math. Sci. Univ. Tokyo, 3, 629-654 (1996) 13. S. Albeverio, W. Karwowski, K. Yasuda, Trace Formula for p-adics, (to be published) 14. N. Koblitz, p-adic Numbers, p-adic Analysis and Zeta-Function, Springer, New York, 2nd ed. (1984) 15. M. Schreckenberg, Long Range Diffusion in Ultrametric Spaces, Z. Phys. B60, 483^88 (1985) 16. S.N. Evans, Local Properties of LeVy Processes on a Totally Disconnected Group, J. Theoret. Probab. 2, 209-259 (1989) 17. V.S. Vladimirov, Generalized Functions over the Field of p-adic Numbers, Russian Math. Surveys 43, 19-64 (1988) 18. A.N. Kochubei, Parabolic Equations over the Field of p-adic Numbers, Math. USSR Izviestiya 39, 1263-1280 (1992) 19. M. Fukushima, Y. Oshima, M. Takeda, Dirichlet Forms and Symmetric Markov Processes, De Gruyter, Berlin 1994 20. S. Albeverio, Z. Zhao, On the Relation between different Construction of Random Walks on p-adic, (to be published) 21. D.A. Hejhal, The Selberg trace Formula for PSL(2,R), Vol. 1,2, Lecture Notes in Mathematics 548 resp. 1001, Springer-Verlag (1976 resp. 1983) 22. H. Kaneko, A class of spatially inhomogeneous Dirichlet spaces on the p-adic number field. Preprint (1999)
68 CHARACTERIZATION OF TEST F U N C T I O N S IN CKS-SPACE NOBUHIRO ASM* Graduate School of Mathematics Nagoya University Nagoya 464-8602 JAPAN IZUMI KUBO Department of Mathematics Faculty of Science Hiroshima University Higashi-Hiroshima 739-8526 JAPAN HUI-HSIUNG KUO t Department of Mathematics Louisiana State University Baton Rouge LA 70803 USA We prove a characterization theorem for the test functions in a CKS-space. Some crucial ideas concerning the growth condition are given.
1
Introduction
Let £ be a real countably-Hilbert space with topology given by a sequence of norms {| • | } £ 1 0 (see n . ) Let £p be the completion of £ with respect t o t h e norm | • | p . Assume the following conditions: (a) There exists a constant 0 < p < 1 such t h a t | • | 0 < p\ ■ |i < • • • < pp\ ■ \p <
(b) For any p > 0, there exists q > p such t h a t the inclusion m a p iq>p : £Q —►
69
By using the Riesz representation theorem to identify £0 with its dual space we get the continuous inclusion maps: £ C £p C So C S'v C £',
p > 0,
where £' and S' are the dual spaces of Sv and S, respectively. Condition (b) says that S is a nuclear space and so S C So C S' is a Gel'fand triple. Let fi be the standard Gaussian measure on S'. Let (L2) denote the Hilbert space of /^-square integrable functions on S'. By the Wiener-Ito theorem each tp in (L2) can be uniquely expressed as oo
oo
V = £/n(/„) = l > ® n : , / n ^ , n=0
fn£Sfn,
(1)
n=0
and the (L 2 )-norm \\
2
\ i/a
IMIo=[$>!|/n| J ■ Now, we describe the spaces of test and generalized functions on the space £' introduced by Cochran et al. in a recent paper 4 . Let {a(n)}£L 0 D e a sequence of numbers satisfying the follow ing conditions: (Al) a(0) = 1 and inf„> 0 a(n) > 0. (A2) l i m n _ H X ) ( 2 i j i ) 1 / B = 0. Let
\ V2
IMIp,a= (Y,nla(n)\fn\2p) \n=0
■
(2)
/
Let [Sp]a — {
[£]a c [Sp]a c (L 2 ) c [SPYa c [S]*a, p > o. If $ G [Sp]*a (the dual space of [£p]a) is represented by $ = E^=o( : ' <8n •'»F" ~>> then its [£p]*a-norm is given by / oo
II*H-»I/«=
,
\ V2
a
E^-i -P
•
^
70
This Gel'fand triple [£]a C (L 2 ) C [£\*a is called the CKS-space associated with a sequence {a(n)}£L 0 of numbers satisfying the above conditions (Al) and (A2). Several characterization theorems for generalized functions in [£]^ have been proved in the paper 4 . However, no characterization theorem for test functions in [£]a is given. The purpose of the present paper is to prove such a theorem. In addition we will mention some crucial ideas in order to get a complete description of the characterization theorems for test and general ized functions in our ongoing research collaboration project. We remark that similar results have been obtained by Gannoun et al. 5 . 2
Characterization theorems
For £ e £c (the complexification of £,) the renormalized exponential function :e(''t~*: is defined by oo
:e<,
^
:=
£^<:-®n:^n^-
(4)
n=0
For any p > 0, we have l|:c<-' I - > :|| P .a = G a ( | ^ ) 1 / 2 ,
(5)
where Ga is the exponential generating function of the sequence {a(n)}^_ 0 , i.e., Ga(z) = £
^ * » ,
Z € C.
n=0
By condition (A2) of the sequence {a(n)}£L 0 , the function Ga is an entire function. Equation (5) implies that :e^' ,a:_+ : is a test function in [£]a for any
Se£c. The 5-transform of a generalized function $ in [£]£ is the function S $ defined on £c by S*(fl = «*,:<:<■•*-:»,
£e£e,
where ((•, •)) is the bilinear pairing of [£]* and [£]aWe state three conditions on the sequence {a(n)}^L 0 as follows: (Bl) l i m s u p ^ ( ^ i n f r > 0 ^ - f
" < oo.
71 (B2) The sequence 7(71) = ^ffi,n
> 0, is log-concave, i.e.,
7(n)7(n + 2) < 7 ( n + l ) 2 ,
Vn > 0.
(B3) The sequence {a(n)}£L 0 is log-convex, i.e., a(n)a{n + 2) > a ( n + l ) 2 ,
Vn > 0.
Condition (Bl) is used in the characterization theorem for generalized functions. By Corollary 4.4 in 4 condition (B2) implies condition (Bl). Con dition (B3) implies condition (B2) to be defined below. The following characterization theorem for generalized functions in [£]£, has been proved in 4 . T h e o r e m 2.1 Suppose the sequence {a(n)}^L0 satisfies conditions (Al) and (A2). (I) Let $ G [€]*a. Then the S-transform F = S$ of $ satisfies the conditions: (1) For any £, n in £c, the function F(z£ + n) is an entire function of z € C. (2) There exist constants K,a,p>
0 such that
\F(0\
£££,
(6)
(II) Conversely, assume that condition (Bl) holds and let F be a function on Sc satisfying the above conditions (1) and (2). Then there exists a unique $ G [£]* such that F = S$. We mention that condition (2) is actually equivalent to the condition: there exist constants K, p > 0 such that \F{0\
S€£C.
This fact can be easily checked by using the fact that |£| p < pQ~p\^\q for any q > p. Having the constant a in condition (2) is for convenience to check whether a given function F satisfies the condition. By this theorem, if we assume condition (Bl), then conditions (1) and (2) are necessary and sufficient for a function F defined on £e to be the Stransform of a generalized function in [S\*a. As mentioned above, condition (B2) implies condition (Bl). Thus under condition (B2), conditions (1) and (2) are also necessary and sufficient for F to be the 5-transform of a general ized function in [£]*.
72
For the characterization theorem for test functions in [£]Q, we need to define the exponential generating function of the sequence { j^y}^=o : OO
j
Moreover, we need the corresponding conditions (A2) (Bl) (B2) for the sequence(^}^o: (A2) l i m „ - + o c ( ^ l ) 1 / r i = 0. (Bl) l i m s u p ^ ^ (n!a(n) inf r>0 % ^ ) '
/ n
< oo.
(B2) The sequence {^7^y}n^=o ' s log-concave. It follows from condition (A2) that the exponential generating function G\/a is an entire function. Note that by condition (Al), a(n) > c*o for all n, where Q 0 = inf„>o a(n). Hence by the Stirling formula l/n
n\a{n) J
/
<
t
-:— \n\aoJ
\ l/n
/
~
j
\a0
\ '/"
; \j2-nn)
e
n
> 0,
as n -> oo.
This shows that condition (Al) implies condition (A2). On the other we see t n a t hand, by applying Corollary 4.4 in 4 to the sequence {^y}^Lo 1)J condition (B2) implies condition (Bl). Moreover, it is easy to check that condition (B3) implies condition (B2). Now, we study the characterization theorem for test functions in [£]aFirst we prove a lemma. L e m m a 2.2 Assume that condition (Bl) holds and let F be a function on £c satisfying the conditions: (1) For any £, n in £c, the function F(z£ + 77) is an entire function of z G C. (2) There exist constants K,a,p>0
such that
\F(0\
(e£c.
Then there exists a unique ip € [£]l, such that F = S
^— )
< 1
(8)
73
and
ML < £ £ (»'*(») W % ^ ) (-2Hv„ll^)nn=0 ^
0)
'
Proof. We can modify the proof of Theorem 8.9 in n . Use the analyticity in condition (1) to get the expansion oo
n=0
Then use the Cauchy formula and condition (2) to show that for any £ i , . . . , £n in [Sp]a and any R > 0: l(/n,6®"-®€n-^|<^
^
—
Kl|-p"-|£n|-p.
Make a change of variables r = an2R2 and use the inequality n n < n!e n /v / 27r (see page 357 in n ) and then take the infimum over r > 0 to get
K/»,{,S-SC- - I2 < £«V« ( i n f % ^ ) l&lV-KnlV Now, suppose G [0,p] satisfies the condition in Equation (8). Then by similar arguments as in the proofs of Theorems 8.2 and 8.9 in n we can derive l/n|, <
2?ra
e
^nt o
r„
J||tp,,||HS.
Therefore oo
|M|2,Q = ^n!a(n)|/„| 2 n=0
n=0 ^
'
Note that the series co nverges because of the condition in Equation (8). ■ Theorem 2.3 Suppose the sequence {a(n)}£L 0 satisfies conditions (Al). (I) Let if E [£}a- Then the S-transform F = Sip of
Z€£e.
(10)
74
(II) Conversely, assume that condition (Bl) holds and let F be a function on £c satisfying the above conditions (1) and (2). Then there exists a unique ¥ G [£]a such that F = Sip. We remark that condition (2) is actually equivalent to the condition: for any constant p > 0 there exists a constant K > 0 such that \F(0\
(e£c.
This fact can be easily checked by using the fact that |£|_, < p ? _ p | £ j _ p for a^y Q > P- Having the constant a in condition (2) is for convenience to check whether a given function F satisfies the condition. Proof. Let
= ((:e^:,v)),
£ € £C.
Since
|F(0|
Hence
|F(OI
<
^\S\-P-
Therefore, we obtain
\F(t)\ < IMI,,aG1/a(|f|2_,)1/2 < IMI,.aG1/a(a|f|2_p)1/2. To prove the converse, assume that condition (Bl) holds and let F be a function on £c satisfying conditions (1) and (2). Let q > 0 be any given number. Choose a, p > 0 such that
(
C (r) \ '^n nla{n) inf — — — 1 < 1. r>0
T
)
This inequality can be achieved in two ways: (1) choose any a > 0 and then use the fact that limp-joo ||i P l ,||//s = 0, (2) choose any p > 0 such that i Pi9 is a Hilbert-Schmidt operator and then choose a sufficiently small number a > 0. (For the first way we can choose a — I and this is exactly the fact mentioned in the above remark.) With the chosen a and p, use condition (2) to get a constant K such that the inequality in Equation (10) holds. Then we apply
75
Lemma 2.2 to get the inequality in Equation (9) so that ||y|| 9 ,a < oo. Hence
Examples and comments Four conditions
In the definition of CKS-space and the characterization theorems of general ized and test functions we have assumed several conditions on the sequence {<*(n)}£°=0: (Al), (A2), (Bl), (B2), (B3), (A2), (Bl), (B2). Recall that we have the f ollowing implications: (A1)=*(A2),
(B2)=»(B1),
(B2)=>(B1),
(B3) = » (B2).
Taking these implications into account we will consider below the four con ditions: (Al), (A2), (B2), (B3). 3.2
Examples
We give three examples corresponding to the Hida-Xubo-Takenaka, Kondratiev-Streit, and CKS-spaces. Example 3.1 For the Hida-Kubo-Takenaka space {£) C (L 2 ) C {£)* (see 6 9 10 l3 ,) the sequence is given by a(n) = 1. Obviously, this sequence satisfies the above four conditions. The corresponding exponential generating functions are Gtt(r)=er,
G1/Q(r)=eT.
Thus the growth conditions in Equations (6) and (10) can be stated as
|F(0I < Kea^l,
\F(0\ < Kea^'-'.
Theorems 2.1 and 2.3 are due to Potthoff-Streit 14 and (uc-Potthoff-Streit , respectively. Example 3.2 For the Kondratiev-Streit space (£)& C (L 2 ) C (£)*p (see 7 8 n ,) the sequence is given by a(n) = (n!)^. It is easy to check that this sequence satisfies the above four conditions. The corresponding exponential generating functions are 12
oo
n=0
1
v
';
oo
n=0
.. v
';
76
But from page 358 in u and Lemma 7.1 (page 61) in 4 we have the inequalities: exp (1 - / 3 ) r ^ |
< 2 /3 exp [(1 - / 3 ) 2 ^ 3 r ^ l .
(11)
On the other hand, from page 358 in n and the same argument as in the proof of Lemma 7.1 in4 we can derive the inequalities: 2-0exp
(l + 0)2~&r^]
< exp [(1 + / 3 ) r ^ l .
(12)
The inequalities in Equations (11) and (12) imply that the growth conditions in Theorems 2.1 and 2.3 are respectively equivalent to the conditions: • There exist constants K,a,p>0
such that
|F(0|
Z€£e.
• For any constants a, p > 0 there exists a constant K > 0 such that \F(0\
t e Sc.
These two inequalities are the growth conditions used by Kondratiev and Streit in7 8 (see also Theorems 8.2 and 8.10 in11.) Example 3.3 (Bell's numbers) For each integer k > 2, consider the k-th iterated exponential function exp t (z) = exp(exp(- • • (exp(z)))). This function is entire and so it has the power series expansion expk(z) =
^
Bk(n) nn ^—-^-z .
n=0
The k-th order Bell's numbers {bk(n)}%L0 are defined by
Obviously, this sequence {bk{n)}%L0 satisfies conditions (Al) and (A2). It has been shown recently in J that this sequence satisfies conditions (B2) and (B3). The corresponding exponential generating functions are given by G,,Ar) = Z^ob^r»
= 2%$,
Gift(--)=E:=os4yrn-
(13)
(14)
The function expk(r) in Equation (13) can be used to give the growth condition in Theorem 2.1 for generalized functions, i.e.,
77
• There exist constants K,a,p>
0 such that
|F(OI
Z6£c
On the other hand, we cannot express the sum of the series in Equation (14) as an elementary function. Thus the growth condition (2) in Theorem 2.3 is very hard, if not impossible, to verify. Hence it is desirable to find similar inequalities as those in Equation (12) for the sequence {bk(n)}^=0. 3.3
General growth order
Being motivated by Examples 3.2 and 3.3, we consider the question: What are the possible functions U and u such that the growth conditions in Theorems 2.1 and 2.3 can be respectively stated as follows? • There exist constants K, a,p > 0 such that \F(Q\
Z ££c.
• For any constants a, p > 0 there exists a constant K > 0 such that
\F(0\
£ e£c.
The answer to this question will be given in our forthcoming papers 2 . In particular, when a(n) = bk{n), condition (2) in Theorem 2.3 can be replaced by the following growth condition: 3
• For any constants a, p > 0 there exists a constant K > 0 such that \F(0\ < Kexp [2 J a | e
p
log*-i (a^| 2 _ p ) 1 ,
£ G £e,
where logj, j > 1, is the function denned by l o g ^ i ) = log(max{x,e}),
log^i) = logxOog^^x)),
j > 2.
References 1. Asai, N., Kubo, I., and Kuo, H.-H.: Bell numbers, log-concavity, and logconvexity; in: Classical and Quantum White Noise (A special volume in honor of Hida's 70th birthday,) L. Accardi et al. (eds.) Kluwer Academic Publishers (1999) 2. Asai, N., Kubo, I., and Kuo, H.-H.: Log-concavity and growth order in white noise analysis; Preprint (1999)
78
3. Asai, N., Kubo, I., and Kuo, H.-H.: General characterization theorems and intrinsic topologies in white noise analysis; Preprint (1999) 4. Cochran, W. G., Kuo, H.-H., and Sengupta, A.: A new class of white noise generalized functions; Infinite Dimensional Analysis, Quantum Probability and Related Topics 1 (1998) 43-67 5. Gannoun, R., Hachaichi, R., Ouerdiane, H., and Rezgui, A.: Un Thoreme de Dualite Entre Espaces de Fonctions Holomorphes a Croissance Exponentiele; BiBoS preprint, no. 829 (1998) 6. Hida, T., Kuo, H.-H., Potthoff, J., and Streit, L.: White Noise: An Infinite Dimensional Calculus. Kluwer Academic Publishers, 1993 7. Kondratiev, Yu. G. and Streit, L.: A remark about a norm estimate for white noise distributions; Ukrainian Math. J. 44 (1992) 832-835 8. Kondratiev, Yu. G. and Streit, L.: Spaces of white noise distributions: Constructions, Descriptions, Applications. I; Reports on Math. Phys. 33 (1993) 341-366 9. Kubo, I. and Takenaka, S.: Calculus on Gaussian white noise I; Proc. Japan Acad. 56A (1980) 376-380 10. Kubo, I. and Takenaka, S.: Calculus on Gaussian white noise II; Proc. Japan Acad. 56A (1980) 411-416 11. Kuo, H.-H.: White Noise Distribution Theory. CRC Press, 1996 12. Kuo, H.-H., Potthoff, J., and Streit, L.: A characterization of white noise test functional; Nagoya Math. J. 121 (1991) 185-194 13. Obata, N.: White Noise Calculus and Fock Space. Lecture Notes in Math. 1577, Springer-Verlag, 1994 14. Potthoff, J. and Streit, L.: A characterization of Hida distributions; J. Fund. Anal. 101 (1991) 212-229
79
N O N L I N E A R LIE ALGEBRAS IN Q U A N T U M PHYSICS A N D THEIR INTEREST IN Q U A N T U M FIELD THEORY J. BECKERS' Theoretical and Mathematical Physics, Institute of Physics (B5) University of Liege, B-4000 Liege 1 (Belgium) Solutions of the soliton type and their stability conditions obviously appear through polynomial interactions in toy models enlightening relativistic scalar field theories. We show that differential realizations of operators characterizing non linear Lie algebras connected with deformations of the angular momentum algebra are useful in the study of such polynomial interactions in quantum field theory.
A Ludwig, en reference a notre premiere rencontre lors du "Workshop" de I'Universite de Bujumbura (Burundi, octobre 1987) ! Onze ans, deja ! The use and interest of spectrum generating algebras [1] in quantum problems are very well known : in particular, they lead to solvable (or quasisolvable) Schrodinger equations and facilitates the determination of the eigen values of the spectrum and their associated eigenfunctions. The method has been often exploited, when, in general it is subtended by linear Lie alge bras such as, for example, the real forms sl(2,U) and su(2,C) of the simple Lie algebra A\ in the Cartan classification [2]. There are no reasons to ex clude nonlinear Lie algebras which have been studied very recently [3,4]. In fact, here we want to exploit the irreducible representations of such nonlin ear sl(2,3R)-Lie algebras in order to add further information in quantum field theory in the particularly well studied context of the scalar case. As a typical work in this scalar field theory, let us point out the remarkable Christ and Lee reference [5] leading to solitonic solutions but without any reference to a spectrum generating algebra and let us see how a (non-linear) cubic (Lie) algebra - the so-called Higgs algebra [6] - can help us to get more general solutions. In terms of the scalar field <j>(x,t) and its derivatives
80
(l+l)-dimensional space, the Lagrangian density takes the form C{4>,u
(l)
where V{<j>) refers to the interaction potential characterizing specific toy mod els very often called <j>4 or ^-models more specifically. Through the principle of least action in the stationary context and by varying
z = f{x)
(2)
(when we restrict ourselves to scalar potentials seen as even functions of (j>), simply becomes d2
d
F(z)=u,2F(z)
.
(3)
Here Pk(z) obviously refers to polynomials of degree k in the variable z. Such an equation (3) is thus equivalent to a Schrodinger equation associated with a specific scalar field theory. Its general solutions are welcome through con siderations on nonlinear Lie algebras as we propose to show in the following. Let us first consider the linear sl{2, 8t)-algebra given by the commutation relations [J0,J±) = ±J±
,
(4)
[J+,J-]=2J0
,
(5)
where Jo is the diagonal operator and J± the scaling ones as is well known and, secondly, the nonlinear Higgs algebra [6] given by the commutation relations (4) but with eq. (5) replaced by [J+,J-] = 2Jo + 80Jg
(6)
where (3 refers to a possible family of such nonlinear algebras and can also be interpreted as a deformation parameter. This algebra (4), (6) is recognized as a specific W-algebra [7] and belongs to the so-called category of polynomial deformations already studied [3,8] with respect, in particular, to the determi nation of its irreducible representations through an angular momentum basis [3] or a monomial one [8], the latter one asking obviously for a realization of the operators (Jo,J±) as differential operators.
81
In terms of a single variable x and its powers N — 1,2,3,..., we have proposed [8] differential operators on the forms N
J + = x F(D)
dN , J_ = G{D) _
1 , J 0 = _ (D + c)
where D is the dilation operator x — with the properties ax dN N N [D,x ] =Nx , = » & ■ dxN
(7)
<8»
when we take account of the usual Heisenberg commutation relations. If we ask for Schrodingerlike Hamiltonian operators expressed in terms of generators belonging to (nonlinear) spectrum generating algebras and if we want to priviledge the category of Higgs algebras, we have noticed [8] that we have two families of realizations for N = 1 and 2 only respectively given in terms of real parameters (c\, c2, ..., C5) and (di, d2, ^3, dt) by J+ = x(ciD2 + c2D + cs) , J-=c4D
+ c5 , J0 = D + c
(9)
and 2
2
J+ = x (diD
+ diD + d3) , J-=d4
d2 1 -^s , J0 = -(D
+ c) .
(10)
Both families lead to a commutation relation of the type
[j+,j-] = f(i,j0,JZ,JZ)
,
(ii)
including obviously the Higgs character (6) and leading to Schrodingerlike Hamiltonians given by H = ai J+ + a2J- + o 3 J 0 + a 4 Jo + a 5
(12)
where a\, a2, ..., 05 are parameters chosen in such a way that this if-operator is selfadjoint as required. In particular, the above realization (9) then leads to a corresponding Hamiltonian
H^fW^
+^Ws
+ ^W
<13)
where the superscripts refer to the jV = 1-case and where P3(1)(x) = AlX3 + Bxx2 + Cix , P2(1)(x) = DlX2 + Eix + 1, Pll\x)
= Fix + G{U)
Here is the evident connection with the stability condition (3) and its Hamil tonian intimately connected with our information (13) and (14).
82
We thus conclude that nonlinear deformations of the angular momentum algebra sl(2,3?) have to deal with polynomial interactions of special interest in scalar field theories, these (nonlinear) algebras playing the role of (nonlinear) spectrum generating algebras for such developments. Let us end this discussion by considering the example of a 0 6 -model lead ing to a Hamiltonian in the N = 2-case : BW=Pi2\z)~i
+ Pi2)(z)±+Fi%)
(15)
with P4(z) = 4z(z - A){z - B)(z - C) ,
(16)
P3(z) = 2 [4z3 - 3z2(A + B + Q + 2z(AB + AC + BC) - ABC] (17) and P2(z) = -15z2 + 6z(A + B + Q + AB + AC + BC
(18)
6
so that the equation (3) is completely determined in this 0 -model. We thus get families of solutions of the corresponding Heun equation [9] and it is easy to show that we obtain in that way more general (solitonic) solutions in comparison with the old results coming from the Christ and Lee developments [5]. In particular, the solution F(z) = (A- zf2
(, - * ± £ ) , J2 = 2(S - Cf ,
(19)
exists not only for A = B = 1 and C = —\ [5] but also if A = 2(B + C) in general, etc.
References 1. "Dynamical Groups and Spectrum generating Algebras", A. Barut, A. Bohm and Y. Ne'eman, Eds., World Scientific (Singapore, 1989) 2. J.F. Cornwell, "Group Theory in Physics, an Introduction", Academic Press, London (1997) 3. B. Abdesselam, J. Beckers, A. Chakrabarti and N. Debergh, J. Phys. A 29 (1996) 3075 and references therein 4. N. Debergh, J. Phys. A30 (1997) 5239 5. N.H. Christ and T.D. Lee, Phys. Rev. D 1 2 (1975) 1606 6. P.W. Higgs, J. Phys. A12 (1979) 309
83
7. F. Barbarin, E. Ragoucy and P. Sorba, Nucl. Phys. B442 (1995) 425 8. J. Beckers, Y. Brihaye and N. Debergh, "On realizations of nonlinear Lie algebras by differential operators" (University of Liege preprint, August 1998) 9. A. Erdelyi (Editor), "Higher Transcendental Functions", McGraw Hill, New York, 1955).
84
C O N T I N U O U S PERCOLATION: T R E E A P P R O X I M A T I O N A N D MOST P R O B A B L E C L U S T E R S
PH. BLANCHARD Fakultat fur physik, Theoretische Physik and BiBos, Universitat Bielefeld, Universitatsstrasse, 25, D-33615, Bielefeld, GERMANY. E-mail: blanchard@physik. uni-bielefeld. de G. F. DELL'ANTONIO Dipartimento di Matematica, Universita di Roma, "La Sapienza", Piazzale Aldo Moro 5, 00185, Roma, Italy. E-mail: [email protected] D. GANDOLFO Centre de Physique Theorique, CNRS, Luminy, F-13288 Marseille Cedex 9, and PhyMat, DSpartement de Mathimatiques, UTV, F-83957 La Garde, France. E-mail: [email protected] M. SIRUGUE-COLLIN Centre de Physique Theorique, CNRS, Luminy, F-13288 Marseille Cedex 9, and UFR Sciences de la matiere, Universite de Provence, 3 Place Victor Hugo, 13001 Marseille, France. E-mail: [email protected] A new approach based on tree approximation is derived in order to estimate the critical parameters of continuous percolation of overlapping disks in R 2 when the centres of the disks are Poisson distributed. In the case of directed percolation, a lower bound is found for the critical density. The cluster structure near the percolation threshold is analyzed.
1
Introduction
Percolation theory on regular lattices has already a long standing and success ful history. On the contrary percolation in continuous systems and on random graphs has received less attention in spite of its great interest in applications. Indeed continuous percolation has been used to model such diverse phenomena as conductivity in disordered semiconductors, permeability of porous media, fracture network of rocks, spread of Aids epidemics ! and even quark-gluon plasma formation in QCD gauge theories 2 . We would like to stress that our approach here will refer to continuous systems without any introduction of lattice approximation. Let us recall the
85
main features of continuous percolation, i.e. modeling a random medium — in fact randomly distributed objects in homogeneous media — by using the features of stochastic geometry. The basic facts are quite elementary: a great number N of "small" objects are thrown at random in a large volume V; their distribution follows a Poisson law if the mean density p = y is sufficiently low. So let {xt}ie/cN be the points of a standard Poisson process with intensity p in R d , d > 2. We choose a geometrical object P with a distinguished point xo and a given orientation and for each i £ I we place a copy of P with xo s n. We could also consider objects of random shape and size but in the following we will only be concerned with the simplest case. Denoting Pi the object centered in Xi we are interested in the description of clusters of overlapping objects and particularly in the possible existence of an infinite cluster "spanning" the space domain. Let us declare two objects Pi and Pj adjacent if Pi f] Pj ^ 0. We say that two objects (or the two corresponding points Xi and Xj) are connected, Pi <—> Pj (or Xi <—> Xj), if there exists a chain of adjacent objects (or points) connecting these two objects (or these two points). A cluster is a maximal set {Pi}iGJciN (or {xi}i£j) of connected objects (or points). The aim of continuous percolation theory is to study the sizes and shapes of clusters for specified values of p and typical objects P. As in the case of lattice percolation in a infinite domain, the existence of an "infinite incipient" cluster is linked to the long range connectivity property of the random system. The study of the phase transition associated with the appearance of an infinite cluster is one of the main problem of percolation theory. Objects as diverse as disks, squares, needles in R 2 , spheres, cubes, ellipsoids, in R 3 have been used for typical models (see e.g. 5 ' 1 3 ).
Figure 1.
86
For general results on continuous percolation see 3 and also 4 ' 6 ' 7 . For recent results concerning the scaling properties of the infinite percolating cluster see n 12 ' . Percolation theory in continuous media is naturally related to classical problems of stochastic geometry, see e.g. 8 . Let us now state the framework in which we want to develop our approach. It follows from the scaling property of the Poisson process that the statistical properties of the system are invariant under the scaling transformations: p —» Xp,
v —► Xv,
VA €J0, oo[
where v is the volume of the elementary objects. Due to this property, the simplest invariant, dimensionless parameter we can use in order to describe the behaviour of a percolation model is the socalled filling factor 7 defined by 77 = pv. We can better understand the physical meaning of the parameter 77 by considering N objects of volume v thrown at random in a domain of volume V. In the limit of low density, the fraction of space not covered by these
\N
(V -
objects is lim ( ——— I . In the limit V —> 00, JV —> 00, TT = po < 00, we V—too \
V
J
have
(—y^-J - exP (~vV)
= exp -7?
( )-
Then the fraction of space (p occupied by the randomly distributed objects is clearly given by
Notice that (pc is in some rough sense an indicator of how the "incipient spanning cluster" is branched, at least if one makes the assumption that there is only one such cluster. In order to find the main characteristics of the cluster distribution one defines the following quantities for 77 near T7C • the probability Px for a disk to belong to an infinite cluster,
87
• the correlation length £, maximum size above which the clusters are ex ponentially scarce, • the average cluster size Sci (mean size of the clusters containing one given disk ), Currently accepted wisdom tell us that there should exist a set of critical exponents describing the behaviour of these quantities near the percolation threshold, i.e. there should exist positive constants /3,7, v such that the fol lowing asymptotic behaviour holds: P 0 0 (T ? )
« (i - - )
>
v>Vc
Z(v) <* \v-Vc \~" Sci{rj) oc I r\-i]c I" 7 In R 2 for different types of objects (squares, sticks, . . . ) the percolation threshold r?c changes but the critical exponents are the same as those found for lattice percolation in dimension d = 2, i.e.
fl-A
-t
-H
^~36' V 3' 7 18" The same description can be used for spheres percolation in R 3 , in this case one has T7(d==3) = p^irR3. The corresponding numerical values are in this case: 7]c w 0.35 (
p w 0.41,
v ss 0.90,
7 w 1.8.
In R 3 , as in R 2 , lattice percolation and continuous percolation belong to the same universality class of critical phenomena: the class of random (uncorrelated) percolation. Let us just remark that in R the continuous percolation problem is trivial, clearly we have
88
• hopping conduction in doped semi-conductors, • localization of electrons in random potential generated by disordered sys tems, • mathematical biology, epidemiology, • deconfinement of strongly interacting matter at high density (transition to a plasma of deconfined, coloured quarks and gluons). In spite of its theoretical and practical importance, continuous percolation is essentially an open field for mathematical physics. Some exact general results have already been obtained (see e.g. 11>12) but almost all what is known in practice concerning continuous percolation has been obtained from numerical simulations (mainly from Monte-Carlo simulations or direct connectedness expansion) or by using heuristic arguments 4 . However no realistic model has been constructed which allows to have even a tentative answer to the problems which are of practical importance: • size and shape of typical clusters, • percolation threshold and the large scale phenomena such as the existence of an infinite cluster, • behaviour of the percolation process near the threshold: power laws, critical exponents, • generalization to others types of percolating objects (sticks, ellipsoids, squares, cubes, . . . ) , • problems tied to computer simulation e.g. the determination of the proper window in which to perform computations, • connection with scaling theory, renormalization group, universality, frac tal properties, • study of the properties of spanning probabilities on compact regions
\s,s}d. (see however 11,12 for the last two points). In fact, as we shall give examples in the following, we can think of dif ferent basic techniques in order to build more realistic models. Besides the general mathematical framework which is related to the general theory of "coverage processes" 10 or the relation between continuous percolation mod els and bond percolation on Vorono'i lattices, some techniques which appear relatively "natural" are:
89 • put in evidence geometrical and probabilistic constraints which give an insight on the behaviour of the system. • use the connection with the theory of branching processes by constructing a tree approximation (via, e.g. the method of generations). • use the results and experience of bond and site percolation on the lattice Zd, i.e. establish a relation between the continuous percolation problem on Rrf with an ordinary percolation process on Zd (long range percola tion). • use methods of statistical mechanics and relate phase transition in con tinuous Potts model with the existence of an infinite cluster in a corre sponding continuous percolation process . . . Concerning the last point, we would like to state some facts and results dealing with the random cluster representation of a Potts spin system when the spin positions are Poisson distributed in R d . This problem was initialy introduced by Klein 15 as an attempt to generalize Fortuin and Kasteleyn transformation 16 for Potts model in the case of systems in continuous spaces. Closely related are the studies of the Widom-Rowlinson model 17>18, (see also 19 ). Random cluster measures are natural generalizations of percolation, Ising and Potts models. Let Go, = {Vw,Eu) be the finite graph associated to some realization w of a Poisson distribution of density p in a finite domain A C R d . With E w = { 1 , 2 , . . . , q}v" and Q, = {0,1} E " we consider the probability measure TPu,(
K 1 -P( e ))<Me),o + P( e )<W),i *»(«)] , f f £ ^ , w ' e £2'
where p(e) = 1 — exp [—/37(e)] > 0 with inverse temperature j3 and it \ in i\ / 1 if | Xj - X j |< 2R J(e) = J ( | x i - x j | ) = | ( ) o t h e ; w i g e and <Ja(e) = Saiaj,We = (xi,Xj) = (i,j) € E. The first marginal of Pw(<7,w) leads to the distribution of the continuous Potts model nA(
-x-exp
-/3
£
J(e)(l-6aiai)
which reduces, when q = 2, to the Ising model with spins positioned on R rf , where (i, j) £ E means any edge for which \xt —Xj |< 2R.
90 The second marginal is the random cluster measure $ w
( ' ) = F " n [p(e)w'(e) <j -p( e )) i _ w ' ( e ) ] 9fc(a,,(e)) A
k u
e€E„
e
where q ( ^ " is the number of open components of u' (connected clusters in the sense that two points are connected iff [ x, — Xj \< 1R). For q = 1 we recover the continuous percolation model. This generalization of the continuous percolation model allows a geometric representation of the Potts model. The interested reader is encouraged to consult 20 . Our purpose here is to develop a model for continuous percolation based on tree approximation. 2 2.1
Model of generations Presentation
Consider the simplest model of percolating spheres (radius R, volume v
Figure 2. Notice that a path is a collection of overlapping spheres but the order in the sequence may not be uniquely defined. The percolation threshold pc can be denned as the infimum over L of the values p{L) of the density such that, for any box of characteristic length L, there exists, with probability one, a
91
spanning cluster, i.e. a cluster which connects two opposite sides of the box. In the limit of large volume, where translational invariance of the Poisson law holds, we can choose any point 0 in the percolation cluster and consider all the paths leaving O. The set of points belonging to such a graph can then be divided in generations Gk, (fc = 1,2,...), with Go = {O} in the following way: xk £ Gk iff {0,Xi,... ,xk,... ,xk+i,...} being on the same path, then necessarily
I *k -xk±i I Xk-Xk±i
|< 2R |> 2i?,Vi>2
Figure 3. This definition of generations is unique if the graph has no loop. Due to the Poisson distribution, for any sphere of volume vj the mean number of neighbours is given by: JE(#neighbours) = pv^l ,
p being the mean density of centres of sphere.
Then, for any k sufficiently large, we can define Fk by: E(#Gfc+1|#Gfc),FfcE(#Gfc) Fk is some geometrical factor which depends on the definition of generations. It is an index of the amount in which the graph "folds" if one passes from one generation to the next. In this tree approximation the critical value pc is defined by the relation pc{Fk) = 1 , where (Fk) is a suitable average of Fk for different characterizations of generations. We expect (Fk) to be independent of k in the asymptotic limit. Define (F) = lim (Fk), then pc is defined by k—♦oo
PcVd = % =
2.2
Vd — .
Directed percolation
This method of generations turns out to be particularly well suited to study the directed percolation problem of disks in R 2 (characterized by a mean direction vector e).
92 This model is defined by simply adding to the previous conditions (*), the additional constraint: (**) For a path (0,xi that
(xfc+i - x f c ) . e > 0 .
€ G i , x 2 € G 2 , . . . ) (see Fig. 4) the condition (*) implies / xi 6 Ci
\ x2£U(6)
=
C2\(C1nC2).
Figure 4. Moreover, the condition (**) implies x2 £ U($)f]PA = Ur(9) where P A is the half-plane limited by A (xi € A , e orthogonal to A). Then using some easy geometrical arguments and obvious symmetry simplifications, it is easy to calculate the mean value (U) of Ur{6). The same is true for any sequence of 3 successive points. In this way we can identify (U) and F. We obtain F = -;pfl2
and
r)c = 1.21.
93 To our knowledge, there was till now no known result concerning this type of percolating model which can be thought to model, for example, percolation of liquid in porous rocks under the influence of the gravitational field. Numeri cal simulations based on a new algorithm for cluster statistics in continuous percolation systems confirm the above result whose details will be developed in a forthcoming paper. We can point out that this result is not too far from the known numerical result for ordinary (undirected) percolation (TJC = 1.18). The case of directed percolation in R 3 can be handled in the same way. The corresponding value for the critical parameter is found to be r\c = 0.47 to be compared with r)c = 0.35 for ordinary continuous percolation. On a quite general basis, one of the interesting problems is to find the structure of the most probable types of clusters near the percolation threshold (p
3
The A:-cluster model
In 4 Alon, Balberg and Drory proposed a new heuristic percolation criterion for continuous systems, based on the comparison of two fundamental lengths of the system, namely the average bounding length Zi (defined as the mean distance between two connected centres) and the average distance lk between two centres each of which have at least k neighbours. They postulated that percolation, which can be regarded as the condition that the system be macroscopically connected, occurs when the condition lk = 2lx is fulfilled for some definite k depending on the space dimension. The mean distance lk of the centres of two fc-clusters is approximated using the effective density pk of such clusters which is given by the basic Poisson law :
■ exp(-B)
Pk
i=i
where
B — p2dVd
J
The results they obtained (k = 4 in dimension 2 and k = 2 in dimension 3), are in very good agreement with the numerical simulations. A better insight concerning the typical clusters near the percolation threshold can be obtained using a probabilistic model. First let us give some results on the cluster shapes.
94 3.1
Most probable cluster shapes d
In R we consider spheres with radius R, thrown at random with uniform density p. Their centres are distributed according to a homogeneous Poisson law so that the probability P n ( ^ ) to have n points in a volume V is IP„(V) = ^ p e x p ( - p V ) Two points xi and X2 are neighbours iff | x\ - x
Figure 5. The volume of the empty region is minimal if the centres of the spheres almost coincide, but this event has very small probabiUty. Nevertheless, for the same volume of overlapping spheres, the empty volume is minimal if the spheres are in a compact shape, so to say in a sphere of radius r 0 . Then there is a balance between the probabilities of the two following events (see Fig. 6): • to have n centres in a sphere of radius ro, • to have no centre in an annulus }ro, TQ + 2R).
Figure 6.
95 The probability to have a cluster of size n and radius r$ surrounded by a corona of depth 2R is (with TT^ the volume of the unit ball in dimension d): P(„, ro) =
KF
J>
exp(-p7rdr0d) exp [-fmd((r0
+ 2R)d - rft]
This probability is maximal for ro = fo such that: f0(f0 + 2R) d-\
P*d
For d = 2, we obtain r~o(n) / n K 2 R ' = - l + A /l + pwR For a given density (and a given filling factor rj) r"o(n) can be taken to be the radius of the most probable clusters with n spheres. 3.2
Structure of percolating clusters
To find out whether some types of clusters (with a fixed number of elements) have more prominence than others at the percolating threshold, we now intro duce a probabilistic model of percolating clusters. For this we use as building blocks the most probable clusters introduced above. Let O be the centre of a sphere in R d , O is an occupied point. We consider a spherical cluster i.e. an occupied point 0\ surrounded by k occupied points Ai,i = l,...,k, at the same distance IR from 0\ and uniformly distributed around 0\. Take two such clusters {0\,Ai\02,Bj} such as O1O2 = SR
OtAi = OiB, Oi02 = 5R
IR
6 and / are dimensionless parameters. Figure 7. and let Mk(l,8)R be the minimum of the mean square distance between one point of the set {Oi, A{} and one point of the set {O2, Bj}. We shall stipulate
96 that the two clusters percolate if Mk(l,5)R fixed I, a critical value Sc(k) for <S Mk(l,Sc(k)) The set of curves Mk(l,S/l) and 6c(k) satisfies
is less than 2R. This gives, for =2
can be obtained by numerical simulations
M fc (lA(fc)/i) = 2/J
(1)
The proportion of points with fc-neighbours decreases with k. They have a density Xk and if LkR is the mean distance between them we can write (LkR)dXk
« 1
(2)
Xk is given by the Poisson law, i.e. Xk — P^r exp(—a) and a = pTT(i(2R)d = 2dr\ being the mean number of neighbours of one occupied point. Percolation is supposed to appear first when Sc(k) « Lk.
(3)
We focus first on the case d = 2. Using dimensionless variables, the minimum of the mean squared distance between two elements of two different k-clusters is given by the set of curves fk(x) =
\Mk(l,x)}2
which depend only on one parameter x = f and are obtained numerically. If we choose for I the radius of the most probable cluster with fc + 1 elements, I is a function of k and rj:
h(v) = - i + ^ i +
^ -
Moreover the Poisson law can be put in the form:
Then the relations (1), (2), (3) lead to a system of two coupled equations:
{h{x)=^z
\ hk(v) = ,a[!„(„)]» For each value of fc we can determine the critical value T]c(k) given by the lowest intersecting point of the curves
Fk(v) = fk—I
4
lk(v)2\
97 and
4
F
11/2
"fc(77) = [wrm. The percolation critical value r\c is just given by T]C =mmr)c(k)
(4)
K
The value fco for which TJC minimum is attained in (4) can be regarded as a measure of the size of the building blocks of the percolation cluster. The results we obtain with these prescriptions are summarized below. 3.3
Results
• Dimension d = 2 — the curves do not intersect for fc < 3, are almost tangent for fc = 4, but intersect clearly for k w 5. — t)c(k) is almost stationary around k = 5 and very near the value obtained by numerical simulations TJC = 1.18. Moreover the heuristic hypothesis of Alon, Bradberg and Drory, is in reasonable agreement we the result we have obtained. However the percolation seems not to distinguish between fc-clusters for k between 4 and 6. — For k > 6 T]c(k) increases regularly and behaves roughly as ^(k + 1 ) for k —> oo. • Dimension d = 3 The situation is rather different. The clusters are smaller and there is no solution with one type of cluster, compatible with the known physical values of the percolation problem. This may be seen as evidence that in dimension 3 the percolating cluster is more structured and is composed of several types of "building blocks" placed according to an optimal strategy. In order to understand this structure we have to consider more collective phenomena, replacing Xk by the probability of clusters of size greater than or equal to k. Then for I w 0.75 and k = 2 we obtain r\c ~ 0.32 to compare with rjc « 0.35 obtained by numerical simulations.
98
Acknowledgments The BiBoS research center (Bielefeld-Bonn-Stochastic) of Bielefeld University, the Centre de Physique Theorique, CNRS, Marseille and the Mathematical Institute Guido Castelnuovo, Universita La Sapienza, Romal are gratefully acknowledged for kind hospitality and financial support. G.F.D.A. acknowl edges the generous support of the Alexander Von Humboldt Foundation. References 1. Ph. Blanchard, Modelling BJV/AIDS, in Fractals in Biology and Medicine, T. F. Nonnenmacher, G. A. Losa, E. R. Weibel Eds, Birkhaiiser, (1994). 2. H. Satz, Nucl. Phys. A 642, 130 (1998). 3. I. Balberg, Philos. Mag. B 55, 991 (1987). 4. U. Alon, I. Balberg and A. Drory, Phys. Rev. A 42, 4634 (1990). 5. E. J. Garboczi, K. A. Snyder, J. F. Douglas and M. F. Thorpe, Phys. Rev. E 52, 819 (1995). 6. U. Alon, I. Balberg, B. Berkowitz and A. Drory, Phys. Rev. A 43, 6604 (1991). 7. M. B. Isichenko, Rev. Mod. Phys. 64, 4, 961 (1992). 8. D. Stoyan, W. S. Kendall and J. Mecke, Stochastic Geometry and Appli cations, Akademie-Verlag Berlin (1989). 9. V. K. S. Shante and S. Kirkpatrick, Adv. Phys. 20, 325 (1971). 10. P. Hall, Introduction to the Theory of Coverage Processes, Wiley Series in Probability and Mathematical Physics (1988). 11. M. Aizenman, Nucl. Phys. B , 551 (1997). 12. M. Aizenman, A. Burchard, C. M. Newman and D. B. Wilson, to appear in: Random Structures and Algorithms, (1999). 13. G. E. Pike and C. H. Seager, Phys. Rev. Lett. B 10, 1421 (1984). 14. E. T. Gawlinski and H. E. Stanley, J. Phys. A 10, 205 (1977). 15. W. Klein, Phys. Rev. B 26,5, 2677 (1982). 16. P. W. Kasteleyn and C. M. Fortuin, J. Phys. Soc. J. (suppl) 26, 11 (1969). 17. D. Ruelle, Phys. Rev. Lett. 27, 1040 (1971). 18. L. Chayes and R. Kotecky, Comm. Math. Phys. 172, 551 (1995). 19. O. Haggstrom, Mark. Proc. Rel. Fields 4, 275 (1998). 20. G. R. Grimmet, Lectures on probability theory and statistics, Ecole d'ee de Probabilites de Saint Flour, (XXV), Lectures Notes in Mathematics, M. T. Barlow, D. Nualart Eds. Spinger, (1998).
99 R I G G E D HILBERT SPACE R E S O N A N C E S A N D TIME A S Y M M E T R I C Q U A N T U M MECHANICS* A. BOHM AND H. KALDASS University of Texas at Austin Physics Department Austin, TX 78712 E-mail: [email protected] The Rigged Hilbert Space (RHS) theory of resonance scattering and decay is re viewed and contrasted with the standard Hilbert space (HS) theory of quantum mechanics. The main difference is in the choice of boundary conditions. Whereas the conventional theory allows for the in-states <j>+ and the out-states (observables) rl>~ of the S-matrix elements (ip~,(p+) = (V>ou',S0,n) any elements of the HS H, {ip~} = {
1
Introduction
I am happy to be here to celebrate Ludwig Streit. My relation with Ludwig goes back to 1967 when we overlapped as post-docs at Syracuse University. I still remember that at Syracuse Ludwig gave me his unsolicited advise, that I did not heed, which had severe repercussions. He was already at that time much more clever than I. The other thing that I remember often is that sometime in the 1980s Ludwig worked on Gamow vectors and organized a meeting on resonances 1,2. I think the motivation for the meeting was that nobody really knew what resonance states were. At that time my paper on Gamow kets was out al ready 3 and I had written many chapters in a monograph 4 on that subject, but Ludwig did not invite me to that meeting. Then there is Ludwig's interest in the Rigged Hilbert Space 5 . I had been advocating its use in quantum physics for many years, but was not very "Based on a talk delivered at the International Conference on Mathematical Physics and Stochastic Analysis in honor of the 60th birthday of Professor Ludwig Streit.
100
successful. When I found out that Ludwig also was using it, I gained hope again that ultimately the Rigged Hilbert Space will make it into physics. Resonances and Rigged Hilbert Space are the subjects that I want to discuss here. Resonances and decaying states can really not be understood as au tonomous elementary particles in Hilbert space quantum mechanics because the Hilbert space mathematics does not allow state vectors characterized by both an energy ER, and a lifetime r (or a width T = ft/7*)- This is in con trast to the way experimentalists analyze their data and list their results as Breit-Wigner peak value and width (ER,T) for resonances (large values of Y/ER) and as (Ed,T) for decaying states if the (Breit-Wigner or Lorentzian) line shape cannot be resolved but the decay rate can be fitted to an exponen tial (small values of ^ - ) - Since experimental data are finite in number and there are always experimental tolerances and interference with background, one will never be able to say that the experiment has precisely established an ideal Breit-Wigner or an exact exponential. However, Breit-Wigner en ergy distribution and exponential time evolution have been observed so many times for such an enormous number of different systems in all areas of physics and chemistry that one can safely take them as the defining signature of a quasistable state and attribute observed experimental deviations to the ever present background and calculated mathematical deviations to an unjustified mathematical idealization. For the mathematical physicists this means that they have to provide the suitable mathematical idealization. How and why to do that is the content of our discussion here. 2
The fundamental calculational tools of quantum mechanics
In quantum theory one has states denoted by p or W , or by <j> for a pure state p = \(j>)(\,
(1)
and one has observables denoted by A(= A1), A, P (P2 =P),
or by V, if P = \^)(tf>\, for properties.
(2)
The vectors
101
(or 4>) is thus experimentally denned ("determined") by the preparation ap paratus. The quantum physical observables are observed or registered by a registration apparatus, e.g. a detector. The observable A (or ip) or A are thus experimentally defined by the registration apparatus. In experiments with quantum systems one measures ratios of large inte gers Ni/N or N(t)/N, e.g. as ratios of detector counts of the z-th detector JVj and counts of all detectors N or as ratios of detector counts N(t) in the time interval between t = 0 and t = t and in "all" time N = N(oo). This ratio of large numbers is interpreted as probability, e.g. as probability V{Pi) for a property Pi jj*l>(Pi)
(3a)
or as probability for the observable A(() at a time t
M«p(A(0)
(3b)
where the observables P* or A are experimentally given ("defined") by the detector. For a more general observable
X>P<
(4)
one obtains the average value (of the eigenvalues a*) finite
J.J
oo
'=1 finite
,=1
»,
(3c)
oo
The symbol « denotes the association of the experimentally measured quan tity on the left hand side with the theoretically calculated quantity on the right hand side. In quantum theory the probability of an observable A in the state W at time t is calculated as V(t) = V(\(t))
= Tr(A(*)W0) = Tr{AoW{t)).
(5a)
If the state is pure, given by the state vector <j>, and if the observable is a property given by the one-dimensional projector \tp)(ip\, or given by the "observable" vector V, then
P(0 = KVW0>l2 = KiK0WI2-
(5b)
102
The trace in (5) is calculated using any basis system of the space $; either any discrete basis
#^=53|t)<»|0;
(6a)
i
or any continuous basis (Dirac basis vector expansion) $ 9 0 = fd\\X)(X\4>)
(6b)
or any basis system consisting of discrete and continuous (generalized) eigen vectors of a complete system of commuting observables. Thus the trace is given by e.g. Tr(AW) = ^2(i\AW\i)
or Tr(AW) = f dX{X\AW\X).
(7a)
•*
i
For the special case (5b) the probability for the observably ip in the state
2
W)M*WI) = KV#)I
£<#)<#>
(7b)
i=0
or by (7c) if one uses a continuous basis of Dirac kets |A). The time evolution in (5a) and (5b) (dynamics of the quantum system) is given by the Hamiltonian operator H of the system; either as ^
= i[H,W(t))
(8a);
i h ^ - = H<j>{t)
(8b)
(j)[t = 0) = 0o ( Schroedinger picture), or by : ^
= -i[^,A(0]
(8c);
i h
<^
=
-
H m
.
(8d)
(Heisenberg picture) None of the above equations is mathematically precise until we define the space $ , the kets |A) or the integration dX in (6b) and (7c), and specify the initial-boundary conditions
103
3
Empirical reasons for time asymmetric quantum mechanics
The time t in (5) and in the operators A{t) and W(t) is usually allowed to take positive and negative values, i.e. one chooses time symmetric boundary conditions for the Schroedinger and Heisenberg equations of motion (8). This choice is not compelling from an empirical point of view for the following reason. An obvious expression of causality is that a state WQ (or <j>0) has to be prepared first before an observable A(t) or rp(t) can be measured in it. If one calls t = 0 the time at which the preparation of the state W (or <j>) is completed then the time translation in A(t) (or ij)(t)) in (5) makes physically sense only for t > 0. The registration counts N(t) in (3b), by which V(t) is measured, can be taken in the future t > 0, but not in the past. Thus, Ti(A(t)W) will have an experimental counterpart for t > 0 but not for t < 0. The experimental data for V(t) would give that V(t) = Tr(A(t)W)
« ^
for t> 0
(9a)
and V{t) = Tr(A{t)W) « 0 for t < 0
(9b)
because if the detector would click before the state is prepared we would discount this click as noise. Often, such as for stationary states and/or time independent observables, it does not matter at which time V(t) = Tr(A(t)W) = Ti(AW(t)) is calcu lated. But one cannot make it a general principle that time evolution of the observable A(t) (or of the state W(t)) must go into both directions. In fact the most natural description of the experimental situation (9) would be to require a time ordering in the probability formula V(t) = Tr(A(0W(0))
t> t0 = 0
(10)
and admit in the mathematical description only time translations of observ ables A(t) relative to the state W(to) by an amount t-to > 0. This would not only reflect the experimental situation, which forbids time translation of the registration apparatus relative to the preparation apparatus to a time t < t0. It would also incorporate the notion of causality into the mathematical theory on the operator level. The time translation is quite different from other transformations of the space-time symmetry group, e.g. rotations and space-translations. These sym metry transformations of space-time are experimentally realized as transfor mations of the registration apparatus (detector) relative to the preparation
104
apparatus (accelerator). Whereas space translations and rotations of the ap paratuses can be performed back and forth, the time at which the detector is activated can only be shifted into the future not backward past the time of preparation. Thus time translations of the apparatuses form only a semigroup whereas rotations and translations of the apparatuses form a group. Therefore it is natural to represent space translations and rotations by unitary groups of operators in the space of physical states. But it is unjustified to assume that time translations are also always represented by group operators in the space of states, i.e. to assume that in quantum mechanics only states with reversible time evolution exist. When we make the choices that give a precise mathematical meaning to the vectors <j> and ip and the operators W, A, A, A* in (1) and (2) we should therefore not restrict ourselves to those mathematical assumptions that dictate a unitary group evolution. 4
The Hilbert space idealization has reversible time evolution
Of the vectors <j>, tp we have so far assumed that they fulfill the mathemat ical axioms of a linear space with scalar product (
105
in (7a) or other topological notions. The equality w between experimental and mathematical quantities has a meaning only within certain experimental errors and for large numbers N. Therefore, one has to make some—more or less—arbitrary mathematical idealizations in order to obtain complete math ematical structures, like linear topological spaces and topological algebras of operators. One such idealization is the Hilbert space (HS). This is obtained by ad joining to the linear scalar product space $ the limit element of infinite con verging (Cauchy) sequences, with the Hilbert space convergence (or Hilbert space topology TU) defined by
(11)
where the norm ||^>|| is defined by the scalar product |>|| = \J(<j>,
(12)
are not Riemann but are Lebesgue integrals. This means the values of the functions V'(A) = (A|V>) at a particular point (or at all rational numbers) are not defined, which in turn means that the Dirac kets |A) cannot be defined. The Lebesgue integration also makes the interpretation of the probability density |(.E|V>)|2 as the energy resolution of the detector for the observable \ip){ip\ rather unintuitive, at least for those «/> 6 % that do not have a smooth function in the class of Lebesgue integrable functions {(E\ip)} which represent In the (complete) Hilbert space one can define self-adjointness and give the precise meaning of "spectral resolution" to equations like (4). One can make the hypothesis that observables are (not just hermitian or symmetric but) self-adjoint operators and one can prove existence theorems. One of these existence theorems is the Gleason theorem 7 which states that if V{Pi) is the function on the set of projectors {Pi} which fulfills the axioms of probabilities 1 then there exists a positive trace class operator (density x
For V(Pi) or 'P(A) to be a probability it has to fulfill the axioms of probability theory : V(P)>0,
-P(1) = 1, ([PU Pj] =
V{Pi + Pj) = 0,PiPi=0).
V{Pi)+V{Pj),
106
operator) p in U such that V{Pi) = Tr{Pip) = Vp{Pl). This operator p thus defines the state. Since the converse (for any positive trace class operator p, Vp(Pi) = Tr(Pip) fulfills the axioms of probability theory ') is simple to see, one was led to the conclusion 8 that there is one to one correspondence between quantum physical states £> density operators p
(13a)
and in particular between pure quantum physical states o
vectors 4> (up to a phase) in H
(13b)
Another existence theorem (Stone-von Neumann operator calculus 9 ) as serts that the Cauchy problem in quantum mechanics, e.g. in the form of the Schroedinger equation (8b) with the initial condition
-oo<<
(14a)
where 4>(t) depends continuously on the initial condition. This means U^t + r) = t/ f (<)£/* ( r ) , ^^-<j>={-iHYU^(t)4>
UH-t)
= (Ul)-1(t)=U(t)
= U\t){-iHY4>
for 4> eD(H).
(15) (16)
The same result holds for the time evolution of the observable vector ip in H xl>{t) = U(t)4>0 = eiHtip0
- co < t < oo.
(14b)
In terms of the operators W and A this is given by W{t) = U\t)W0U{t)
= e-iHtW0eiHt
\{t) = eiHtA0e~iHt
- oo < t < +00
-oo
(17a) (17b)
2 T h e operator £/'(£) is defined in all of H by the Stone-von Neumann calculus as U^(t) = / ^ ° e~'EtdP(E) not by the exponential series
n=0
which is defined (converges) only on a dense subspace A C D C 7i called analytic vectors.
107
The conclusion that one draws from these two existence theorems is the following : The choice to describe states in the HS theory is very restricted and can only be given by a density operator (positive definite trace-class) and (for a pure state) by a vector <j> 6 K. The time evolution of these states must be the time symmetric reversible group evolution (14). There are no states in the HS quantum mechanics of closed systems (which fulfill (8a) or (8b)) which have asymmetric (irreversible) time evolution 3 . The evolution of the quantum mechanical observables is also time symmetric, i.e. the observables A'(t - to) can evolve relative to the state W(t0) by an arbitrary amount t - t0 > 0 as well as t - t 0 < 0. For every E/(<2 — D (or more general state WD) which has been created at any finite time to / — °° (which we call to = 0) and then decays into decay products. This mathematical consequence is not surprising if one keeps in mind that in the HS theory the t-evolution-where t is the relative time between state and observable-is given by a reversible unitary group. One cannot just impose on the group evolution an arbitrary condition like causality, chopping off one half of the theory, and expect that what remains is still a consistent theory. The reversibility of quantum mechanics in HS and the violation of the causality principle is a consequence of the mathematical idealization given by the topology of the HS, i.e. by (11). It is not a consequence of the fundamental hypothesis of quantum physics as given by the interpretation (5) and by the dynamical equation (8). We shall therefore return to the quantum mechanical Cauchy problem (8) and modify the boundary conditions
Irreversibility in conventional quantum theory is considered to be "non-quantum mechan ical" and always thought of as being due to external influences upon "open" systems. It is described by an additional term on the right hand side of (8a) which is not given by the commutator with H of the quantum system.
108
since we have already a scalar product ( , ) in the linear space $ (and we need the scalar product to calculate such physical quantities like the probabilities \(<j>, ifr)\2), Banach space completion is no more an option. 5
The Rigged Hilbert Space idealization has irreversible semigroup evolution and Gamow vectors with exponential decay
We shall complete the linear scalar product space $ into a locally convex nuclear space with a topology stronger than the Hilbert space topology given by the scalar product. Specifically we shall choose a countable Hilbert space where the meaning of convergence, i.e. the topology T* is defined by a count able number of scalar products ( , ) p , p = 0 , 1 , 2 , . . . where ( , ) p= o = ( > ) is the scalar product of the HS. Convergence with respect to T*, <j>„ A 0, means :
for every p = 0 , 1 , 2 , . . . . This topology is stronger than T « . The countable number of scalar products, i.e. the topology T*, is usually chosen such that the algebra of observables for the physical system under consideration becomes an algebra of r$ -continuous operators. If we denote the "^-completion of the linear space $ again by $ then we have the two complete topological spaces $ C % with $ dense in %. Taking in addition the space $ x of r*-continuous antilinear functionals F(4>) on # , and the space H* of r«-continuous functionals f(h) = (h, f) which are given by the scalar product, we obtain the Gelfand triplet or Rigged Hilbert Space (RHS) 10 $CH
= H* C $ x .
(19)
We shall use the Dirac notation for the r*-continuous functionals F(<j>) = (4>\F), because F(0) is an extension of the scalar product {h,f) to those F € $ x which are not in H. In these RHS's (one for each kind of quantum physical system) Dirac's formalism of kets (with a continuous set of eigenvalues) and the continuous basis vector expansion (6b) attain a mathematical meaning and the integrals in (7) are Riemann integrals. These RHS's also allow for time-asymmetric solutions of the quantum mechanical Cauchy problem (8). In the RHS formulation one can choose different subspaces of V. to distin guish between states and observables. We call $_ the space that describes the
109 states (called in-states in the scattering experiment) prepared by preparation apparatuses (e.g. accelerator). We call $ + the space that describes the ob servables (called out-states in scattering theory) registered by the registration apparatuses (e.g. detector). The two subspaces $_ and $+ are not disjoint. The HS formulation, in contrast, does not allow for this mathematical dis tinction into separate subspaces of states and observables. Thus there is one Hilbert space % and for each quantum mechanical (scattering) system two dense subspaces $ T and therewith two RHS's <J>_ C H C $ ! with the physical interpretation as in-states and
(20a)
$+ C H C $+ with the physical interpretation as out-observables.
(20b)
Mathematically $ T are defined by their realization as function spaces for their energy wavefunctions :
>+ e $_ «• (+E\4>+) eSr\H2_|R+ il>-e*+&(-E\i>-)€Sr\Hl\R+
(2ia) (2ib)
where S denotes the Schwartz space of functions and "H2_, %\ denotes Hardy class functions in the lower and upper complex plane, respectively. (Here complex half-planes refer to the second Riemann sheet of the S-matrix). The ± in {±E\ refers to the ±ie in the Lippmann-Schwinger equation for the eigenkets oiH = H0 + V. The interpretation (20) of the mathematical spaces (21) can be inferred from the preparation -> registration arrow of time n . In addition to the vectors
(Hxp- \ER - tr/2-> = (i,- \H* \ER - tr/2-> =
(ER-iT/2)(ip-\ER-iT/2-)
for all xj)~ 6 $+ (the space of observed decay products). Here ER represents the resonance energy and T the width of the Breit-Wigner energy distribu tion. The superscript " in \ER - iT/2~) is inherited from the kets of the Lippmann-Schwinger equation, the subscripts on the corresponding space $.£ from the mathematicians' notation for Hardy class functions (21). Scattering theory in physics and Hardy class functions in mathematics were developed independently of each other. Except for this discrepancy in the notation for
110
the labels +", the Hardy class spaces $_ provide an excellent mathematical image for prepared states {
{ t )
= e-iHXt\ER-iT/2-)
= e~iE-le-rt/2\ER
-
iT/2~),
for t > 0 only. The Gamow kets are solutions of the Schroedinger equation (8b) but do not fulfill the Hilbert space boundary condition. Instead they fulfill the time asymmetric boundary condition ipG(t = 0) £ $ + . Whereas the first part of (23a) can be formally verified from (22), the derivation of the time asym metry t > 0 is highly non-trivial and requires specific properties of Hardy class functions. The semigroup eZ'H l is only defined for positive values of the time, t > 0, (H* is the operator H^ extended into 4>x defined by the first equality of (22)). There is another semigroup eZtH l,t<0 and another Gamow ket tj)G = \ER + iT/2+) G $ * with the asymmetric time evolution jG(t)
= e~_iHXt\ER+ir/2+) = e-iB^e+r^\ER for t < 0 only .
+
iV/2+),
The Gamow vectors are defined from the pole term of the analytically con tinued 5-matrix at the resonance position at zR = ER — iT/2 (and at zR = ER + iF/2 for (23b)) in the second Riemann sheet. From this one obtains, using the Hardy class property, the Breit-Wigner energy distribution for their wave function 3 ' 13 {-E\i,G) = iJfrbr——± £, - (iSR -
^ - , - 0 0 / / < E < +oo .
(24)
l-^)
The variable E extends over the physical values (upper rim of positive real axis first sheet = lower rim of positive real axis second sheet) and from — oo// to 0 in the second sheet. This is an idealized Breit-Wigner in contrast to the standard Breit-Wigner for which 0 < E < oo. The results (22) and (23) are derived from the pole term definition using the properties of the Hardy spaces (21) 3 ' 13 . If there are iV* resonances in the system, each occurring as a pole of the j - t h partial 5-matrix at the positions zRi = ERi — iIV2, then one obtains N Gamow vectors ipG.
111
The Gamow vectors %l)f are members of a "complex" basis vector expan sion 13 . In place of the well known Dirac basis system expansion (6b) given for the Hamiltonian H by /•+00
4>+ = / dE\E+)(+E\^) (25) Jo (where a discrete sum over bound states has been ignored), every state vector (f>+ £ $ _ can be expanded as
dE\E+){+E\<j>+)
(26)
(where —OOJI indicates that the integration along the negative real axis (or other contours) is in the second Riemann sheet of the 5-matrix). The "com plex" basis vector expansion (26) is rigorous. This allows us to mathemati cally isolate the exponentially decaying states tpf. It also allows us an easy approximation by omitting the background integral in (26), and just using N +
* =!>?>*'
ci = -<^?|0+).
(27)
t=i
Then one obtains the "effective" theories with finite complex Hamiltonian. For instance, for the K\ - K% meson system with TV = 2,
+ ^bL
(28) 14
and one obtains the Lee-Oehme-Yang theory . The finite dimensional ap proximations (27) have been successfully applied to many areas of physics, in particular to nuclear physics 15 ' 2 , which shows that to isolate the Gamow states can be a good approximation. Since (23a) implies the exponential law for the decay rate 6 Vv(t) = Tve~rt the width T of the Breit-Wigner distribution (24) and the lifetime fulfill the exact relation r = p. This has not been obtained before as an exact, precisely derived relation, though it has always been assumed on the basis of some "approximate derivations" 16 . Gamow vectors are ideally suited to describe resonances (the pair (23a) and (23b)) in a scattering process or quasistable particles (23a) that decay. Like the Dirac-Lippmann-Schwinger kets | S ± ) , from which they are con structed, they do not describe interaction-free in- or out- asymptotic states. In a theory that allows only asymptotic particles, they are therefore not ad mitted. Gamow states have all the properties that heuristically the unstable
112
states need to possess. In addition and unintended they give rise to an asym metric time evolution, which may not have been wanted but is in agreement with the empirical principle of causality. Acknowledgement We gratefully acknowledge valuable support of the Welch Foundation for the preparation of this paper and kind hospitality of Rui Vilela Mendes in Lisbon. References 1. Resonances-Models and Phenomena, S. Albeverio, L. S. Ferreira and L. Streit [Eds], Springer Lecture Notes in Physics 211, (Berlin, 1984). 2. M. Baldo, L. S. Ferreira and L. Streit, Nucl. Phys. A 467, 44 (1987); M. Baldo, L. S. Ferreira and L. Streit, Phys. Rev. C 36, 1743 (1987). 3. A. Bohm, J. Math. Phys. 22, 2813 (1981); Lett. Math. Phys. 3, 455 (1978). 4. A. Bohm, Quantum Mechanics, 3rd ed., sections XVIII.6-9 and XX.3, (Springer-Verlag, New York, 1994). 5. L. Streit in Stochastic Analysis and Applications in Physics, A. I. Car doso, et al. [Eds.], (Klumer, Dordrecht, 1994), page 415; M. de Faria and L. Streit in Stochastic Analysis on Infinite Dimensional Spaces, H. Kunita and H. H. Kuo [Eds.], (Longman, 1994), page 52. 6. A. Bohm and N. L. Harshmann in Irreversibility and Causality, p. 225, Sect. 7.4; A. Bohm, H. D. Doebner and P. Kielanowski [Eds.], (Springer, Berlin, 1998). 7. A. M. Gleason, J. Math. 6, 885 (1957). 8. J. von Neumann, Mathematische Grundlagen der Quantentheorie, (Springer, Berlin, 1931) (English translation by R. T. Beyer) (Prince ton University Press, Princeton, 1955). 9. M. H. Stone, Ann. of Math. 33, 643 (1932). 10. A. Bohm and M. Gadella, Dirac Kets, Gamow Vectors and Gel'fand Triplets, Lecture Notes in Physics 348, (Springer-Verlag, Berlin, 1989). 11. A. Bohm, I. Antoniou and P. Kielanowski, Phys. Lett. A 189, 442 (1994); A. Bohm, I Antoniou and P. Kielanowski, J. Math. Phys. 36, 2593 (1995). 12. E. P. Wigner, Symmetries and Reflections, (Indiana University Press, Bloomington, 1967), page 38. 13. A. Bohm, S. Maxson, M. Loewe and M. Gadella, Physica A 236, 485 (1997). 14. T. D. Lee, R. Oehme and C. N. Yang, Phys. Rev. 106, 340 (1957).
113
15. L. S. Ferreira.in Resonances, E. Brandas [Ed], Lecture Notes in Physics 325, (Springer, 1989), p. 201; C. Mahaux, \ p. 139; 0 . I. Tolstikhin, V. N. Ostrovsky and H. Nakamura, Phys. Rev. 58, 2077 (1998); V. I. Kukulin, V. M. Krasnopolsky and J. Horacek, Theory of Resonances, (Kluwer Acad. Publ., 1989). 16. e.g. M. L. Goldberger and K. M. Watson, Collision Theory, (Wiley, New York, 1964); Newton, Scattering Theory of Waves and Particles, 2nded., (Springer-Verlag, 1982), chapter 8.
114 Q-ISING N E U R A L N E T W O R K D Y N A M I C S : A COMPARATIVE R E V I E W OF VARIOUS A R C H I T E C T U R E S D. BOLLE AND G. JONGEN Instituut voor Theoretische Fysica, K.U. Leuven, B-3001 Leuven, Belgium E-mail: {desire.bolle, greetje.jongen}@fys.kuleuven.ac.be. G. M. SHIM Department of Physics, Chungnam National University Yuseong, Taejon 305-764, R-O. Korea E-mail: [email protected]. This contribution reviews the parallel dynamics of Q-Ising neural networks for var ious architectures: extremely diluted asymmetric, layered feedforward, extremely diluted symmetric, and fully connected. Using a probabilistic signal-to-noise ratio analysis, taking into account all feedback correlations, which are strongly depen dent upon these architectures the evolution of the distribution of the local field is found. This leads to a recursive scheme determining the complete time evolution of the order parameters of the network. Arbitrary Q and mainly zero tempera ture are considered. For the asymmetrically diluted and the layered feedforward network a closed-form solution is obtained while for the symmetrically diluted and fully connected architecture the feedback correlations prevent such a closed-form solution. For these symmetric networks equilibrium fixed-point equations can be derived under certain conditions on the noise in the system. They are the same as those obtained in a thermodynamic replica-symmetric mean-field theory approach.
1
Introduction
Artificial neural networks have been widely applied to memorize and retrieve information. During the last few years there has been considerable interest in neural networks with multistate neurons (see * and references cited therein) in order to function as associative memories for gray-toned or coloured patterns or to allow for a more complicated internal structure in the retrieval, e.g., a distinction between background and patterns. Here we review the dynamics of so-called Q-Ising neural networks (see the references in 2 ) for arbitrary Q. They are built from Q-Ising spin-glasses 3,4 with couplings defined in terms of patterns through a learning rule. For Q = 2 one finds back the Hopfield model 5 ' 6 , for Q -» oo one has an analog network (see 7 and references therein). One of the aims of these networks is to memorize a number of patterns and find them back as attractors of the
115
retrieval process. Consequently these networks are also interesting from the point of view of dynamical systems. Besides a learning rule one also needs to specify an architecture indicating how the spins (=neurons) are connected with each other. Several architectures have been studied in the literature for different purposes. From a practical point of view mostly perceptrons or, more general, feedforward layered net works are used since a very long time (see, e.g., 8 for a history). Hopfield 5 6 ' studied a fully connected network with symmetric couplings because it satisfies the detailed balance principle and hence a Hamiltonian can be de fined. Asymmetrically diluted models 9 were used because their dynamics can be solved exactly and because they can learn us something about the loss of information content when some of the synaptic couplings break down. In this contribution we review the study of the parallel dynamics of these types of network using a probabilistic approach (see, e.g., 1 0 ' H ) . In more detail, employing a signal-to-noise ratio analysis based on the law of large numbers (LLN) and the central limit theorem (CLT) we derive the evolution of the distribution of the local field at every time step. This allows us to obtain a recursive scheme for the evolution of the relevant order parameters in the system being, in general, the main overlap for the condensed pattern, the mean of the neuron activities and the variance of the residual overlap responsible for the intrinsic noise in the dynamics of the main overlap (sometimes called the width parameter). The details of this approach depend in an essential way on the architecture because different temporal correlations are possible. For extremely diluted asymmetric and layered feedforward architectures recursion relations have been obtained in closed form directly for the relevant order parameters 9 . 12 . 13 . 14 . This has been possible because in these types of networks there are no feedback correlations as time progresses. As a technical consequence the local field contains only Gaussian noise leading to an explicit solution. For the parallel dynamics of networks with symmetric connections, how ever, things are quite different Mo.n Even for extremely diluted versions of these systems 15>16'17 feedback correlations become essential from the second time step onwards, complicating the dynamics in a nontrivial way. There fore, explicit results concerning the time evolution of the order parameters for these models have to be obtained indirectly by starting from the distribution of the local field. Technically speaking, both for the symmetrically diluted and fully connected architectures the local field contains both a discrete and a normally distributed part. The difference between the diluted and fully connected models is that the discrete part at a certain time t does not involve the spins at all previous times t-l,t — 2,... up to 0 but only the spins at time
116
step t — 1. But in both cases the discrete part prevents a closed-form solution of the dynamics for the relevant order parameters. Nevertheless, the devel opment of a recursive scheme is possible in order to calculate their complete time evolution. In this way a comparative discussion of the parallel dynamics at zero temperature for the various architectures specified above is possible. Finally, by requiring the local field to become time-independent implying that some correlations between its Gaussian and discrete noise parts are ne glected we can obtain fixed-point equations for the order parameters. It turns out that they are equivalent to the fixed-point equations obtained through a thermodynamic replica-symmetric mean-field theory approach. At this point we remark that we do not aim for complete rigour in our derivations. From the point of view of rigorous mathematics, the Hopfield model and, in general, spin-glass theory is recognized to be an extremely difficult, if not imposible, field. For a recent overview of the modest results obtained, mostly concerning thermodynamics, we refer to 18 . The rest of this contribution is organized as follows. In Section 2 we introduce the model, its dynamics and the Hamming distance as a macro scopic measure for the retrieval quality. In Section 3 we use the probabilistic approach in order to derive a recursive scheme for the evolution of the distribu tion of the local field, leading to recursion relations for the order parameters. The differences between the various architectures are outlined. We do not aim for complete rigour and mostly concentrate on zero temperature. In Sec tion 4 we discuss the evolution of the system to fixed-point attractors. Some concluding remarks are given in Section 5.
2
Q-Ising neural networks
Consider a neural network A consisting of AT neurons which can take values Oi from a discrete set S — {-1 = s\ < s 2 < . . . < SQ = +1}. The p patterns to be stored in this network are supposed to be a collection of independent and identically distributed random variables (i.i.d.r.v.), {£? 6 S}, fi £ V = { 1 , . . . ,p} and i € A, with zero mean, E[^] = 0, and variance A = Var[£f]. The latter is a measure for the activity of the patterns. We remark that for simplicity we have taken the patterns and the neurons out of the same set of variables but this is no essential restriction. Given the configuration a\(t) = {
hi(crA(t)) = Y/Jij(t)oj(t) ;6A
(1)
117
with Jij the synaptic coupling from neuron j to neuron i. In the sequel we write the shorthand notation /iA,i(t) = hi(
J D
« = ^£tftf
J c
Z = 4iZGG
■^(0 = ^ £ with the {cij = 0,1},i,j Pr{aj - x} = (1 - C/N)SXt0 symmetric dilution, and c^ for asymmetric dilution. At zero temperature all rule
for
**>''
for
<*>•
^ + 1)^(0,
W (3) (4)
€ A chosen to be i.i.d.r.v. with distribution + (C/N)8Xti and satisfying Cy — Cji, cu = 0 for and Cji statistically independent (with ca — 0) neurons are updated in parallel according to the
<Ti(t) -> <Ji{t + l)=sk:
mine i [s|o- A (t)] = ti[sk\
(5)
We remark that this rule is the zero temperature limit T = / 3 _ 1 —> 0 of the stochastic parallel spin-flip dynamics defined by the transition probabilities « . , , , , > ,
c
o
U
^M _
exP[-/3£i(Sfc|
Pi{(Ji{t + l) = sk €S\
= = . E 8 € 5 e x Pl-/ 3 € <0 s l < T Mt))] Here the energy potential €i[s|crA] is defined by 19 uls\
f(.,
(6)
(7)
118
where 6 > 0 is the gain parameter of the system. The updating rule (5) is equivalent to using a gain function g 6 (),
g,,(hAM) Q
gb(x) = £
*k P [b{sk+i +sk)-x}-6
[b(sk + 8k-t) - x]]
(8)
t=i
with so = - c o and SQ+1 = +oo. For finite Q, this gain function g6(-) is a step function. The gain parameter b controls the average slope of g6(-)In order to measure the retrieval quality of the system one can use the Hamming distance between a stored pattern and the microscopic state of the network
i€A
This introduces the main overlap and the arithmetic mean of the neuron activities
igA
t€A
We remark that for Q = 2 the variance of the patterns A = 1, and the neuron activity a\(t) = 1. 3 3.1
Solving the dynamics Correlations
We first discuss some of the geometric properties of the various architectures which are particularly relevant for the understanding of their long-time dy namic behaviour. For a fully connected architecture there are two main sources of strong correlations between the neurons complicating the dynamical evolution : feed back loops and the common ancestor problem 20 . Feedback loops occur when in the course of the time evolution, e.g., the following string of connections is possible: i —y j —>■ k —>• i. We remark that architectures with symmetric connections always have these feedback loops. In the absence of these loops the network functions in fact as a layered system, i.e., only feedforward con nections are possible. But in this layered architecture common ancestors are still present when, e.g., for the sites i and j there are sites in the foregoing time steps that have a connection with both i and j .
119 In extremely diluted asymmetric architecture these sources of correlations are absent. This class of neural networks was introduced in connection with Q = 2-Ising models 9 . We recall that the couplings are then given by eq. (2) and that in the limit N —> co two important properties of this network are essential 9 ' 2 1 . The first property is the high asymmetry of the connections, viz. Pr{ C l J = c Jt } = ( § )
,
Pr{cy = l A C i i = 0 } = ^ l - ^ .
(11)
Therefore, the number of symmetric connections in the infinite configuration c = {cij}, i, j ^ i € N is finite with probability one, i.e. almost all connections of the graph GN(C) = {(i,j) ■ Cy; = l,i, j ^ i € N} are directed : Cy ^ Cj{. The second property in the limit of extreme dilution is the directed lo cal Cayley-tree structure of the graph Gn(c). By the arguments above the probability FJ. ' (c) that k connections are directed towards a given site i 6 A is
where jf/ = {cji = l , j £ A \ i } is the in-tree for i and |T> m, | its cardinality. This probability is equal to Pr{& = |l\ ( o u t ) | = \{Cij = 1, j € A \ i}\} for connections directed outward a given site i € A. In the limit of extreme dilution we get a Poisson distribution : lim FlA)(C) = ~e-c.
(13)
Hence, the mean value of the number of in (out) connections for any site i E A is E[\T^in)ioui)\] = C. The probability that two sites i and i' have site j as a common ancestor is obviously equal to C/N. From £[17^ '|] = C it follows that after t time steps the cardinality of the cluster of ancestors for site i will be of the order of Cl. The same is valid for site i'. Therefore, the probability that the sites i and i' have disjoint clusters of ancestors approaches (1 - Cl/N)c' ~ exp(-C 2 7A0 for N » 1. So we find that in the limit of extreme dilution : (i) Almost all (i.e. with probability 1) feedback loops in GN(C) are eliminated, (ii) With probability 1 any finite number of neurons have disjoint clusters of ancestors. So we first dilute the system by taking N -v oo and then we take the limit C -* oo in order to get infinite average connectivity allowing to store infinitely many patterns p.
120 This implies that for this asymmetrically diluted model at any given time step t all spins are uncorrelated and, hence, the first step dynamics describes the full time evolution of the network. For the symmetrically diluted model the architecture is still a local Cayley-tree but no longer directed and in the limit N —> oo the probabil ity that the number of connections Ti = {j £ A|cy = 1} giving information to the the site i 6 A is still a Poisson distribution with mean C = £[|7i|]. However, at time t the spins are no longer uncorrelated causing a feedback from t > 2 onwards I 7 . In order to solve the dynamics we start with a discussion of the first time step dynamics, the form of which is independent of the architecture. 3.2
First time step
Consider a fully connected network. Suppose that the initial configuration of the network {a-j(0)},z G A, is a collection of i.i.d.r.v. with mean E[
ml>0.
(14)
This pattern is said to be condensed. By the law of large numbers (LLN) one gets for the main overlap and the activity at t = 0 m^O) = Jim m\(0) £
^ [ ^ ( 0 ) ] = mj
o(0) = lim o A (0) = E[
(15) (16)
where the convergence is in probability 22 . In order to obtain the configuration at t = 1 we first have to calculate the local field (1) at t = 0. To do this we employ the signal-to-noise ratio analysis (see, e.g., 10 ' 12 ). Recalling the learning rule (3) we separate the part containing the condensed pattern, i.e., the signal, from the rest, i.e., the noise to arrive at
M-A(0)) = ^
£ ^(0) + V 5 ^ E &~m £ tf"i(o)
(17) where a = p/N. The properties of the initial configurations (14)-(16) assure us that the summation in the first term on the r.h.s of (17) converges in the limit N -¥ oo to
^^E^W"-1^)ieA\t
(18)
121
The first term £?m1 (0) is independent of the second term on the r.h.s of (17). This second term contains the influence of the non-condensed patterns causing the intrinsic noise in the dynamics of the main overlap. In view of this we define the residual overlap r " ( 0 = lim r£(<) = lim - ± = £ # o , - ( « )
/i£P\{l}.
(19)
jGA
Applying the CLT to this second term in (17) we find
„"5.,/f £ ««*«> = J S . ^ p E ^
£ ^ (0)(20)
= >/a-V(0,XZJ(0))
(21)
where the quantity M(0, V) represents a Gaussian random variable with mean 0 and variance V and where D(Q) = Var[r'i(0)] = a(0). Thus we see that in fact the variance of this residual overlap, i.e., D(t) is the relevant quantity characterising the intrinsic noise. In conclusion, in the limit N -> oo the local field is the sum of two independent random variables, i.e. hi(0)=
lim h A i i ( 0 ) i £ r o 1 ( 0 ) + >/aA/r(0,a(0)).
(22)
For a more rigourous discussion of the first time step for the underlying spinglass model we refer to 23 . At this point we note that the structure (22) of the distribution of the local field at time zero - signal plus Gaussian noise is typical for all architectures discussed here because the correlations caused by the dynamics only appear for t > 1. Some technical details are different for the various architectures. The first change in details that has to be made is an adaptation of the sum over the sites j to A for the layered feedforward architecture and to Tj, the part of the tree connected to neuron i, in the diluted architectures. The second change is that for the diluted architectures an additional limit C -> oo has to be taken besides the N -> oo limit. So in the thermodynamic limit C, N -> oo all averages will have to be taken over the treelike structure, viz. jj £ < G A -> £ £ i 6 7 v • Furthermore a = p/N has to be replaced by a = p/C. 3.3
Recursive dynamical scheme
The key question is then how these quantities evolve in time under the parallel dynamics specified before. For a general time step we find from the LLN in
122
the limit C, N —► oo for the main overlap and the activity (10) m\t
+ l) £
l^\{ai{t
+ \))0)),
a(t + l) £
(({*i(t + l))}))
(23)
with the thermal average denned as
( / ( . ( . + i))>,=V
/w
T^; ( ' ) ^T ) l
<M>
E a e s exp[±/?<7(M0 - for)] where hi(t) = lim^^oo h,\ti(t). In the above ((•)) denotes the average both over the distribution of the embedded patterns {ff} and the initial configurations {<7j(0)}. The average over the latter is hidden in an average over the local field through the updating rule (8). In the sequel we focus on zero temperature. Then, using eq. (8) these formula reduce to
m^t+l)
P
^ jUlgMt)))),
a(t + l) ^
((gUhiit)))) ■
(25)
As seen already in the first time step, we have to study carefully the influ ence of the non-condensed patterns causing the intrinsic noise in the dynamics of the main overlap. The method used to obtain these order parameters is then to calculate the distribution of the local field as a function of time. In order to determine the structure of the local field we have to concentrate on the evolution of the residual overlap. The details of this calculation are very technical and depend on the precise correlations in the system and hence on the architecture of the network as discussed before. For these technical de tails we refer to the relevant literature 1.2.12,14,15 } j e r e w e g j v e g^ extensive discussion of the results obtained. In general, the distribution of the local field at time t + 1 is given by M< +1) = tlml(t
+1) +Af(0,aa(t
+1)) + X(t)[F(hi{t) -gml(t))
+ Baai{t)} (26) where F and B are binary coefficients given below, which depend on the specific architecture. From this it is clear that the local field at time t consists out of a discrete part and a normally distributed part, viz. hi(t) = Mi(t)+M(0,V(t))
(27)
where Mi(t) satisfies the recursion relation M{(t + 1) = X(t)[F(Mi(t)
- Zlml(t))
+ Baai(t)} + ^ m 1 ^ + 1)
(28)
and where V(t) = aAD{t) with D(t) itself given by the recursion relation D(t + 1) = ^