Mathematical Physics, Analysis and Geometry 3: 1–31, 2000. © 2000 Kluwer Academic Publishers. Printed in the Netherlands.
1
Asymptotic Distribution of Eigenvalues of Weakly Dilute Wishart Matrices A. KHORUNZHY1 and G. J. RODGERS2
1 Institute for Low Temperature Physics, Kharkov 310164, Ukraine.
e-mail:
[email protected] 2 Department of Mathematics and Statistics, Brunel University, Uxbridge, Middlesex UB8 3PH, U.K. e-mail:
[email protected] (Received: 30 September 1998; in final form: 18 October 1999) Abstract. We study the eigenvalue distribution of large random matrices that are randomly diluted. We consider two random matrix ensembles that in the pure (nondilute) case have a limiting eigenvalue distribution with a singular component at the origin. These include the Wishart random matrix ensemble and Gaussian random matrices with correlated entries. Our results show that the singularity in the eigenvalue distribution is rather unstable under dilution and that even weak dilution destroys it. Mathematics Subject Classifications (2000): 15A52, 82B44. Key words: matrices, random, dilute, Wishart.
1. Introduction Random matrices of large dimensions play a central role in a number of theoretical physics applications, such as statistical nuclear physics, solid state physics, statistical mechanics, including neural network theory and quantum field theory (see, e.g., the monographs and reviews [8 – 10, 13, 24, 27]). In this work, most interest is attached to the various ensembles of random matrices whose entries are all of the same order of magnitude. This corresponds to the situation when the elements of a system all strongly interact with one another. In the last decade, however, systems in which some of the links between the different elements are broken have been studied in a variety of applications. This effect is particularly important in neural network theory where the total number of neurons is several orders of magnitude greater than the average number of connections per neuron [2, 5, 13]. Such matrices known as dilute (or sparse) ones are also important in other applications, such as the theory of random graphs and linear programming [21].
2 1.1.
A. KHORUNZHY AND G. J. RODGERS
STRONG DILUTION AND SEMICIRCLE LAW
In papers [17, 19] we have studied the limiting eigenvalue distribution of large random matrices that are strongly diluted. These are determined as the N-dimensional matrices that have, on average, pN nonzero elements per row and 1 pN N as N → ∞. We proved that under natural conditions the limiting eigenvalue distribution of strongly dilute random matrices exists and coincides with the Wigner’s famous semicircle law [32]. The semicircle law is also valid when the entries of the dilute random matrix are statistically dependent random variables [16, 20]. This case is of special interest in applications (see, for instance, [1, 7, 10, 13]). In the pure (nondilute) regime these matrices have singularities in the eigenvalue distribution. The strong dilution removes this singularity because the density of the semicircle distribution is bounded. It should be noted that the Wigner’s semicircle distribution is typical for large random matrices with jointly independent entries. Therefore we have conjectured that the semicircle law arises in the ensembles of [16, 20] because the strong dilution eliminates the statistical dependence between random matrix entries. The same reasoning can explain the disappearance of the singularity in the eigenvalue distribution. However, the last conjecture is not true. 1.2.
WEAKLY DILUTE WISHART MATRICES
To study the transition to the semicircle law under dilution, we pass to the case of weakly dilute random matrices. This means that we are now interested in the asymptotic regime when pN = qN, q > 0 as N → ∞. In this case the statistical dependence between random matrix entries persists, provided it exists in the pure (nondilute) ensemble. We consider two random matrix ensembles with different types of statistical dependence between the entries. These are the Wishart random matrices HN and Gaussian random matrices AN with correlated entries. The first ensemble is widely known in multivariate statistical analysis (see, e.q., [1]). Recent applications of these matrices are related with the the theory of disordered spin systems of statistical mechanics [14, 26, 24] and learning algorithms of memory models of neural network theory [2, 13, 14]. The entries {HN (x, y)} are statistically dependent (but uncorrelated) random variables. The degree of the dependence between HN (x, y) and HN (s, t) does not relate to the ‘distance’ |x − s| + |y − t|. In contrast, correlations between matrix elements AN (x, y) and AN (s, t) in the second ensemble we consider decay when the distance between them increases. Due to this property, {AN } can be regarded as the ensemble intermediate betweem random matrix models with strongly correlated entries (see, e.g., [7, 10]) and random matrices with independent entries. 1.3.
MAIN RESULTS AND STRUCTURE OF ARTICLE
We study the limit of the normalized eigenvalue counting function of weakly dilute real symmetric matrices HN and AN as N → ∞. We derive explicit equations for
EIGENVALUES OF WEAKLY DILUTE WISHART MATRICES
3
the Stieltjes transform of the limiting distribution functions and determine recurrent equalities for their moments. Basing on these relations, we study the properties of the limiting eigenvalue distributions. We show that both these distributions are different from the semicircle law. We prove that, nevertheless, the singularity disappears from the limiting eigenvalue density and that this happens for arbitrary values of q < 1. Thus, our principal conclusion is that the singularity of the eigenvalue distribution is rather unstable under dilution and is destroyed even when this dilution is weak. To complete this introductory section, let us note that our results can be regarded as generalizations of the statements proved for strongly dilute random matrices in [19] and [20]. In this paper we use the technique developed in [20]. However, the case of weak dilution studied here is more complicated and requires more accurate analysis. The paper is organised as follows. In Section 2 we present our main results for the Gaussian random matrices AN with correlated entries. In Section 3 we consider the weak dilution of the Wishart matrices HN . We prove the existence of the limiting eigenvalue distributions σ (i), i = 1, 2 in the limit N → ∞ (Theorems 2.1 and 3.1). At the second part of each of these sections we formulate Theorems 2.2 and 3.2 concerning the properties of the respective σ (i) . In Sections 4 and 5 we derive the main equations that determine the eigenvalue distributions of AN and HN , respectively, as N → ∞. Section 6 contains the proofs of Theorems 2.2 and 3.2. We also present there the recurrent relations for the moments of σ (i) and make conclusions about the support of the measure dσ (i) . In Section 7 we give a summary of our results.
2. Gaussian Random Matrices with Correlated Entries Let us consider N × N symmetric random matrices 1 AN (x, y) = √ a(x, y), N
x, y = 1, . . . , N,
(2.1)
where random variables {a(x, y), x 6 y, x, y ∈ N} have a joint Gaussian distribution. We assume that {a(x, y)} satisfy the following conditions: Ea(x, y) = 0, Ea(x, y)a(s, t) = V (x − s)V (y − t) + V (x − t)V (y − s),
(2.2a) (2.2b)
where the sign E represents the mathematical expectation with respect to the measure generated by the a(x, y)’s and V (x) is a nonrandom function such that V (−x) = V (x) and V is nonnegatively defined. Then the right-hand side satisfies conditions for the covariance of random variables (see Lemma 4.5 at the end of Sec-
4
A. KHORUNZHY AND G. J. RODGERS
tion 4). The eigenvalue distribution of the random matrix ensemble (2.1), (2.2), where V (x) satisfies the condition X |V (x)| ≡ Vm < ∞, (2.3) x
was studied in [18]. This case is known as weakly correlated random variables. Indeed, condition (2.3) implies decay of the correlations between random matrix entries AN (x, y) and AN (s, t) that are situated far enough from each other in the matrix. 2.1.
ENSEMBLE AND MAIN EQUATIONS
In this section we consider the ensemble of real symmetric random matrices 1 (q) AN = √ a(x, y)πxy , N
x, y = 1, . . . , N,
(2.4)
where a(x, y) are the same as in (2.1) and the random variables {πxy , x 6 y} are both independent between themselves and independent from {a(x, y)}. We assume that πyx = πxy and the random variables have the common probability distribution 1 1, with probability q, (2.5) πxy = √ q 0, with probability 1 − q. Our main result concerns the normalized eigenvalue counting function (NCF) of (q) AN given by the formula (q) 6 λ N −1 , (2.6) σ λ; AN = # λ(N) j (q)
6 · · · 6 λ(N) where λ(N) 1 N are eigenvalues of AN . THEOREM 2.1. Assume that (2.3) holds. Then (q)
(i) given q ∈ (0, 1) the NCF σ (λ; AN ) weakly converges in probability as N → ∞ to a nonrandom function σq (λ); R (ii) the Stieltjes transform fq (z) = (λ − z)−1 dσq (λ), Imz 6= 0 can be found from the the system of equations Z 1 g˜q (p; z) dp, (2.7a) fq (z) = 0
g˜q (p; z) =
− z − (1 − q)v fq (z) − q V˜ (p)
Z
2
0
where V˜ (p) =
X x∈Z
V (x)E{2π ipx},
1
V˜ (r)g˜q (r; z) dr
−1 , (2.7b)
EIGENVALUES OF WEAKLY DILUTE WISHART MATRICES
and
Z v ≡ V (0) =
1
5
V˜ (p) dp;
0
(iii) system (2.7) is uniquely solvable in the class 0 of functions g(p; z), p ∈ (0, 1), z ∈ C± analytical in this region and such that Im g(p; z)Im z > 0, z ∈ C± .
(2.8)
Remarks. (1) Here and below we mean by the weak convergence of nonnegative nondecreasing functions σ (λ; AN ) the weak convergence of the corresponding measures Z ∞ Z ∞ ϕ(λ) dσ (λ; AN ) = ϕ(λ) dσ (λ), ϕ ∈ C0∞ (R). lim N→∞
−∞
−∞
Generally the convergence of integrals can be regarded as convergence in average, in probability or with probability 1. (2) Each function gˆR∈ 0 determines a nonnegative nondecreasing function σˆ (λ) such that [11] g(z) ˆ = (λ − z)−1 dσˆ (λ) and Z 1 b Img(µ ˆ + iη) dµ. (2.9) σˆ (a) − σˆ (b) = lim η↓0 π a Relation (2.9), known as the inversion formula for the Stieltjes transform, is valid for all a, b such that σˆ is continuous at these points. Theorem 2.1 is proved in Section 4. Basing on (2.7), one can study the properties of σq (λ). 2.2.
LIMITING EIGENVALUE DISTRIBUTION
To discuss the consequences of Theorem 2.1, let us note first that relations (2.7) considered with q = 0 can be reduced to the equation f0 (z) =
1 . −z − v 2 f0 (z)
(2.10)
This equation is uniquely solvable and determines the Wigner semicircle distribution σ0 (λ) [32] with the density √ 1 4v 2 − λ2 , if |λ| 6 2v, 0 %0 (λ) ≡ σ0 (λ) = (2.11) 2π v 0, otherwise. This observation shows that Theorem 2.1 generalizes the results of paper [20], where the eigenvalue distribution of the ensemble (2.4) has been studied under con(N) that take values N 1/2 p −1/2 dition that random variables πxy are replaced by πˆ xy with probability p/N and 0 with probability 1 − p/N. Then the limit p, N → ∞
6
A. KHORUNZHY AND G. J. RODGERS
considered in [20] corresponds to subsequent limiting transitions N → ∞ and q → 0. Another limiting transition q → 1 in (2.7) leads to equations Z 1 g1 (p; z) dp, f1 (z) = 0
g1 (p; z) =
− z − V˜ (p)
Z
V˜ (r)g1 (r; z) dr
−1 .
This system has been derived in [18] the the limit f1 (z) of the Stieltjes transforms of the NCF σ (λ; AN ) of the ensemble determined by (2.1), (2.2), and (2.3) (see also [3, 4] for others and more general ensembles). It should be noted that if Z 1 dp = ∞, (2.12) 0 V˜ (p) then corresponding to f1 (z), measure σ1 (dλ) has an atom at the origin. Indeed, one can easily derive from (2.7) and (2.9) that if (2.12) holds, then lim Imf1 (iε) = ∞ ε↓0
(see also [4]). Relation (2.12) is known in stochastic analysis as the interpolation condition for the infinite sequence of Gaussian random variables {γ (x), x ∈ Z} with zero average Eγ (x) = 0 and covariance Eγ (x)γ (y) = V (x − y) [22]. If (2.12) holds, then the sequence {γ (x)} can be regarded as insufficiently random. Apparently, random dilution makes the sequence of such random variables ‘more disordered’. This observation can be regarded as a heuristic explanation of the following proposition. THEOREM 2.2. If q < 1, then the density %q (λ) = σq0 (λ) exists and is bounded √ everywhere by 1/(π v 1 − q). Theorem 2.2 is proved in Section R 6. There we also derive and analyse recurrent relations for the moments Lk = λk dσq (λ). Basing on these relations, we study the support of the measure dσq (λ). 3. Weak Dilution of Wishart Matrices In this section we study the eigenvalue distribution of symmetric matrices (q) HN (x, y)
N 1 X = ξµ (x)ξµ (y)πxy , N µ=1
x, y = 1, . . . , N,
(3.1)
where {ξµ (x), x, µ ∈ N} are independent random variables having joint Gaussian distribution. We assume that these random variables satisfy conditions Eξµ (x) = 0,
Eξµ (x)ξν (y) = δxy δµν ,
(3.2)
7
EIGENVALUES OF WEAKLY DILUTE WISHART MATRICES
where δxy is the Kronecker δ-symbol; 1, if x = y, δxy = 0, if x 6= y.
We assume also that πxy = πyx and {πxy , x 6 y are i.i.d. random variables (independent of {Hµ}) that have probability distribution (2.5). Thus, (3.1) represents the weak dilution of random matrices m 1 X ξµ (x)ξµ (y), HN,m (x, y) = N µ=1
x, y = 1, . . . , N
(3.3)
known since 30s in the multivariate statistical analysis as the Wishart matrices [1]. Being at present of considerable importance in this field, the ensemble (3.3) is extensively studied in the statistical mechanics of the disordered spin systems (see, e.g., [6, 26, 31] for rigorous results). Another important application of (3.3) is related with the neural network theory, where HN are used as the interation matrix of learning algorithms modelling auto-associative memory. In this approach, N-dimensional vectors ξµ are regarded as the patterns to be memorised by the system [13]. Dilution versions of (3.3) are important in this field of applications as the models that can be tuned to give more precise correspondence with real systems (see, e.g., [2]). These models are mostly studied in the regime of strong dilution [5, 30]. The following statement concerns the normalized eigenvalue counting function (2.6) of of weakly dilute random matrices (3.1), (3.2). (q)
THEOREM 3.1. For each fixed q ∈ (0, 1) the NCF σ (λ; HN ) converges in the limit N, m → ∞, m/N → c > 0 to a nonrandom function σq,c . The Stieltjes transform fq,c (z) of σq,c (λ) satisfies equation −1 √ cu2 q 4 fq,c (z) = − z − cu (1 − q)fq,c + . (3.4) √ 1 + u2 qfq,c (z) This equation is uniquely solvable in the class of functions 0 determined in Theorem 2.1 and satisfying (2.8). We prove Theorem 3.1 in Section 5. Regarding (3.4), one can easily observe that, in complete analogy with (2.7), this equation determines a ‘mixture’ of two equations: the one for the semicircle distribution with q = 0 (cf. with (2.10)) f0,c (z) =
1 −z − cu4 f0,c (z)
and the equation (q = 1) −1 cu2 f1,c (z) = − z + 1 + u2 f1,c (z)
(3.5)
8
A. KHORUNZHY AND G. J. RODGERS
derived in [23] for the Stieltjes transform of σ (λ; HN,m ) in the limit N, m → ∞, m/N → c > 0. Corresponding to (3.5) eigenvalue distribution has the density given by the formula [23] q 2 1 dσ1,c (λ) = [1 − c]+ δ(λ) + 4cu2 − λ − (c + 1)u2 , (3.6) 2 dλ 2π u λ where [x]+ = max(0, x) and δ(x) is the Dirac delta function. Let us stress that if c < 1, then the density of σ1,c (λ) has the singular component at the origin. The following statement shows that this singularity disappears in the weak dilution regime. THEOREM 3.2. If q < 1 then√the density of σq,c (λ) determined by fq,c (z) (3.4) is bounded from above by 1/u2 c(1 − q) for all c > 0. We prove Theorem 3.2 in Section 6. In this section we also derive recurrent R relations for the moments Lk = λk dσq,c (λ). 4. Proof of Theorem 2.1 In this section we use the resolvent method developed in a series of papers (see, e.g., [18]) and improved in [15]. This method is based on the derivation of relations for the moments of the normalized trace of the resolvent of random matrix AN 1 1 X fN (z) ≡ Tr GN (z) = GN (x, x; z), GN (z) = (AN − z)−1 . (4.1) N N x=1 The important and in certain sense characteristic property of random matrices is that EfN (z) converges as N → ∞ to the variable f (z) and one can derive closed equations for it. The variance of fN (z) vanishes as N → ∞ that means that the normalized trace (4.1) is the self-averaging random variable. The case of weakly diluted random matrices AN,q (2.7) is more complicated. The average EGN (x, y) is expressed in terms of the generalized ‘trace’ N 1 X TN (x, y) = πxr G(r, s)Vrs πsy , N r,s=1
(4.2)
where we denoted G(x, y) ≡ GN (z) = (AN,q − z)−1 (q)
and Vxy ≡ V (x − y). Relations for the limit of EGN (x, y) involve the limit of the average ETN (x, y). Our main observation is that this pseudo-trace TN (x, y) is also the selfaveraging variable as N → ∞. We obtain expression for t (x, y) = limN→∞ ETN (x, y).
EIGENVALUES OF WEAKLY DILUTE WISHART MATRICES
9
In contrast with the strong dilution, matrix t (x, y) is nonzero and its entries take two different values depending on whether x = y or not. These two limits t1 and t2 involve explicitly parameter q and t1 − t2 vanishes for q → 1. This determines the difference between equations that we obtain for weakly dilute random matrices and those derived for the corresponding pure (nondilute) ensemble. To make the derivation self-consistent, we recall the basic elements of the method developed in [15, 18]. Given two symmetric (or Hermitian) operators H and Hˆ acting in the same ˆ = (Hˆ − z)−1 space, the resolvent identity holds for G = (H − z)−1 and G ˆ = −G(H − Hˆ )G. ˆ G−G
(4.3)
This identity leads to two important observations; that the dependence of the resolvent on the random matrix can found explicitly and that this dependence is expressed in terms of the resolvent (see formulas (4.7) and (4.9)). These two properties make the resolvent approach a fairly powerful tool in the spectral theory of random matrices. We prove Theorem 2.1 by showing the convergence of the trace fN,q (z) = (q) −1 N Tr GN (z); lim EhfN,q (z)i = fq,V˜ (z),
(4.4)
lim E |fN,q (z) − EhfN,q (z)i|2
(4.5)
N→∞
and N→∞
for all z ∈ 3q = z ∈ C, |Imz| > 2Vm q −1 + 1 ,
(4.6)
where we denoted by h·i the mathematical expectation with respect to the measure generated by random variables {πxy }. On this way, we derive equation for fq,V˜ (z) that determines the limiting eigenvalue distribution. Taking into account that the normalized trace of the resolvent (4.1) is the Stieltjes transform of the NCF, we conclude that relations (4.4) and (4.5) imply the weak convergence in probability of σ (λ; AN,q ), in the limit N → ∞, to a nonrandom function σq (λ; V˜ ). This can be proved by the usual arguments of the theory of Herglotz functions. The reasonong is based on weak compactness of the family σ (λ; AN,q ) and the Helly theorem [11] (see, e.g., [18] for more details). Thus, relations (4.4) and (4.5) prove items (i) and (ii) of Theorem 2.1. Item (iii) is proved in Lemma 4.4 (see the end of this section). We split the remaining part of this section into three subsections. In the first one we derive main relations that lead to the equation for fq,V˜ (z). In the second subsection we derive relations leading to the proof of (4.5). The third subsection contains the proofs of the auxiliary facts and estimates.
10 4.1.
A. KHORUNZHY AND G. J. RODGERS
DERIVATION OF MAIN RELATIONS
Let us consider identity (4.3) with H = AN,q and Hˆ = 0; X −1/2 EG(x, s)a(s, y)π(s, y) , EhG(x, y)i = ζ δxy − ζ N
(4.7)
s
where ζ = −z−1 . In (4.7) and below, we omit the varable z and subscripts q, N and do not indicate limits of summations, if no confusion can arise. To compute the average EG(x, s)a(s, y), we use the following elementary facts (see also [18]). It is related to the Gaussian random vector γ = (γ1 , . . . , γk ) with zero average: Eγj F (γ1 , . . . , γk ) =
k X l=1
∂F , Eγj γl E ∂γl
(4.8)
where F is a nonrandom function such that all integrals in (4.8) exist. This formula can be proved by using the integration by parts technique. We will also use the formula that is a direct consequence of identity (4.3); ∂G(x, s) 1 = − √ G(x, p)G(r, s)πpr . ∂a(p, r) N
(4.9)
Now we can write that 1 X EG(x, s)a(s, y) = − √ Vps Vry + Vpy Vrs EG(x, p)G(r, s)πpr . N x,p Substituting this relation into (4.7), we obtain equality X
EhG(x, y)i = ζ δxy + ζ EhG(x, p)T (p, y)iVpy + ζ E ψN(1) (x, y) , (4.10) p
where T is determined in (4.2) and we denoted ψN(1) (x, y) =
1 X G(x, p)πpr G(r, s)Vps πsy Vry . N p,r,s
(4.11)
In the last part of this section we prove that ψN(1) vanishes in the limit N → ∞ (see Lemma 4.1). Turning back to (4.7), we can write for the average EhG(x, y)i ≡ g(x, y) the following relation X
g(x, p)t (x, y)Vpy + ζ E ψN(1) (x, y) + g(x, y) = ζ δxy + ζ p
+
ζ ψN(2)(x, y),
(4.12)
EIGENVALUES OF WEAKLY DILUTE WISHART MATRICES
11
where t (x, y) ≡ EhT (x, y)i and X {EhG(x, y)T (x, y)i − g(x, y)t (x, y)}Vpy . ψN(2) (x, y) = p
Given a random variable γ with finite mathematical expectation, let us introduce the centered random variable γ 0 ≡ γ − Eγ . We will also use denotation [γ ]0 for more complex expressions. In what follows (see Subsection 4.2), we prove that
(4.13) lim sup E |G0 (x, y)|2 = 0 for all z ∈ 3q . N→∞ x,y
Let us note that the last inequality of (4.11) implies that supx,y |T (x, y)| 6 Vm q −1 η−1 . This estimate together with (4.13) leads to relation lim sup ψN(2)(x, y) = 0. (4.14) N→∞ x,y
It should be noted also that (4.13) implies (4.5). To derive the final equation (2.7b) from (4.12), it remains to compute the limit of the average t (x, y) = EhT (x, y)i. Basing on (4.2), one can write that EhT (x, y)i =
1 X hπxr πsy ig(r, s)Vrs + ψN(3) (x, y), N r,s
(4.15)
where the term ψN(3) (x, y) =
1 X
E πxr G0 (r, s)πsy Vrs N r,s
vanishes due to the self-averaging property (4.13). Examining the first term on the right-hand side of (4.15), we observe that it depends only on whether x = y: 1 X t , if x = y, hπxr πsy ig(r, s)Vrs = 1 t N r,s 2 , if x 6= y, where t1 = (1 − q)gV ¯ 0 + q g¯V ,
t2 = q g¯V ,
and N 1 X g¯ = g(x, x), N x=1
N 1 X g¯V = g(x, y)Vxy . N x,y=1
12
A. KHORUNZHY AND G. J. RODGERS
Let us consider the equation g(x, ˆ y) = ζ δxy
N X ˆ ˆ ˆ + ζ t1 − t2 g(x, ˆ y)V (x, y) + ζ t2 g(x, ˆ p)Vpy ,
(4.16)
p=1
where x, y = 1, . . . , N and tˆi are determined by the same formulas as ti with g(x, y) replaced by g(x, ˆ y). It is not hard to prove that (4.16) has a unique solution for z ∈ 3q (4.6) (see Lemma 4.3 at the end of this section). Also it is easy to see that N 1 X g(x, x) − g(x, ˆ x) = 0. N→∞ N x=1
lim
(4.17)
Some elementary calculations based on the finite-difference form of Vxy = V (x − y) show that Equation (4.16) leads in the limit N → ∞ to Equation (2.7b) and therefore N 1 X g(x, ˆ x) = fq,V˜ (z), lim N→∞ N x=1
where fq,V˜ (z) is given by (2.7a). Relation (4.4) is proved. 4.2.
SELFAVERAGING PROPERTY
Let us prove relation (4.13) that obviously implies (4.5). We consider, at the same time with G(x, y) ≡ G(x, y; z) the resolvent G0 (x, y) ≡ G(x, y; z0 ), and study the average
(4.18) SN (x, y) = E G0 (x, y)G0 (x, y) . Loosely speaking, the main idea is to derive relation of the form SN (x, y) = BS(N) (x, y) + 8N (x, y) where B is certain expression involving SN and term 8N vanishes as N → ∞. The crucial observation used in this approach is that BS(N) can be estimated by SN itself multiplied by coefficients depending on η = |Imz|−1 (see, e.g., [15]). This lead to relations (4.13) provided η is large enough. 4.2.1. Selfaveraging of S(x, y) Applying (4.3) to the last factor of (4.18), we obtain equality X
S(x, y) = −ζ 0 N −1/2 E G0 (x, y)G0 (x, t)a(t, y)πty , t 0
0
where ζ = −1/z . Using (4.8) and (4.9), we derive relation X
E G0 (x, y)G0 (x, p)T 0 (p, y) Vpy + 0N (x, y). S(x, y) = ζ 0 p
(4.19)
13
EIGENVALUES OF WEAKLY DILUTE WISHART MATRICES
Here we denoted
0N (x, y) = ζ 0 E G0 (x, y)φN0 (x, y) + 1 X G(x, p)πpr G(r, y)G0 (x, t)πt,y (Vpt Vry + Vpy Vrt ), + N p,r,t
where φN0 (x, y) is given by (4.11) with G replaced by G0 . Using identity
E ξ10 ξ2 ξ3 = E ξ10 ξ2 Ehξ3 i + E ξ10 ξ2 ξ30 , we can rewrite (4.19) in the form X
E [G(x, y)]0 T 0 (p, y) E G0 (x, p) Vpy + S(x, y) = ζ 0 p
+ ζ0
X
E [G(x, y)]0 [G0 (x, y)]0 T 0 (p, y) Vpy + ζ 0 0N (x, y). (4.20) p
Repeating the reasoning used to estimate (4.11) (see Lemma 4.1), one can easily show that for z, z0 ∈ 3q (4.6) |0N | ≡ sup |0N (x, y)| = O N −1/2 + N −1/2 |SN | , x,y
where |SN | ≡ sup |SN (x, y)|. x,y
Substituting this inequality into (4.20), we obtain that for z0 = z¯ |SN |2 6 |SN ||UN |η−2 Vm + q −1 η−2 Vm2 |SN |2 + η−1 |0N |,
(4.21)
where
1/2 |UN | ≡ sup E |T 0 (x, y)|2 . x,y
Now it is clear that (4.21) implies (4.13) provided the estimate |UN | = o(1) as N → ∞
(4.22)
is true. 4.2.2. Selfaveraging of T (x, y) To prove (4.23), we treat the average
0 1 X 0 0 0 UN (x, y) ≡ E T (x, y)T (x, y) = E T (x, y) πxr G (r, s)Vrs πsy N r,s
14
A. KHORUNZHY AND G. J. RODGERS
by the same procedure as is used to study SN (x, y). We apply to G0 (r, s) the resolvent identity (4.3) and use (4.8) and (4.9) to compute the mathematical expectations. As a result, we obtain relations
UN (x, y) = ζ 0 E T 0 (x, y)ρ(x, y) V0 + X 0 −1 0 0 0 + ζ N E T (x, y) πxr Vrs G (r, p)T (p, s)Vps πs,y + p,r,s
+
ζ 0 θN(1) (x, y),
(4.23)
where we have denoted 1 X ρ(x, y) = πxr πry N r and θN(1) (x, y) =
1 X πxα G(α, l)πlj G(j, β)πβy × N 2 α,β,p,r,s × Vpl Vsj + Vpj Vsl πxr G0 (r, p)πps Vrs πsy .
Standard computations (see, e.g., Lemma 4.1) show that 4 (1) θ ≡ sup θ (1) (x, y) 6 2Vm . N N q 3 η3 x,y
Let us rewrite (4.23) in the following form
UN (x, y) = ζ 0 E T 0 (x, y)ρ(x, y) V0 + X + ζ 0 N −1 E T 0 (x, y) πxr Vrs G0 (r, p)[T 0 (p, s)]0 Vps πs,y + p,r,s
X 0 −1 0 0 0 0 + ζ N E T (x, y) πxr Vrs [G ] (r, p)Vps πs,y iEhT (p, s) + p,r,s
X 0 −1 0 πxr Vrs Vps πs,y EhG0 (r, p)iEhT 0 (p, s)i + + ζ N E T (x, y) p,r,s
+
ζ 0 θN(1) (x, y).
Then using a-priori estimate |T (x, y)| 6 Vm /(qη) and taking z0 = z¯ , we derive inequality |UN |2 6
0 2 1/2 vVm Vm2 sup E ρ (x, y) + |UN |2 + qη2 x,y qη2 1/2 X Vm3 −1 0 [πxr πsy ] Vrs + + 2 2 |UN | |SN | + |UN | N q η r,s + η−1 θN(1) .
(4.24)
EIGENVALUES OF WEAKLY DILUTE WISHART MATRICES
15
Now we use elementary estimates 2 1/2
= O N −1/2 sup E ρ 0 (x, y) x,y
and
X 0 2 1 πxr Vrs πsy = O N −1 E N r,s
and derive from (4.24) that |UN |2 6
Vm2 Vm3 2 |U | + |UN ||SN | + O N −1/2 . N 2 2 2 qη q η
This inequality regarded jointly with (4.21) for z ∈ 3q (4.6) implies (4.22). Relation (4.5) is derived. We complete this section with the outline of the proofs of the following auxiliary statements. 4.3.
PROOF OF THE AUXILIARY FACTS
LEMMA 4.1. Relation
sup |φN (x, y)| = O Vm2 η−2 q −1 N −1/2 .
(4.25)
x,y
holds with probability 1. Proof. Taking into account that |πxy | < q −1/2 with probability 1, one can write the inequality X (1) ψ (x, y) 6 1 |G(x, p)|Vps G(r, s)Vry |. N qN p,r,s Let us look at |G(x, p)| ≡ Gx (p) and |G(s, r)| ≡ Gr (s) for given fixed x and r as vectors in l 2 (N) and observe that the norm of the linear operator Vˆ in l 2 (N) with the kernel Vˆ (p − s) = |Vps | is bounded by Vm (2.3). Then using the estimate X |G(x, p)|2 = |Gex |2 6 |G|2 6 η−2 , (4.26) p
where we introduced the unit vectors ex with components ex (j ) = δxj , we obtain that 2 1/2 Vm X X |G(s, r)| · |Vry | |φN (x, y)| 6 qηN s r 1/2 Vm2 X X 2 6 |G(s, r)| . qηN s r
16
A. KHORUNZHY AND G. J. RODGERS
This estimate, together with inequalities (4.26) leads us to (4.25).
2
LEMMA 4.2. Equation (4.16) has a unique solution in the class G of matrices {g(x, y; z), x, y = 1, . . . , N} such that (4.27) sup g(x, y; z) 6 2η−1 , η = |Imz| > 2Vm . x,y
Proof. The proof is based on the use of the method of subsequent approximations. We introduce the sequence of matrices {g (k) (x, y)}, k = 1, 2, . . . by the relations ζ = −z−1 , g (k+1) (x, y) = ζ δxy + ζ t2(k) g (k) V (x, y) + + ζ t1(k) − t2(k) g (k) (x, y)V (x, y),
g (1) (x, y) = ζ δxy ,
where t1(k) = V0 N −1 Tr g (k) + q N −1 Tr g (k)V − V0 N −1 Tr g (k) and t2(k) = qN −1 Tr g (k) V . It is easy to see that if g (k) satisfies (4.27), then g (k+1) also satisfies (4.27). Then simple computation shows that there exists such α(z) that 1k+1 ≡ sup g (k+1) (x, y) − g (k) (x, y) < α(z)1k x,y
and |α(z)| < 1 for z ∈ 1q . This completes the proof of the Lemma 4.2.
2
LEMMA 4.3. Relation (4.17) holds for z ∈ 3q (4.6). Proof. Subtracting (4.17) from (4.15), we obtain for the difference δg (x, y) = g(x, y) − b g (x, y) relation g V (x, y) + ζ t2 δg V (x, y) + δg (x, y) = ζ δt(2) b + ζ δt(1) − δt(2) b g (x, y)V (x, y) + ζ t1 − t2 δg (x, y)V (x, y) + + ζ ψN0 (x, y), (4.28) where δg (x, y) = g(x, y) − b g (x, y), δt(1) = N −1 Trδg + q N −1 Tr δg V − V0 N −1 Trδg , δt(2) = qN −1 Trδg V
17
EIGENVALUES OF WEAKLY DILUTE WISHART MATRICES
and
sup x,y ψN0 (x, y) = O(1) as N → ∞.
(4.29)
Now we regard (4.28) as the matrix relation δg = Lζ,V δg + ζ δt(2)b g V + ζ δt(1) − δt(2) g(x, y)V (x, y) + ζ ψN0 , where L is a linear operator X (Lζ,V δ)(x, y) = ζ t2 δ(x, p)V (p, y) + ζ(t1 − t2 )δ(x, y)V (x, y). p
It is easy to show that for large enough values of |Imz| there exists β(z) < 1 such that kLζ,V k1 < β(z),
where the norm k · 1 is determined as
(4.30)
kδk1 = sup |δ(x, y)|. x,y
A priori estimates for g(x, y) and ti |g(x, y)| 6 η−1 ,
|ti | 6 Vm η−1
and estimates (4.27) and (4.29) allow one to deduce from (4.30) that kδg k1 = o(1)
as N → ∞. 2
This proves Lemma 4.3.
LEMMA 4.4. Item (iii) of Theorem 2.1 is true. Proof. We introduce a sequence {g (k), k = 0, 1, 2, . . .} of functions g (k) (p; z), p ∈ (0, 1), z ∈ C by the formulas g (0) (p; z) = −1/z and Z g (k+1) (p; z) = −z − (1 − q)v 2 g (k) (r; z) dr − q V˜ (p) × Z ×
V˜ (r)g (k) (r; z) dr
−1 .
It is easy to verify that if g (k) (p; z) satisfies conditions (2.8), then g (k+1) (p; z) also does. Next, it is not hard to deduce that inequality sup g (k) (p; z) 6 2η−1 , z ∈ 3q (4.31) p
implies the same for g (k+1) (p; z).
18
A. KHORUNZHY AND G. J. RODGERS
The next step is to show that the sequence {g (k) } is the Cauchy one that determines, in the limit k → ∞, the function γ (p; z) ∈ 0 that satisfies (2.7b). This function satisfies (4.31) that proves uniqueness of the solution of (2.7b). 2 LEMMA 4.5. Due to positivity of Vxy = V (x − y) the matrix M(x, y; s, t) = Vxs Vyt + Vxt Vys is also positively determined on the vectors ψ with complex components ψ(x, y), x, y = 1, 2, . . . , n; n X
M(x, y; s, t)ψ(s, t)ψ(x, y) > 0.
(4.32)
x,y,s,t =1
Proof. Let us consider the Gaussian random variables {γ (x), x ∈ N} such that Eγ (x) = 0,
Eγ (x)γ (y) = V (x − y).
Then one can derive with the help of (4.8) that Eγ (x)γ (y)γ (s)γ (t) = M(x, y; s, t). Therefore the right-hand side of (4.32) can be rewritten in the follpwing form X 2 n X n Eγ (x)γ (y)γ (s)γ (t)ψ(s, t)ψ(x, y) = E γ (x)γ (y)ψ(x, y) . x,y,s,t =1
This proves the lemma.
x,y=1
2
5. Proof of Theorem 3.1 To prove Theorem 3.1, we again use the resolvent approach of Section 3. However (q) the present case of HN is more complicated because of the bilinear structure of H (x, y) (3.3) with respect to the random variables {ξ µ (x)}. We modify the general (q) (q) scheme and regard the resolvent GN (x, y; z) of matrices HN (3.1) first as a function of the random variables {ξ µ (x) and derive equations for the mathematical (q) expectation EGN . Then we prove the selfaveragenessin the limit N → ∞ of the (q) random variable GN (x, y; z) as a function of random variables {πxy }. As before, our tools are the resolvent identity (4.3) and the version of (4.8) applied for jointly independent Gaussian random variables {γj } ∂F . E γj F (γ1 , . . . , γk ) = Eγj2 E ∂γj
(5.1)
19
EIGENVALUES OF WEAKLY DILUTE WISHART MATRICES (q)
We will also use the consequence of (4.3) applied for GN (we omit the super- and (q) subscripts in GN ): 1 X ∂G(x, s) = − G(x, r)G(t, s) + G(x, t)G(r, s) ξ ρ (r)πrt . ρ ∂ξ (t) N r
5.1.
(5.2)
DERIVATION OF MAIN RELATIONS
We start with the resolvent identity (4.3) written in the form EhG(x, y; z)i = ζ δxy − ζ EhM(x, y; z)i,
ζ = −z−1 ,
where we denoted M(x, y) =
1 X G(x, s)ξ µ (s)ξ µ (y)πsy . N s,µ
Here and below we omit the variable z when it is not important. To compute the mathematical expectation of M(x, y), we can use (5.1) Taking into account (5.2), we obtain the following relation EhM(x, y)i =
mu2
E G(x, y)πyy − N u2 X
E G(x, y)πyt H (t, s)G(t, s)πsy + ϑN(1), − N s,t
(5.3)
where H is the same as in (3.3) and ϑN(1) =
u2 X
E G(x, t)πty H (t, s)G(y, s)πsy . N s,t
(5.4)
In Lemma 5.1 (see the second part of this section) we prove that the variable ϑN(1) vanishes in the limit N → ∞. The √proof is based on the observation that all the moments of the random variables NH (t, s) and H (t, t) are bounded for all N. We also use elementary inequality (4.26). Let us introduce variable 1 X πxs H (s, t)G(s, t)πty (5.5) L(x, y) = N s,t that resembles the generalized trace T (x, y) (4.2). Taking into account that hπyy = q and denoting cN = mN −1 , we rewrite (5.3) in the form EhM(x, y)i = cN qu2 EhG(x, y)i − u2 EhG(x, y)iEhL(y, y)i − − ϑN(1)(x, y) + ϑN(2) (x, y),
(5.6)
20
A. KHORUNZHY AND G. J. RODGERS
where
ϑN(2) (x, y) = cN u2 E G0 (x, y)πyy + u2 E G0 (x, y)L0 (y, y)
(5.7)
and the subscripts 0 denote the centered random variables. In Lemma 5.2 we prove that ϑN(2) vanishes in the limit N → ∞. Let us note that Equation (5.6) is similar to Equation (4.12). Following the same ideas as of Section 4,we derive relations for the variable EhL(x, y) . To do this, we employ once more (5.1) and obtain relation (cf. (4.15)) X
E πxs G(s, s)πsy − EhL(x, y)i = cN u2 s
u X
− E L(x, t)G(t, t)πty + ϑN(2) , N t 2
(5.8)
where ϑN(3) = −
u2 X
E πxs H (s, t)G(s, t)πsr G(r, s)πty . 2 N s,t,r
(5.9)
Denoting lN (x, y) = EhL(x, y)i and gN (x, y) = EhG(x, y)i and taking into account definition (2.8), we can rewrite (5.6) and (5.7) as the system of relations √ gN (x, y) = ζ δxy − ζ cu2 qgN (x, y) + ζ u2 gN (x, y)lN (y, y) + + ϑN(1)(x, y) + ϑN(2)(x, y), (5.10a) √ X 2 u q lN (x, t)gN (t, t) + lN (x, y) = cu2 gˆN δxy + (1 − δxy )q − N t + ϑN(3) + ϑN(4), (5.10b) P −1 where gˆN = N x gN (x, x) and u2 X 0 u2 X
E πxs G0 (s, s)πsy − 2 E L (x, t)G(t, t)πty − ϑN(4) (x, y) = cN N s N t −
0 u2 X
E L(x, t) E G (t, t)π . ty N2 t
In Lemma 5.2 we prove that sup ϑ (i) (x, y) = o(1) x,y
N
as N → ∞, i = 3, 4.
Let us consider the system of matrix equations √ RN (x, y) = ζ δxy − ζ cu2 qRN (x, y) + ζ u2 RN (x, y)SN (y, y), SN (x, y) = cu2 RN (x, y) δxy + (1 − δxy )q − √ N u2 q X SN (x, t)RN (t, t). − N t =1
(5.11)
(5.12a)
(5.12b)
EIGENVALUES OF WEAKLY DILUTE WISHART MATRICES
21
The subscript N indicates that we consider N-dimensional matrices. It is not hard to √ show that this system is uniquely solvable provided z ∈ 3q (4.6) with Vm = 2u c. See the proof in Lemma 5.3 at the end of this section. Under this condition it is not hard to show that vanishing of terms ϑN(i) , i = 1, 2, 3, 4 leads to relation N 1 X EhgN (x, x)i − RN (x, x) = 0. N→∞ N x=1
lim
(5.13)
This can be proved by the same reasoning as used Lemma 4.3 of the previous section. The next observation is that the system (5.12) admits the solution such that the diagonal elements RN = RN (x, x) do not depend on x and SN (x, y) = δxy SN(1) + (1 − δxy SN(2) Therefore we conclude that (5.10) is equivalent to the scalar system √ RN = ζ − ζ cu2 qRN + ζ u2 RN SN(1) , (5.14a) 1 (1) N − 1 (2) √ (5.14b) S + SN , SN(1) = cu2 RN − u2 qRN N N N 1 (1) N − 1 (2) (2) 2 2√ SN = cqu RN − u qRN (5.14c) S + SN N N N that naturally has the unique solution (Lemma 5.3). Now elementary computations show that RN satisfies the same Equation (3.4) that determines fq,c (z). Then we come to the conclusion that lim gˆN − fq (z) = 0. (5.15) N→∞
P Since the normalized trace gˆN = N −1 x gN (x, x) represents the Stieltjes transform of certain measure and lim N→∞ gˆN is unique, then the Helly theorem and (5.15) imply that fq,c (z) also represents the Stieltjes transform of certain measure σq,c (λ). Thus, to complete the proof of Theorem 5.1, it remains to show that 2 X 1 N 0 (5.16) GN (x, x) = 0. lim N→∞ E N x=1 According to the general scheme, relations (5.15) and (5.16) imply convergence in probability of the measure σ (λ; HN(π)) to the limit σq,c (λ). Relation (5.16) is proved in Lemma 5.2. 5.2.
PROOF OF THE AUXILIARY FACTS
LEMMA 5.1. Variables (5.7) and (5.9) admit the following estimate for all z ∈ 3q (4.6) sup ϑ (i) (x, y) = O N −1/2 as N → ∞ x,y
N
for i = 1, 3.
22
A. KHORUNZHY AND G. J. RODGERS
Proof. We can estimate |ϑN(1) (x, y)| by the sum u2 X u2 X Eh|G(x, t)H (t, s)G(y, s)|i + Eh|G(x, s)H (s, s)G(y, s)|i.(5.17) qN s6=t qN s For the first term of this sum we have inequality u2 X Eh|G(x, t)H (t, s)G(y, s)|i qN s6=t 6
2 1/2 √ u2 X
2 2 1/2 E |G(x, t)| |G(y, s)| NH (t, s) . qN 3/2 s6=t
√ Let us note that all the moments of the random variables N H (t, s) and H (t, t) are bounded for all N and do not depend on x and y. Taking into account that 1/2 X 1/2 1 X
E |G(x, t)|2 |G(y, s)|2 6E |G(x, t)|2 |G(y, s)|2 N s6=t s6=t and a priori estimate (4.26), we deduce that the first term of (5.17) is of the order O(N −1/2 ) as N → ∞. The second one is also of the same order of magnitude. Let us consider the case of i = 3. As in the previous case, we can write inequality (3) ϑ (x, y) 6 N
XX
√ u2 NH (r, t) + E |G(t, s)G(s, r)| · 3/2 5/2 q N r6=t s +
u2 η X Eh|G(t, s)| · |H (t, t)|i. q 3/2 N 2 s,t
Now it is easy to see that estimates (5.4) and (5.5) complete the proof of the lemma. 2 LEMMA 5.2. Regarding the centered random variables G0 (x, y) = G(x, y) − EG(x, y)
and
L0 (x, y) = L(x, y) − EL(x, y),
the following relations hold in the limit N → ∞ 2
sup E G0 (x, x) = o(1),
(5.18)
x
and 2
sup E L0 (x, y) = o(1). x,y
(5.19)
23
EIGENVALUES OF WEAKLY DILUTE WISHART MATRICES
Remark. Relations (5.18) reflect the selfaveraging property of the random variables G(x, x) and L(x, y). It is easy to see that (5.18) and (5.19) imply estimates (5.11). Let us also note that (5.18) implies relation (5.16). Proof. We prove (5.18) using again the resolvent identity (4.3) and relations (5.1) and (5.2). According to the definition of the centered random variable, we can write relation
0 1 X 0 (5.20) πxr H (r, t)G(r, t)πty . E L (x, y)L(x, y) = E L (x, y) N r,t Let us apply relation (4.4) to the last average in (5.20). With the help of (5.2) we obtain equality
mu2 0 E L0 (x, y)L(x, y) = E L (x, y)πxt G(t, t)πty − N 1 X 2 0 0 − u E L (x, y) L(x, t)G(t, t)L (t, y) + N t + φN(1) (x, y) + φN(2) (x, y), where φN(1) (x, y)
(5.21)
1 X 0 = −E L (x, y) 2 πxt G(t, s)πsr H (t, r)G(r, s)πsy N r,s,t
and φN(2)(x, y) can be estimated by inequality (2) φ (x, y) 6 N
1 X Eh|G(l, p)[H (x, t) + H (l, t)]G(t, p)|i + q 2 N 2 l,p,t + +
X 1 Eh|G(k, l)H (r, t)H (k, p)G(r, p)G(t, l)|i + q 3 N 3 k,l,p,r,t 1 q3N 3
X
Eh|G(k, r)H (r, t)H (k, p)G(l, p)G(t, l)|i.
k,l,p,r,t
It is not hard to find, with the help of the estimates (5.4) and (5.5), that 2 1/2 u4 η2 sup φN(1) (x, y) 6 L0 √ . x,y q N A little more cumbersome computations lead to the estimate (2) 1 sup φN (x, y) = O 3 3/2 . q N x,y
(5.22)
(5.23)
24
A. KHORUNZHY AND G. J. RODGERS
Taking into account the identity 1 X 0 L (x, y) L(x, t)G(t, t)πty N t 1 X 0 0 = L (x, y) L(x, t) G(t, t)πty + N t
1 X 0 G(t, t)πty L(x, t) , + L (x, y) N t we derive from (5.21) inequality
0 L (x, y)L(x, y) 6
2 1/2
η mu2 0 2 1/2 sup G(x, y) + |PN (x, y)| + √ L0 (x, y) N q t 2 1/2 u4 η2 √ + η q sup |QN (x, t)| + √ L0 (x, y) + t q N + φ (3)(x, y) |ψN (x, y)|, N
(5.24)
where
X 0 1 πxt G(t, t)πty , PN (x, y) = L N t
and 2 1/2
G(t, t)0 2 1/2. QN (x, y, t) = L0 (x, y)G(t, t) 6 L0 (x, y) ˆ that are The term ψN contains the terms arising when one passes from L0 G to Lˆ 0 G 0 independent from random variables {πxy } and back to L G, after averaging over πty . In what follows, we derive inequalitites
PN (x, y) 6 √η L0 (x, y) 2 1/2 + √η L0 (x, y) 2 1/2 sup L0 (x, t) 2 1/2 + q t N (4) (5.25) + η sup QN (x, t) + φ (x, y), t
and
N
2 √
0
G (t, t) 2 6 mηu q G0 (t, t) 2 + η G0 (t, t) 2 h|L(x, x)|i + N
+ η2 G0 (t, t)L0 (t, t) + 8ηqu2 N −1/2 ,
where the variable |φN(4)(x, y)| is of the order O(N −1/2 ) as N → ∞.
(5.26)
EIGENVALUES OF WEAKLY DILUTE WISHART MATRICES
25
Then elementary calculations show that (5.24), (5.25) and (5.26) imply (5.18) and therefore (5.19). Let us consider PN (x, y). Using the resolvent identity (2.10), we can write that 1 X PN (x, y) = ζ¯ E L0 πxt πty − N t X 0 1 ¯ − ζE L πxt G(t, k)H (k, t)πkt πty , (5.27) N t,k where we denoted ζ¯ = −1/z and L0 ≡ L0 (x, y). It is easy to derive that 0 2 1/2 X
0 2 1/2 1 X E L0 1 πxt πty 6 L πxt πty N t N t 1 2 1/2 6 √ L0 . N Turning to the second term in the right-hand side of (5.27), we use (5.2) to compute the average. Then we obtain X 0 1 πxt G(t, k)H (k, t)πkt πty E L N t,k X 0 m =E L 2 πxt G(t, t)πkt πty − N t,k X 0 1 −E L 2 πxt G(t, t)H (t, t)πkt πty + φN(5) (x, y), N t,k where φ (5) N (x, y), in common with (5.24) and (5.25), includes the terms of the order N −1/2 h|L0 |2 i1/2 and those that are of the order O(N −3/2 ). Now, repeating arguments used to derive expressions (5.22) and (5.23) we arrive at (5.19). Let us derive (5.18). Using the resolvent identity (4.3) and formula (5.2), we obtain the equality
m
E G0 (x, x)G(x, x) = −ζ¯ u2 E G0 (x, x)G(x, x)πxx + N
+ ζ¯ E G0 (x, x)G(x, x)L(x, x) + ζ¯ X 0 E G (x, x)G(x, t)H (s, t)πt x G(s, x)πsx + + N s,t +
2ζ¯ X
E G(x, t)H (s, t)πt x G(x, x)G(x, s)πsx . N s,t
Then (5.18) easily follows from (5.19). Lemma is proved.
2
26
A. KHORUNZHY AND G. J. RODGERS
LEMMA 5.3. The system of equations (5.12) is uniquely solvable and (5.15) holds. Proof. Let us introduce recurrence relations √ RN(k+1)(x, y) = ζ δxy − ζ cu2 qRN(k)(x, y) + ζ u2 RN(k) (x, y)SN(k) (y, y), u2 √ X (k) SN(k+1) (x, y) = cu2 RN(k) (x, y) δxy + q 1 − δxy − q SN (x, t)RN(k) (t, t) N t for k > 0 and RN(0) (x, y) = SN(0) (x, y) = 0. Then one can easily show that for all z ∈ 31 (2.9b)
(k)
R ≡ sup R (k)(x, y) 6 2η−1 (5.28) N N x,y
and the same relation holds for RN(k). Then elementary computations show that the differences 1k+1 (x, y) ≡ RN(k+1)(x, y) − RN(k) (x, y) and δk+1 (x, y) ≡ SN(k+1) (x, y) − SN(k) (x, y) are such that
k k1k+1 k 6 4(c + 1)u2 η−k−1 k10 k
and
k kδk+1 k 6 4(c + 1)u2 η−k−1 kδ0 k.
This proves the first statement of the lemma. Using the estimates (5.28), it is easy to prove (5.15). Lemma is proved. 2
6. Properties of the Eigenvalue Distribution In this section we prove Theorems 2.2 and 3.2. The reasonings based on the inversion formula (2.9) are similar in both cases and use Equations (2.7) and (3.4), respectively. Proof of Theorem 3.2. Let us denote I (z) = Im fq,c (z) and R(z) = Re fq,c (z). It is easy to derive from (3.6) that √ u2 qI I 4 2√ = Im z + cu (1 − q)I + cu q . √ I 2 + R2 (1 + u2 qR)2 + u4 qI 2 Taking into account inequality I (λ + iε) > 0, we derive the following estimate 1 > ε + cu4 (1 − q)I (λ + iε) I (λ + iε)
(6.1)
27
EIGENVALUES OF WEAKLY DILUTE WISHART MATRICES
for all values of ε > 0 and λ. Using (2.9), one can easily deduce from (6.1) the estimate mentioned in Theorem 3.2. 2 Proof of Theorem 2.2. Taking into account that g˜q (p; z) posesses property (2.8), one can easily derive from (2.7b) inequality Im g˜q (p; λ + iε) 6
1 ε + (1 −
q)v 2 Im f
q (λ
+ iε)
.
Integrating both parts of this relation over p, one obtains inequality [Im fq (λ + iε)]2 6
1 . (1 − q)v 2 2
This estimate added by formula (2.9) completes the proof.
We continue the study of the distributions σq (λ) and σq,c (λ) and derive from Equations (2.7) and (3.4), respectively, relations for the moments Z Z Mk = λk dσq (λ) and Lk = λk dσq,c (λ). Let us note first that item (iii) of Theorem 2.1 implies that g˜q (p; z) can be represented in the form g˜q (p; z) = −
∞ X mk (p) k=0
zk+1
.
(6.2)
R Then (2.7a) implies that Mk = mk (p) dp. R Writing down similar to (6.2) expressions for fq (z), g˜q (r; z)V˜ (r) dr and their products, one can easily derive from (2.7b) the system m0 (p) = 1,
m1 (p) = 0, X mk+1 (p) = (1 − q)v 2 m(0) i mj (p) + q i+j =k−1
X
˜ m(1) i mj (p)V (p),
i+j =k−1
R where m(l) mi (r)V˜ l (r) dr, l = 0, 1 and the sum is taken over i, j > 0. This i = system can be reduced to the system Z (l) (6.3a) m0 = Vl ≡ V˜ l (r) dr, m(l) 1 = 0, X X (l) (l+1) 2 m(0) m(1) . (6.3b) m(l) i mj + q i mj k+1 = (1 − q)v i+j =k−1
i+j =k−1
THEOREM 6.1. (i) Distribution σq (λ) is even in the sence that M2k+1 = 0, ∀k ∈ N;
28
A. KHORUNZHY AND G. J. RODGERS
(ii) the support of the measure dσq (λ) is bounded; p supp dσq (λ) ⊂ (−2l, 2l), l = v[(1 − q)v + qVm ].
(6.4)
Proof. Item (i) obviously follows from (6.3). We start the proof of item (ii) with the following observation. Since matrix V with enries V (x − y) is positively determined, then obviously m(l) 0 > 0. Now it is (l) (l) easy to show that if ms > 0, then ms+2 > 0. The second step is the following statement. PROPOSITION 6.1. For all s > 0 and l > 0 6 Vm m(l) m(l+1) s s .
(6.5)
Proof. Relation (6.5) obviously holds for s = 0. It is easy to deduce from the form of (6.3b) that if (6.5) is valid for m(l) s with s 6 t then (6.5) is true also for . 2 m(l) t +1 Using Proposition 6.1, we can derive from (6.3b) inequality X (1) (l) m(l) (1 − q)v 2 m(0) mj . k+1 6 i + qVm mi
(6.6)
i+j =k−1
Now let us introduce the numbers m ˆ (l) k by the following recurrent relations Z V˜ l (r) dr, m ˆ (l) (6.7a) m ˆ (l) 0 = Vl ≡ 1 = 0, X (l) (1 − q)v 2 m ˆ (0) ˆ (1) m ˆj . (6.7b) m ˆ (l) i + qVm m i k+1 = i+j =k−1
It follows from Proposition 6.1 and (6.6) that ˆ (l) m(l) k 6m k .
(6.8)
Let us determine the support of the measure dσˆ q (λ) given by its moments m(0) k . To do this, we introduce the function ∞ X m(l) . gˆ (z) = − zk+1 k=0 (l)
It is not hard to derive from (6.7) that it satisfies the system of equations m(l) 0 gˆ (z) = , −z − (1 − q)v 2 gˆ (0) (z) − qVm gˆ (1) (z) (l)
l = 0, 1, 2, . . . .
(6.9)
ˆ (0)(z), m(1) It follows from (6.9) that gˆ (1)(z) = m(1) 0 g 0 ≡ v and then (cf. (2.10)) gˆ (0) (z) =
1 −z − [(1 −
q)v 2
+ qvVm ]gˆ (0) (z)
.
(6.10)
EIGENVALUES OF WEAKLY DILUTE WISHART MATRICES
29
It is clear that dσˆ q (λ) determined by its Stieltjes transform gˆ (0) (z) (6.10) is the semicircle distribution (2.11) with v 2 replaced by (1 − q)v 2 + qvVm . Thus, its support is given by the interval from the right-hand side of (6.4). Relation (6.8) implies that supp dσq (λ) ⊂ supp dσˆ q (λ). 2
This observation completes the proof of Theorem 6.1. We finish this section with relation for the moments Lk . Using representation fq,c (z) = −
∞ X Lk , k+1 z k=0
one can easily derive from (3.6) the system L0 = 1, Lk+1
L1 = bc,
L2 = ac + (1 + c)bL1 , X X = (c − 1)bLk + b Li Lj + ac Li Lj − − abc
X
i+j =k
Li Lj Ll ,
i+j =k−1
for k > 2,
(6.6)
i+j +l=k−2
√ where a = u4 (1 − q) and b = u2 q. Analysis of this system is more complicated than that of (6.3). 7. Summary We have studied the eigenvalue distribution of weakly diluted N × N random matrices in the limit N → ∞. The basic ensemble represents dilution of the Wishart random matrices HN which are widely known in multivariate statistical analysis. The second ensemble is the dilute version of the Gaussian random matrices AN with weakly correlated entries. In the pure (nondilute) case both of these ensembles, {HN } and {AN }, have limiting eigenvalue distributions that can have a singular component at the origin. We derived explicit equations determining the limiting eigenvalue distributions of the dilute versions of {HN } and {AN }. We showed that in the case of weak dilution, when each matrix row contains, on average, qN, 0 < q < 1 nonzero entries, the density of the eigenvalue distribution is bounded by const.(1 − q)−1 . Thus, we can conject that in general the singularities (if any) in the spectra of random matrices are rather unstable and disappear when the dependence between matrix entries is waived. Taking into account our results, one can say that this perturbation can be fairly weak, as it is represented by the weak random dilution.
30
A. KHORUNZHY AND G. J. RODGERS
Acknowledgements A.K. would like to thank the Department of Mathematical Sciences at Brunel University for hospitality while this work was performed and both authors are grateful to the Royal Society, London for the financial support.
References 1. 2. 3. 4.
5. 6.
7. 8. 9. 10. 11. 12. 13.
14. 15. 16. 17.
18.
19.
Anderson, T. W.: An Introduciton to Multivariate Statistical Analysis, John Wiley, New York, 1984. Amit, D. J.: Modeling Brain Function, Cambridge University Press, Cambridge, 1989. Boutet de Monvel, A. and Khorunzhy, A.: On the norm and eigenvalue distribution of large random matrices, Ann. Probab. 27(2) (1999), 913–944. Boutet de Monvel, A., Khorunzhy, A. and Vasilchuk, V.: Limiting eigenvalue distribution of eigenvalues of random matrices with correlated entries, Markov Proc. Rel. Fields 2 (1996), 607–636. Bovier, A. and Gayrard, V.: Rigorous results on the thermodynamics of the dilute Hopfield model, J. Stat. Phys. 69(1–2) (1993), 597–627. Bovier, A. and Gayrard, V.: The Hopfield model as a generalized random mean field model, In: A. Bovier and P. Picco (eds), Mathematical Aspects of Spin Glasses and Neural Networks, Progr. Probab. 41, Birkhäuser, Boston, 1997, pp. 3–89. Brouwer, P. W. and Beenakker, C. W. J.: Diagrammatic method of integration over the unitary group, J. Math. Phys. 37(10) (1996), 4904–4934. Brody, T. A., Flores, J., French, J. B., Mello, P. A., Pandey, A. and Wong, S. S. M.: Randommatrix physics: Spectrum and strength fluctuations, Rev. Modern Phys. 53(3) (1981), 385–479. Crisanti, A., Paladin, G. and Vulpiani, A.: Products of Random Matrices in Statistical Physics, Springer Ser. Solid-State 104, Springer-Verlag, Berlin, 1993. Di Francesco, P., Ginsparg, P. and Zinn-Justin, J.: 2D gravity and random matrices, Phys. Rep. 254 (1995), 1–133. Donoghue, W.: Monotone Matrix Function and Analytic Continuation, Grundlehren Math. Wiss. 207, Springer-Verlag, New York, 1974. Girko, V.: An Introduction to Statistical Analysis of Random Arrays, VSP, Utrecht, 1998. Hertz, J. A., Krogh, A. and Palmer, R. G.: Introduction to the Theory of Neural Computations, Santa Fe Inst. Stud. Sci. Complexity Lecture Notes, I, Addison-Wesley, Redwood City, CA, 1991. Hopfield, J. J.: Neural networks and physical systems with emergemt collective computational abilities, Proc. Natl. Acad. Sci. USA 79 (1982), 2554–2558. Khorunzhy, A.: Eigenvalue distribution of large random matrices with correpated entries, Mat. Fiz. Anal. Geom. 3(1–2) (1996), 80–101. Khorunzhy, A.: On dilute unitary random matrices, J. Phys. A: Math. Gen. 31 (1998), 4773– 4784. Khorunzhy, A., Khoruzhenko, B., Pastur, L. and Shcherbina, M.: Large-n limit in statistical mechanics and the spectral theory of disordered systems, In: C. Domb and J. L. Lebowitz (eds), Phase Transitions and Critical Phenomena 15, Academic Press, New York, 1992, pp. 67–237. Khorunzhy, A. and Pastur, L.: On the eigenvalue distribution of the deformed Wigner ensemble of random matrices, In: V. A. Marchenko (ed.), Spectral Operator Theory and Related Topics, Adv. Sov. Math. 19, Amer. Math. Soc., Providence, RI, 1994, pp. 97–127. Khorunzhy, A. and Rodgers, G. J.: Eignevalue distribution of large dilute random matrices, J. Math. Phys. 38 (1997), 3300–3320.
EIGENVALUES OF WEAKLY DILUTE WISHART MATRICES
20. 21. 22. 23. 24. 25. 26. 27. 28. 29. 30. 31. 32.
31
Khorunzhy, A. and Rodgers, G. J.: On the Wigner law in dilute random matrices, Rep. Math. Phys. 42 (1998), 297–319. Kolchin, V. F.: Random Graphs, Cambridge University Press, Cambridge, 1999. Lamperti, L.: Stochastic Processes: A Survey of the Mathematical Theory, Appl. Math. Sci. 23, Springer-Verlag, New York, 1977. Marchenko, V. A. and Pastur, L. A.: Eignevalue distribution of certain ensembleof random matrices, Mat. Sbornik 72, 507–536 [English translation: Math. USSR-Sb 1 (1967), 457–483]. Mézard, M., Parisi, G. and Virasoro, M. A.: Spin-Glass Theory and Beyond, World Scientific Lecture Notes in Phys. 9, World Scientific, Singapore, 1987. Mirlin, A. D. and Fyodorov, Ya. V.: Universality of level correlations of sparse random matrices, J. Phys. A: Math. Gen. 24(10) (1991), 2273–2286. Pastur, L. A. and Figotin, A.: Exactly solvable model of a spin glass, Soviet J. Low Temp. Phys. 3 (1977), 378–383. Porter, C. E. (ed.): Statistical Theories of Spectra: Fluctuations, Academic Press, New York, 1965. Rodgers, G. J. and Bray, A. J.: Density of states fo dilute random matrix, Phys. Rev. B 37(7) (1988), 3557–3562. Rodgers, G. J. and De Dominicis, C.: Density of states of sparse random matrices, J. Phys. A: Math. Gen. 23(9) (1990), 1567–1573. Stariolo, D. A., Curado, E. M. F. and Tamarit, F. A.: Eigenvalue distribution of dilute Hopfield model, J. Phys. A: Math. Gen. 29(15) (1996), 4733–4739. Talagrand, M.: Rigorous results for the Hopfield model with many patterns, Probab. Theory Related Fields 110 (1998), 177–276. Wigner, E.: Characteristic vectors of bordered matrices with infinite dimensions, Ann. Math. 62 (1955), 548–564.
Mathematical Physics, Analysis and Geometry 3: 33–47, 2000. © 2000 Kluwer Academic Publishers. Printed in the Netherlands.
33
On Ground-Traveling Waves for the Generalized Kadomtsev–Petviashvili Equations A. PANKOV1 and K. PFLÜGER2
1 Department of Mathematics, Vinnitsa State Pedagogical University, Ukraine.
e-mail:
[email protected] 2 Institut für Mathematik I, Freie Universität Berlin, Germany. e-mail:
[email protected]
(Received: 18 March 1999; in final form: 18 October 1999) Abstract. As a continuation of our previous work, we improve some results on convergence of periodic KP traveling waves to solitary ones as the period goes to infinity. In addition, we present some qualitative properties of such waves, as well as nonexistence results, in the case of general nonlinearities. We suggest an approach which does not use any scaling argument. Mathematics Subject Classifications (2000): 35Q53, 35B10, 35A35, 35A15. Key words: generalized Kadomtsev–Petviashvili equation, traveling waves, variational methods.
1. Introduction Kadomtsev–Petviashvili (KP) equations, both original and generalized, appear in the theory of weakly nonlinear dispersive waves [7]. They read ut + uξ ξ ξ + f (u)ξ + εvy = 0,
vξ = uy
(1)
or, eliminating v, (ut + uξ ξ ξ + f (u)ξ )ξ + εuyy = 0.
(2)
More precisely, these are KP-I equations if ε = −1, and KP-II equations if ε = +1. The original KP equations correspond to the case f (u) = 12 u2 , form a completely integrable Hamiltonian system, and were studied extensively by means of algebrogeometrical methods (see, e.g., [8]). There are also a number of papers dealing with more general equations (1) or (2), mainly in the case of pure power nonlinearity: [1, 3 – 6, 10, 17, 20, 23], to mention a few. In particular, solitary traveling waves were studied [1, 4 – 6, 10, 17, 23]. Here we consider mainly the case of KP-I equations. Remark that KP-II equations do not possess traveling waves at all (see [5] for the case of solitary waves and pure power nonlinearity, and Section 4 for the general case). The present paper is a direct continuation of our previous work [17]. It concerns the existence of ground-traveling waves, both periodic and solitary, and the limit
34
A. PANKOV AND K. PFLÜGER
behavior of periodic waves, as the period goes to infinity. Corresponding equations for traveling waves read −cux + uxxx + f (u)x + εvy = 0,
vx = uy
(3)
and (−cux + uxxx + f (u)x )x + εuyy = 0,
(4)
respectively. Here x = ξ − ct, c > 0 is the wave speed. In [17], among other results we have proved that k-periodic in x ground waves converge to a solitary ground wave in a very strong sense (Theorem 5 of that paper). Unfortunately, that result does not cover the case of the original KP equation, but includes the case f (u) = u3 . The first aim of this paper is to extend the results of Theorems 4 and 5, [17], in order to include nonlinearities like f (u) = |u|p−1 , 2 < p < 6. This will be done in Section 2. Our second goal is to discuss, in Section 3, some qualitative properties of KP traveling waves: symmetry, continuity, and rate of decay. As for continuity and decay properties, we follow very closely the paper [6] and point out only the main differences. On the contrary, our proof of symmetry with respect to y-variable relies on a quite different variational characterization of ground waves and permits us to treat the case of nonhomogeneous nonlinearity. Finally, in Section 4, we discuss the nonexistence of traveling waves, both solitary and periodic, for general nonlinearities. All the assumptions we impose here are satisfied for the nonlinearities f (u) = c|u|p−1
and
f (u) = c|u|p−2 u +
k X
ci |u|pi −2 u,
i=1
with c, ci > 0,
2 < p < 6,
2 < pi < p.
Unfortunately, Assumptions (N) and (N1) below are not satisfied for a very interesting nonlinearity f (u) = u2 − u3 which appears in some physical models [18]. In addition, let us remark that it is natural to look at (k, l)-periodic traveling wave solutions with respect to (x, y), as well as l-periodic waves with respect to y which are decaying in x. All the results below, except of Theorem 2, have their straightforward counterparts in both these cases. The case of double periodic waves is even simpler, since the corresponding functional satisfies the Palais–Smale condition. However, as we already mentioned in our previous paper [17], our technique does not work, at least directly, when we try to study the behavior of such traveling waves as (k, l) → ∞ (respectively, l → ∞). 2. Ground Waves Denote by F (u) = assumptions:
Ru 0
f (t) dt the primitive function of f . We make the following
35
GROUND-TRAVELING WAVES AND KP EQUATIONS
(1) (2) (3) (4)
f ∈ C(R), f (0) = 0; |f (u)| 6 C(1 + |u|p−1 ), 2 < p < 6, andR f (u) = o(|u|) as u → 0; there exists ϕ ∈ C0∞ (R2 ) such that 1/λ2 R2 F (λϕx ) → +∞, as λ → +∞; there exists µ > 2 such that µF (u) 6 uf (u) for all u ∈ R.
Remark that if F (u) > 0 for all u 6= 0, then Assumption (3) follows from (4). Let Qk = (−k/2, k/2) × R, 0 < k 6 ∞. We set Z x −1 Dx,k u(x, y) = u(s, y) ds, k ∈ (0, ∞]. (5) −k/2
We shall simply write Dx−1 in the case k = ∞. Define the Hilbert space Xk as the completion of {ϕx : ϕ ∈ Ck∞ }, where Ck∞ is the space of smooth functions on R2 which are k-periodic in x and have finite support in y, with respect to the norm 1/2 kukk = (u, u)k , Z −1 −1 (u, v)k = ux vx + Dx,k uy · Dx,k vy + cuv. Qk
Similarly, X = X∞ is the completion of {ϕx : ϕ ∈ C0∞ (R2 )} with respect to the 1/2 −1 is well-defined on norm kuk = kuk∞ = (u, u)1/2 = (u, u)∞ . The operator Dx,k the space Xk , k ∈ (0, ∞]. For k-periodic traveling waves, k ∈ (0, ∞), Equation (4) may be written in the form [17] −2 uyy + cu − f (u))x = 0. (−uxx + Dx,k
(6)
Solitary waves are solutions of the same Equation (6), with k = ∞. The action functional associated with (6) reads [17] Z 1 2 Jk (u) = kukk − F (u); (7) 2 Qk Jk is of the class C 1 on Xk . We consider weak solutions of (6), i.e. critical points of Jk in Xk . Now let us consider the so-called Nehari functional Z 0 2 uf (u), (8) Ik (u) = hJk (u), ui = kukk − Qk
and the Nehari manifold Sk = {u ∈ Xk : Ik (u) = 0, u 6= 0}. All traveling wave solutions lie in the corresponding Nehari manifold and we will find ground waves, i.e. solutions with minimal action among all nontrivial solutions, solving the following minimization problem: mk = inf{Jk (u) : u ∈ Sk }.
(9)
36
A. PANKOV AND K. PFLÜGER
Remark that
Z
Jk (u) = Qk
1 uf (u) − F (u), 2
u ∈ Sk .
(10)
In what follows, we will omit the subscript k if k = ∞ and write simply J, I, . . . . Throughout this section, in addition to Assumptions (1)–(4), we impose the following one (N) For any u ∈ L2 (R2 ) such that Z uf (u) > 0, R2
the function of t Z t −1 uf (tu) R2
is strictly increasing on (0, +∞). In the proof of Theorem 1, [17], we have considered the Mountain Pass Values ck for Jk and proved that they are uniformly bounded from below and above by positive constants. More precisely, ck = inf max Jk (γ (t)), γ ∈0k t ∈[0,1]
where 0k = {γ ∈ C([0, 1], Xk ) : γ (0) = 0, Jk (γ (1)) < 0}. Here we have defined 0k in a slightly different way than in [17], but it does not effect on the value of ck . Consider also another minimax value ck0 = inf+ sup Jk (tv), v∈Xk t >0
where Xk+
Z = v ∈ Xk :
F (v) > 0 .
Qk
Due to Assumption (4), Xk+ 6= ∅. LEMMA 1. For every v ∈ Xk+ , there exists a unique tk = tk (v) such that tk v ∈ Sk , Jk (tk v) = max Jk (tv), t >0
and tk (v) depends continuously on v ∈ Xk+ .
GROUND-TRAVELING WAVES AND KP EQUATIONS
37
Proof. Assumption (4) implies that Z vf (v) > 0 Qk
for any v ∈ Xk+ . Therefore, due to Assumption (N), the function Z d 2 2 −1 vf (tv) Jk (tv) = Ik (tv) = t kvkk − t dt Qk vanishes at only one point tk = tk (v) > 0. Equation (10) and Assumption (4) imply that Jk is positive on Sk . Since Jk (0) = 0, we see that tk is a point of maximum for 2 Jk (tv). Continuity of tk (v) is easy to verify. LEMMA 2. ck = ck0 = mk . Proof. Since uf (u) is subquadratic at 0 and the quadratic part of Jk is positive defined, we see that Ik (v) > 0 in a neighborhood of the origin, except of 0. Hence, Ik (γ (t)) > 0, γ ∈ 0k , for small t > 0. Due to Assumption (4), for v ∈ Xk+ we have Z Z F (v) > kvk2k − µ F (v) 2Jk (v) = kvk2k − 2 Qk Qk Z 2 vf (v) = Ik (v). > kvkk − 2 Qk
Hence, Ik (γ (1)) < 0. Therefore, γ (t) crosses Sk and this implies that ck > mk . By Assumption (4), for any v ∈ Xk+ we have F (tv) > αt µ , α > 0, if t > 0 is large enough. This implies that Jk (tv) < 0 for every v ∈ Xk+ and sufficiently large t > 0. Hence, the half-axis {tv : t > 0} generates in a natural way an element of 0k . This implies the inequality ck 6 ck0 . R Now let v ∈ Sk . By the definition of Ik , σ = Qk vf (v) > 0, and (N) implies that Z Z Z d F (tv) = vf (tv) > t −1 vf (tv) > σ > 0 dt Qk Qk Qk provided t > 1. Hence, for t > 0 large enough Z F (tv) > 0. Qk
By definitions of ck0 and mk , we see that ck0 = mk .
2
THEOREM 1. Assume Assumptions (1)–(4) and (N) are fulfilled. Then, for any k ∈ (0, ∞), there exists a minimizer uk ∈ Sk of (9) which is a critical point of Jk . Moreover, Jk (uk ) = mk is bounded from above and below by positive constants independent on k.
38
A. PANKOV AND K. PFLÜGER
Proof. In the proof of Theorem 1, [17], it is shown that there exists a Palais– Smale sequence uk,n ∈ Xk at the level ck , i.e. Jk0 (uk,n ) → 0,
Jk (uk,n ) → ck p
as n → ∞. Moreover, uk,n → uk weakly in Xk and strongly in Lloc (R2 ), where uk ∈ Xk is a nontrivial solution of (6). Therefore, Ik (uk,n ) = hJk0 (uk,n ), uk,n i → 0 and
Z Jk (uk,n ) −
1 I (u ) 2 k k,n
= Qk
1 u f (uk,n ) 2 k,n
− F (uk,n ) → ck .
Due to Assumption (4), the integrand here is nonnegative and, since uk,n → uk in L2loc (R2 ), we have Z 1 u f (uk ) − F (uk ) 6 ck . 2 k Qk
However, uk is a nontrivial solution, hence, uk ∈ Sk . Therefore, we deduce from (10) that Z 1 u f (uk ) − F (uk ) > mk . Jk (uk ) = 2 k Qk
Now Lemma 2 implies that Jk (uk ) = mk and uk is a ground-wave solution. The last statement of the theorem follows immediately from Lemma 2 and 2 uniform estimates for ck . Remark 1. The Nehari variational principle suggested in [13] was used successfully in many papers (see, e.g., [2, 9, 14 – 17, 22]). In all these papers, except [16], the geometry of Nehari manifold is simple enough: it is a bounded surface without boundary around the origin, like a sphere. In the case we consider here, the picture is different: Sk may look like a sphere if, e.g., f (u) = |u|p−2 u, and may be unbounded if, e.g., f (u) = |u|p−1 . Nevertheless, in any case, Sk separates the origin and the domain of negative values of Jk , which is sufficient for our purpose. In [16], such a manifold is also unbounded in general, but there we have used different arguments. Now we are going to study the behavior of uk , as k → ∞. Recall the definition of cut-off operators Pk : Xk → X, [17]. Let χk ∈ C0∞ (R) be a nonnegative function such that χk (x) = 1 for x ∈ [−k/2, k/2], χk (x) = 0 for |x| > (k + 1)/2, and |χk0 |, |χk00 | 6 C0 , with some constant C0 > 0. We set −1 Pk u(x, y) = [χk (x)Dx,k u(x, y)]x .
39
GROUND-TRAVELING WAVES AND KP EQUATIONS
THEOREM 2. Assume that Assumptions (1)–(4) and (N) are satisfied. Let uk ∈ Xk be a sequence of ground-wave solutions. Then there exists a nontrivial ground wave u ∈ X and a sequence of vectors ζk ∈ R2 such that, along a subsequence, Pk uk (· + ζk ) → u weakly in X. If, in addition, |f (u + v) − f (u)| 6 C(1 + |u|p−2 + |v|p−2 )|v|,
v ∈ R,
(11)
then, along the same subsequence, lim kuk (· + ζk ) − ukk = 0.
k→∞
Proof. By Theorem 2, [17], there is a nontrivial solution u ∈ X such that Pk uk (· + ζk ) → u weakly in X for some ζk ∈ R2 (along a subsequence). Let us prove that u is a ground wave, i.e. J (u) = inf{J (v) : v ∈ S} = m. First of all, for any v ∈ S and any ε > 0, there exist kε and vk ∈ Sk such that Jk (vk ) 6 J (v) + ε,
k > kε .
Indeed, since J and I are continuous, we can find ϕk ∈ C0∞ (Qk ) such that ηk = Dx ϕk → v in X and, hence, J (ηk ) → J (v),
I (ηk ) → I (v) = 0.
Since I (v) = 0 and v 6= 0, we have Z vf (v) = kvk2 > 0. Qk
R Hence, Qk ηk f (ηk ) > 0 for k large enough. Due to (N), there exists τk > 0 such that I (τk ηk ) = 0 and τk → 1. Let vk be a unique k-periodic function which coincides with τk ηk on Qk . Then Jk (vk ) = J (τk ηk ) 6 J (v) + ε provided k is large enough. In particular, we have lim supk→∞ mk 6 m. Now, exactly as in the proof of Theorem 5, [17], we see that lim infk→∞ mk > J (u) > m. Hence, m = J (u) and u is a ground wave solution. The second part of the theorem follows from Theorem 3, [17], exactly as at the end of the proof of Theorem 5, [17]. 2
3. Qualitative Properties of Traveling Waves Now we are going to study such properties of KP traveling waves as symmetry, regularity, and decay. We start with the following:
40
A. PANKOV AND K. PFLÜGER
LEMMA 3. Suppose that Assumptions (1) and (2) are satisfied. In the case 2 < p 6 5, assume, in addition, that f ∈ C 2 (R) and |f (j ) (u)| 6 C(1 + |u|p−1−j ),
j = 1, 2,
u 6= 0.
(12)
Then any traveling wave is continuous. Moreover, any solitary (resp. periodic) wave tends to zero as (x, y) → ∞ (resp. y → ∞). Proof. For such a wave u ∈ Xk , we have −cvxx − vyy + vxxxx = f (u)xx = gxx . Let
Z (Fk,x h)(ξ ) =
(13)
k/2
h(x) exp(−iξ x) dx −k/2
be the Fourier transform if k = ∞ (then we simply write Fx ), and the sequence of Fourier coefficients if k < ∞. In the last case, ξ ∈ (2π/k)Z. Now we get from (13) Fk,x Fy u = p(ξ1 , ξ2 )(Fk,x Fy g),
(14)
where p(ξ ) = p(ξ1 , ξ2 ) =
ξ12 , cξ12 + ξ14 + ξ22
ξ1 and ξ2 are dual variables to x and y, respectively. If k = ∞, there is nothing to do. In the case 5 < p < 6, one needs only to repeat the proof of Theorem 1.1, [6], which does not use any particular property of power nonlinearity, except of its growth rate. In the case 2 < p 6 5 the arguments from the proof of Lemma 4.1, [5], work and just here assumption (12) is needed. Now we explain how to cover the case of periodic waves. Recall the following Lizorkin theorem [11]. Let p(ξ ), ξ ∈ Rn , be of the class C n for |ξj | > 0, j = 1, . . . , n. Assume that k1 ∂kp ξ · · · ξ kn 6 M, n 1 ∂ξ1k1 · · · ∂ξnkn with kj = 0 or 1, k = k1 +· · · +kn = 0, 1, . . . , n. Then p(ξ ) is a Fourier multiplier on Lr (Rn ), 1 < r < ∞. We rewrite now (14) as follows: Fk,x u = Fy−1 [p(ξ1 , ξ2 )Fy Fk,x g] = P (ξ1 )g, where P (ξ1 ) is the operator Fy−1 p(ξ1 , ·)Fy for any fixed ξ1 . It is easy to verify that P (ξ1 ) ∈ L(Lr (Ry )), the space of bounded linear operators in Lr (Ry ). Moreover, due to the Lizorkin theorem, p(ξ ) is a multiplier in Lr (R2 ). Hence, so is it for P (ξ1 ) in the space Lr (Rx , Lr (Ry )) = Lr (R2 ). It is not difficult to verify that
GROUND-TRAVELING WAVES AND KP EQUATIONS
41
P (ξ1 ) depends continuously on ξ1 with respect to the norm in L(Lr (Ry )) at any point ξ1 6= 0. Therefore, by Theorem 3.8 of Ch. 7, [21], we see that P (ξ1 ) is also a multiplier in the space Lr ((−k/2, k/2), Lr (Ry ))0 = Lr (Qk )0 considered as the space of k-periodic in x functions. The subscript 0 means that for functions from this space Fk,x u vanishes at ξ1 = 0. Since p(0, ξ2 ) = 0, the corresponding multiplier vanishes on {u ∈ Lr (Qk ) : Fk,x u = 0 if ξ1 6= 0} and, hence, is a bounded operator on the entire space Lr (Qk ). In fact, we need here an extension of that theorem for operator-valued multipliers which may be discontinuous at the point 0. However, in this case, the proof presented in [21] works without any change. To complete the proof in the case 5 < p < 6, we can now use the same reiteration argument as in [6]. In the case 2 < p 6 5, again one needs to invoke the arguments of the proof of Theorem 4.1, [5]. Here we have to apply the remark on operator-valued multipliers to p1 (ξ ) = ξ12 p(ξ ), p2 (ξ ) = ξ2 p(ξ ), as well as to p(ξ ) itself. 2 We also need the following additional assumption: R (N1) f ∈ C 1 (R) and, for any v ∈ L2 (R2 ) such that R2 f (v)v > 0, we have Z Z f (v)v < f 0 (v)v 2 R
R2
R2
f (tv)v > 0 ∀t > 0. R Calculating the derivative of t −1 R2 f (tv)v, we see that (N1) implies (N). Let us introduce the functional Z 1 f (v)v − F (v) , v ∈ Xk . Lk (v) = 2 and
R2
Qk
As we have seen, Lk = Jk on Sk and Lk (v) > 0, ∀v ∈ Xk . LEMMA 4. Under R Assumption (N1), Lk (tv) is a strictly increasing function of t > 0, provided Qk f (v)v > 0. Proof. It follows immediately from the following elementary identity Z Z 1 d Lk (tv) = f 0 (tv)t 2 v − f (tv)tv . 2 dt 2t Qk Qk We also need the following dual characterization of ground-traveling waves: LEMMA 5. Suppose Assumptions (1)–(4) and (N1) are satisfied. For nonzero u ∈ Xk , k ∈ (0, ∞], the following statements are equivalent:
42
A. PANKOV AND K. PFLÜGER
(i) u is a ground wave, (ii) Ik (u) = 0 and Lk (u) = mk = inf{Lk (v) : v ∈ Sk }, (iii) Ik (u) = 0 = sup{Ik (v) : v ∈ Xk , Lk (v) = mk }. Proof. Implication (i) ⇒ (ii) is proved in Section 2. To prove (ii) ⇒ (i) assume that u ∈ Xk satisfies (ii). Since Jk = Lk on Sk , there exists a Lagrange multiplier λ such that λIk0 (u) = Jk0 (u). Then λhIk0 (u), ui = hJk0 (u), ui = Ik (u) = 0. On the other hand, hIk0 (u), ui
=
Z −
0
Z
f (u)u − f (u)u Qk Z Z f (u)u − f 0 (u)u2 = 2Ik (u) + Qk Qk Z Z = f (u)u − f 0 (u)u2 . 2kuk2k
2
Qk
R
Qk
Qk
However, Qk f (v)v > 0 on Sk and, due to (N1), hIk0 (u), ui < 0. Therefore, λ = 0 and u is a ground wave. Now let us prove (ii) ⇒ (iii). For u as in (ii), Ik (u)R= 0. Assume that there is v ∈ Xk such that Lk (v) = mk and Ik (v) < 0. Then Qk f (v)v > 0 and there exists t0 ∈ (0, 1) such that Ik (t0 v) = 0. By Lemma 4, Lk (t0 v) < Lk (v) = mk , which is impossible. Finally, we prove (iii) ⇒ (ii). Let u ∈ XR k satisfies (iii). Then, Lk (u) > mk . Assume that Lk (u) > mk . Again we have Qk f (u)u > 0. By Lemma 4, there exists t0 ∈ (0, 1) such that Lk (t0 u) = mk . However, Ik (t0 u) > 0 and this contradicts (iii). 2 Now we are ready to prove the symmetry property for all kinds of ground waves we consider. As in [6], we use the approach suggested in [12] (see also [22]). THEOREM 3. In addition to Assumptions (1)–(4) and (N1), suppose that f ∈ C 2 (R). In the case 2 < p 6 5 assume also that inequality (12) is fulfilled. Then any ground wave u ∈ Xk , k ∈ (0, ∞], is symmetric with respect to some line 1 = {(x, y) ∈ R2 : y = b}. Proof. Choose b in such a way that Z Z 1 1 mk , f (v)v − F (v) = f (v)v − F (v) = 2 2 2 1+ ∩Qk 1− ∩Qk where 1+ and 1− are the corresponding upper and lower half-planes. Let u± be a symmetric (with respect to 1) function such that u± = u on 1± . Then u± ∈ Xk and Lk (u± ) = Lk (u) = mk .
GROUND-TRAVELING WAVES AND KP EQUATIONS
43
By Lemma 5, Ik (u± ) 6 0. On the other hand, Ik (u+ ) + Ik (u− ) = 2Ik (u) = 0. Using Lemma 5, we conclude that u± is a ground wave. To conclude that u± = u and, hence, complete the proof, it is sufficient to use the same unique continuation result as in [6], and just here we need the assumption f ∈ C 2 (R) and Lemma 3. Remark that a periodic version (with 5 = 1± ) of unique continuation Theorem A.1, [6], can be proved exactly as that theorem itself. 2 In addition, we formulate the following direct generalization of results of [6] for decay of solitary waves. THEOREM 4. Suppose Assumptions (1) and (2) to be satisfied. Let u ∈ Xk , k ∈ (0, ∞], be a traveling wave. If k = ∞, then r 2 u ∈ L∞ (R2 ),
r 2 = x2 + y2.
If 0 < k < ∞, then y 2 u ∈ L∞ (Qk ). The proof is essentially the same as in [6]. In the case k < ∞, one needs only to use the partially periodic Fourier transform as in Lemma 3. Let us remark that the classical rational KP-solitons decay exactly as r−2 . On the other hand, in [24] a family of traveling waves is constructed for the original KP-I equations which are periodic in x and decay exponentially fast with respect to the transverse variable. Thus, it seems that the statement of Theorem 4 is not exact in the case k < ∞, while it is so in the case k = ∞. Also unknown are Zaitsev’s ground-wave solutions. 4. On Nonexistence of Traveling Waves In this section we turn to general KP equations (3), with ε = ±1, and discuss the nonexistence problem. We use the same approach as in [5]. However, the case of periodic waves is more involved (see the proof of Lemma 6). Here we consider traveling waves belonging to the space −1 uyy ∈ L2 (Qk ), f (u)u ∈ L1 (Qk )} Yk = {u ∈ Xk : u ∈ H 1 (Qk ), uxx , Dx,k
if k < ∞, and −1 Y = Y∞ = {u ∈ X : u ∈ H 1 (R2 ), uxx , Dx,k uyy ∈ L2loc (R2 ), f (u)u ∈ L1 (R2 )}.
First, we collect some useful identities. LEMMA 6. Suppose that f satisfies Assumptions (1) and (2). Let u ∈ Yk , k ∈ (0, ∞], be a solution of Equations (3). Then
44
A. PANKOV AND K. PFLÜGER
Z
v2 c 2 3 2 u + ux + ε − uf (u) + F (u) = 0, 2 2 Qk 2 Z v2 c 2 1 2 u + ux + ε + F (u) = 0, 2 2 Qk 2 Z [cu2 + u2x − εv 2 − f (u)u] = 0.
(15) (16) (17)
Qk
Proof. First, we remark that, for any k, (17) is an extention of the case ε = ±1 of Ik (u) = 0 stated in Section 2. Therefore, we concentrate on (15) and (16) only. In the case of solitary waves (k = ∞), the calculations carried out in the proof of Theorem 1.1, [5], work equally well for general nonlinearities. Therefore, we look at periodic waves (k < ∞). Fixed κ ∈ (0, 1), let ϕT ∈ C0∞ (R) be a nonnegative function such that ϕT = 1 on [−T /2, T /2], ϕT (x) = 0 if |x| > (T + T κ )/2, and ϕ (j ) (x) 6 Cj /|x|j , j = 1, 2, . . ., if T /2 6 |x| 6 (T + T κ )/2 (the construction of such a function will be given later on). Multiplying the first equation (3) by xϕT u and integrating over R2 , after a number of integrations by parts, we get Z Z Z Z 3 c 2 ϕT u − ϕT uf (u) + ϕT F (u) + ϕT u2x + 2 2 Z Z Z Z 1 1 2 0 2 0 ϕT v + xϕT u − xϕT uf (u) + xϕT0 F (u)+ +ε 2 2 Z Z Z Z 3 1 0 00 0 2 xϕT ux + ε xϕT0 v 2 = 0. + 2 ϕT uux + xϕT uux + 2 2 Dividing the last identity by T , we are going to pass to the limit as T → ∞. First, we point out that here the integrals containing ϕT are taken over QT ∪ Q0T ∪ Q00T = QT ∪ {(T /2, (T + T κ )/2) × R} ∪ {(−(T + T κ )/2, −T /2) × R}, while those containing ϕT0 and ϕT00 are over Q0T ∪ Q00T . Moreover, ϕT = 1 on QT . Now, let g ∈ L1loc (R2 ) be a function which is k-periodic in x. Then, it is easy to verify that Z Z 1 g= g. lim T →∞ T Q Qk T Next, dueRto the properties of ϕT , all the integrals over Q0T can be estimated from above by Q0 g, with a nonnegative k-periodic in x function g ∈ L1loc (R2 ). Now T Z Z 1 Tκ +1 g6 g. T Q0T T Qk
GROUND-TRAVELING WAVES AND KP EQUATIONS
45
This justifies the passage to the limit and gives rise to (15). Identity (16) can be proved exactly as (2.8), [5], with the only change: take the cut-off functions χj depending on y only. Now we construct the function ϕT . Fix ε > 0 and let ( 1 if x 6 T /2, g(x) = 1 − log(x/T ) if T /2 < x 6 (T + T κ )/2 + ε, 0 if x > (T + T κ )/2 + ε. ∞ We R choose a nonnegative function h ∈ C0 (R) such that supp h ⊂ (0, ε) and h = 1, and set Z ϕT (x) = ϕ˜T (|x|). ϕ˜T (x) = h(x − t)g(t) dt,
For this function, it is easy to verify all the properties we need.
2
THEOREM 5. Suppose that f ∈ C(R) satisfies Assumption (4). Then there is no nontrivial traveling wave u ∈ Yk , k ∈ (0, ∞], provided ε = +1, or ε = −1 and µ > 6. Proof. Adding (15), (16) and subtracting (17), we get Z Z 2 ux = −2ε v2. Qk
Qk
This rules out the case ε = +1. In the case ε = −1 (KP-I equations) the last identity together with (15) and (17), respectively, implies Z c 2 5 2 u + v − f (u)u + F (u) = 0 2 Qk 2 and
Z [cu2 + 3v 2 − f (u)u] = 0. Qk
Eliminating v, we get Z Z 2c u2 = [6F (u) − f (u)u]. Qk
Qk
If µ > 6, we have Z Z 2 2c u 6 [µF (u) − f (u)u] 6 0. Qk
Qk
Hence, u = 0 and we conclude.
2
46
A. PANKOV AND K. PFLÜGER
Acknowledgements This work was carried out during the visit of the first author (A.P.) to the Institut für Mathematik, Humboldt Universität Berlin (September–December, 1998) under the support of Deutsche Forschungsgemeinschaft. A.P. is very grateful to K. Gröger for his kind hospitality and a lot of stimulating discussions. The authors thank the anonimous referee for information on paper [24]. References 1.
Ablowitz, M. J., Segur, H. and Wang, X. P.: Wave collapse and instability of solitary waves of a generalized Kadomtsev–Petviashvili equation, Physica D 78(3–4) (1994), 241–265. 2. Bartsch, Th. and Willem, M.: Infinitely many radial solutions of a semilinear elliptic problem on RN , Arch. Rat. Mech. Anal. 124 (1993), 261–276. 3. Bourgin, J.: On the Cauchy problem for the Kadomtsev–Petviashvili equation, Geom. Funct. Anal. 3(4) (1993), 315–341. 4. De Bouard, A. and Saut, J.-C.: Remarks on the stability of generalized KP solitary waves, Contemp. Math. 200 (1996), 75–84. 5. De Bouard, A. and Saut, J.-C.: Solitary waves of generalized Kadomtsev–Petviashvili equations, Ann. Inst. H. Poincaré Anal. Non Linéaire 14 (1997), 211–236. 6. De Bouard, A. and Saut, J.-C.: Symmetry and decay of the generalized Kadomtsev–Petviashvili solitary waves, SIAM J. Math. Anal. 28 (1997), 1064–1085. 7. Kadomtsev, B. B. and Petviashvili, V. I.: On stability of waves in weakly dispersive media, Soviet Phys. Dokl. 15 (1970), 539–541, transl. from Dokl. AN SSSR 192 (1970), 753–756. 8. Krichever, I. M. and Novikov, S. P.: Holomorphic bundles over algebraic curves and nonlinear equations, Russ. Math. Surv. 35(6) (1980), 53–79, transl. from Uspekhi Mat. Nauk 35(6) (1980), 47–68. 9. Lions, P.-L.: The concentration-compactness method in the calculus of variations. The locally compact case. I, II, Ann. Inst. H. Poincaré Anal. Non Linéaire 1 (1984), 109–145, 223–283. 10. Liu Yue and Wang, X. P.: Nonlinear stability of solitary waves of a generalized Kadomtsev– Petviashvili equation, Comm. Math. Phys. 183 (1997), 253–266. 11. Lizorkin, P. I.: Multipliers of Fourier integrals, Proc. Steklov Inst. Math. 89 (1967), 269–290. 12. Lopes, O.: A constrained minimization problem with integrals on the entier space, Bol. Soc. Brasil Mat. (N.S.) 25 (1994), 77–92. 13. Nehari, Z.: On a class of nonlinear second-order differential equations, Trans. Amer. Math. Soc. 95 (1960), 101–123. 14. Pankov, A. A.: Semilinear elliptic equations in Rn with nonstabilizing coefficients, Ukrainian Math. J. 41(9) (1989), 1075–1078, transl. from Ukrain. Mat. Zh. 41(9) (1989), 1247–1251. 15. Pankov, A. A.: On positive solutions of nonlinear elliptic equations on whole space, Soviet Math. Dokl. 44 (1991), 337–341, transl. from Dokl. AN SSSR 319(6) (1991), 1318–1321. 16. Pankov, A. A. and Pflüger, K.: On a semilinear Schrödinger equation with periodic potential, Nonlinear Anal. 33 (1998), 593–609. 17. Pankov, A. A. and Pflüger, K.: Periodic and solitary traveling waves for the generalized Kadomtsev–Petviashvili equation, Math. Meth. Appl. Sci. 22 (1999), 733–752. 18. Pouget, J.: Stability of nonlinear structures in a lattice model for phase transformations in alloys, Phys. Rev. B 46 (1992), 10554–10562. 19. Rabinowitz, P. H.: Minimax Methods in Critical Point Theory with Applications to Differential Equations, Regional Conf. Ser. Math. 65, Amer. Math. Soc., Providence, 1986. 20. Saut, J.-C.: Remarks on the generalized Kadomtsev–Petviashvili equation, Indiana Univ. Math. J. 42 (1993), 1011–1026.
GROUND-TRAVELING WAVES AND KP EQUATIONS
21. 22. 23. 24.
47
Stein, E. M. and Weiss, G.: Introduction to Fourier Analysis on Euclidian Spaces, Princeton Univ. Press, Princeton, 1971. Willem, M.: Minimax Methods, Birkhäuser, Boston, 1996. Willem, M.: On the generalized Kadomtsev–Petviashvili equation, Rapp. Semin. Math. Louvain Nov. Ser. 245–260 (1996), 213–222. Zaitsev, A. A.: On formation of nonlinear stationary waves by means of superposition of solitons, Dokl. Akad. Nauk SSSR 272 (1983), 583–587 (in Russian).
Mathematical Physics, Analysis and Geometry 3: 49–74, 2000. © 2000 Kluwer Academic Publishers. Printed in the Netherlands.
49
Pole Dynamics for Elliptic Solutions of the Korteweg–deVries Equation BERNARD DECONINCK1 and HARVEY SEGUR2
1 Department of Applied Mathematics, Box 352420, University of Washington, Seattle, Washington,
98195, U.S.A. 2 Department of Applied Mathematics, University of Colorado, Boulder, CO 80309-0526, U.S.A. (Received: 21 April 1999; in final form: 24 December 1999) Abstract. The real, nonsingular elliptic solutions of the Korteweg–de Vries equation are studied through the time dynamics of their poles in the complex plane. The dynamics of these poles is governed by a dynamical system with a constraint. This constraint is solvable for any finite number of poles located in the fundamental domain of the elliptic function, often in many different ways. Special consideration is given to those elliptic solutions that have a real nonsingular soliton limit. Mathematics Subject Classifications (2000): 34M05, 35A20, 35Q53, 37K10, 37K20. Key words: KdV equation, elliptic, finite gap solutions, pole dynamics, Calogero–Moser.
1. Introduction In 1974, Kruskal [15] considered the interaction of solitons governed by the Korteweg–de Vries equation (KdV), ut = 6uux + uxxx .
(1.1)
Each KdV soliton is defined by a meromorphic function in the complex x-plane (i.e., sech2 k(x − x0 )), so Kruskal [15] suggested that the interaction of two or more solitons could be understood in terms of the dynamics of the poles of these meromorphic functions in the complex x-plane, where the poles move according to a force law deduced from (1.1). This was followed by the work of Thickstun [17] who considered the case of two solitons in great detail. Following a different line of thought, Airault, McKean and Moser [2] studied rational and elliptic solutions of the KdV equation. An elliptic solution of KdV is by definition a solution of the KdV equation that is doubly periodic and meromorphic in the complex x-plane, for all time. Note that the soliton case is an intermediary case between the elliptic and the rational case. It was treated as such in [2]. Airault, McKean and Moser [2] approached these elliptic solutions and their degenerate limits through the motion of their poles xi (t) in the complex x-plane. In particular, they looked for elliptic KdV solutions of the form u(x, t) = −2
N X i=1
℘ (x − xi (t)).
(1.2)
50
BERNARD DECONINCK AND HARVEY SEGUR
Here ℘ (z) denotes the Weierstrass elliptic function. It can be defined by its meromorphic expansion X 1 1 1 − , (1.3) ℘ (z) = 2 + 2 2 z (z + 2mω + 2nω ) (2mω + 2nω ) 1 2 1 2 (m,n)6=(0,0) with ω1 /ω2 not real.? More properties of the Weierstrass function will be given as they are needed. It is shown in [2] that the dynamics of the poles xi (t) is governed by the dynamical system x˙i = 12
N X
℘ (xi − xj ),
i = 1, 2, . . . , N,
(1.4a)
j =1,j 6=i
(the dot denotes differentiation with respect to time) with the invariant constraint N X
℘ 0 (xi − xj ) = 0,
i = 1, 2, . . . , N.
(1.4b)
j =1,j 6=i
Here the prime denotes differentiation with respect to the argument. The solutions (1.2) generalize an elliptic solution given earlier by Dubrovin and Novikov [7], corresponding to the case N = 3. These authors also recall the Lamé–Ince potentials [13] u(x) = −g(g + 1)℘ (x),
(1.5)
which are the simplest g-gap potentials of the stationary Schrödinger equation ∂ 2ψ + u(x)ψ = λψ. ∂x 2
(1.6)
The remarkable connection between the KdV equation and the stationary Schrödinger equation has been known since the work of Gardner, Greene, Kruskal and Miura [9]. Dubrovin and Novikov show [7] that the (N = 3)-solution discussed in [7] is a 2-gap solution of the KdV equation with a 2-gap Lamé–Ince potential as initial condition. If one considers the rational limit of the solution (1.2) (i.e., the limit in which the Weierstrass function ℘ (z) reduces to 1/z2 ), then the constraint (1.4b) is solvable only for a triangular number of poles, N=
n(n + 1) , 2
(1.7)
? In this paper it is always assumed that ω is real and ω is imaginary. This is necessary to 1 2
ensure reality of the KdV solution (1.2) when x is restricted to the real x-axis. Other considerations for reality of the elliptic KdV solutions will be discussed in Section 4.
POLE DYNAMICS FOR ELLIPTIC SOLUTIONS OF THE KdV EQUATION
51
for any positive integer n [2]. Notice that the Lamé–Ince potentials are given by g(g + 1)/2 times an N = 1 potential. Based on these observations, it was conjectured in [2] that also in the elliptic case given by (1.2), the constraint (1.4b) is solvable only for a triangular number N, ‘or very nearly so’. From the moment it appeared this conjecture was known not to hold, because it already fails in the soliton case, where the Weierstrass function degenerates to hyperbolic functions. This failure of the conjecture easily follows from the work of Thickstun [17]. A further understanding of the elliptic case had to wait until 1988, when Verdier provided more explicit examples of elliptic potentials of the Schrödinger equation [21]. Subsequently, Treibich and Verdier demonstrated that 4 X gi (gi + 1) ℘ (x − x0 − ωi ) u(x) = −2 2 i=1
(1.8)
(ω3 = ω1 + ω2 , ω4 = 0, the gi are positive integers) are finite-gap potentials of the stationary Schrödinger equation (1.6) and, hence, result in elliptic solutions of the KdV equation [19, 18, 20]. The potentials of Treibich and Verdier were generalized by Gesztesy and Weikard [10, 11]. They showed that any elliptic finite-gap potential of the stationary Schrödinger equation (1.6) can be represented in the form u(x) = −2
M X gi (gi + 1) ℘ (x − αi ), 2 i=1
(1.9)
for some M and positive integers gi . Notice that this formula coincides with (1.2) if all the gi are 1. The focus of this paper is the constrained dynamical system (1.4a–b). We return to the ideas put forth by Kruskal [15] and Thickstun [17]. This allows us to derive the system (1.4a–b) in a context which is more general than [2]: there it was obtained as a system describing a class of special solutions of the KdV equation. Here, it is shown that any meromorphic solution of the KdV equation which is doubly periodic in x is of the form (1.2). Hence the consideration of solutions of the form (1.2) and the system of equations (1.4a–b) leads to all elliptic solutions of the KdV equation. Simultaneously, some of the results of Gesztesy and Weikard [10, 11] are recovered here. Because of the connection between the KdV equation and the Schrödinger equation, any potential of Gesztesy and Weikard can be used as an initial condition for the KdV equation, which determines any time dependence of the parameters of the elliptic KdV solution with that initial condition. Our approach demonstrates which parameters in the solutions (1.9) are time independent and which are time dependent. The following conclusions are obtained in this paper:
52
BERNARD DECONINCK AND HARVEY SEGUR
• All finite-gap? elliptic solutions of the KdV equation are of the form (1.2), with all xi (t) distinct for almost all times (see below). In other words, if u(x, t) is a finite-gap KdV solution that is doubly periodic in the complex x-plane, then u necessarily has the form (1.2) except at isolated instants of time. • Any number N 6= 2 of xj is allowed in (1.2). This is trivially true, since it follows from the work of Thickstun [17] that it is true in the soliton limit, which is a special case of the elliptic case. As a consequence, the constraint (1.4b) is solvable for any positive integer?? N 6= 2. In our numerical method, the elliptic case is viewed as a deformation of the soliton limit of the system (1.4a–b). This viewpoint is useful because it provides good initial guesses for many of the numerical solutions of (1.4b) in Section 5. This deformation concept does not lead to all elliptic solutions of the KdV equation, but only to those that have nonsingular soliton limits. In particular, we are unable to find the solutions corresponding to the Treibich–Verdier potentials (1.8) in this way; for these solutions it is necessary to find an initial guess by some other means. In Section (5.2), a solution corresponding to one particular Treibich–Verdier potential is discussed. It is the only solution discussed in this paper that does not have a nonsingular soliton limit. Its inclusion allows us to point out some differences with the other examples in a very concrete way. • If |ω1 /ω2| 1, then for a given N > 4, nonequivalent configurations satisfying the constraint exist that do not flow into each other under the KdV flow and which cannot be translated into each other. To the best of our knowledge this is a new result. • The xi are allowed to coincide, but only in triangular numbers: if some of the xi coincide at a certain time tc , then gi (gi + 1)/2 of them coincide at that time tc . At this time tc , the solution can be represented in the form (1.9) with not all gi = 1. Such times tc are referred to as collision times and the poles are said to collide at the collision time. Before and after each collision time all xi are distinct, hence pole collisions are isolated events. At the collision times, the dynamical system (1.4a) is not valid. The dynamics of the poles at the collision times is easily determined directly from the KdV equation. Gesztesy and Weikard [10, 11] demonstrate that (1.9) are elliptic finite-gap potentials of the Schrödinger equation. These potentials generalize to solutions of the KdV equation, but this requires the gi to be nonsmooth functions of time. Only at the collision times tc are the gi not all one. Furthermore, at all times but the collision times, the number of parameters αi (which are time P g dependent) is N = M i=1 i (gi +1)/2. This shows that the generalization from [10, 11] to solutions of the KdV equation is nontrivial. ? See Section 3. ?? The constraint (1.4b) is solvable for N = 2 [2]. As noted there, the corresponding solution
reduces to a solution for N = 1, with smaller periods. This case is therefore trivial and is disregarded.
POLE DYNAMICS FOR ELLIPTIC SOLUTIONS OF THE KdV EQUATION
53
In the terminology of Chapter 7 of [3], the poles are in general position if all gi are equal to one. Otherwise, if not all gi = 1, the poles are said to be in special position. We conclude that at the collision times the poles are in special position. Otherwise they are in general position. • The solutions discussed here are finite-gap potentials of the stationary Schrödinger equation (1.6), with t treated as a parameter. In other words, each solution specifies a one-parameter family of finite-gap potentials of the stationary Schrödinger equation. It follows from our methods that to obtain a g-gap potential that corresponds to a nonsingular soliton potential, one needs at least N = g(g + 1)/2 poles xj . The Lamé–Ince potentials show that this lower bound is sharp. If we consider potentials that do not have soliton limits (such as the Treibich–Verdier potentials (1.8)) then it may be possible to violate this lower bound. • In Section 5, we present an explicit solution of the form (1.2) with N = 4. Notice that if in (1.8) all gi are one, the solution is reducible to a solution with N = 1 and smaller periods. To see this it is convenient to draw the pole configuration corresponding to (1.8) in the complex plane. This is actually true, even if gi is not one, but all gi are equal. In that case, (1.8) reduces to a Lamé–Ince potential (1.5). Unlike any of the Treibich–Verdier solutions, the (N = 4)-solution, presented in Section 5, has a nonsingular soliton limit. The first five conclusions are all discussed in Section 4. In Section 2, the results of Kruskal [15] and Thickstun [17] for the soliton solutions of the KdV equation are reviewed, but they are obtained from a point of view that is closer to the approach we present in Section 3 for the elliptic solutions of the KdV equation. Finally, in Section 5, some explicit examples are given, including illustrations of the motion of the poles in the complex x-plane. 2. The Soliton Case: Hyperbolic Functions In this section, the results of Kruskal [15] and Thickstun [17] for the dynamics of poles of soliton solutions are discussed from a point of view that will allow us to generalize to the periodic case. Consider the one-soliton solution of the KdV equation u(x, t) = 2k 2 sech2 k(x + 4k 2 t − ϕ).
(2.1)
Here k is a positive parameter (the wave number of the soliton) determining the speed and the amplitude of the one-soliton solution. Using the meromorphic expansion [12] ∞ X 1 1 2x cosech = 2 T T (x + inπ T )2 n=−∞
(2.2)
54
BERNARD DECONINCK AND HARVEY SEGUR
(uniformly convergent except at the points x = inπ T ), one easily obtains the following meromorphic expansion for the one-soliton solution of the KdV equation: u(x, t) = −2k 2
∞ X n=−∞
(k(x +
4k 2 t
1 . − ϕ) + i π2 + inπ )2
(2.3)
From this expression, one easily finds that the locations of the poles of the onesoliton solution of the KdV equation for all time are given by iπ 1 2 xn = ϕ − 4k t − n+ . (2.4) k 2 This motion is illustrated in Figure 1a. Notice that the locations of the poles are symmetric with respect to the real x-axis. This is a consequence of the reality of the solution (2.1). In order for a solution to be real it is necessary and sufficient that if xn (t) is a pole, then so is xn∗ (t), where ∗ denotes the complex conjugate.? The closest distance between any two poles is d = π/k and is constant both along the vertical line Re(x) = ϕ − 4k 2 t and in time. Note that the poles are moving to the left. This is a consequence of the form of the KdV equation (1.1), which has time reversed, compared to the version Kruskal [15] and Thickstun [17] used. Since two solitons of the KdV equation cannot move with the same speed, a two-soliton solution of the KdV equation asymtotically appears as the sum of two one-soliton solutions which are well-separated: the higher-amplitude soliton, which is faster, is to the right of the smaller-amplitude soliton as t → −∞. As t → ∞, the higher-amplitude soliton is to the left of the smaller-amplitude soliton. Hence as t → −∞, the pole configuration of a two-soliton solution with wave numbers k1 and k2 is as in Figure 1b. In this limit, the two-soliton solution is a sum of two one-soliton solutions. Each results in a vertical line of equispaced poles, with interpolar distance respectively d1 = π/k1 and d2 = π/k2 . As long as the solitons are well-separated, these poles move in approximately straight lines, parallel to the real axis, with respective velocities v1 = −4k12 and v2 = −4k22 . Since |v1 | > |v2 |, the solitons interact eventually. This interacting results in nonstraight line motion of the poles. After the interaction, the situation is as in Figure 1b, but with the two lines of poles interchanged. Thickstun [17] considered the case where k1 and k2 are rationally related, so k1 /k2 = p/q, where p and q are positive integers. In this case, one can define D = pd1 = qd2 . The complex x-plane is now divided into an infinite number of equal strips, parallel to the real x-axis, each of height D. The real x-axis is usually taken to be the base of such a strip. It is easy to show [17] that the motion of the poles in one strip is repeated in every strip. Hence, one is left studying the motion of a finite number N (= p + q) of poles in the fundamental strip, whose base is the ? The vertical line of poles can be rotated arbitrarily. The expression (2.3) still results in a solution
of the KdV equation, but it is no longer real. Again, we will only consider real, nonsingular solutions, when restricted to the real x-axis.
POLE DYNAMICS FOR ELLIPTIC SOLUTIONS OF THE KdV EQUATION
55
(a)
(b) Figure 1. (a) The motion of the poles of a one-soliton solution in the complex plane. (b) The asymptotic motion of the poles of a two-soliton solution.
56
BERNARD DECONINCK AND HARVEY SEGUR
real x-axis. Thickstun examined this motion by analyzing the exact expression for a two-soliton solution of the KdV equation. Any two-soliton solution is expressible as [1] u(x, t) = 2∂x2 ln τ (x, t).
(2.5)
It follows from this formula that the poles of u(x, t) are the zeros of τ (x, t) if τ (x, t) is entire in x. Then the Weierstrass Factorization Theorem [5] gives a factorization for τ (x, t): ∞ Y x x/xk 1− e . (2.6) τ (x, t) = C xk k=1 Since only the second logarithmic derivative of this function is relevant, the constant C is not important. If the solution is periodic in the imaginary x-direction, this is rewritten as N Y ∞ Y x 1− ex/(xn +ilD) , (2.7) τ (x, t) = C x + ilD n n=1 l=−∞ where the first product runs over the poles in the fundamental strip. The second product runs over all strips. Using the uniform convergence of (2.7), u(x, t) = −2
N X ∞ X
1 , (x − xn − ilD)2 n=1 l=−∞
(2.8)
which, using (2.2), is rewritten as N π(x − xn ) π2 X cosech2 u(x, t) = −2 2 D n=1 D
= −2
p+q k2 (x − xn ) k22 X cosech2 , 2 q n=1 q
(2.9)
where the pole locations xn depend on time: xn = xn (t). One recovers the onesoliton solution (2.1) easily, by equating k1 = 0, p = 0, q = 1. Equation (2.9) essentially expresses a two-soliton solution as a linear superposition of N onesoliton solutions with nonlinearly interacting phases. Note that the first equality in (2.9) is valid for arbitrary soliton solutions that are periodic in x with period iD. This is the case for a g-soliton solution if its wavenumbers ki , i = 1, 2, . . . , g are all commensurable: (k1 : k2 : . . . : kg ) = (p1 : p2 : . . . : pg ), for positive distinct integers pi , i = 1, 2, . . . , g which have no overall common integer factor. The total number of poles in a strip is then N = p1 + p2 + · · · + pg . In obtaining (2.8) and (2.9), we have deviated from Thickstun’s approach [17] to an approach that is generalized to the elliptic case of the next section in a straightforward way.
POLE DYNAMICS FOR ELLIPTIC SOLUTIONS OF THE KdV EQUATION
57
Next, we derive the dynamics imposed on the poles xn (t) by the KdV equation. This is conveniently done by substituting (2.8) into (1.1) and examining the behaviour near one of the poles: x = xn + . This results in several singular terms as → 0, corresponding to negative powers of . The dynamics of the poles is then determined by the vanishing of the coefficients of these negative powers and the zeroth power. This results in only two nontrivial equations, obtained at order −3 and −2 respectively: ∞ X 1 1 x˙n = 12 + 12 , 2 (xk − xn − ilD) (−ilD)2 k=1,k6=n l=−∞ l6=0,l=−∞ N X
0=
N X
∞ X
∞ X
1 , (xk − xn − ilD)3 k=1,k6=n l=−∞
(2.10a)
(2.10b)
for n = 1, 2, . . . , N. Using (2.2) and its derivative, x˙n = −4 0=
N π2 π2 X π(xk − xn ) , + 12 cosech2 2 2 D D k=1,k6=n D
N X k=1,k6=n
cosech2
π(xk − xn ) π(xk − xn ) coth , D D
(2.11a)
(2.11b)
for n = 1, 2, . . . , N. Hence the dynamics of the poles xn (t) is determined by (2.11a). This dynamics is constrained by the equations (2.11b). These constraint equations (2.11b) are invariant under the flow of (2.11a). This follows from a direct calculation. Remarks • Since the KdV equation has two-soliton solutions for any ratio of the wavenumbers k1 /k2 6= 1, the constraint (2.11b) is solvable for any value of N, excluding N = 2, which can only be obtained by p = q = 1, resulting in equal wavenumbers k1 and k2 . • In particular, it follows that for almost all times (i.e., all times except collision times) the minimum number of poles in a fundamental strip required to obtain a g-soliton solution is N = 1 + 2 + 3 + · · · + g = g(g + 1)/2, corresponding to a g-soliton solution with wavenumbers which are related as (k1 : k2 : . . . : kg−1 : kg ) = (g : g − 1 : . . . : 2 : 1). • Equating k1 = 0, p = 0, q = 1, one obtains from (2.11a) x˙1 = −4k22 , corresponding to the dynamics of the one-soliton solution. The asymptotic behavior of the poles of a two-soliton solution also follows from (2.11a): from the separation of the poles into distinct vertical lines, it follows from (1.4a) that the velocity of these vertical lines is given by the one-soliton velocity for
58
BERNARD DECONINCK AND HARVEY SEGUR
each line, as expected. This result follows from easy algebraic manipulation and the identity p−1 p2 − 1 X 2 nπ , (2.12) cosec = 3 p n=1 valid for any integer p > 1. • A full analysis of the interaction of the poles for the case of any two-soliton solution with k1 /k2 = p/q is given in [17].
3. The Elliptic Case Consider the quasiperiodic finite-gap solutions of the KdV equation with g phases [14] u(x, t) = 2∂x2 ln θg (kx + ωt + φ|B), where θg (z|B) =
X
exp
1 2
(3.1)
m · B · m + im · z ,
(3.2)
m∈Z g
a hyperelliptic Riemann theta function of genus g. The g × g real Riemann matrix (i.e., symmetric and negative definite) B originates from a hyperelliptic Riemann surface with only one point at infinity. Furthermore, k, ω and φ are g-dimensional vectors. The derivation of equations (2.9), (2.11a) and (2.11b) is easily generalized to the case where the solution is periodic not only in the imaginary x-direction, but also in the real x-direction: u(x + L1 , t) = u(x, t) = u(x + iL2 , t).
(3.3)
This divides the complex x-plane into an array of rectangular domains, each of size L1 × L2 . One of these domains, called the fundamental domain S, is conveniently placed in the lower left corner of the first quadrant of the x-plane. The theta function has the property [6] 1 θg (z + iB · M + 2π N|B) = θg (z|B) exp − M · B · M + iM · z , (3.4) 2
for any pair of g-component integer vectors M, N. This expression is useful to determine conditions on the wavevector k and on the Riemann matrix B in order for u(x, t), given by (3.1), to satisfy (3.3): ∃N0 , M0 ∈ Z g : kL1 = 2π N0 , kL2 = B · M0 .
(3.5)
POLE DYNAMICS FOR ELLIPTIC SOLUTIONS OF THE KdV EQUATION
59
These results are now used to determine the number of poles N of u(x, t) in the fundamental domain. The poles of (3.1) are given by the zeros of ϑ(x, t) = θg (kx+ ωt + φ|B), regarded as a function of x: I 1 N= d ln ϑ(x, t) 2π i ∂S Z L1 Z iL2 1 d ln ϑ(x, t) + d ln ϑ(x + L1 , t)− = 2π i 0 0 ! Z L1 Z iL2 − d ln ϑ(x + iL2 , t) − d ln ϑ(x, t) 0
Z
0 L1
1 ϑ(x, t) d ln 2π i 0 ϑ(x + iL2 , t) Z L1 using (3.4) 1 = d(−ixM0 · k) 2π i 0 L1 = −M0 · N0 = − M0 · B · M0 . 2π L2 using (3.4)
=
(3.6)
The first equality of (3.6) confirms that N is an integer. The second equality shows that N is positive, by the negative-definiteness of B. We now proceed to determine the dynamical system satisfied by the motion of the N poles of u(x, t) in the fundamental domain S. Again, the poles of u(x, t) are the zeros of ϑ(x, t). Furthermore, zeros of ϑ(x, t) result in double poles of u(x, t), as in the hyperbolic case. The Weierstrass Factorization theorem [5] gives the following form for ϑ(x, t): x x2 Y x x + 2 cx 2 /2 1− e k 2xk , (3.7) ϑ(x, t) = e xk k where the product runs over all poles xk . The additional exponential factors, as compared to (2.6), are required because the poles now appear in a bi-infinite sequence: both in the vertical and horizontal directions. These exponential factors ensure uniform convergence of the product. The parameter c is allowed to depend on time. It determines the behavior of ϑ(x, t) as x approaches infinity in the complex x-plane [6]. Using (3.3), this is rewritten as ∞ N ∞ Y Y Y x 2 1− × ϑ(x, t) = exp(cx /2) xn + mL1 + ilL2 n=1 m=−∞ l=−∞ x2 x × exp . (3.8) + xn + mL1 + ilL2 2(xn + mL1 + ilL2 )2 The first product runs over the number of poles (N) in the fundamental domain, the second and third products result in all translations of the fundamental domain.
60
BERNARD DECONINCK AND HARVEY SEGUR
From the uniform convergence of (3.8), u(x, t) = 2c − 2
∞ N ∞ X X X n=1 m=−∞ l=−∞
1 − (x − xn − mL1 − ilL2 )2 1 − . (xn + mL1 + ilL2 )2
(3.9)
Using the definition of the Weierstrass function (1.3), this is rewritten as u(x, t) = 2c − 2
N X
℘ (x − xn ) + 2
n=1
N X
℘ (xn ),
(3.10)
j =1
where the periods of the Weierstrass function are given by 2ω1 = L1 , 2ω2 = iL2 . Define c˜ = 2c + 2
N X
℘ (xn ).
(3.11)
n=1
The dynamics of the poles xn = xn (t) is determined by substitution of (3.10) or (3.9) into the KdV equation and expanding in powers of for x near a pole: x = xk + . Equating the coefficients of −3 , −2 and 0 to zero result in N X
x˙n = 12
℘ (xj − xn ),
(3.12a)
j =1,j 6=n N X
0=
℘ 0 (xj − xn ),
(3.12b)
j =1,j 6=n
,c˙˜ = 0 ⇐⇒ c(t) ˜ =α=0
(3.12c)
for n = 1, 2, . . . , N. (The constant α can always be removed by a Galilean shift, so it is equated to zero, without loss of generality.) The constraints (3.12b) are invariant under the flow, as can be checked by direct calculation. Notice that (3.12a–b) are identical to the equations obtained by Airault, McKean and Moser [2]. These equations are obtained here in greater generality: any solution (3.1) that is doubly periodic in the x-plane gives rise to a system (3.12a–b). This allows us to reach the conclusions stated in the next section. Remarks • In the limit L1 → ∞, Equations (3.12a–b) reduce to (2.11a–b). This limit is most conveniently obtained from the Poisson representation of the
POLE DYNAMICS FOR ELLIPTIC SOLUTIONS OF THE KdV EQUATION
61
Weierstrass function: ℘ (x) =
2
πx 1 + + cosech2 3 L2 ! ∞ X 2 π 2 nπ L1 cosech . (3.13) + (x + nL1 ) − cosech L L 2 2 n=−∞,n6=0 π L2
This representation is obtained from (1.3) by working out the summation in the vertical direction. It gives the Weierstrass function as a sum of exponentially localized terms, hence few terms have important contributions in the fundamental domain. A Poisson expansion for ℘ 0 (x) is obtained from differentiating (3.13) term by term with respect to x. • Define the one-phase theta function θ1 (z, q) [12]: ∞ X 2 θ1 (z, q) = 2 (−1)n q (n+1/2) sin(2n + 1)z,
(3.14)
n=0
with |q| < 1. If L2 < L1 , then the relationship ℘ (z) = a − ∂x2 ln θ1 (π z/L1, iL2 /L1 ), with a a constant [12], allows us to rewrite (1.2) as u(x, t) = aˆ +
2∂x2
N Y L2 π , ln θ1 (x − xj (t)), i L1 L1 j =1
(3.15)
with aˆ = −2aN. Hence, for the doubly-periodic solutions of the KdV equation of the form (3.1), it is possible to rewrite the g-phase theta function as a product of N 1-phase theta functions, with nonlinearly interacting phases. Note that this does not imply that the g-phase theta function appearing in (3.1) is reducible. Reducible theta functions do not give rise to solutions of the KdV equation [6]. • By taking another time derivative of (3.12a) and using (3.12b), one obtains x¨n = −(12)
2
N X
℘ 0 (xj − xn )℘ (xj − xn ).
(3.16)
j =1,j 6=n
It is known that this system of differential equations is Hamiltonian [4], with Hamiltonian
62
BERNARD DECONINCK AND HARVEY SEGUR N N N 1 X 2 (12)2 X X H= p + ℘ 2 (xk − xj ), 2 k=1 k 2 k=1 j =1,j 6=k
(3.17)
and canonical variables {xk , pk = x˙k }. A second Hamiltonian structure for the equations (3.16) is given in [4]. A Lax representation for the system (3.12a–c) is also given there. This Lax representation is a direct consequence of the law of addition of the Weierstrass function [12]. It is unknown to us whether a Hamiltonian structure exists for the constrained first-order dynamical system (3.12a–b). The Hamiltonian structure (3.17) shows that the system (3.12a–b) is a (constrained) member of the elliptic Calogero–Moser hierarchy [3].
4. Discussion of the Dynamics In this section, the constrained dynamical system (3.12a–c) is discussed. In particular, the assertions made in the introduction are validated here. For reality of the KdV solution (3.1) when x is restricted to the real line, it is necessary and sufficient that if xj (t) appears, then so does xj∗ (t). Because the Weierstrass function is a meromorphic function of its argument that is real valued on the real line, this reality constraint is invariant under the dynamics (3.12a). As a consequence, the distribution of the poles in the fundamental domain S is symmetric with respect to the horizontal centerline of S. Poles are allowed on the centerline. Most of what follows is valid for both real KdV solutions? and KdV solutions that are not real, but we restrict our attention to real KdV solutions. 4.1.
ALL FINITE - GAP ELLIPTIC SOLUTIONS OF THE KDV EQUATION ARE OF THE FORM (1.2) , UP TO A CONSTANT
A straightforward singularity analysis of the KdV equation [2] shows that any algebraic singularity of a solution of the KdV equation is of the type u(x, t) = −2/(x − α(t))2 + O(x − α), for almost all times t. At isolated times tc , the leading order coefficient is not necessarily −2. It can be of the form −g(g +1) (see below), but the exponent of the leading term is always −2. Hence, an elliptic function ansatz for u(x, t) can only have second-order poles and with the substitution u(x, t) = 2∂x2 ln ϑ(x, t) gives rise to a Weierstrass expansion of the form (3.8), with an arbitrary prefactor exp(c(x, t)), for an arbitrary function c(x, t), entire in x. Substitution of this ansatz in the KdV equation then determines that cxx is doubly periodic and meromorphic in x. The only such c is a constant. Hence, all finite-gap elliptic solutions of the KdV equation are of the form (1.2). ? ‘Real KdV solution’ refers to a solution of the KdV equation which is real when x is restricted to the real x-axis.
POLE DYNAMICS FOR ELLIPTIC SOLUTIONS OF THE KdV EQUATION
4.2.
63
IF L1 /L2 1, ANY NUMBER OF POLES N 6= 2 IN THE FUNDAMENTAL DOMAIN IS ALLOWED
This immediately follows from the soliton case, which is a special case of the elliptic case. Thickstun’s results [17] already implied that any value of N 6= 2 occurs. We do not provide a direct proof for finite values of L1 . On the other hand, the numerical evidence presented in Section 5 supports this statement. We have already argued that the equations (2.9) and (2.11a–b) are obtained from (1.2) and (3.12a–b) in the limit L1 → ∞. On the other hand, (3.13) allows us to rewrite each term in (1.2) as an infinite sum of solitons, each of which is localized in a different real-period interval of the solution. In the limit L1 → ∞, only two terms in each one of these series remain: the constant term and a onesoliton term. This gives rise to the soliton limit of the solution (2.9). The terms that vanish as L1 → ∞ are then regarded as a deformation of the soliton limit. Similar deformations are valid at the level of the dynamical system and the constraints. The success of the numerical method prompts us to formulate the following CONJECTURE. Every nonsingular soliton solution that is periodic in ix with period iL2 is obtained as the limit in which L1 → ∞ of an elliptic solution with periods iL2 and L1 . At this point, it is appropriate to remark that if one is interested in elliptic solutions of the Kadomtsev–Petviashvili (KP) equation, ∂x (−ut + 6uux + uxxx ) + 3σ 2 uyy = 0
(4.1)
(with parameter σ ), then the Equations (1.2), (3.12a–b) are replaced by [4] u(x, y, t) = −2
N X
℘ (x − xn (y, t)),
(4.2a)
n=1
2 N X ∂xn 2 ∂xn = 3σ + 12 ℘ (xj − xn ), ∂t ∂y j =1,j 6=n σ2
N X ∂ 2 xn = −16 ℘ 0 (xj − xn ). ∂y 2 j =1,j 6=n
(4.2b)
(4.2c)
This clarifies the appearance of the constraint (3.12b) on the motion of the poles of elliptic solutions of the KdV equation, where the poles are independent of y. For y-independent solutions, the KP equation reduces to the KdV equation, and Equations (4.2a–c) reduce to (1.2), (3.12a–b), forcing the poles to remain on the invariant manifold defined by (3.12b). For the KP equation, no such constraint exists and the number of poles N in the fundamental domain can be any integer, not equal to two.
64 4.3.
BERNARD DECONINCK AND HARVEY SEGUR
FOR ANY N > 4, NONEQUIVALENT CONFIGURATIONS EXIST, FOR L1 /L2 SUFFICIENTLY LARGE
Consider the asymptotic behavior for t → −∞ of limL1 →∞ u(x, t). In this soliton limit, as t → −∞, the poles are collected in groups corresponding to one-soliton solutions. In this section, two configurations are called nonequivalent if the above asymptotic behavior results in two different groupings of the poles. For N = 3, all configurations are equivalent to one configuration. In the limit L1 → ∞, this configuration corresponds to the two-soliton case with k1 : k2 = 2 : 1. This configuration is discussed in Section 5.1. For N = 4, all configurations are again equivalent to one configuration. This configuration corresponds to the two-soliton case with k1 : k2 = 3 : 1. Recall that k1 and k2 are not allowed to be equal, hence a configuration with two poles to the left and two poles to the right does not exist. Another way of expressing that only one configuration exists is that N = 4 can only be decomposed in one way as the sum of distinct positive integers without common factor (> 1), namely as N = 3 + 1. Again, all N = 4 configurations are equivalent. This configuration is discussed in Section 5.2. That section also discusses another example of an N = 4 potential which does not have a nonsingular soliton limit. This potential is a special case of one of the Treibich–Verdier potentials (1.8). Any integer N > 4 can be written as a sum of distinct positive integers without number of terms in the m-th overall common factor in more than one way.? Let theP m decomposition of N be denoted as Nm , then N = N i=1 ni , with the ni distinct and having no overall common factor. This configuration corresponds to the Nm soliton case with wavenumber ratios (k1 : k2 : . . . : kNm ) = (n1 : n2 : . . . : nNm ). A solution with these wave numbers has Nm phases and is an Nm -soliton solution. Hence for any N > 4 there exist at least as many different configurations as there are decompositions of N into distinct positive integers, without overall common factor. These configurations need not have the same number of phases. Two nonequivalent configurations corresponding to N = 5 are discussed in Section 5.3. 4.4.
THE POLES ONLY COLLIDE IN TRIANGULAR NUMBERS
A collision of poles is a local process in which only the colliding poles play a significant role. The analysis of the collisions is identical to that of the rational and the soliton cases because close to the collision point, the Weierstrass function reduces to 1/x 2 . Kruskal [15] already noticed that the poles do not collide in pairs, but triple collisions do occur. In fact, any triangular number of poles can participate in a collision, in which case the solution of the KdV equation at the collision time tc , nearby the collision point xc is given by u(x, tc ) = −g(g + 1)/(x − xc )2 . ? 5 = 1 + 4 = 2 + 3, N = 1 + (N − 1) = 1 + 2 + (N − 3), for N > 5.
POLE DYNAMICS FOR ELLIPTIC SOLUTIONS OF THE KdV EQUATION
65
Asymtotically near the collision point xc , before the collision, i.e., t < tc , the poles lie on the vertices of a regular polygon [2] with g(g + 1)/2 vertices. For t > tc , the poles emanate from the collision points, again forming a regular polygon with g(g + 1)/2 vertices [2]. If g(g + 1)/2 is even, this polygon is identical to the polygon before the collision. If g(g + 1)/2 is odd, the polygon is rotated around the collision point by 2π/g(g + 1) radians. Of all these collision types, the one where three poles collide (corresponding to g = 2) is generic. It is the one observed in the examples illustrated in Section 5. Since the poles only collide in triangular numbers, it is possible that at any given (1.9), with not all gi = 1. time tc the solution of the KdV equation has the form P At almost every other time t, such a solution has N = M i=1 gi (gi + 1)/2 distinct poles. 4.5.
THE SOLUTIONS (1.2) ARE FINITE - GAP POTENTIALS OF THE STATIONARY SCHRÖDINGER EQUATION (1.6)
By construction the solutions (1.2) are periodic in x because they are obtained as a Weierstrass factorization of the theta function appearing in (3.1), upon which we have imposed the double periodicity. Hence the solutions (1.2) are finite-gap potentials of the Schrödinger operator. In [11], another proof of this can be found. For solutions that are elliptic deformations of the nonsingular solitons of Section 2, more can be said: an elliptic deformation of a g-soliton solution is a ggap potential of the Schrödinger equation. The reasoning is as follows: we already know that any elliptic deformation results in a finite-gap potential of the Schrödinger equation. On the other hand, any finite gap potential of the Schrödinger equation is of the form (3.1). The soliton limit of such a finite-gap solution with g-phases is a g-soliton solution [3]. Hence the number of phases of an elliptic deformation of a g-soliton solution is equal to g. This limit is the soliton limit of the periodic solutions, in which the fundamental domain reduces to the fundamental strip. In order to have a g-soliton solution of the KdV equation, we remarked in Section 2 that at least N = g(g + 1)/2 = 1+2+· · ·+g poles are required in the fundamental strip. Hence, this many poles are required in the fundamental domain to obtain a g-gap potential of the Schrödinger equation that is an elliptic deformation of a nonsingular soliton solution. 5. Examples In this section, some explicit examples of elliptic solutions of the KdV equation are discussed. These are illustrated with figures displaying the motion of the poles in the fundamental domain. Other figures display the solution of the KdV equation u(x, t) as a function of x and t. All these figures were obtained from numerical solutions of the corresponding constrained dynamical system. In all cases, the constrained dynamical system was solved using a projection method: the dynamical
66
BERNARD DECONINCK AND HARVEY SEGUR
system (3.12a) is used to evolve the system for some time. Subsequently, the new solution is projected onto the constraints (3.12b) to correct numerical errors, after which the process repeats. In all examples given, L1 = 4 and L2 = π . This seems to indicate that one can wander far away from L1 /L2 1 and still obtain soliton-like elliptic solutions of the KdV equation. This is not surprising as (3.13) indicates that as perturbation parameter on the soliton case one should use = exp(−2π L1 /L2 ). For the values given above, this gives = 0.00034.
5.1.
THE SOLUTION OF DUBROVIN AND NOVIKOV [7]: N = 3
Dubrovin and Novikov [7] integrated the KdV equation with the Lamé–Ince potential u(x, 0) = −6℘ (x − xc ) as initial condition. They found the solution to be elliptic for all time, with N = 3. They gave explicit formulae for the solution, which they remarked was probably the simplest two-gap solution of the KdV equation. The dynamics of the poles in the fundamental domain is displayed in Figure 2a. Figure 2b displays the corresponding two-phase solution of the KdV equation. Animations of the behavior of the poles and of u(x, t) as t changes are also available at http://amath-www.colorado.edu/appm/other/kp/papers. Notice the solitonlike interactions of the two phases in the solution. In terms of the classification of Lax [16], these are interactions of type (c) (i.e., u(x, t) has only one maximum while the larger wave overtakes the smaller wave). From Figure 2a, it appears that the Dubrovin–Novikov solution is periodic in time. This was indeed proven by Ènol’skii [8]. For this specific solution only one of the three constraint equations is independent: since the derivative of the Weierstrass function is odd, the sum of the constraints is zero. Furthermore, labelling the three poles by x1 , x2 and x3 , for reality x2 = x1∗ + iL2 and x3 is on the centerline. Hence, the second constraint is the complex conjugate of the first constraint. The constraints (3.12b) reduce to the single equation ℘ 0 (x1 − x1∗ ) + ℘ 0 (x1 − x3 ) = 0.
(5.1)
This equation was solved numerically to provide the initial condition shown in Figure 2a. The initial guess required for the application of Newton’s method is based on the knowledge of the soliton limit. In that case two poles on the right represent a faster soliton, one pole on the left represents the slower soliton. The periodic case is not that different: the vertical line of poles with the smallest vertical distance between poles has poles closer to the real x-axis than the others and correspond to the wave crest with the highest amplitude, as seen in Figure 2b. We refer to the Dubrovin–Novikov solution as a (2 : 1)-solution because of the natural separation of the poles in a group of 2 poles (x1 , x2 ) and a single pole (x3 ).
POLE DYNAMICS FOR ELLIPTIC SOLUTIONS OF THE KdV EQUATION
67
(a)
(b) Figure 2. The solution of Dubrovin and Novikov, with L1 = 4 and L2 = π. (a) The motion of the poles in the fundamental domain. The initial position of the poles is indicated by the black dots. The arrows denote the motion of the poles. (b) The KdV solution u(x, t).
Equating x1 = x3 + and only considering the singular terms of (5.1), it is possible to examine the location of the poles close to a collision points xc . With ℘ 0 (x) = −2/x 3 in this limit and = r + ii , one finds r3 − 3i2 r = 0,
9i3 − 3r2 i = 0.
(5.2)
This set of equations has three √ solutions, √ corresponding to the three distances between the poles: r /i ∈ {0, 3, − 3}. This allows for two triangular configurations of the poles: an equilateral triangle pointing left of the collision point and one pointing right.
68
BERNARD DECONINCK AND HARVEY SEGUR
Using the dynamical system (3.12a) in the same way and only retaining singular terms results in ˙r = −
3 r2 − i2 + 3 2 , 4i2 r2 + i2
˙i =
6r i r2 + i2
2 .
(5.3)
Since the constraints (3.12b) are invariant under the flow (3.12a), the solutions to (5.2) give invariant directions of the system (5.3). Along these invariant directions, one obtains ordinary differential equations for the motion of the poles as they approach the collision point. It follows from these equations that the poles approach the collision point xc with infinite velocity. Integrating the equations with initial condition (tc ) = 0 gives =
35/6 √ ( 3 + i)(tc − t)1/3 . 2
(5.4)
Using the three branches of (tc − t)1/3 results in the dynamics of each edge of the equilateral triangle. If t < tc this triangle is pointing left, for t > tc it is pointing right. 5.2. N = 4:
AN ELLIPTIC DEFORMATION AND A TREIBICH – VERDIER SOLUTION
The next solution we discuss has 4 poles in the fundamental domain and is an elliptic deformation of a soliton solution. In the limit L1 → ∞, this solution corresponds to a two-soliton solution with wavenumber ratio k1 /k2 = 3/1, so this solution is refered to as a (3 : 1)-solution. The motion of the poles in the fundamental domain is displayed in Figure 3. Corresponding to the given wavenumber ratio, the amplitude ratio of the two phases present in the solution is roughly k12 /k22 = 9/1. As a consequence, the form of u(x, t) is not very illuminating and it has been omitted. Animations with the time dependence of both the positions of the poles and of u(x, t) are again available at http://amath-www.colorado.edu/appm/other/kp/papers. Note that the poles of the (3 : 1)-solution do not collide. This is in agreement with the results of Thickstun [17] who outlined which configurations lead to collisions and which do not, in the hyperbolic case. As mentioned before, the examination of collision behavior is essentially local and no differences appear among the rational, hyperbolic and elliptic cases. Another configuration with N = 4 exists. Consider the potential u(x, t = tc ) = −2℘ (x − x0 ) − 6℘ (x − x0 − ω1 ),
(5.5)
with x0 on the centerline. This is a Treibich–Verdier potential, obtained from (1.9) with M = 2, g1 = 1, g2 = 2, α1 = x0 and α2 = x0 + ω1 . It is referred to as a Treibich–Verdier potential because the position of the poles is given in terms of the
POLE DYNAMICS FOR ELLIPTIC SOLUTIONS OF THE KdV EQUATION
69
Figure 3. N = 4: the pole dynamics of a (3 : 1)-solution. The initial positions of the four poles are indicated. The arrows on the centerline indicate that the poles there move in both directions.
periods of the Weierstrass function, as Pin (1.8). Also, it can be obtained from (1.8) as a degenerate case. As before N = 2i=1 gi (gi +1)/2 = 4, hence for all times that are not collision times, this solution has 4 distinct poles in the fundamental domain. The time t = tc is a collision time. Immediately after the collision time t = tc , the 3 poles located at x0 + ω1 separate, as in the Dubrovin–Novikov solution, along an equilateral triangle. The result appears to be a three-phase solution. However, it is known that the potential (5.5) is a two-gap potential of the Schrödinger equation and its hyperelliptic Riemann surface is given explicitly in [3]. This solution is not an elliptic deformation of a nonsingular soliton solution and the separation into different phases does not make sense. This is also seen from the following argument: if, for a fixed time which is not a collision time, we attempt to take the limit as L1 = 2ω1 → ∞, the poles seem to separate in three distinct solitons with respective wave numbers (k1 : k2 : k3 ) = (2 : 1 : 1). Such a nonsingular soliton solution does not exist for the KdV equation and the separation into different phases does not make sense. The dynamics of the poles is illustrated in Figure 4a. The corresponding KdV solution is shown in Figure 4b. The dynamics of the poles illustrated in Figure 4a exhibits behavior that appears qualitatively different from any other solution discussed here. The trajectories traced out by the motion of the poles in the fundamental domain appear to have singular points (cusps), away from the collision points. Upon closer investigation, these ‘cusps’ are only a figment of the resolution of the plot and the poles trace out a regular curve as a function of time, away from the collision times. Exactly why the global pole dynamics of the Treibich–Verdier potential (5.5) under the KdV flow appears so different from the pole dynamics of elliptic deformations of soliton solutions of the KdV equation is an open problem. Another question one may ask is whether similar behavior is observed for other solutions originating from Treibich–Verdier potentials.
70
BERNARD DECONINCK AND HARVEY SEGUR
(a)
(b) Figure 4. An N = 4, M = 2 Treibich–Verdier solution, with L1 = 4 and L2 = π. (a) The motion of the poles in the fundamental domain. The initial position of the poles is indicated by the black dots. The initial time t = 0 was chosen different from the collision times tc . The arrows denote the motion of the poles. (b) The KdV solution u(x, t).
5.3. N = 5:
TWO DIFFERENT POSSIBILITIES
For N = 5, two soliton configurations are possible, and corresponding to each of these is an elliptic solution. The first solution is a (4 : 1)-solution. The second solution is a (3 : 2)-solution. The (4 : 1)-solution offers no new pole-dynamics: initially 1 pole is located on the centerline, at the left in the fundamental domain. The other poles are located at the right of the fundamental domain, symmetric with respect to the centerline. The three poles closest to the centerline interact as the (2 : 1)-solution. The two outer
POLE DYNAMICS FOR ELLIPTIC SOLUTIONS OF THE KdV EQUATION
71
(a)
(b)
(c) Figure 5. N = 5: (a) The pole dynamics of a (4 : 1)-solution in the fundamental domain. (b) The pole dynamics of a (3 : 2)-solution in the fundamental domain. (c) The KdV solution u(x, t) corresponding to the pole dynamics in (b). In (a) and (b), the initial positions of the poles are indicated. Both solutions are quasiperiodic in time.
72
BERNARD DECONINCK AND HARVEY SEGUR
Figure 6. The pole dynamics of a (3 : 2 : 1)-solution. The black dots mark the initial position of the poles; the grey dots mark the position of the poles at t = 0.4.
poles behave as the two outer poles of the (3 : 1)-solution. The pole dynamics of the (4 : 1)-solution is displayed in Figure 5a. The (3 : 2)-solution is more interesting. It is displayed in Figure 5c, together with the motion of the poles in the fundamental domain 5b. Again, the two crests of u(x, t) interact in a soliton-like manner. In Lax’s classification [16], this is an interaction of type (a), where at every time two maxima are observed. Figure 5b only displays the motion of the poles for a short time, in order not to clutter the picture. The motion of the poles is presumably quasiperiodic in time, as is the case for the (4 : 1)-solution. It appears that the two poles above (or below) the middle line of the fundamental domain share a common trajectory. It is an open problem to establish whether or not this is the case.
5.4. N = 6: TWO
DIFFERENT POSSIBILITIES . A THREE - PHASE SOLUTION
For N = 6, two distinct pole configurations are possible. The first one corresponds to a (5 : 1)-solution and results in a two-gap potential of the Schrödinger equation. It essentially behaves as the (3 : 1)-solution with two more poles added, which also behave as the outer poles of the (3 : 1)-solution. The second configuration is a (3 : 2 : 1)-solution, which limits to a three-soliton solution with wavenumber ratio (k1 : k2 : k3 ) = (3 : 2 : 1). This elliptic solution is a three-phase solution of the KdV equation. The amplitude ratio of the (3 : 2 : 1)-solution is (9 : 4 : 1), which explains why the third phase is hard to notice in Figure 7. Animations of the pole dynamics and of the time dependence of the (3 : 2 : 1)-solution are available at http://amath-www.colorado.edu/appm/other/kp/papers.
POLE DYNAMICS FOR ELLIPTIC SOLUTIONS OF THE KdV EQUATION
73
Figure 7. A (3 : 2 : 1)-solution of the KdV equation, corresponding to the pole dynamics in Figure 6.
Acknowledgements The authors acknowledge useful discussions with B. A. Dubrovin, S. P. Novikov, C. Schober, A. Treibich and A. P. Veselov. This work was carried out at the University of Colorado and the Mathematical Sciences Research Institute. It was supported in part by NSF grants DMS 9731097 and DMS-9701755.
References 1. 2.
3.
4. 5.
Ablowitz, M. J. and Segur, H.: Solitons and the Inverse Scattering Transform, SIAM, Philadelphia, PA, 1981. Airault, H., McKean, H. P., and Moser, J.: Rational and elliptic solutions of the Korteweg– de Vries equation and a related many-body problem, Comm. Pure Appl. Math. 30(1) (1977), 95–148. Belokolos, E. D., Bobenko, A. I., Enol’skii, V. Z., Its, A. R., and Matveev, V. B.: AlgebroGeometric Approach to Nonlinear Integrable Problems, Springer Ser. Nonlinear Dynam., Springer-Verlag, Berlin, 1994. Chudnovs’ki, D. V. and Chudnovs’ki, G. V.: Pole expansions of nonlinear partial differential equations, Nuovo Cimento B (11) 40(2) (1977), 339–353. Conway, J. B.: Functions of One Complex Variable, 2nd edn, Springer-Verlag, New York, 1978.
74 6.
BERNARD DECONINCK AND HARVEY SEGUR
Dubrovin, B. A.: Theta functions and nonlinear equations, Russian Math. Surveys 36(2) (1981), 11–80. 7. Dubrovin, B. A. and Novikov, S. P.: Periodic and conditionally periodic analogs of the manysoliton solutions of the Korteweg–de Vries equation, Soviet Phys. JETP 40 (1975), 1058–1063. 8. Ènols’kii, V. Z.: On solutions in elliptic functions of integrable nonlinear equations associated with two-zone Lamé potentials, Soviet Math. Dokl. 30 (1984), 394–397. 9. Gardner, C. S., Greene, J. M., Kruskal, M. D., and Miura, R. M.: Method for solving the Korteweg–de Vries equation, Phys. Rev. Lett. 19 (1967), 1095–1097. 10. Gesztesy, F. and Weikard, R.: On Picard potentials, Differential Integral Equations 8(6) (1995), 1453–1476. 11. Gesztesy, F. and Weikard, R.: Picard potentials and Hill’s equation on a torus, Acta Math. 176(1) (1996), 73–107. 12. Gradshteyn, I. S. and Ryzhik, I. M.: Table of Integrals, Series, and Products, 5th edn, Translation edited and with a preface by Alan Jeffrey, Academic Press, Boston, MA, 1994. 13. Ince, E. L.: Further investigations into the periodic Lamé functions, Proc. Roy. Soc. Edinburgh 60 (1940), 83–99. 14. Its, A. R. and Matveev, V. B.: Schrödinger operators with the finite-band spectrum and the Nsoliton solutions of the Korteweg–de Vries equation, Theoret. and Math. Phys. 23(1) (1976), 343–355. 15. Kruskal, M. D.: The Korteweg–de Vries equation and related evolution equations, Lectures in Appl. Math. 15, Amer. Math. Soc. Providence, 1974, pp. 61–83. 16. Lax, P. D.: Integrals of nonlinear equations of evolution and solitary waves, Comm. Pure Appl. Math. 21 (1968), 467–490. 17. Thickstun, W. R.: A system of particles equivalent to solitons, J. Math. Anal. Appl. 55(2) (1976), 335–346. 18. Treibich, A. and Verdier, J.-L.: Revêtements tangentiels et sommes de 4 nombres triangulaires, C.R. Acad. Sci. Paris Sér. I Math. 311(1) (1990), 51–54. 19. Treibich, A. and Verdier, J.-L.: Solitons elliptiques, In: The Grothendieck Festschrift, Vol. III, With an appendix by J. Oesterlé, Birkhäuser, Boston, 1990, pp. 437–480. 20. Treibich, A. and Verdier, J.-L.: Revêtements exceptionnels et sommes de 4 nombres triangulaires, Duke Math. J. 68(2) (1992), 217–236. 21. Verdier, J.-L.: New elliptic solitons, In: Algebraic Analysis, Vol. II, Academic Press, Boston, MA, 1988, pp. 901–910.
Mathematical Physics, Analysis and Geometry 3: 75–89, 2000. © 2000 Kluwer Academic Publishers. Printed in the Netherlands.
75
On Bianchi and Bäcklund Transformations of Two-Dimensional Surfaces in E 4 YU. AMINOV? and A. SYM Institute of Theoretical Physics of Warsaw University, Hoza street 69, 00-681 Warsaw, Poland (Received: 22 July 1999) Abstract. In this paper we discuss the existence and properties of the Bianchi transformations for pseudospherical surfaces in E 4 . The results of the paper show that the theory of Bianchi transformations in the discussed case is essentially different from the well-known case of pseudospherical surfaces in E 3 (in general n-manifolds of constant and negative curvature in E 2n−1 ). Mathematics Subject Classifications (2000): 53A05, 53A25, 37K35. Key words: Bianchi transformations, Bäcklund transformations, differential geometry, Gauss curvature, pseudospherical surface.
1. Historical Remarks These remarks serve as an introduction to the fascinating history of both Bianchi transformations and their generalization known as Bäcklund transformations. Simultaneously, we fix a terminology of the problem following the tradition of classical differential geometry. To begin, we recall the classical notion of an evolute of a surface S in E 3 [1]. Consider a family of all normals to S. In general, there exist two uniquely defined surfaces, say F and F¯ , such that each normal to S is tangent to both surfaces F and F¯ . Then F together with F¯ is an evolute of S. Luigi Bianchi in his 1879 habilitation thesis [2] introduced a notion of ‘complementary surfaces’. There are exactly two pieces F and F¯ of an evolute of some surface in E 3 . In this way, for Bianchi, the primary objects are F and F¯ , while S is a secondary one. Indeed, at the very beginning of his thesis he formulated the main problem as follows: “It is particularly interesting to consider complementary surfaces to a surface of constant negative curvatures”. In order to solve this problem, Bianchi made use of the following observation [1]: The tangents to curves of geodesic foliation of surface F are normal to an infinity of parallel surfaces. ? On leave of absence from B. I. Verkin Institute for Low Temperature of NAN of the Ukraine, 47 Lenin Ave, 310164, Kharkov, Ukraine.
76
YU. AMINOV AND A. SYM
Let F be a pseudospherical surface of Gaussian curvature K = −b2 = const. F can be always equipped with coordinates (x, y) such that the metric form of F is equal to ds 2 = e2by dx 2 + dy 2 .
(1)
The system (x, y) is called an ‘horocyclic system’ since the curves y = const are horocycles. Simultaneously, the lines x = const are geodesics. Consider now a set of all tangents to the curves of the geodesic foliation of F defined by geodesics x = const. Relying on the earlier work of Weingarten [3], Bianchi pointed out (without proof) that the complementary surface F¯ is also pseudospherical of the same Gaussian curvature K¯ = −b2 . The corresponding analytical expression is very simple. Let r = r(x, y) be a position vector to F . A position vector r¯ = r¯ (x, y) to F¯ is given by 1 r¯ = r − ry . b
(2)
In fact, (2) is the Example 4 in [1], p. 290. To summarize, one can speak of the well-defined transformation of the pseudospherical (K = −b2 ) surface F into the pseudospherical surface F¯ of the same curvature. After Darboux [4], we call F → F¯ the ‘Bianchi transformation’. In 1883, A. V. Bäcklund published a paper, [5], which contains a generalization of the Bianchi transformation. This is still a transformation between two pseudospherical surfaces of the same Gaussian curvature. This paper is an essentially analytical one. Bäcklund represented F (F¯ ) in the Monge form z = f (x, y)(z0 = f 0 (x 0 , y 0 )) and replaced the original Bianchi transformation by a system of four equations relating (x, y, z, p, q) and (x 0 , y 0 , z0 , p 0 , q 0 ), where p = zx , q = zy , etc. Next, he formulated a generalization of this system containing a real parameter which defines an angle between normals to the corresponding surfaces at the corresponding points. Within the theory of pseudospherical surfaces in E 3 , this generalization is called the ‘Bäcklund transformation’. Bäcklund paper [5] is based on his earlier paper [6]. In any case, the Bäcklund transformation of [5] gave rise to further developments which were finally completed by the French school of J. Clairin and E. Goursat as a theory of Bäcklund transformations [7]. In a remarkable paper, Bianchi [8] reformulated the Bäcklund transformation in a purely geometric setting. In the same paper, he pointed out, for the first time the possibility, of interpreting Bäcklund’s results within Kummer’s theory of congruences. We recall that according to the classical terminology, a congruence in E 3 is a 2-parameter family of straight lines. Indeed, the Bianchi transformation, defines, for instance, a congruence: this is a 2-parametric family of tangent lines defined by ry . The important point is that F and F¯ can be now interpreted as the so-called focal surfaces of the congruence. This idea is fully developed in [9] published in 1887. For some other historical and modern aspects of this topic, see [10].
ON BIANCHI AND BÄCKLUND TRANSFORMATIONS
77
2. Modern Developments Surprisingly enough, the topic of Bäcklund transformations has been out fashion for many years. A strong resurgence in interest in them has been evident since the creation of the theory of solitons (the theory of completely integrable systems of partial differential equations) [11]. In certain respects, a series of papers by G. L. Lamb [12, 13] and H. Wahlquist and F. Estabrook [14, 15] is a direct continuation of the work produced by the school of Clairin and Goursat. We have already pointed out that the Bianchi transformation (‘complementary transformation’ in Bianchi’s terminology) is a particular case of the Bäcklund transformation for pseudospherical surfaces corresponding to the special choice of Bäcklund parameters. This is why – despite its geometrical beauty – not much of attention was paid to this transformation in the past. There is, for instance, a discussion of Bianchi transformations by G. Darboux [4]. A generalization of the Bianchi transformation (with replacements: ‘pseudospherical surface’ by ‘an isometric immersion of the domain of the n-dimensional Lobachevski space’ and E 3 by E 2n−1 ) was formulated and proved in 1978 [16], almost one hundred years after Bianchi [2]. On the other hand, a similar generalization of the Bäcklund transformation for pseudospherical surfaces was presented by Tenenblat and Terng in 1979 [18], see also [19]. Their approach is based on Bianchi’s formulation of the Bäcklund transformations in terms of congruences. Paper [16] initiated further investigations concerning the generalized Bianchi transformations [17]. Similarly, papers [18, 19] are a point of departure for further studies [20]. 3. The Questions To the best of our knowledge, investigation of the Bäcklund (Bianchi) transformations for isometric immersions of n-dimensional Lobachevski space into E N with N > 2n − 1 has never been undertaken. In this paper we attempt to answer some questions related to the Bianchi transformations for isometric immersions of domains of Lobachevski plane L2 into E 4 . We start with L2 equipped with the horocyclic system of coordinates (x, y). The metric of L2 in these coordinates is given by ds 2 = e2y dx 2 + dy 2 .
(3)
Note that (3) is in fact (1) with b = 1, (K = −1). Coordinate curves x = const are geodesics while curves y = const are of the geodesic curvature equal to 1. F 2 denotes some isometric immersion of the domain of L2 into E 4 . r = r(x, y) stands for a position vector to F 2 . Guided by formula (2), we put r¯ = r − ry . A surface in E 4 defined by its position vector r¯ is called a Bianchi transform of F 2 and is denoted by F¯ 2 while a map F 2 → F¯ 2 is called a Bianchi transformation in ¯ E 4 . The Gauss curvature of F¯ 2 is denoted by K.
78
YU. AMINOV AND A. SYM
A few questions can be asked here. QUESTION 1. When, in the definition above, E 4 is replaced by E 3 , we are assured that F¯ 2 is a pseudospherical surface as well. Does the same hold in the discussed case (K¯ = −1?)? QUESTION 2. Is the vector ry tangent to F¯ 2 as well? Soon we shall see that the answers to the above-mentioned questions are negative. Hence, one can weakes these two questions. QUESTION 3. Does a pseudospherical surface F 2 in E 4 exist (the case F 2 ⊂ E 3 is excluded) for which F¯ 2 is pseudospherical as well? QUESTION 4. Does a pseudospherical surface F 2 in E 4 exist such that the vector ry is tangent to F¯ 2 as well? In this case, we say that the vector field ry on F 2 enjoys a bi-tangency property. For simplicity, we also say that F 2 enjoys a bi-tangency property. QUESTION 5. exist in E 4 ?
How many pseudospherical surfaces with bi-tangency property
This question in turn inspires the following general question. QUESTION 6. Consider now an arbitrary surface F 2 in E 4 . Let a be a field of unit tangent vectors to F 2 . Let F¯ 2 be given by its position vector r¯ = r +ha, where h = h(u1 , u2 ) is a function on F 2 . The question is: under which conditions does the vector field ha exist on F 2 with the bi-tangency property? It is evident that this transformation r → r¯ is a generalization of the Bäcklund transformation. 4. The Answers The answer to Questions 1 and 2 is given by the following theorem: THEOREM 1. There exist pseudospherical surfaces F 2 in E 4 such that its Bianchi transformation has K¯ 6= −1. For surfaces F 2 ⊂ E 4 in general position, the vector ry is not tangent to F¯ 2 . If the bi-tangency property is fulfilled, then the Bianchi transformation has its remarkable property as in E 3 : THEOREM 2. If F 2 enjoys the bi-tangency property, then K¯ = −1 (F¯ 2 is a pseudospherical surface as well).
ON BIANCHI AND BÄCKLUND TRANSFORMATIONS
79
Theorem 2 answers Question 3. The answer to Questions 4 and 5 is given by the following theorem: THEOREM 3. The set of all pseudospherical surfaces in E 4 with a bi-tangency property is arbitrarily of four functions, each with a single variable. Consider now Question 6. A partial answer (a necessary condition) to Question 6 follows below (Theorem 4). First of all, we recall that the set of all m-dimensional subspaces of n-dimensional Euclidean space through fixed point O (called a Grassmann manifold and denoted by Gm,n ) is a Riemannian space [25, 26]. The corresponding sectional curvature for the element of the area tangent to the Grassmann image of F 2 is denoted by KG . We recall that the Grassmann image of F 2 is elliptic (parabolic, hyperbolic) if and only if KG > 1, (KG = 0, KG < 1). THEOREM 4. (1) If the Grassmann image of F 2 is elliptic, then the vector field a tangent to both F 2 and F¯ 2 never exists. (2) If the Grassmann image is hyperbolic, then at every point x ∈ F 2 there exist at most two directions of the vector field tangent to both F 2 and F¯ 2 . One can complete Theorem 4 by the following theorem: THEOREM 5. Suppose F 2 ⊂ E 4 is of a flat normal connection and its Gaussian curvature K 6= 0. If, additionally, there exists a vector field ha with constant lenght h and tangent to regular surface F¯ 2 , then necessarily F 2 ⊂ E 3 . If F 2 6⊂ E 3 , then for every field of principal directions τi , it is possible to define the function h – by putting it equal to the minus radius of the geodesic curvature of the orthogonal curvature line – such that the vector field hτi will be tangent to F 2 and F¯ 2 .
5. The Proofs PROOF OF THEOREM 1
We put x = y1 , y = y2 . Select any field of the normalized normal vectors nσ , σ = 1, 2. Correspondingly, Lσij are the coefficients of the second fundamental forms. We denote the first and second derivatives of r by ri and rij . The Gauss decompositions for F 2 imply j
r¯i = ri − r2i = ri − 02i rj − Lσ2i nσ . 1 2 Since 021 = 1 and 022 = 0, we have
r¯1 = −Lσ21 nσ ,
r¯2 = r2 − Lσ22 nσ .
80
YU. AMINOV AND A. SYM
Then we obtain the formula for the metric of F¯2 . d s¯2 = (¯r1 dy1 + r¯2 dy2 )2 = (−Lσ21 nσ dy1 + (r2 − Lσ22 nσ )dy2 )2 2 X = (Lσ12 dy1 + Lσ22 dy2 )2 + dy22 .
(4)
σ =1
Now we make an additional assumption: the normal connection of F 2 is flat. Hence, one can choose the field n1 , n2 with the torsion coefficients µασ |i = 0. Fortunately, this implies an existence of two functions φσ = φσ (y1 , y2 ) such that Lσ12 dy1 + Lσ22 dy2 = e−y2 dφσ .
(5)
In other words, (5) is equivalent to ∂Lσ12 ey2 ∂Lσ22 ey2 − = 0. ∂y2 ∂y1
(6)
But the Codazzi equations for F 2 have the form Lσ12,2 − Lσ22,1 = 0 or, in detail, ∂Lσ12 ∂Lσ22 − + Lσ12 = 0. ∂y2 ∂y1 Therefore, Equations (5), (6) are true. The metric d¯s 2 of F¯ 2 can be written as d¯s 2 = e−2y2 (dφ12 + dφ22 ) + dy22 .
(7)
If φ1 , φ2 , y2 were independent coordinates, (6) could serve as a metric of the threedimensional Lobachevski space of the curvature equal to −1. This means that, apart from F¯2 ⊂ E 4 , we can consider some immersion of its metric into L3 . This resulting surface is denoted by M 2 ⊂ L3 . The Gauss equation for F 2 reads L111 L122 − (L112 )2 + L211 L222 − (L212 )2 = −e2y2 .
(8)
The assumption of the flat normal connection is expressed by e2y2 (L111 L112 − L112 L211 ) + (L112 L222 − L122 L221 ) = 0.
(9)
We can solve (8) and (9) with respect to L111 and L211 and the resulting expressions can be substituted into the Codazzi equations: Lσ11,2 − Lσ12,1 = 0, σ = 1, 2. Let P = (e2y − e−2y (u2x + vx2 ))ux + vy (ux vy − uy vx ), Q = (e2y − e−2y (u2x + vx2 ))vx + uy (uy vx − ux vy ), where we have put φ1 = u, φ2 = v. Finally, we obtain the following system: ∂ P + uxx e−2y = −uy , (10) ∂y ux uy + vx vy ∂ Q + vxx e−2y = −vy . (11) ∂y ux uy + vx vy
ON BIANCHI AND BÄCKLUND TRANSFORMATIONS
81
The important point is that the whole system of Gauss–Codazzi–Ricci equations in the discussed case reduces to system (10)–(11). This system is of the form uxx = A(y, ux , uy , vx , vy , uxy , uyy , vxy , vyy ),
(12)
vxx = B(y, ux , uy , vx , vy , uxy , uyy , vxy , vyy ),
(13)
where A and B are analytical functions of their variables. As initial data, we can select the following functions: u(0, y), ux (0, y), v(0, y), vx (0, y). We can certainly, use these functions as parameters labelling isometric immersions of L2 into E 4 . The arbitrariness is four functions of a single argument. After these preliminary remarks, we go over the proof of Theorem 1. In fact, we present two proofs (one indirect and one direct). The indirect proof. Let us consider the surface M 2 ⊂ L3 . For surfaces in L3 , ¯ eL ) is the Gauss (extrinsic) curvature of M 2 . we have K¯ = KeL − 1, where K(K L Obviously K¯ = −1 iff Ke = 0. As a convenient model of L3 we take an open ball D ⊂ E 3 of the unit radius equipped with the metric dsl2 =
(ldl)2 + (1 − l 2 )(dl)2 , (1 − l 2 )2
(14)
where l is a position vector starting at the center of the ball and ending at the point of D. In this way, any surface M 2 ⊂ D can be interpreted either as a Lobachevskian surface or as a Euclidean one. Sidorov [21] in particular, derived the following formula relating KeL (extrinsic curvature of M 2 ⊂ L3 ) to K E (Gauss curvature of M 2 ⊂ E3) 1 − l2 2 L E Ke = K , 1 − (ln)2 where n is a unit normal to M 2 ⊂ E 3 . This formula implies that KeL = 0 iff K E = 0. Now the set of all developable (K E = 0) surfaces is parametrized by two functions of a single argument. In other words, of the set of all surfaces F¯ 2 of the Gaussian curvature = −1 the arbitrariness is smaller than the arbitrariness of all isometric immersions of L2 into E 4 with a flat normal connection. The direct proof. Consider again metric (4) rewritten as follows: d¯s 2 = e−2y (du2 + dv 2 ) + dy 2 . The two-dimensional metric ds02 = e−2y du2 + dy 2
82
YU. AMINOV AND A. SYM
is of the Gaussian curvature equal to −1 as well. The equations of the immersion read u = u(x, y),
v = v(x, y).
We want to compute the value of K¯ at a given point P0 , say x = 0, y = 0. Let us assume that vx (0, 0) = vy (0, 0) = 0 at P0 . According to the Frobenius formula for the Gaussian curvature [22], one can write Eyy − 2Fxy + Gxx + 9(E, F, . . . , Ex , . . . , Gy ), K¯ = − 2(EG − F 2 ) where we have E = e−2y (u2x + vx2 ),
F = e−2y (ux uy + vx vy ),
G = e−2y (u2y + vy2 ) + 1. We point out that all expressions for the first derivatives of E, F and G at point P0 do not contain the derivatives of the function v. At point P0 we have 2 Eyy − Fxy + Gxx = −2(vxx vyy − vxy ) + 3(ux , uy , . . . , uyy ).
EG − F 2 = u2x . Moreover, the expression of 9 at P0 does not involve the derivatives of v. The Gaussian curvature at P0 is given by K¯ = −1 +
2 vxx vyy − vxy
u2x
.
As initial data we select v(0, y) = 0,
vx (0, y) = y,
ux (0, 0) 6= 0,
Thus, at point P0 vyy = 0,
vxy = 1,
vx = vy = 0.
So, finally, we obtain at P0 1 K¯ = −1 − 2 = 6 −1. ux Theorem 1 is proved in a direct way.
uy (0, 0) 6= 0.
ON BIANCHI AND BÄCKLUND TRANSFORMATIONS
83
PROOF OF THEOREM 2
Let ξ1 and ξ2 denote two fields of normalized normals to F¯ 2 . We select ξ1 and ξ2 as follows. It is not difficult to show that e−y2 ry1 is orthonormal to F¯ 2 . Hence, we put ξ1 = e−y2 ry1 , ξ2 = λ[¯ry1 r¯y2 ry1 ] (skew product in E 4 ), and λ = |[· · ·]|−1 is a normalized factor. Trilinearity of the skew product implies ξ2 = λ(L112[n1 ry2 ry1 ] + L212 [n2 ry2 ry1 ] + (L122 L212 − L112 L222 )[n1 n2 ry1 ]).
(15)
This in turn implies (ry2 ξ2 ) = (L122 L212 − L112 L222 )(ry2 n1 n2 ry1 ) = 0, where the second factor of the the RHS is a four-dimensional ‘mixed product’. For regular surfaces the ‘mixed product’ is nonzero everywhere. The vector ry2 is tangent to F¯ 2 iff (ry2 , ξ2 ) = 0, or iff L122 L212 − L112 L222 = 0.
(16)
If so, then (15) reduces to −L2 n1 + L112 n2 . ξ2 = q 12 (L112 )2 + (L212 )2
(17)
(17) implies that normal planes to F 2 and F¯ 2 at the corresponding point intersect along a line generated by ξ2 . It is natural to choose ξ2 = n1 (L112 = 0), while (16) implies L122 L212 = 0.
(18)
The case L212 = 0 is excluded (F¯ 2 is not regular). So we are left with L112 = L122 = 0. The Codazzi equation reduces to ∂L212 ∂L222 − + L212 = 0. ∂y2 ∂y1 Thus L212 dy1 + L222 dy2 = e−y2 dφ, and the metric of F¯ 2 assumes the form d¯s 2 = e−2y2 dφ 2 + dy22 .
(19)
84
YU. AMINOV AND A. SYM
Certainly, the Gauss curvature of (19) is equal to −1. Theorem 2 has been proved. PROOF OF THEOREM 3
We recall (see the previous proof) that in the discussed case L112 = L122 = 0, L212 = e−y2 φy1 ,
L222
−y2
=e
φy2 .
(20) (21)
Thanks to (20) and (21) one of the Codazzi equations is satisfied. Certainly, the other Codazzi equation is to be satisfied L112,2 − L122,1 = µ21|2 L212 − µ21|1 L222 ,
(22)
where the µ coefficients are called ‘torsion coefficients’ [23]. From (20) it follows that the right-hand side of (22) vanishes. We can put µ21|2 = λL222,
µ21|1 = λL212 ,
(23)
where λ = λ(y1 , y2 ) is an unknown factor. The Gauss equation L211 L222 − (L212 )2 = −e2y2
(24)
implies L211 =
e−2y2 φy21 − e2y2 e−y2 φy2
.
Combining the Codazzi equation L211,2 − L212,1 = µ12|2 L111 − µ12|1 L212 = µ12|2 L111 and (23) we obtain ∂ e−2y2 φy21 − e2y2 − e−2y2 φy1 y1 = −λφy2 L111 e−2y2 . ∂y2 φy2
(25)
By using (23) and (24) we can rewrite the last Codazzi equation L111,2 − L112,1 = µ21|2 L211 − µ21|1 L212 as ∂L111 − L111 = λ[L211 L222 − (L212 )2 ] = −λe2y2 . ∂y2
(26)
(25) is equivalent to ∂ (L1 e−y2 ) = −λey2 . ∂y2 11
(27)
ON BIANCHI AND BÄCKLUND TRANSFORMATIONS
85
Finally we consider the (single) Ricci equation µ21|2,1 − µ21|1,2 + g ll (L2l2 L1l1 − L2l1 L1l2 ) = 0, where g ll stands for contravariant components of the metric tensor. By using (23) we obtain ∂ ∂ (λe−y2 φy2 ) − (λe−y2 φy1 ) + e−3y2 L111 φy1 = 0. (28) ∂y1 ∂y2 (27) gives λ which we insert into (25) and (28). Let θ stand for L111 . We conclude the calculation of the current proof with the following statement: the system of Gauss–Codazzi–Ricci equations in the discussed case reduces to the system of two equations for two functions φ(y1 , y2 ) and θ(y1 , y2 ) of the form ∂ e−2y2 φy21 − e2y2 ∂θ − e−2y2 φy1 y1 = e−3y2 φy2 θ , ∂y2 φy2 ∂y2 ∂ ∂e−y2 θ ∂e−y2 θ ∂ − e−2y2 φy2 + e−2y2 φy1 + e−3y2 φy1 θ = 0. ∂y1 ∂y2 ∂y2 ∂y2 Rewrite these equations in the brief form φy2 y2 = A(y2 , θ, φy1 , φy2 , θy2 , φy1 y1 , φy1 y2 ), θy2 y2 = B(y2 , θ, φy1 , φy2 θy1 , θy2 , φy1 y2 , θy1 y2 ). As usual φ(y1 , 0), φy2 (y1 , 0), θ(y1 , 0) and θy2 (y1 , 0) denote the initial conditions. Thus, the arbitrariness of isometric immersions of L2 into E 4 (no flat normal condition is assumed) with the bi-tangency property is of four functions of a single argument. Moreover, we can select the initial functions in such a way that the Gauss torsion κ0 := g ll (L1l1 L2l2 − L1l2 L2l1 ) 6= 0 and hence the discussed immersion of L2 essentially lies in E 4 not in E 3 . PROOF OF THEOREM 4
Let F 2 ⊂ E 4 be any regular surface and (u1 , u2 ) any orthogonal coordinates on it. The metric of F 2 reads ds 2 = Edu21 + Gdu22 .
(29)
We write r¯ = r + ha,
(30)
where r = r(u1 , u2 ) is a point vector to F 2 , while a is a field tangent to F 2 of unit lenght: |a| = 1, r¯ is a position vector to a surface F¯ 2 . Instead of (29) we can write r¯ = r + h(cos γ τ1 + sin γ τ2 ),
86
YU. AMINOV AND A. SYM
where τi (i = 1, 2) are normalized coordinate vectors. Let b = − sin γ τ1 +cos γ τ2 . We recall that our aim is to state necessary conditions for ha to be tangent to F¯ 2 as well. To this end we compute cos γ Lσi1 sin γ Lσi2 + √ nσ , √ (31) r¯ui = rui + hui a + hbAi + h E G where A1 =
∂γ 1 ∂E − √ , ∂u1 2 EG ∂u2
A2 =
∂γ 1 ∂G + √ . ∂u2 2 EG ∂u1
(32)
We simplify the right-hand side of (31) as r¯u1 = a1 τ1 + a2 τ2 + c1 n1 + c2 n2 , r¯u2 = b1 τ1 + b2 τ2 + d1 n1 + d2 n2 ,
(33) (34)
where ai , . . . , di are some coefficients. The vector a is a linear combination of the vectors r¯u1 , r¯u2 iff the following matrix a2 c1 c2 a1 b2 d1 d2 D = b1 (35) cos γ sin γ 0 0 has the rank D 6 2. Hence c1 d2 − c2 d1 = 0.
(36)
In the left-hand side of (36) we replace ci and di by their explicit forms given in terms of the coefficients of I I 1 and I I 2 . We arrive at cos2 γ 1 2 sin γ cos γ 1 2 (L11 L12 − L112 L211 ) + √ (L11 L22 − L122 L211 ) + E EG sin2 γ 1 2 (L12 L22 − L122 L212 ) = 0. + (37) G So far we have not specified orthogonal coordinates u1 , u2 . Now we select them as those defined by the indicatrix of the normal curvature of F 2 ⊂ E 4 [24]. We recall it is an ellipse in our case. If n1 and n2 are parallel to its axes then L111 = E(α + a),
L211 = Eβ, √ = 0, L212 = EGb, = G(α − a), L222 = Gβ,
(38)
L112 L122
(39) (40)
where a and b are lengths of semi-axes while α and β are coordinates of the ellipse center. Substituting the expressions (38)–(40) into (37) gives cos2 γ b(α + a) + 2 cos γ sin γ aβ − sin2 γ b(α − a) = 0.
(41)
ON BIANCHI AND BÄCKLUND TRANSFORMATIONS
87
This quadratic equation has two roots iff a 2 β 2 + α 2 b2 − a 2 b2 > 0.
(42)
To conclude our proof, let us recall the formula for the Gaussian curvature KG of the Grassmann image of F 2 in G2,4 expressed in terms of the Gaussian curvature K of F 2 [25, 26] KG =
K 2 + 4a 2 b2 . K 2 + 4(a 2 β 2 + b2 α 2 )
(43)
It is evident that (41) implies KG < 1. According to the definition given in [25, 26] the Grassmann image of F 2 is hyperbolic. This ends the proof of Theorem 4. PROOF OF THEOREM 5
In this proof we employ all the notations of the previous proof. The assumption of a flat normal connection implies b = 0. Thus the ellipse of normal curvature is degenerated and Equation (40) reduces to sin γ cos γ aβ = 0.
(44)
One should consider three cases: (1) sin γ = 0 (cos γ = 0 is treated similarly), (2) a = 0, and (3) β = 0. In case (1) a = τ1 and we have r¯ = r + hτ1 , r¯u1 = a1 τ1 + a2 τ2 + c1 n1 + c2 n2 , √ r¯u2 = hu2 τ1 + ( G + hA2 )τ2 .
(45)
¯2 Let ξi (i = 1, 2) be normalized √ normal vectors to F . Consider at first the case h = const and denote b2 = G + hA2 . But (¯ru2 ξi ) = 0 implies b2 = 0 or (τ2 ξi ) = 0. If b2 = 0, hu2 = 0, then F¯ 2 is necessarily degenerated. So, let b2 6= 0. Now (τ2 ξi ) = 0 when combined with (τ1 ξi ) = (aξi ) = 0 gives an identity of normal spaces to F 2 and F¯ 2 at the corresponding points. Thus c1 = c2 = 0 and therefore c2 = L211 = β = 0. By the result of [27] we know that the identities β = 0, b = 0 and inequality K 6= 0 imply F 2 ⊂ E 3 . In case (2) or a = 0 and b = 0 implies F 2 is a standard sphere S 2 ⊂ E 3 . Case (3) has been already discussed above. The consideration above and (44) imply: if F 2 ⊂ E 4 is of a flat normal connection, F 2 6⊂ E 3 , K 6= 0 and if the field ha enjoys the bi-tangency property, then a must be necessarily a field of principal directions. As we have proved above, h must be some nonconstant function on F 2 . Now we shall prove that in general it is possible to select the function h in such a way that the field hτi will enjoy the bi-tangency property.
88
YU. AMINOV AND A. SYM
Consider, for example, the transformation r¯ = r + hτ1 . We recall formula (44). To determine the function h, we put √ √ 1 ∂G G + hA2 = G + h √ = 0. 2 EG ∂u1 This equation can be rewritten in the following simple form 1+h
1 = 0, ρgv
where 1/ρgv denotes the geodesic curvature of the curvature line u = const. So, if the geodesic curvature of the orthogonal curvature line is not equal to zero, then we can take −h equal to the radius of the geodesic curvature of this line. If this geodesic curvature is not constant, then hu2 6= 0, r¯u2 6= 0 and F¯ 2 is a regular surface. In this case, matrix D has rank = 2. Therefore field ha is tangent to both F 2 and F¯ 2 . Acknowledgement Work partially supported by the Polish Committee of Scientific Researches (KBN grant 2 PO3B 185 09). References 1.
Eisenhart, L. P.: A Treatise on the Differential Geometry of Curves and Surfaces, Ginn, New York, 1909. 2. Bianchi, L.: Ricerche sulle superficie a curvatura costante esulle elicoidi, Ann. Sci. Norm. Sup. Pisa (1) 2 (1879), 285–340. 3. Weingarten, J.: Ueber eine Klasse auf einander abwickelbarer Flachen, Crelle J. 59 (1861), 382–390. 4. Darboux, G.: Leçons sur la théorie generale des surfaces, vol. III, Gauthier-Villars, Paris, 1894, ch. XII. 5. Bäcklund, A. V.: Om ytor med konstant negativ krokning, Lunds Univ. Arsskr. 19 (1883), 1–41. 6. Bäcklund, A. V.: Zur Theorie der partiellen Differentialgleichung erster Ordnung, Mat. Ann. 17 (1880), 285–328. 7. Forsyth, A. R.: Theory of Differential Equations, Vol. VI, Dover Publications, New York, 1959. 8. Bianchi, L.: Sopra i sistemi tripli ortogonali di Weingarten, Ann. di Mat. (2) 13 (1885), 177– 234. 9. Bianchi, L.: Sui sistemi doppiamente infiniti di raggi, Ann. di Mat. (2) 15 (1887–1888), 161– 172. 10. Prus, R. and Sym, A.: Rectilinear congruences and Bäcklund transformations: roots of the soliton theory, In: D. Wojcik and J. Cie´sli´nski (eds), Nonlinearity and Geometry, PWN (Polish Scientific Publishers), Warsaw, 1998. 11. Ablowitz, M. J. and Clarkson, P. A.: Solitons, Nonlinear Evolution Equations and Inverse Scattering, Cambridge University Press, Cambridge, 1991. 12. Lamb, G., Jr.: Bäcklund transformations at the turn of the century, In: R. M. Miura (ed.), Bäcklund Transformations, Lecture Notes in Math. 515, Springer, New York, 1976.
ON BIANCHI AND BÄCKLUND TRANSFORMATIONS
13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26. 27.
89
Lamb, G. L., Jr.: Bäcklund transformations for certain nonlinear evolution equations, J. Math. Phys. 15 (1974), 2157–2165. Wahlquist, H. D. and Estabrook, F. B.: Bäcklund transformations for solutions of the Korteweg– de Vries equation, Phys. Rev. Lett. 31 (1973), 1386–1390. Wahlquist, H. D. and Estabrook, F. B.: Prolongation structures of nonlinear evolution equations, J. Math. Phys. 16 (1975), 1–7. Aminov, Yu.: Bianchi transform for the domain of multidimensional Lobachevsky space, Ukrain. Geom. Sb. 21 (1978), 3–5 (in Russian). Masal’tzev, L.: Pseudospherical Bianchi congruencies in E 2n−1 , Math. Phys. Anal. Geom. 1(3/4) (1994), 505–512. Tenenblat, K. and Terng, C.-L.: A higher dimension generalization of the sine-Gordon equation and its Bäcklund transformation, Bull. Amer. Math. Soc. (N. S.) 1 (1979), 589–599. Tenenblat, K. and Terng, C.-L.: Bäcklund theorem for n-dimensional submanifolds of R 2n−1 , Ann. of Math. 111 (1980), 477–490. Tenenblat, K.: Bäcklund’s theorem for submanifolds of space form and a generalized wave equation, Bol. Soc. Brasil Mat. 16 (1985), 67–92. Sidorov, L.: Some properties of surfaces of negative extrinsic curvature in the Lobachevsky space, Mat. Zametki 4 (1968), 165–169. Blaschke, W.: Einfuhrung in die Differentialgeometrie, Springer-Verlag, Berlin, 1950. Eisenhart, L. P.: Riemannian Geometry, Princeton University Press, Princeton. Cartan, E.: Léçons sur la géométrie des espaces de Riemann, 2nd edn, Gauthier-Villars, Paris, 1946. Aminov, Yu.: Grassmann transform of a two-dimensional surface in four-dimensional Euclidean space, Ukrain. Geom. Sb. 23 (1980), 3–16. Aminov, Yu.: Determination of a surface in 4-dimensional Euclidean space by means of its Grassmann image, Mat. USSR Sb. 45 (1983), 155–167. Aminov, Yu.: Torsion of two-dimensional surfaces in Euclidean spaces, Ukrain. Geom. Sb. 17 (1975), 3–14.
Mathematical Physics, Analysis and Geometry 3: 91–100, 2000. © 2000 Kluwer Academic Publishers. Printed in the Netherlands.
91
The Ground State of Certain Coulomb Systems and Feynman–Kac Exponentials AMÉDÉE DEBIARD and BERNARD GAVEAU Université Pierre et Marie Curie (Paris 6), Mathématiques, T-46, 5ème étage, Boîte 172, 4 place Jussieu, 75252 Paris Cedex 05, France (Received: 10 January 2000) Abstract. We give a lower bound for the ground state energy of certain Coulomb Hamiltonians using the Feynman–Kac formula. We show that this bound is very precise for two electron atoms. Mathematics Subject Classifications (2000): 58J70, 60J65, 81S40. Key words: Coulomb systems, Brownian motion, ground state.
1. Introduction In this short note, we show that it is possible to use the Feynman–Kac formula to obtain a very sharp (indeed, almost exact) lower bound for the ground state of a two-electron atom, using only the basic facts of Brownian motion theory and elementary calculus. The starting point is the Feynman–Kac formula which provides a way of writing the solution of the heat equation with a potential, using a Wiener expectation on the Brownian motion. A consequence is the formula for the upper bound of the spectrum of 12 1−V in terms of the asymptotic estimate of such Wiener expectations (see [1] for the original reference, [2, 3] for a recent review of the subject and, more recently, [4] for another method of derivation which can be generalized to other situations). The estimation of such Wiener expectations have been the subject of many works (see [5 – 7], and their references). We thus obtain an upper estimate of the spectrum of 12 1−V , or a lower bound for the spectrum of the Hamiltonian H . The usual variational method (Rayleigh–Ritz) would provide an upper bound of the ground state of H . We obtain, in the special case of a Coulomb potential, an upper bound of the Wiener expectation, which uses the fact that the Coulomb potential is homogeneous of degree −1. That fact allows us to reduce the problem to an ergodic theorem for the spherical Brownian motion. For the twoelectron system, the absolute ground state is also the physical ground state. The estimation that we obtain in this case can be compared to the experimental value and is surprisingly sharp (1.2% of error), indicating the power of the Feynman–Kac formula and path integrals.
92
´ EE ´ DEBIARD AND BERNARD GAVEAU AMED
In Section 2, we recall the basic facts about the Feynman–Kac formula and in Section 3, we specialize to the case of the Coulomb system. Section 4 gives the main estimate, Section 5 gives the final result. The Appendix contains the calculations of various integrals on a sphere. 2. The Feynman–Kac Formula for the Ground State We consider the operator H = − 12 1 + V where 1 is the standard Laplace operator in an N-dimensional space RN and V is a real function, such that H is essentially a self-adjoint operator. We consider the Cauchy problem ∂ψ (2.1) = −H ψ, ψ|t =0 = ψ0 . ∂t This can be solved, using the Feynman–Kac formula as follows. We consider the N-dimensional standard Brownian motion b(t) = (b1 (t), . . . , bN (t)), where bj (t) are independent one-dimensional standard Brownian motions. Then, if x ∈ RN , the solution of Equation (2.1) is Z t ψ(x, t) = E exp − V (b(s)) ds ψ0 (b(t)) b(0) = x , (2.2) 0
where E{. . . | b(0) = x} is the conditional expectation on Brownian paths, starting from x at time t = 0. We refer to [1 – 3] and also to [4] for a different proof. Because H is essentially a self-adjoint operator, it has a spectral decomposition and its spectrum is bounded from below. Call 3 an upper bound of the spectrum of −H . If ψ0 is not orthogonal to the generalized eigenfunction of eigenvalue 3, then ψ(x, t) ∼ C exp(3t)
(2.3)
and comparing Equations (2.2) and (2.3), we deduce Kac’s formula for the upper bound 3: Z t 1 (2.4) V (b(s)) ds ψ0 (b(t)) b(0) = x . 3 = lim log E exp − t →∞ t 0 We shall apply this formula to the situation where 3 is an eigenvalue of −H , the corresponding eigenfunction is integrable (and, as usual, square integrable), and it does not change sign, so that we can use any ψ0 > 0, bounded. This last hypothesis is normally fulfilled for standard Schrödinger Hamiltonians because the ground state does not change sign. 3. N Electrons Atom Hamiltonian → → rN ) in the Coulomb We consider the situation of N electrons (positions − r1 , . . . , − field of a nucleus of charge Ze. In atomic units, the Hamiltonian is 1X → → H =− 1i + V (− r1 , . . . , − rN ), 2 i=1 N
(3.1)
THE GROUND STATE OF CERTAIN COULOMB SYSTEMS
93
→ ri and where 1i is the three-dimensional Laplace operator for the variables − → → V (− r1 , . . . , − rN ) = −
N X Z i=1
ri
X
+
16i<j 6N
1 . → → |− ri − − rj |
(3.2)
The Hamiltonian H satisfies all the hypotheses of the previous section. We shall apply Equation (2.4) with V given as the Coulomb potential Equation (3.2) and with ψ0 = exp(−ϕ), where ϕ is a real function such that ψ0 is bounded. Our problem is, then, to find the large time behavior of the quantity Q(N, Z, t) Z t → → → V (− r1 (s), . . . , − rN (s)) ds − ϕ(− r1 (t), . . . = E exp − 0 − (0) → → − − → . . . , rN (t)) ri (0) = ri ,
(3.3)
→ → where − r1 (s), . . . , − rN (s) are N independent three-dimensional Brownian motions. 4. Transformation of Q(N, Z, t) To obtain a bound of Q, we shall use a method related to the one used in [7] for compact manifolds. Here, the fact that the Coulomb potential is homogeneous of degree −1 replaces the compactness. We define → → ϕ(− r1 , . . . , − rN ) =
N X j =1
→ Z− rj −
1 X − → |→ ri − − rj |. 2 16i<j 6N
(4.1)
It is easy to see that, using 12 1r = 1r , X N 1 1i ϕ = −V . 2 i=1
(4.2)
Now, we use Itô’s formula for the Brownian motion (see, e.g., [8]), under the form → → ϕ(− r1 (t), . . . , − rN (t)) → − → = ϕ( r1 (0), . . . , − rN (0)) + Z tX N − → → → − (∇j ϕ)(− r1 (s), . . . , − rN (s)) · d→ rj (s) + + 0 j =1
Z +
0
t
! N 1 X → → r1 (s), . . . , − 1j ϕ(− rN (s)) ds. 2 j =1
(4.3)
94
´ EE ´ DEBIARD AND BERNARD GAVEAU AMED
Equations (4.2) and (4.3) allow us to rewrite the Feynman–Kac exponential inside Equation (3.3) as Z t → − − → → − − → V r1 (s), . . . , rN (s) ds − ϕ r1 (t), . . . , rN (t) exp − 0
Z tX N − → → → − = exp −ϕ(− r1 (0), . . . , − rN (0)) exp − rj (s) . ∇j ϕ · d →
(4.4)
0 j =1
We use the exponential martingale (see [8]). Given a N-dimensional Brownian mo→ − → − tion b (s) and a vector-valued random function f (s, ω) (ω being the simple path − → of the Brownian motion, s the time) such that f is nonanticipating, the quantity Z t Z 1 t − − → → − →2 Mt ≡ exp f (s, ω) d b (s) − | f | ds (4.5) 2 0 0 is a martingale and, in particular, its expectation is equal to 1. We shall rewrite the exponential of the stochastic integral in Equation (4.4) as Z tX N − → − → ∇j ϕ · d rj (s) = F · G exp − 0 j =1
with F and G given as Z tX N N pX − − → − → 2 → − → ∇j ϕ( r (s)) · d rj (s) − |∇j ϕ| ds , F = exp − 2 j =1 0 j =1
p G = exp 2
Z tX N − → 2 |∇j ϕ| ds 0 j =1
and p is a positive number. Then Q can be rewritten Q(N, Z, t) = exp −ϕ(rEj (0)) E(F G)
(4.6)
and we apply the Hölder–Young inequality with (1/p) + (1/q) = 1 1/p
E(F G) 6 E(F p )
1/q
E(Gq )
to the expectation in Equation (4.6). Thus, E(F p ) is of the type E(Mt ) for a certain Mt , as in Equation (4.5), with → − → − − → − → → f = ( fj ), fj = −p (∇j ϕ)(− r (s))
95
THE GROUND STATE OF CERTAIN COULOMB SYSTEMS
so that E(Mt ) = 1 and, thus, Q(N, Z, t)
1/q Z N −→ − 2 pq t X − → → (0) → − 6 exp −ϕ(r ) E exp ∇j ϕ r (s) ds rj (0) = rj . (4.7) 2 0 j =1 (0)
P − → 2 Now, N j =1 |∇j ϕ| is a function homogeneous of degree 0 so that it is a function on the (3N −1)-dimensional unit sphere in R3N and we can use the ergodic theorem on the sphere [9] to get * N + Z N X − pq pq t X − → − 2 → ∇j ϕ → (4.8) r (s) ds ∼ |∇j ϕ|2 t, 2 0 j =1 2 j =1
where h · i is the average on the (3N − 1)-dimensional sphere. Then, from Equations (4.7)–(4.8), * N + 1 p X − → 2 lim log Q(N, Z, t) 6 |∇j ϕ| . t →∞ t 2 j =1 This upper bound is valid for any p > 1, noticing that q has disappeared and so, finally, we have the upper bound: + * N 1 1 X − → 3 = lim log Q(N, Z, t) 6 |∇j ϕ|2 ≡ 8. (4.9) t →∞ t 2 j =1 This estimate gives an upper bound of the spectrum of −H = 12 1 − V , and so a lower bound of the ground state of the Coulomb Hamiltonian. 5. Estimation of 3 We need to calculate the second member of Equation (4.9), namely the quantity 8 defined there. We have → → − → rj − − rj 1X − ri − → − ∇j ϕ = Z → − − → rj 2 i6=j | rj − ri | and so 2 N N − → → − X − X X → r r 1 − r − → 2 j j i Z |∇j ϕ| = → − → − r −2 | r − r | j j i i6=j j =1 j =1 → ri N(N − 1) N X − = NZ 2 + + 4 2 i,j 6=1 ri i6=j
→ − rj + r j
→ − → − → X → ri − − ri rj − rj − − , + NZ → → → |− ri − − rj | |− rj | |→ ri | i<j
96
´ EE ´ DEBIARD AND BERNARD GAVEAU AMED
where, in the last summation, each couple (i, j ) is counted once. Moreover, calling → → ri and − rj , one has θij the angle between − − → − → → → rj ri ri + rj ri − − rj − − = − (1 − cos θij ) → − − → → − → − → → | ri − rj | | rj | | ri | | ri − − rj | so that, finally, using hcos θij i = 0, * N + 1 X − → 2 8 = |∇j ϕ| 2 j =1 N N −1 N(N − 1) r1 + r2 2 = Z + − Z − (1 − cos θ12 ) . → 2 4 4 |→ r1 − − r2 |
(5.1)
We obtain a trivial upper bound, saying that r1 + r2 > 1, − → → | r1 − − r2 |
hcos θ12 i = 0,
namely 86
N − 1 (N − 1)Z N Z2 + − . 2 4 2
In the Appendix, it is proved that 4 r1 + r2 (1 − cos θ12 ) = , → − → − | r1 − r2 | π
(5.2)
(5.3)
so that
N − 1 2(N − 1)Z N 2 Z + − . 8= 2 4 π
(5.4)
For N = 2, the Pauli principle states that the physical ground state is indeed the absolute ground state, while this is no more correct for N > 3. For N = 2, we can restore the units (see Appendix), and obtain − from Equation (5.2) 8 6 88 eV; − from the more exact Equation (5.4) 8 ' 80 eV to be compared with the actual value of the physical ground state 79 eV. See, e.g., [10] or [11].
97
THE GROUND STATE OF CERTAIN COULOMB SYSTEMS
Appendix: Calculation of Spherical Integrals In this Appendix we calculate the spherical integral r1 + r2 I= − (1 − cos θ12 ) → |→ r1 − − r2 |
(A.1)
and prove Equation (5.3). In particular, we prove that this is independent of N. (I)
COORDINATES ON THE SPHERE
Call
q 2 2 → − → r1 + − r2 , q 2 2 → → r3 + . . . + − r 00 = − rN , p R = r 02 + r 002 . r0 =
→ In the space − r1 , we use the traditional polar angles (θ1 , ϕ1 ) and write → d− r1 = r12 dr1 sin θ1 dθ1 dϕ1 . → → r1 , namely angles In the space − r2 , we use polar angles around the axis defined by − (θ12 , ϕ12 ) → d− r2 = r22 dr2 sin θ12 dθ12 dϕ12 . Call r1 = r 0 cos γ , r2 = r 0 sin γ (0 6 γ 6 π/2), dr1 dr2 = r 0 dr 0 dγ . → → rN ), we use any polar coordinate system and write Moreover, in the space (− r3 , . . . , − − → → → rN = r 003N−7 dr 00 dσ (− r3 , . . . , − rN ), d− r3 . . . d→ where dσ is the spherical volume element of the unit sphere in R3N−6 . Finally, we write r 0 = R cos ρ,
r 00 = R sin ρ
(0 6 ρ 6 π/2).
In this way, we obtain → → → r2 . . . d− rN d− r1 d− 3N−1 =R dR cos5 ρ(sin ρ)3N−7 dρ cos2 γ sin2 γ dγ × → → r3 , . . . , − rN ) × sin θ1 dθ1 dϕ1 sin θ12 dθ12 dϕ12 dσ (−
98
´ EE ´ DEBIARD AND BERNARD GAVEAU AMED
and thus the spherical volume of the unit sphere in R3N → → dσ (− r1 , . . . , − rN ) 2 = cos γ sin2 γ dγ sin θ1 dθ1 dϕ1 sin θ12 dθ12 dϕ12 × → → × cos5 ρ(sin ρ)3N−7 dρ dσ (− r3 , . . . , − rN ).
(A.2)
The function to be integrated is r1 + r2 cos γ + sin γ (1 − cos θ12 ) = √ (1 − cos θ12 ). − → → − | r1 − r2 | 1 − sin 2γ cos θ12
(A.3)
→ → It is independent of − r3 , . . . , − rN by definition, but also on ρ, θ1 , ϕ1 , ϕ12 , so that the average value is the average value with respect to the measure dσ = cos2 γ sin2 γ dγ · sin θ12 dθ12 , π 06γ 6 , 0 6 θ12 6 π. 2
( II )
(A.4)
CALCULATION OF THE INTEGRAL
First of all we calculate Z Z π/2 Z π π (sin 2γ )2 dσ = dγ sin θ12 dθ12 = . 4 8 0 0
(A.5)
Then we need to calculate the integral of Equation (A.3), namely abbreviating θ = θ12 , Z Z π 1 π/2 cos γ + sin γ 2 J ≡ (sin 2γ ) dγ sin θ dθ √ (1 − cos θ) (A.6) 4 0 1 − sin 2γ cos θ 0 so that I=
8J . π
For convenience, we give details of the calculation: Z Z 1 1 π/4 cos γ + sin γ 2 J = (1 − u). (sin 2γ ) dγ du √ 2 0 1 − (sin 2γ )u −1 Thus Z 1 1−u du √ 1 − (sin 2γ )u −1 p p 2 1 − sin 2γ − 1 + sin 2γ + =− sin 2γ
(A.7)
(A.8)
99
THE GROUND STATE OF CERTAIN COULOMB SYSTEMS
Z 1 1 −(sin 2γ )u + √ du sin 2γ −1 1 − (sin 2γ )u 2 =− (1 − sin 2γ )1/2 − (1 + sin 2γ )1/2 + sin 2γ 2 (1 − sin 2γ )1/2 − (1 + sin 2γ )1/2 − + 2 sin 2γ 2 − (1 − sin 2γ )3/2 − (1 + sin 2γ )3/2 . 2 3 sin 2γ But
π , 4 (1 + sin 2γ )3/2 − (1 − sin 2γ )3/2 = (2 sin γ )(1 + 2 cos2 γ ), Z 1 2 sin γ 1−u 2 − . du √ = cos γ 3 cos2 γ 1 − (sin 2γ )u −1 (1 + sin 2γ )1/2 − (1 − sin 2γ )1/2 = 2 sin γ
06γ 6
From Equations (A.8)–(A.9), we deduce Z 4 π/4 2 J = sin γ (4 cos2 γ + 2 sin γ cos γ − 1) dγ = 3 0
(A.9)
1 2
because Z π/4 π 4 sin2 γ cos2 γ dγ = , 8 0 Z π/4 1 2 sin3 γ cos γ dγ = , 8 0 Z π/4 π 1 sin2 γ dγ = − , 8 4 0 so that I = 4/π . ( III )
RESTORING THE UNITS
In the unit of this article, the exact ground state of an hydrogenoïd atom of charge Z would be Z 2 /2. For Z 2 = 1 (hydrogen), we obtain 13.5 eV, so 1 unit corresponds to 27 eV, from which we deduce the numerical values given in this article. Acknowledgement We thank the referee for comments.
100
´ EE ´ DEBIARD AND BERNARD GAVEAU AMED
References 1.
Kac, M.: Probability and Related Topics in the Physical Sciences, Interscience, New York, 1989. 2. Kac, M.: Integration in Function Spaces and Some of its Applications, Lezioni Fermiane, Pisa, 1980. 3. Schulman, L. S.: Techniques and Applications of Path Integration, Wiley, New York, 1981. 4. Gaveau, B. and Schulman, L. S.: Grassmann-valued processes for the Weyl and the Dirac equations, Phys. Rev. D 36 (1987), 1135–1140. 5. Berthier, A.-M. and Gaveau, B.: Convergence des exponentielles de Kac et applications en physique et en géométrie, J. Funct. Anal. 29 (1978), 416–424. 6. Gaveau, B. and Mazet, E.: Divergence des fonctionnelles de Kac et diffusion quantique, Publ. RIMS, Kyoto 18 (1982), 365–377. 7. Gaveau, B.: Estimation des fonctionnelles de Kac sur une variété compacte et première valeur propre de 1 + f , Proc. Japan Acad. Sci. 60 (1985), 361–364. 8. McKean, H. Jr.: Stochastic Integrals, Academic Press, New York, 1969. 9. Itô, K. and McKean, H. Jr.: Diffusion Processes and their Sample Paths, Academic Press, New York, 1964. 10. Levine, I. N.: Quantum Chemistry, Prentice-Hall, Englewood, 1991. 11. Karplus, M. and Porter, R. N.: Atoms and Molecules, Benjamin Cummings, Reading, 1970.
Mathematical Physics, Analysis and Geometry 3: 101–115, 2000. © 2000 Kluwer Academic Publishers. Printed in the Netherlands.
101
Periodic Ground State Configurations in a One-Dimensional Hubbard Model of Statistical Mechanics M. M. KIPNIS Chelyabinsk Pedagogical University, 69 Lenin Ave, Chelyabinsk, 454080, Russia e-mail:
[email protected] (Received: 1 July 1999; in final form: 17 April 2000) Abstract. This paper considers an averaging procedure for the description of a particles arrangement in a Hubbard model with antiferromagnetic interactions. The arrangements are described by the devil’s staircase. Completeness of the staircase is proved. Mathematics Subject Classification (2000): 82B20. Key words: statistical mechanics, Hubbard model, periodic ground state configurations, symbolic dynamics, phenomenon of even two-colouring, averaging procedure, devil’s staircase, completeness.
1. Introduction In the Hubbard model [1, 3, 8, 13, 14] of statistical mechanics, the ground states are described by bilateral sequences (un ) ((un ): Z → {−1; 1}). The values of variable un may be interpreted either as an electron (un = 1) and a hole (un = −1) or as a particle with an up (un = 1) or down (un = −1) spin. The ground states provide the minimum of the formally defined Hamiltonian H : H = −ψ
X i∈Z
ui +
X
γi−j ui uj ,
(1)
i>j ; i,j ∈Z
where γi ((γi ): N → R) may be interpreted as the interaction energy of two particles at the distance of i units; ψ stands for a chemical potential. We shall give a euristical procedure for calculating the minimal arrangement for the Hamiltonian (1) in the spirit of dynamic programming. Let us suppose that the values um (m < n) are constructed and we have no information about um for m > n. Excluding um for m > n from (1) and defining the value of un , we get the energy increment ∞ X 1Hn = un −ψ + γi un−i . i=1
(2)
102
M. M. KIPNIS
To decrease energy according to (2), it is quite natural to provide 1Hn 6 0 by putting ∞ X (3) γi un−i . un = sgn ψ − i=0
Equation (3) will be called a Boolean averaging system. It will be the main object of our investigations. Another method of minimization of Hamiltonian (1) is given in [3]. We may consider that in (3) one selects either un = 1 or un = −1 in an attempt to provide the equality of ψ and a weighted average of the sequence un , un−1 , . . . with weights γ1 , γ2 , . . . That is why we refer to (3) as an averaging system. Equation (3) was introduced by the author [13, 14]. The averaging system (3) describes a variety of systems: relay periodic processes in a sampled-data control system [6, 11]; periodic output in analog-to-digital converter with sigma-delta modulation and leaky integration [4, 5, 14, 18]; the point itineraries in the iterates of some maps on an interval [10, 14, 15]; the rotation of the circle through a rational angle [17]. The common property of these systems is the even mixture of two kinds of objects on a circle in a given proportion, i.e. the even 2-colouring [16, 17, 21]. We treat of two equivalent formal descriptions of the periodic configurations in the above-mentioned systems (all as words in the −+ alphabet): Hubbard configurations [3] and the set J [11]. The frequency of pluses in the periodic configurations is called its rotation number. The rotation numbers are described in a parameter space by the devil’s staircase. The completeness of the devil’s staircase is the main result in the article. It is stated in Section 3 (Theorem 2) and proved in Section 8. Some results of this paper were given earlier without proofs by the author [13, 14]. 2. Periodic Ground State Configurations It is convenient to consider the periodic ground state configurations in Boolean averaging system (3) as words in the −+ alphabet. DEFINITION 1. Let for each i (i ∈ N, 1 6 i 6 p, p ∈ N) εi = − or εi = +. The word ε1 . . . εp is called a periodic ground state configuration in Boolean averaging system (3) for the given ψ, if there exists a bilateral sequence (un ) (n ∈ Z), satisfying Equation (3), such that (1) for each i (1 6 i 6 p) if εi = − then ui = −1; if εi = + then ui = 1; (2) for each integer n un+p = un . Let A be a word in the −+ alphabet. The set of points ψ, such that A is a periodic ground state configuration in (3) for a given ψ, is called the domain of the periodic configuration A.
PERIODIC GROUND STATE CONFIGURATIONS
103
EXAMPLE 1. Let γ1 = 10, γ2 = 9, γ3 = 6, γ4 = 4, and γn = 0 when n > 4. The complete list of periodic ground state configurations (up to shifts and degrees) is −+ for ψ ∈ (0, 3); − − ++ for ψ ∈ (0, 1); − + − + + for ψ ∈ (3, 11); − + + for ψ ∈ (11, 17); − + ++ for ψ ∈ (17, 21); − + + + + for ψ ∈ (21, 29); + for ψ > 29, and symmetrically for ψ < 0 (interchanging plusses and minuses). We have here non-uniqueness of periodic configurations: there exist simultaneously two configurations −+ and − − ++ for ψ ∈ (−1, 1). DEFINITION 2. For the word A the ratio of the number of plusses to the total number of letters is called the rotation number of A. The rotation number of A is denoted as ω(A). For example, ω(− + − + +) = 3/5. In the following definition [a] stands for the integer part of a. For each word A in the −+ alphabet the symbol A0 denotes the empty word. For the nonnegative integer n An+1 = An A. The word B is called the shift of A if there exist words C, D such that A = CD and B = DC. DEFINITION 3 [3]. The word A in −+ alphabet is called a Hubbard configuration, if in each of its shifts B there are either [i/ω(A)] or [i/ω(A)] − 1 letters between every plus and ith plus on the right-hand side of it. Besides, the words −n (n > 0), −n + (n > 0) and their shifts are also called Hubbard configurations. EXAMPLE 2. The words −+, −+−++, −−+−+ are Hubbard configurations, while the word − − ++ is not.
3. Statement of the Main Result Now we are ready to connect the periodic ground state configurations in averaging procedure (3) and the Hubbard configurations. THEOREM 1 [13]. Let for each i (1 6 i < ∞) γi > γi+1 > 0, let the series P∞ i=1 γi be convergent and (γi ) be convex (i.e. γi+1 < (γi + γi+2 )/2 for i > 1). Then the word A is a periodic configuration in Boolean averaging system (3) if and only if it is a Hubbard configuration. We avoid giving the proof of Theorem 1 because in Theorem 2 which comes later, we’ll deal with the wider class of interaction functions, free from the condition of convexity. In addition, results, similar to Theorem 1, were stated earlier [3, 8] for systems, which are not described by Equation (3), but also originate from a Hamiltonian minimization problem. The dependence of rotation numbers of the periodic ground state configurations on the values of ψ is described by the devil’s staircase [3, 14]: it is an increasing function, whose derivative equals to zero almost everywhere. It is a discontinuous function, but it may be extended to a continuous function. The case of nonconvex (γi ) had not been investigated until the author’s paper [13]. It is considered in the following theorem:
104
M. M. KIPNIS
THEOREM 2. If for each i (1 6 i < ∞) γi > γi+1 > 0, and the series is convergent, then:
P∞ i=1
γi
(1) Every Hubbard configuration is the periodic ground state configuration in Boolean averaging system (3) with some value of ψ. (2) Let periodic ground state configuration A in system (3) not be a Hubbard configuration. Let B be a Hubbard configuration with the same rotation number as A. Then the domain of A is a proper subset of the B domain. (3) Let P be the set of values of ψ such that there exist no periodic ground state configurations in Boolean averaging system (3) with a given ψ. Then the set P is a Cantor perfect set of the Lebesgue zero measure. Theorem 2 is our main result. It is proved in Section 8. When (γi ) is nonconvex, some ψ values may correspond to more than one periodic ground state configuration (see Example 1). However, Theorem 2 states the preservation of the devil’s staircase in this case. Part (3) of Theorem 2 means the completeness of the staircase. 4. The Point Itineraries in the Maps of Interval in Itself Consider the variant of Boolean averaging procedure (3) with the exponential weight sequence (γn ): γn = e−nα (eα − 1) (α > 0). If we replace (3) by two equations un = sgn σn ; σn = ψ − from (4): −α e σ + (ψ − 1)(1 − e−α ), if σn > 0; σn+1 = −α n e σn + (ψ + 1)(1 − e−α ), if σn < 0.
P∞ i=1
(4) γi un−i , we get
(5)
Map (5) is a piecewise linear discontinuous transformation of the line in itself. The orbits of transformation (5) were investigated earlier [15]. Write + (plus) when σn > 0 and − (minus) when σn < 0. Then the periodic trajectories in map (5) are described by words in the −+ alphabet. By Theorem 1, the periodic ground state configurations in map (5) are Hubbard configurations and vice versa. Besides, by Theorem 2 the devil’s staircase associated with map (5) is complete. Hence, Theorem 2 gives a new proof of the result obtained earlier [10, 18]. Transformation (5) is associated with selfsimilar structures [14] and deterministic chaos [11]. Changing the variables in Equation (5) by means of equations σn = 2(δn + ω − 1); ψ = 2ω − 1, and letting α go to zero (5) gives the map δ + ω − 1, if δn + ω > 1; δn+1 = n (6) if δn + ω < 1, δn + ω, where ω is a rotation number, 0 6 ω 6 1. Map (6) is the classical model of circle rotation through the angle ω. Write + (plus) if δn ∈ [0, ω) and − (minus)
PERIODIC GROUND STATE CONFIGURATIONS
105
otherwise. If ω is rational, then the sequence of letters −, + is periodic and it forms the so-called Sturmian chain [17]. The latter are equivalent to the Hubbard configurations. 5. Even 2-Colouring and the Set J The even 2-colouring problem has been widely discussed [2, 3, 16, 19, 21]. The problem is how to distribute a objects of the first kind (say, pluses) and b objects of the second kind (say, minuses) on a circle as evenly as possible. In the definition of the Hubbard configuration (Section 2), we found the requirement of the even distribution of plusses in words. The second description of the words in the −+ alphabet with the even distribution of plusses is given by the linear ordered set J [11]. To define J , we define the sequence of finite linear ordered sets Jn , each having words in the −+ alphabet as components. DEFINITION 4. (1) J0 = (−, +). The order in J0 : − < +. (2) Assume Jn be defined. Let the words A and B be arbitrary components of Jn , let A < B and let there be no C in Jn , such that A < C < B. Then the words A, B and AB are the components of Jn+1 and A < AB < B in Jn+1 . There are no other components in Jn+1 . DEFINITION 5. J is the linear ordered set consisting of all components of every set Jn (n > 0) in the order induced by the order in the sets Jn . The sets Jn are constructed by the operations of concatenation and insertion. J0 consists of two words: − (minus) and + (plus). We get J1 by inserting the concatenation of the − (minus) and + (plus) between the aforesaid words: J1 = (−, −+, +). Further, we insert the concatenation of the neighboring words between the neighbors: J2 = (−, − − +, −+, − + +, +), J3 = (−, − − −+, − − +, − − + − +, −+, − + − + +, − + +, − + ++, +), and so on. The sets Jn are the redefinitions of the Farey tree [7, 19]. The two definitions mentioned above are equivalent in a certain sense. THEOREM 3 [12]. For each word A in −+ alphabet the following two assertions are equivalent: (1) A is a Hubbard configuration (2) A is a shift of a nonzero degree of some component of J . EXAMPLE 3. The word − + − + + − + − + + − + + is simultaneously a Hubbard configuration and a component of J (namely, J5 ).
106
M. M. KIPNIS
6. Properties of the Words in J The number of the minuses and plusses in the word A we denote by q− (A) and q+ (A) accordingly. We call the word L the left neighbor of the word R in Jn , if L and R are components of Jn , L < R and there exist no M in Jn , such that L < M < R. For example, the word − + − + + is the left neighbour of − + + in J3 . If L is the left neighbor of R in Jn , then L and R are called the left and the right predecessors of the word LR accordingly. If L is the left neighbour of R in Jn (n > 0), then either R is the right predecessor of L, or L is the left predecessor of R. (For example, the word − + + is the right predecessor of − + − + +.) This dichotomy will be used in the forthcoming lemmas. LEMMA 1. If L is the left neighbor of R in Jn , then q− (L)q+ (R) − q+ (L)q− (R) = 1.
(7)
Proof (by induction on n). The case n = 0 is evident. Induction Step. Let the assertion be valid for n. Let L be the left neighbor of R in Jn+1 . Case 1: L is the left predecessor of R. Then L is a component of Jn and there exists a word R1 , such that L is the left neighbor of R1 in Jn and LR1 = R. By induction hypothesis q− (L)q+ (R1 ) − q+ (L)q− (R1 ) = 1. From the latter equality and equalities q− (R) = q− (R1 ) + q− (L) and q+ (R) = q+ (R1 ) + q+ (L), we get (7). Case 2: R is the right predecessor of L. The proof is similar. Lemma 1 is proved. 2 LEMMA 2. For each two pairs of nonnegative integers (a, b) and (c, d) if ad − bc = 1, then there exist the words L, R and the integer n, such that L is the left neighbor of R in Jn and q− (L) = a, q+ (L) = b, q− (R) = c, q+ (R) = d. Proof (by induction on b + d). The case b + d = 1: since ad − bc = 1, we get b = 0, d = 1. Put L = −, R = +, n = 0. Induction Step. Suppose b + d > 1 and for each quadruple (a1 , b1 , c1 , d1 ) if b1 + d1 < b + d, then the assertion is true. Case 1: b = d. Since ad − bc = 1, we get b(a − c) = 1, hence b = d = 1, a = c + 1. Put L = −a +; R = −c +; then L is the left neighbor of R in Ja . Case 2: b > d. Since ad−bc = 1, we have a > c. Then a−c > 0 and b−d > 0. Take two pairs (a − c, b − d) and (c, d). We have (a − c)d − (b − d)c = 1. The sum of the second components of the pairs in question is less than b + d, hence by the induction hypothesis, there exist two words L1 , R1 and a natural number n, such that L1 is the left neighbor of R1 in Jn , and q− (L1 ) = a − c, q+ (L1 ) = b − d, q− (R1 ) = c, q+ (R1 ) = d. Then the word L1 R1 is the left neighbor of R1 in Jn+1 . Put L = L1 R1 , R = R1 and we are done. Case 3: b < d. The proof is similar to that of case 2. Lemma 2 is proved. 2
107
PERIODIC GROUND STATE CONFIGURATIONS
THEOREM 4. (1) If A is a word in J , then q− (A) and q+ (A) are coprime numbers. (2) If A < B in J , then ω(A) < ω(B) (recall that ω(A) is the rotation number, ω(A) = q+ (A)/(q− (A) + q+ (A))). (3) For each pair of coprime natural numbers (a, b) there exists the only word A in J , such that q− (A) = a, q+ (A) = b. Proof. (1) Part (1) is an evident consequence of Lemma 1. (2) Let A < B in J , then there exists n, such that A and B are in Jn . If A is the left neighbor of B, then the proof is attained by Lemma 1; otherwise there exist the words Ai (1 6 i 6 m) in J , such that A is the left neighbor of A1 , Ai is the left neighbor of Ai+1 (1 6 i 6 m − 1), Am is the left neighbor of B. By Lemma 1 for every i (1 6 i 6 m − 1) ω(A) < ω(Ai ) < ω(Ai+1 ) < ω(B), as required. (3) Let a, b be coprime numbers. Construct numbers c, d, such that ad−bc = 1, and by Lemma 2 we get the word A, such that A is in J and q− (A) = a, q+ (A) = b. The uniqueness of the word A is followed from part (2) of Theorem 4 by the linear ordering of J . Theorem 4 is proved. 2 EXAMPLE 4 (to Theorem 4). (1) The word A = − + − + − + + is a component of J . We have q− (A) = 3, q+ (A) = 4, and 3, 4 are coprime numbers. (2) A = − + − + − + + < − + − + + = B in J , and ω(A) = 4/7 < ω(B) = 3/5. (3) There exists a unique word A in J , such that q− (A) = 11, q+ (A) = 71. The word A is (−+6 )2 − +7 (− +6 −+7 )4 .
7. Domain of the Periodic Ground State Configurations in Boolean Averaging System Let us introduce certain notations. Define the function sg: {−, +} → {−1, 1} by the equations sg(−) = −1, sg(+) = 1. Given a sequence γ = (γn ) and integers P j j j, p (1 6 j 6 p) we define Kp (γ ) by the equation Kp (γ ) = ∞ m=0 γj +mp . j +p j For each integer j we put Kp (γ ) = Kp (γ ). Let A = ε1 . . . εp , where εi = − or εi = + (1 6 i 6 p). Define F (γ , A) =
p X
Kpp−i (γ ) sg εi
=
p X
i=1
Kpi (γ ) sg εp−i .
i=1
Let Shiftp(A) (respectively, Shiftn(A)) stand for the set of the shifts of A, which end with the letter + (plus) (respectively, with − (minus)). THEOREM 5 (Domain Theorem). Let Shiftp(A) 6= φ and Shiftn(A) 6= φ. Then the word A is a periodic ground state configuration in Boolean averaging system (3) with the given ψ, if and only if min
B∈Shiftn(A)
F (γ , B) > ψ >
max
B∈Shiftp(A)
F (γ , B).
(8)
108
M. M. KIPNIS
Proof. (1) Let A = ε1 . . . εp , where εi = − or εi = + (1 6 i 6 p), and A be a periodic ground state configuration in Boolean averaging system (3) with the given ψ. Let (un ) be a bilateral sequence (n ∈ Z) satisfying Equation (3) and the conditions (1), (2) of Definition 1. Then un+mp = sg εn for all integers m and = A, An = 1 6 n 6 p. Denote the various shifts of A: A1 = ε2 . . . εp ε1 , Ap P εn+1 . . . εp ε1 . . . εn (2 6 n 6 p − 1). Introduce the notation σn = ψ − ∞ i=1 γi un−i (1 6 n 6 p). With 1 6 n 6 p we get σn = ψ −
p X
sg εi Kpn−i (γ ) = ψ − F (γ , An ).
i=1
If εn = + (1 6 n 6 p) then by part (1) of Definition 1 σn = ψ − F (γ , An ) > 0; if εn = − (1 6 n 6 p) then σn = ψ − F (γ , An ) < 0. Hence, we get (8). condition (8) be hold. Then put for each integer n σn = ψ − Pp(2) Let the n−i sg ε K i p (γ ) and un = sgn σn . If 1 6 n 6 p, then σn = ψ − F (γ , An ). i=1 By (8) with εn = + (1 6 n 6 p), we have σn = ψ − F (γ , An ) > 0 and un = 1, with εn = − (1 6 n 6 p) we have σn = ψ − F (γ , An ) < 0 and un = −1. The sequence (σn ) is p-periodic by the p-periodicity of Kpi (γ ) as to the superscript, hence un+mp = un (1 6 n 6 p, m ∈ Z). The sequence (un ) satisfies Equation (3) because ψ−
∞ X
sgn σn−i γi = ψ −
p X
i=1
sg εi Kpn−i (γ ) = σn .
i=1
2
Theorem 5 is proved.
COROLLARY 1. The word A in the −+ alphabet is a periodic configuration in Boolean averaging system (3) if and only if min
B∈Shiftn(A)
F (γ , B) >
max
B∈Shiftp(A)
F (γ , B).
8. Proof of Theorem 2 8.1.
SUBSIDIARY ASSERTIONS TO THE PROOF OF THEOREM 2
We need some definitions and lemmas. Let A be a component of J , A 6= −, A 6= +. By + A (resp. by A− ) let us denote the word A where the first (resp., last) letter is replaced by the letter + (plus) (resp., − (minus)). By + A− denote the result of the two operations mentioned above. Define the functions λ and ρ over Jn by recursion on n. DEFINITION 6. (1) For words in J0 : λ(−) = −, λ(+) = 3 (3 is an empty word), ρ(−) = 3, ρ(+) = +.
PERIODIC GROUND STATE CONFIGURATIONS
109
(2) Let λ and ρ be defined in Jn . Let A be a component of Jn+1 and L, R be components of Jn , L and R be the left and the right predecessors of A respectively. Case 1: L is the left predecessor of R. Take λ(A) = λ(LR) = Lλ(R), ρ(A) = ρ(LR) = ρ(R). Case 2: R is the right predecessor of L. Take λ(A) = λ(LR) = λ(L), ρ(A) = ρ(LR) = ρ(L)R. EXAMPLE 5. Let A = − + + − + + − + + + − + + − + + +. Then λ(A) = − + + − + + −, ρ(A) = + + + − + + − + ++. LEMMA 3. If the word A is a component of Jn (n > 1) and L, R are the left and the right predecessors of A respectively, then ρ(A) = + L, λ(A) = R − , ρ(A)λ(A) = + A− . Proof (by induction on n). The case n = 1 is obvious. Induction Step. Suppose the assertion be valid for n. Let A be a component of Jn+1 (n > 1), let L and R be the left and the right predecessors of A, respectively. Case 1: R is the right predecessor of L. Then L = L1 R, where L1 , R are the components of Jn−1 . In this case λ(LR) = λ(L), ρ(LR) = ρ(L)R. Hence, ρ(A) = ρ(L)R = + L1 R = + L; λ(A) = λ(LR) = λ(L) = R − . We’ve used the equalities ρ(L) = + L1 and λ(L) = R − which are valid by the induction hypothesis. Thus, ρ(A)λ(A) = + LR − = + A− . Case 2: L is the left predecessor of R. The proof is similar. Lemma 3 is proved. 2 LEMMA 4. If the word A is a component of J and A 6= −, A 6= +, then + A− ∈ Shiftn(A). Proof follows immediately from Lemma 3 and the evident equality λ(A)ρ(A) = A. Lemma 4 is proved. 2 DEFINITION 7. Define the function g(i, A) for the words A in the −+ alphabet and the integers i (1 6 i 6 q+ (A)). Let us number the positions of the letters in the word A. Let the position of the extreme right letter of A be 0; the other positions of the letters from right to left are the integers from 1 to q− (A) + q+ (A) − 1. If A = B− for some B and 1 6 i 6 q+ (A), then g(i, A) is the number of the position of ith plus from the right edge of A. If A = B+ for some B and 1 6 i < q+ (A), then g(i, A) is the number of the position of (i + 1)th plus from the right edge of A. Besides, if A = B+ and i = q+ (A), then g(i, A) = q− (A) + q+ (A). EXAMPLE 6. If A = + + − + − + −, then g(1, A) = 1, g(2, A) = 3, g(3, A) = 5, g(4, A) = 6. If A = − + − + − + +, then g(1, A) = 1, g(2, A) = 3, g(3, A) = 5, g(4, A) = 7. LEMMA 5. For each word A in Jn (n > 0, A 6= −) and for each natural number m g(i, Am ) = [i + iq− (A)/q+ (A)] (1 6 i 6 q+ (Am )).
110
M. M. KIPNIS
Proof (by induction on n) is being demonstrated for m = 1. The expansion on arbitrary m is obvious. The case n = 1 is evident. Induction Step. Suppose the lemma is valid for n. Let L be the left neighbor of R in Jn . Denote q− (L) = a, q+ (L) = b, q− (R) = c, q+ (R) = d. We have g(i, L) = [i + ia/b] (1 6 i 6 b), g(i, R) = [i + ic/d] (1 6 i 6 d). We aim to show that g(i, LR) = [i + i(a + c)/(b + d)] (1 6 i 6 b + d). In view of ad − bc = 1 (Lemma 1), if 1 6 i 6 d, then 0 < i(a + c)/(b + d) − ia/b = i/(d(b + d)) 6 1/(b + d). Hence, [ic/d] = [i(a + c)/(b + d)] and g(i, LR) = g(i, R) = [i + i(a + c)/(b + d)], as required. If d < i 6 b + d, then g(i, LR) = c + d + g(i − d, L) = c + d + [i − d + (i − d)a/b]. Hence, by 0 < i(a + c)/(b + d) − c − (i − d)a/b = (b + d − i)/(b(b + d)) 6 1/(b + d) (the latter equality is the concequence of ad − bc = 1 as well) we have g(i, LR) = [i + i(a + c)/(b + d)], as required. Lemma 5 is proved. 2 LEMMA 6. For any word A (A 6= −n , A 6= +n , n > 0) in the −+ alphabet there exist the words B ∈ Shiftp(A) and C ∈ Shiftn(A) such that for each i (1 6 i 6 q+ (A)) g(i, B) 6 i + [iq− (A)/q+ (A)]; g(i, C) > i − 1 + iq− (A)/q+ (A). Proof. Given the word A we define the number j by max (g(i, A) − i − iq− (A)/q+ (A)) = g(j, A) − j − j q− (A)/q+ (A). (9)
16i6q+ (A)
Let the word B be a shift of A, such that the j th plus of A is the rightmost letter of B, i.e. ( g(j + k, A) − g(j, A), if 1 6 k 6 q+ (A) − j ; (10) g(k, B) = g(k − q+ (A) + j, A) − g(j, A) + q− (A) + q+ (A), if q+ (A) − j < k 6 q+ (A). Then B ∈ Shiftp(A). When 1 6 k 6 q+ (A) − j , we get by (9) g(j, A) − j − j q− (A)/q+ (A) > g(j + k, A) − (j + k) − (j + k)q− (A)/q+ (A). Hence, by (10) g(k, B) = g(j +k, A)−g(j, A) 6 k+kq− (A)/q+ (A), as required. When q+ (A)− j < k 6 q+ (A), we get by (9) and (10) similarly g(k, B) 6 k + kq− (A)/q+ (A). The word C we construct analogously. Lemma 6 is proved. 2 LEMMA 7. If A is a component of J , A 6= −, A 6= +, and B ∈ Shiftp(A), C ∈ Shiftn(A), then for each i (1 6 i 6 q+ (A)) g(i, B) > g(i, A);
g(i, C) 6 g(i, + A− ).
Proof. By Lemma 5 for each i (1 6 i 6 2q+ (A)) g(i, A2 ) = i + [iq− (A)/q+ (A)].
(11)
PERIODIC GROUND STATE CONFIGURATIONS
111
Let us assume that B ∈ Shiftp(A). Then B is the subword of A2 . Let the last letter of B be in the position g(j, A2 ) in the word A2 . Then for any i (1 6 i 6 q+ (A)) by (11) we get g(i, B) = g(i + j, A2 ) − g(j, A2 ) = i + [(i + j )q− (A)/q+ (A)] − [j q− (A)/q+ (A)] > i + [iq− (A)/q+ (A)] = g(i, A). Let us assume that C ∈ Shiftn(A). If q+ (A) = 1, then the conclusion of the Lemma follows from the inequality g(q+ (A), C) 6 q+ (A) + q− (A) − 1 = g(q+ (A), + A− ). Now, consider the variant q+ (A) > 1. The word C is a subword of A2 . We may assume, without loss of generality, that the first letter to the right of the subword C in the word A2 is the letter + (plus). Let the position of this letter in A2 be g(j, A2 ). Then for any i (1 6 i 6 q+ (A) − 1) by (11) and Definition 7 g(i, C) = g(i + j, A2 ) − g(j, A2 ) − 1 = i − 1 + [(i + j )q− (A)/q+ (A)] − [j q− (A)/q+ (A)] 6 i + [iq− (A)/q+ (A)] = g(i, + A− ). Besides, if i = q+ (A), we have g(i, C) 6 q+ (A) + q− (A) − 1 = g(i, + A− ). Lemma 7 is proved.
2
LEMMA 8. If A, B are components of J and A < B in J , then under the conditions of Theorem 2 F (γ , + A− ) < F (γ , B). Proof. Let the conditions of the lemma be valid. Without loss of generality, we may assume that A is the left neighbor of B in some set Jk (k > 0). Hence, by Lemma 1 q− (A)q+ (B) − q− (B)q+ (A) = 1.
(12)
Then there exist integers m, n (m > 0, n > 0), such that q− (Am ) + q+ (Am ) = q− (B n ) + q+ (B n ) = p. By (12) q+ (Am ) < q+ (B n ). By Lemma 5 if 1 6 i 6 q+ (B n ) then g(i, B n ) = i + [iq− (B)/q+ (B)]. Besides, if 1 6 i 6 q+ (Am ) and i/q+ (A) 6∈ N then by Lemma 5 g(i, (+ A− )m ) > g(i, B n ). Consider the case i/q+ (A) = r ∈ N and 1 6 i 6 q+ (B n ). By Lemma 5 and by (12) g(i, B n ) = i + [iq− (B)/q+ (B)] = rq+ (A) + [rq− (A) − r/q+ (B)] 6 rq+ (A) − 1 + rq− (A) = g(i, (+ A− )m ). So, for each i (1 6 i 6 q+ (Am ) < q+ (B n )) g(i, (+ A− )m ) > g(i, B n ). Then by the equalities F (γ , (+ A− )m ) = F (γ , +A− ) and F (γ , B n ) =
112
M. M. KIPNIS j
F (γ , B) and by the monotony of Kp (γ ) as to superscribe with 1 6 j 6 p (j ∈ Z) we get F (γ , + A− ) = F (γ , (+ A− )m ) q+ (Am )
X
= 2
Kpg(i,(
+ A− )m )
(γ ) − K(γ )
i=1 q+ (B n )
X
< 2
Kpg(i,(
+ A− )m )
(γ ) − K(γ )
i=1 q+ (B n )
6 2
X
n
Kpg(i,B ) (γ ) − K(γ ) = F (γ , B n ) = F (γ , B).
i=1
2
Lemma 8 is proved.
8.2.
PROOF OF PART (1) OF THEOREM 2
Evidently, the word + (plus) is a periodic ground state configuration in Boolean P∞ averaging system (3) with the given ψ if and only if ψ > i=1 γi = K(γ ). For − (minus) we have analogously ψ < −K(γ ). Now, consider the components of J of length more than 1. Let A = ε1 . . . εp , where εi = − or εi = + (1 6 i 6 p). We have F (γ , A) =
p X
X
q+ (A)
Kpp−i (γ ) sg εi
=2
i=1
Kpg(i,A) (γ )
−
i=1
p X
Kpi (γ ).
(13)
i=1 j
Because of the monotonicity of (γi ) we get Kpi (γ ) > Kp (γ ) when 1 6 i < j 6 p. Hence, for the word A of length greater than 1 in J by Lemma 7 F (γ , A) =
max
B∈Shiftp(A)
F (γ , B);
F (γ , + A− ) =
min
B∈Shiftn(A)
F (γ , B);
(14)
(Recall that + A− ∈ Shiftn(A)). By (13) we get F (γ , + A− ) = F (γ , A) + 2(Kpp−1 (γ ) − Kpp (γ ) > F (γ , A).
(15)
Hence by (14), (15) and the corollary of Theorem 5 the domain of A is nonempty. Part (1) of Theorem 2 is proved. 8.3.
PROOF OF PART (2) OF THEOREM 2
Let the periodic ground state configuration A in Boolean averaging system (3) not be a Hubbard configuration. By part (3) of Theorem 4 there exists a word B, such that B is a component of J and q+ (A)q− (B) = q+ (B)q− (A). Then q+ (B n ) =
113
PERIODIC GROUND STATE CONFIGURATIONS
q+ (A), q− (B n ) = q− (A), where n = q+ (A)/q+ (B). By Lemma 6 we’ll find C ∈ Shiftp(A), such that for each i (1 6 i 6 q+ (A)) g(i, C) 6 i + [iq+ (A)/q− (A)] = g(i, B n ) (the latter equality is valid by Lemma 5). Besides, there exists i (1 6 i 6 q+ (A)), such that we have a strict inequality in the latter inequality; otherwise the words C and A would be Hubbard configurations. Hence, X
q+ (A)
F (γ , C) = 2
Kpg(i,C) (γ ) − K(γ ) > F (γ , B n )
i=1
X
q+ (A)
= 2
n
Kpg(i,B ) (γ ) − K(γ ).
i=1
Similarly we’ll find D ∈ Shiftn(A), such that F (γ , D) < F (γ , (+ B − )n ). By Theorem 5 the domain of A is defined by the inequalities F (γ , (+ B − )n ) > F (γ , D) >
min
E∈Shiftn(A) n
F (γ , E) > ψ >
> F (γ , C) > F (γ , B ).
max
E∈Shiftp(A)
F (γ , E) (16)
The domain of B n coincides with that of B and is defined by the inequalities (see Theorem 5 and Equation (14)) F (γ , (+ B − )n ) > ψ > F (γ , B n ).
(17)
In (16) and (17), we proceed from the fact that for any word E in the −+ alphabet and for any integer n (n > 0) F (γ , E n ) = F (γ , E). Comparing the domains of ψ, defined by inequalities (16) and (17), we get what is required. Part (2) of Theorem 2 is proved. 8.4.
PROOF OF PART (3) OF THEOREM 2
By Theorem 5, Lemma 8 and part (1) of Theorem 2 the domain of the periodic ground state configurations in J is a disjoint system of intervals on the ψ axis. By part (2) of Theorem 2 the set P is a complement to the system mentioned above on the ψ axis. The intervals (−∞; −K(γ )) and (K(γ ); ∞) in the ψ axis are the domains of the periodic ground state configurations − (minus) and + (plus) respectively, which are the components of J0 . The interval [−K(γ ), K(γ )] in the ψ axis is being divided in three parts: [−K(γ ); K22 (γ ) − K21 (γ )], (K22 (γ ) − K21 (γ ); K21 (γ ) − K22 (γ )), [K21 (γ ) − K22 (γ ); K(γ )]. The middle part is the domain of the periodic ground state configuration −+ in J1 . Furthermore, the domains of − − + and − + + are being deleted respectively from the left-handed and right-handed intervals mentioned above. Thus we have deleted the domains of the components of J2 . This procedure goes on according to the definition of Jn . So we have constructed the Cantor perfect set P . Let us prove that the total length of the deleted intervals is equal to 2K(γ ). True, any periodic configuration in
114
M. M. KIPNIS
J , consisting of p letters, is located in the ψ axis in the interval of the length p−1 p 2(Kp (γ ) − Kp (γ )) (Theorem 5 and Equations (14), (15)). By Theorem 4 there are ϕ(p) components of J of the length p (ϕ is a number-theoretic Euler function, ϕ(p) is the number of integers k with 1 6 k 6 p and k relatively prime to p). So, P p−1 p we must prove that ∞ p=2 ϕ(p)(Kp (γ ) − Kp (γ )) = K(γ ). Indeed, ∞ X
ϕ(p)(Kpp−1 (γ ) − Kpp (γ ))
p=2
=
∞ X
ϕ(p)
p=2
= γ1 +
∞ X
(γmp+p−1 − γmp+p )
m=0 ∞ X n=2
γn
X
p|(n+1),p>1
ϕ(p) −
X p|n,p>1
X ∞ ϕ(p) = γn = K(γ ). n=1
In this chain P of equalities, we have used the Gauss theorem about the Euler function [7]: p|n,p>1 ϕ(p) = n − 1. Part (3) of Theorem 2 is proved. Acknowledgement The author would like to thank B. Slepchenko and M. Zelikin for useful discussions. References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11.
Bak, P. and Bruinsma, R.: One-dimensional Ising model and the complete devil’s staircase, Phys. Rev. Lett. 49 (1982), 249–251. Bernoulli, J., III.: Sur une nouvelle espèce de calcul, Recueil pour les astronomes (Berlin), Vol. 1, 1772, pp. 255–284. Burkov, S. and Sinay, Ya.: Phase diagrams of one-dimensional lattice models with long-range antiferromagnetic interactions, Russian Math. Surveys 38(4) (1983), 235–257. Delchamps, D.: Nonlinear dynamics of oversampling A-to-D converters, Proc. 32nd IEEE CDC, San-Antonio, 1993. Feely, O. and Chua, L.: The effect of Integrator leak in Sigma-delta modulation, IEEE Trans. Circuits Systems 38 (1991), 1293–1305. Gelig, A. and Churilov, A.: Stability and Oscillations in Nonlinear Pulse-modulated Systems, Birkhäuser, Basel, 1998. Hardy, G. and Wright, E.: Introduction to the Theory of Numbers, Clarendon Press, Oxford, 1976. Hubbard, J.: Generalized Wigner lattices in one dimension and some applications to tetracianoquinodimethane (TCNQ) salts, Phys. Rev. B. 17 (1978), 494–505. Jury, E.: Sampled-data Control Systems, Wiley, New York, 1958, 2nd edn, Krieger, 1977. Kieffer, J. C.: Analysis of dc input response for a class of one-bit feedback encoders, IEEE Trans. Comm. 38(3) (1990), 337–340. Kipnis, M. M.: Symbolic and chaotic dynamics of a pulse-width control system, Soviet Phys. Dokl. 324(2) (1992), 273–276.
PERIODIC GROUND STATE CONFIGURATIONS
12. 13.
14. 15. 16. 17. 18. 19. 20. 21.
115
Kipnis, M. M.: On the formalizations of the even 2-colouring, Proc. Chelyabinsk Pedagogical Univ. Series 4. Natural Sciences 1 (1996), 96–104. Kipnis, M. M.: One-dimensional model of statistical mechanics with the Hubbard Hamiltonian and the interaction function, free from the convexity condition, Phys. Dokl. 336(3) (1994), 316–319. Kipnis, M. M.: Boolean Averaging in a statistical mechanics model and in an analog-to-digital converter, Russian J. Math. Phys. 14(3) (1996), 397–402. Leonov, N. N.: On the pointwise transformation of the line in itself, Izv. Vyssh. Uchebn. Zaved. Radiofiz. 2(6) (1959), 942–956 (in Russian). Markoff, A.: Sur une question de Jean Bernoulli, Mat. Ann. 19 (1882), 27–36. Morse, M. and Hedlund, G.: Symbolic dynamics II: Sturmian trajectories, Amer. J. Math. 62 (1940), 1–42. Park, S. and Gray, R.: Sigma-delta modulation with leaky integration and constant input, IEEE Trans. Inf. Theory 38 (1992), 1512–1533. Rockmore, D., Siegel, R., Tongring, N. and Tresser, C.: An approach to renormalization on n-torus, Chaos 1(1) (1991), 25–30. Siegel, R., Tresser, C. and Zettler, G.: A decoding problem in dynamics and in number theory, Chaos 2(4) (1992), 473–493. Smith, H. J. S.: Note on continued fraction, Messenger Math. VI (1877), 1–14.
Mathematical Physics, Analysis and Geometry 3: 117–138, 2000. © 2000 Kluwer Academic Publishers. Printed in the Netherlands.
117
Polynomial Asymptotic Representation of Subharmonic Functions in a Half-Plane P. AGRANOVICH Institute for Low Temperature Physics, Mathematical Division, Lenin Ave. 47, 61164 Kharkov, Ukraine (Received: 5 August 1999; in final form: 1 May 2000) Abstract. Let u(z) be a subharmonic function in a half-plane such that its Riesz measure is concentrated on the finite system of rays. In the paper the connection between the behavior of u(z) and the distribution of its measure (including boundary measure) is investigated in terms of polynomial asymptotic representations. Mathematics Subject Classifications (2000): 30E15, 31A05. Key words: half-plane, subharmonic function, measure, asymptotic representation.
1. Introduction The relation between the asymptotic distribution of zeros of a holomorphic function and the growth of this function at infinity is one of the most important questions of function theory. We will say that a function f (t), t > 0, has polynomial asymptotics if it can be represented in the following way: f (t) = 11 t ρ1 + 12 t ρ2 + · · · + 1n t ρn + κ(t),
t → ∞,
(1)
where 1j , j = 1, . . . , n, are real constants, 0 < [ρ1 ] < ρn < ρn−1 < · · · < ρ1 , and the last term on the right, i.e., the function κ(t), is small in a certain sense in comparison with the previous term. Similarly, we will understand the expression ‘polynomial asymptotics of a function f (z), z → ∞’. In this case, the coefficients 1j are functions of 0 = arg z and t = |z|. The theory of functions of completely regular growth establishes a close connection between the growth of an entire function and the distribution of its zeros by means of one-term asymptotic representations. This theory was constructed by B. Levin and A. Pflüger simultaneously and independently in the thirties and immediately found intensive use in different parts of mathematics and physics. Later, this theory?? was extended to other classes of functions, including holomorphic functions in a half-plane (N. Govorov [5]). Note that, in the latter case, new ?
? [a] is the integral part of a. ?? An extensive bibliography is in [9].
118
P. AGRANOVICH
effects appeared. They are connected with possible singularity of the holomorphic function on the boundary of the half-plane. In addition to such generalizations of this theory, a number of authors have considered problems relating to the connection between the behavior of a subharmonic function and the growth of its Riesz measure in terms of polynomial asymptotic representation. They have shown that it is not possible to use the methods of the function theory of completely regular growth to answer these questions ( see, for example, [4]). The relation between the existence of polynomial asymptotic representations of a subharmonic function in the plane and its Riesz measure was investigated in [1–3,6,7]. In this paper, we consider this problem for a subharmonic function in a halfplane, where the Riesz masses are concentrated on a finite system of rays. For the formulation of the main results, we need the following notations and definitions. Let Uρ be the class of subharmonic functions in the upper half-plane C+ = {z : Imz > 0}, which are bounded from above on every semi-circle CR = {z : z ∈ C+ , |z| < R}, 0 < R < ∞, and have noninteger order ρ in C+ , and where the Riesz measures are concentrated on the positive ray of the imaginary axis. Recall that the order ρu of a subharmonic function u(z) in C+ is the following value (see, for example [9]): ρu = lim sup r→∞
ln+ max{81 (r), 82 (r)} , ln r
where 81 (r) = sup u(reiθ ) 0<θ<π
and
Z
π
82 (r) =
|u(reiθ )| sin θ dθ. 0
If u(z) ∈ Up , then the distribution of its Riesz measure µu can be characterized by the function τu (r) = µu (Ir ), where Ir = {iy : 0 < y 6 r}. In addition, we will use the so-called “boundary measure” µ∂,u defined as Z µ∂,u (φ) = lim u(x + ih)φ(x) dx, h→+0
where φ(x) is a test function on R and the limit is considered in the distribution space D 0 (R). It is known [9] that this limit exists and that it is a real measure (charge). As was shown in [5], the function u(z) ∈ Uρ1 can be represented in the form (p = [ρ1 ]) Z 1 1 1 u(z) = ImPp (z) − Im dµ∂,u + π −1 t − z
POLYNOMIAL ASYMPTOTIC REPRESENTATION OF SUBHARMONIC FUNCTIONS
119
z z − ln 1 + dτu (t) + + Re ln 1 − it it −1 Z ∞ z z + Re ln 1 − − ln 1 + + it it 1 j p X 1 z z j dτu (t) + − + j it −it j =1 Z 1 1 zp + Im − − · · · − p+1 dµ∂,u (t) t −z t t 1<|t |<∞ = U1 + U˜1 Z
1
for p > 1 or Z 1 1 1 dµ∂,u + Im u(z) = C − π −1 t − z Z 1 z z + − ln 1 + dτu (t) + Re ln 1 − it it 1 Z ∞ z z Re ln 1 − + − ln 1 + dτu (t) + it it 1 Z z + dµ∂,u (t) = U2 + U˜2 Im t (t − z) 1<|t |<∞ for p = 0, where Pp (z) is a polynomial of degree p. Since ρ1 is noninteger and the functions U1 , U2 are harmonic for |z| > 1, lim r −ρ1 |Uj (reiθ )| = 0,
j = 1, 2,
r→∞
uniformly for θ ∈ [0, π ]. Without loss of generality, we further consider only the case p > 1 and assume that τu (t) = 0, t ∈ (0, 1], so Z ∞ z z Re ln 1 − − ln 1 + + u(z) = it it 1 j p X z 1 z j + dτu (t) + − j it −it j =1 Z 1 zp 1 − − · · · − p+1 t dτ˜u (t), Im (2) + t −z t t 1<|t |<∞ where τ˜u (t) := τu+ (t) = −
1 2π
Z 1
t
1 dµ∂,u (x) x
120
P. AGRANOVICH
for t ∈ (1, ∞) and τ˜u (t) :=
τu− (t)
1 = 2π
Z
−1
−r
1 dµ∂,u (x) x
for t ∈ (−∞, −1). All results below will be formulated, without these additional restrictions. The smallness of the function κ(t) in (1) will be expressed in terms of the following definition: DEFINITION. We will say that a function f (t) satisfies the condition (q, ρ) if Z 2T |f (t)|q dt = o(T ρq+1 ), T → ∞, T
where ρ > 0 is any fixed number. Theorem 1, which will be formulated below, shows the existence of polynomial asymptotics of a subharmonic function u(z) if there are polynomial asymptotic representations of the functions τu , τu+ and τu− . THEOREM 1. Let u(z), z ∈ C+ , be a subharmonic function of the class Uρ1 , / Z, and let the support of its Riesz measure be contained in the ray {z : Re z = ρ1 ∈ 0, Im z > 0}. Suppose that τu (r) = τu+ (r) = τu− (r) =
n X
1j r ρj + 8(r),
j =1 n X
δj+ r ρj + γ + (r),
j =1 n X
(3)
δj− r ρj + γ − (r),
j =1
where p = [ρ1 ] < ρn < ρn−1 < · · · < ρ1 and the functions 8(r), γ ± (r) satisfy the condition (q, ρn ) for some fixed q > 1. Then there exists a C0,1 -set E ? such that, in any sector η 6 θ 6 π − η, 0 < η < π/2, u(z) has the representation n X π π r ρj 1j π iθ cos %j θ − − π − cos %j θ − − u(re ) = sin πρ 2 2 j j =1 δj− πρj r ρj δj+ πρj r ρj sin ρj (θ − π ) + sin ρj θ + − sin πρj sin πρj + κ(reiθ ),
(4)
? The set E is called a C 0,1 set if it can be covered by a system of circles Kj = {|z − zj | < rj } P such that limj →∞ R1 j :|zj |
POLYNOMIAL ASYMPTOTIC REPRESENTATION OF SUBHARMONIC FUNCTIONS
121
where lim
z→∞,z∈ /E
|z|ρn κ(z) = 0.
Moreover, if in the conditions (q, ρn ) for the functions 8(r), γ ± (r), the number q is strictly greater than 1, then the functions Z π κ(reiθ ) sin θ dθ sup |κ(reiθ )| and η6θ6π−η
0
satisfy the condition (q, ρn ). Remark 1. The assumption ρn > [ρ1 ] is a natural condition. Indeed, let ρn < [ρ1 ]. It is not difficult to show that one can slightly modify the Riesz measure of the subharmonic function u(z) without changing the asymptotics (3) and the conditions (q, ρn ) for the functions 8(r), γ ± (r) and change the asymptotics of u(z). In a certain sense, the following theorem is the converse to Theorem 1. THEOREM 2. Let u(z), z ∈ C+ , be a function of class Uρ1 , ρ1 ∈ / Z, such that its Riesz measure is concentrated on the positive ray of the imaginary axis. Assume equality (4) holds on the union 0 of the positive ray of the imaginary axis and sectors {0 < arg z < θ1 }, {θ2 < arg z < π }, 0 < θ1 < π/2 < θ2 < π , i.e. n X π π π r ρj 1j iθ cos %j θ − − π − cos %j θ − − u(re ) = sin πρj 2 2 j =1 δj+ πρj r ρj δj− πρj r ρj − sin ρj (θ − π ) + sin ρj θ + κ(reiθ ), sin πρj sin πρj where [ρ1 ] < ρn < ρn−1 < · · · < ρ1 ; 11 > 0 and for some fixed q ∈ (1, ∞) and η > 0 the functions iθ iθ κ˜ 1 (r) := sup |κ(re )|, sup |κ(re )|, |κ(ir)| , θ2 6θ6π−η
η6θ6θ1
and
Z κ˜ 2 (r) :=
θ1
Z
π
κ(re ) sin θ dθ + iθ
κ(reiθ ) sin θ dθ θ2
0
satisfy the condition (q, ρn ). Then the functions τu (r), τu+ and τu− in (2) have the asymptotic representations τu (r) =
n X j =1
1j r ρj + 8(r),
122
P. AGRANOVICH
τu+ (r) =
n X
δj+ r ρj + γ + (r),
j =1
τu− (r)
=
− X
δj− r ρj + γ − (r),
j =1
where the remainder terms 8(r) and γ ± (r) satisfy the condition (q, ρn ). Remark 2. The main difficulties in the proof of this theorem are connected with the presence of singularities of the subharmonic function u(z) on the boundary of the half-plane (see Theorem 3 in Section 2).
2. Proof of Theorem 1 Proof. ? As was noted above we may suppose the function u(z) to have representation (2). Integrating by parts, we obtain p n X 1 − itz 1 z j X1 z j τu (t)|R1 − u(z) = lim Re ln − z + R→∞ 1 + it j it j −it j =1 j =1 Z R 2izp+1 (z sin p π2 − t cos p π2 ) τu (t) dt + − t p+1 (t 2 + z2 ) 1 p+1 Z τ˜u (t) z R p+1 τ˜u (t)|1 + pz dt + + Im p p+1 t (t − z) (t − z) 1<|t |
z it z it
+
p X 1 z j
− j it R |t |=R p X zp+1 1 z j τu (t) +Im (p+1) τ˜u (t) = 0. − j −it t j =1 j =1
(5)
Since τu (1) = 0,
τu+ (1) = 0 and
τu− (−1) = 0,
? Only for simplicity in this paper we will consider the case of two-term asymptotic representa-
tion.
POLYNOMIAL ASYMPTOTIC REPRESENTATION OF SUBHARMONIC FUNCTIONS
123
then the expression in (5) is zero when |t| = 1. Thus Z ∞ z sin p π2 − t cos p π2 ) p+1 τu (t) dt + u(z) = −Re 2iz t p+1 (t 2 + z2 ) 1 Z Z τ˜u (t) τ˜u (t) p+1 p dt + Im z dt + p+1 (t − z) p 2 1<|t |<∞ t 1<|t |<∞ t (t − z) (6) = I1 + I2 . Let us consider the integral I1 . Substituting in I1 the expressions of τu (t) from (3) and integrating the main members with the help of residues, we obtain 2 X π r ρj 1j π π iθ cos ρj θ − − π − cos ρj θ − + I1 (re ) = sin πρj 2 2 j =1 + ψ1 (reiθ ),
(7)
where
Z ψ1 (re ) = Re 2i(reiθ )p+1 iθ
1
∞
t cos p π2 − reiθ sin p π2 8(t) dt . t p+1 (t 2 + r 2 e2iθ )
(8)
To estimate the function ψ1 (reiθ ), let us divide the ray [1, ∞) into intervals [2k , 2k+1 ), k = 0, 1, . . . . Let |z| ∈ [2k , 2k+1 ). We can represent the function ψ1 (z) as the sum of three terms: Z 2k−1 Z 2k+2 Z ∞ (k) (k) (k) +Re +Re = I1,1 + I1,2 + I1,3 , ψ1 (z) = Re 2k−1
1
2k+2
(k) ls taken as the principal value when Re z = 0. where I1,2 (k) (k) Estimates of I1,1 and I1,3 can be obtained with the help of the following lemma:
LEMMA A [7]. Let γ (t) be from the space Lq , q > 1, on any finite R the function −p−1 dt < ∞, where p > 0 is an integer. Suppose also interval of R and 0 |γ (t)|t that, for any T > 0 the function γ (t) satisfies the condition (q, ρ), 0 < p < ρ < p + 1. Then, for any fixed positive numbers k1 < 1 and k2 > 1 the asymptotic estimates Z A γ (t) = o(r ρ−p−1 ), r → ∞, sup dt sup p+1 iθ (t − re ) A∈[0,k1 r] 06θ62π 0 t and
sup
Z sup
B∈[k2 r,∞] 06θ62π
are valid.
∞
B
γ (t) dt = o(r ρ−p−1 ), p+1 iθ t (t − re )
r → ∞,
124
P. AGRANOVICH
From our hypothesis on 8, Lemma A and the fact that ρ2 ∈ ([ρ1 ], ρ1 ), we obtain (k) (k) + I1,3 ) → 0, |z|−ρ2 (I1,1
(9)
when |z| → ∞, uniformly for θ = arg z ∈ [0, π ]. (k) into two terms: Let us divide the expression I1,2 Z 2k+2 t cos p π2 − z sin p π2 (k) p+1 I1,2 = Re z 8(t) dt − t p+2 (z − it) 2k−1 Z 2k+2 t cos p π2 − z sin p π2 8(t) dt −Re zp+1 t p+2 (z + it) 2k−1 (k) (k) + I1,2,2 . = I1,2,1
Since z ∈ C+ , it is easy to see that (k) | = o(|z|ρ2 ), |I1,2,2
|z| → ∞.
(10)
(k) For the estimate of the integral I1,2,1 , we will use the following theorem about the ? Cauchy-type integrals:
THEOREM A. (a) For each q ∈ (1, ∞) there exists a constant Cq < ∞ such that, for each function g(t) ∈ Lq (R) and s ∈ R+ the function Z ∞ g(t) g(s) ˆ = sup dt {z:|Im(z−is)|>|Re(z−is)|} −∞ z − it satisfies the estimate kg(t)k ˆ Lq 6 Cq kgkLq . (b) There also exists a constant C < ∞ such that, for any function g(t) ∈ L1 (R) and any h > 0 mes{s : g(s) ˆ > h} 6 CkgkL1 h−1 . (k) consists of two terms: The integral I1,2,1 Z 2k+2 cos p π2 8(t) dt i1 = zp+1 p+1 (z − it) 2k−1 t
and
Z i2 = zp+2
2k+2
2k−1
sin p π2 8(t) dt. t p+2 (z − it)
To estimate the integral i1 , set cos pπ/28(t)t −(p+1), gk (t) = 0, t ∈ / [2k−1 , 2k+2 ).
t ∈ [2k−1 , 2k+2 ),
? In [2] this theorem was formulated for the half-plane {z : Re z > 0}.
125
POLYNOMIAL ASYMPTOTIC REPRESENTATION OF SUBHARMONIC FUNCTIONS
Since the function 8(t) satisfies the condition (q, ρ2 ), where q > 1 is fixed, we have that q k(ρ −p− q10 ) . ), k → ∞, q 0 = kgk kLq = o(2 2 q −1 From this and part (a) of Theorem A, for q > 1 we obtain Z 2k+2 sup |i1 |q ds 2k−1
6 =
Z
{z:|Im(z−is)|>|Re(z−is)|} ∞
sup
Z (k+1)(p+1) 2
∞
gk (t) q ds dt z − it
−∞ {z:|Im(z−is)|>|Re(z−is)|} −∞ q q (k+1)(p+1)q k(p+1)q 2 kgˆk kLq 6 Cq 2 kgk kLq = o(2k(ρ2 q+1) ),
k → ∞.
Now let {εk } be a sequence of positive numbers. Then the measure of the set k k+1 kρ2 ˜ Ek = s ∈ [2 , 2 ) : sup |i1 | > εk 2 {z:|Im(z−is)|>|Re(z−is)|}
satisfies −q mes E˜ k = εk o(2k ),
k → ∞.
If q = 1, then using part (b) of Theorem A with h = εk 2kρ2 , we obtain that mes E˜ k = εk−1 o(2k ),
k → ∞.
Thus, for any fixed q > 1 −q
mes E˜ k = εk o(2k ),
k → ∞.
Since the integral i2 has similar estimates, we conclude that if q > 1, then q Z 2k+1 (k) sup |I1,2,1 | ds = o(2k(ρ2 q+1) ), k → ∞. 2k−1
{z:|Im(z−is)|>|Re(z−is)|}
and for q > 1 the measure of the set sup Ek = s : [2k , 2k+1 ) :
{z:|Im(z−is)|>|Re(z−is)|}
(11)
(k) |I1,2,1 |
kρ2
> εk 2
−q
is εk o(2k ), k → ∞. If the sequence {εk } tends to zero S sufficiently slowly, then it is easy to see that the relative measure? of the set e = k Ek is zero. By comparing this with estimates (9) and (10) we conclude that if [ z∈ / E := {z : |Im(z − is)| > |Re(z − is)|} s∈e ? The relative measure of a set G ⊂ (0, ∞) is defined to be the limit lim −1 t →∞ t mes(G ∩ (0, t)).
126
P. AGRANOVICH
then r −ρ2 |ψ1 (reiθ )| → 0,
r→∞
uniformly for θ ∈ [0, π ], where ψ1 (z) is defined by (8). The relative S measure of E is zero. Evidently we can assume that the set E is open, so E = Ij where Ij is an interval. Let us consider squares with diagonals Ij , j = 1, . . . , and circumscribe circles around each such square. The union of these circles covers E and it is easy to see that their radii satisfy the condition from the definition of C0,1 -set. So the set E is C0,1 -set. If q > 1, then direct calculation shows that Z 2k+1 (k) iθ q sup |I1,2,1(re )| dr = o(2k(ρ2 q+1) ), k → ∞. 2k−1
θ∈[0, π4 ]∪[ 3π 4 ,π]
From this and (9), (10), (11) we obtain that Z 2k+1 sup |ψ1 (reiθ )|q dr = o(2k(ρ2 q+1) ), 2k−1
k → ∞,
06θ6π
and, hence, the function sup06θ6π |ψ1 (reiθ )| satisfies the condition (q, ρ2 ). Now let us consider the integral I2 (see (6)). From (3), integrating the main terms with the help of residues, we conclude that I2 =
2 X −π δ + ρj r ρj
sin πρj
j =1
+
2 X π δj− ρj r ρj j =1
where
sin πρj
sin ρj (θ − π ) + sin ρj θ + ψ2 (reiθ ),
(12)
Z ∞ Z ∞ γ + (t) γ − (t) p ψ2 (re ) = Im z p p dt + (−1) dt + t p+1 (t − z) t p+1 (t + z) 1 1 Z Z ∞ ∞ γ + (t) γ − (t) p dt + (−1) dt + t p (t − z)2 t p (t + z)2 1 1 4 X p+1 = Im z I2,j (z) . (13) iθ
p+1
j =1
Let η, 0√< η < π/2, be fixed. If z = reiθ , 0 < η 6 θ 6 π − η, then |t − z| > 2 tr sin η/2 and, hence, it is easy to see that |ψ2 (reiθ )|r −ρ2 → 0,
r → ∞.
POLYNOMIAL ASYMPTOTIC REPRESENTATION OF SUBHARMONIC FUNCTIONS
127
From here taking into account (7), (12) and that ψ1 satisfies condition (q, ρ2 ), we obtain (4) with the function κ(reiθ ) = ψ1 (reiθ ) + ψ2 (reiθ ) such that supη6θ6π−η |κ(reiθ )| satisfies condition (q, ρ2 ). For the R πcompletion of the proof of Theorem 1 we have to establish that the function 0 κ(reiθ ) sin θ dθ also satisfies the condition (q, ρ2 ). In fact, q Z 2T Z π κ(reiθ ) sin θ dθ dr T 0 q q Z 2T Z π Z 2T Z π iθ iθ 6 Bq ψ (re ) sin θ dθ dr + ψ (re ) sin θ dθ dr 1 2 T 0 T 0 with some constant Bq > 0. By virtue of the condition (q, ρ2 ) for the function sup06θ6π |ψ1 (reiθ )| it follows that q Z 2T Z π iθ ψ1 (re ) sin θ dθ dr = o(T ρ2 q+1 ), T → ∞. (14) T 0 Also, from Theorem A (the case of the half-plane {z : Re z > 0}), we have (see (13)) q Z 2T Z π dr r p+1 ei(p+1)θ )I2,j (reiθ ) dθ = o(T ρ2 q+1 ), T → ∞, j = 1, 2. 0 T The integrals I2,3 and I2,4 can be estimated in the same way, so we will consider only one of them, for example I2,3 . Let us introduce the function + γ (r) sin θei(p+1)θ , r ∈ [1, ∞), θ ∈ [0, π ], + iθ γ˜ (re ) = 0, otherwise. It is easy to see that γ˜ + (reiθ ) ∈ Lq (C) and kγ˜ + (reiθ )kLq (C) = o(r ρ2 q+1 ), q
Then Z π
r → ∞. Z
sin θr
p+1 i(p+1)θ
e
I2,3 (re ) dθ = r iθ
p+1 C
0
γ˜ + (teiθ ) dλ, t p+1 (t − reiθ )2
where dλ is the Lebesgue measure on C. It is clear that this integral is the Beurling transformation [10] of the function γ˜ + (teiθ )t −(p+1) and, hence, for q > 1 q
kr p+1 I2,3 (reiθ )kLq {T 6r62T } = o(T ρ2 q+2 ),
T → ∞.
128
P. AGRANOVICH
So in view of (13), we have q Z 2T Z π iθ ψ2 (re ) sin θ dθ dr = o(T ρ2 q+1 ), T 0
T → ∞, q > 1. 2
Hence, by virtue of (14), we have finished the proof of Theorem 1. 3. Proof of Theorem 2 For the proof of this statement, we need the following theorem.
/ THEOREM 3. Let u(z), z ∈ C+ , be a subharmonic function of the class Uρ1 , ρ1 ∈ Z, and let its Riesz measure be concentrated on the ray {z : Re z = 0, Im z > 0}. In some sector let Y (θ1 , θ2 ) = {z : θi 6 arg z 6 θ2 , 0 < θ1 < π/2 < θ2 < π } and the function u(z) have the asymptotic representation (4) with the remainder term κ(reiθ ) such that supθ1 6θ6θ2 |κ(reiθ )| satisfies the condition (q, ρn ). Then τu may be represented as τu (t) =
n X
1j t ρj + 8(t),
j =1
where the function 8(t) satisfies the condition (q, ρn ). The proof of this theorem is analogous to the proof of Theorem 6 of [2]. Proof of Theorem 2. First we consider the sector Y1 = {z : θ1 6 arg z 6 π/2}. Under the conformal mapping z → (zeiθ1 )π/(π/2−θ1) (so that θ → ψ ≡ π(θ − θ1 )/((π/2) − θ1 )) the sector Y1 is transformed into the upper half-plane and the function u(z) turns into a function v(z) of order ρ1 /π((π/2) − θ1 ) which is harmonic in the upper half-plane. Without loss of generality we can assume that ρ2 /π((π/2) − θ1 ) is noninteger. Otherwise we will take ρ2 π 0 0 0 /Z θ1 , 0 < θ1 < θ1 , such that − θ1 ∈ π 2 and will do all further reasonings for this θ10 . It is easy to see that the hypotheses of Theorem 2 yield that corresponding to θ = θ1 , θ = π/2 ρj π 2 X π π π 1j t π ( 2 −θ1 ) v(t) = cos ρj θ1 + − cos ρj θ1 − + sin πρ 2 2 j j =1 +
δj+ πρj t
ρj π
( π2 −θ1 )
sin πρj 1 π
+ κ(t π ( 2 −θ1 ) eiθ1 ),
sin ρj (π − θ1 ) + t > 0;
δj− πρj t
ρj π
( π2 −θ1 )
sin πρj
sin ρj θ1 +
(151 )
POLYNOMIAL ASYMPTOTIC REPRESENTATION OF SUBHARMONIC FUNCTIONS
129
and v(−t) =
ρj π 2 X π 1j t π ( 2 −θ1 )
j =1
sin πρj
(cos ρj π − 1) +
ρj π πρj t π ( 2 −θ1 ) π + − + sin ρj (δj + δj ) + sin πρj 2
+ κ(t ( 2 −θ1 ) π ei 2 ), π
π
1
t > 0;
(152 )
Nevanlinna’s formula ([8], p. 34) implies that s X
1 v(z) = ak r sin kθ + π k=1 k
zs+1 v(t)Im s+1 dt + O(1), t (t − z) |t |>1
Z
where
ρ1 π s= − θ1 , π 2
z = reiθ1 .
Substituting expressions (151 ) and (152 ) into this, we obtain v(reiθ1 ) =
s X k=1
ak0 r k sin kθ +
ρj π 2 X π 1j r π ( 2 −θ1 )
π 1 π cos ρj +θ − θ1 + θ1 − + sin πρj 2 2 π j =1 π π 1 − cos ρj θ − θ1 + θ1 − − 2 π 2 ρj π δj+ πρj r π ( 2 −θ1 ) π 1 − sin ρj θ − θ1 + θ1 − π + sin πρj 2 π ρj π δj− πρj r π ( 2 −θ1 ) π 1 + I + O(1), + − θ1 + θ1 sin ρj θ sin πρj 2 π where the remainder I is Z 1 ∞ zs+1 ( π2 −θ1 ) π1 iθ κ(t e )Im s+1 dt+ π 1 t (t − z) Z 1 −1 zs+1 π 1 π dt. + κ(|t|( 2 −θ1 ) π ei 2 )Im s+1 π −∞ t (t − z)
130
P. AGRANOVICH π
Define the function κ(t) ˆ as κ(t ( 2 −θ1 ) π eiθ1 )t −(s+1) when t > 1 and π 1 π κ(|t|( 2 −θ1 ) π ei 2 )t −(s+1) when t 6 −1. Then, by virtue of the condition (q, ρ2 ) for κ we have Z π 1 q |κ(t)| ˆ dt = o(T ρ2 ( 2 −θ1 ) π −s−1)q+1), T → ∞, (16) 1
T 6|t |62T
Let us estimate the remainder term I using (16). For this we represent I as the sum of three integrals: Z s+1 Z Z 1 z dt κ(t)Im ˆ + + I = |z| π 16|t |6 |z|2 t − z 6|t |62|z| 2|z|6|t |6∞ 2 = I1 + I2 + I3 . From Theorem A (for the case of the half-plane {z : Re z > 0}) it follows that Z 2T ρ2 π sup |I2 (reiθ )|q dr = o(T π ( 2 −θ1 )q+1 ), T → ∞, (17) T
06θ6π
Since s < ρ2 /π( π2 − θ1 ), then by virtue of Lemma A ρ2 π
|z|− π ( 2 −θ1 ) (|I1 | + |I3 |) → 0, when |z| → ∞. From this and (17) we conclude that Z 2T ρ2 π sup |I (reiθ )|q dr = o(T π ( 2 −θ1 )q+1 ), T → ∞, T
(18)
06θ6π
P 0 k Moreover, it is clear that we can adjoin the sum m k=1 ak r sin kθ to the remainder term I without changing its estimate. Returning to the original function u(z), we deduce that for θ1 6 θ 6 π/2 2 X π π π r ρj 1j − cos ρj θ − − cos ρj θ + u(re ) = sin πρj 2 2 j =1 iθ
−
δj+ πρj r ρj sin πρj
sin ρj (θ − π ) +
δj− πρj r ρj sin πρj
sin ρj θ + κ1 (reiθ ), (191 )
and by virtue of (18) the function supθ1 6θ6 π |κ1 (reiθ )| satisfies the condition 2 (q, ρ2 ). Now let us consider the sector Y2 = {z : π/2 6 arg z 6 θ2 } and repeat the reasonings which we have made for the sector Y1 . Then we obtain that, in Y2 , 2 X 3π π π r ρj 1j − cos ρj θ − − cos ρj θ − u(re ) = sin πρj 2 2 j =1 iθ
POLYNOMIAL ASYMPTOTIC REPRESENTATION OF SUBHARMONIC FUNCTIONS
−
δj+ πρj r ρj sin πρj
sin ρj (θ − π ) +
δj− πρj r ρj sin πρj
131
sin ρj θ + κ2 (reiθ ), (192 )
and the function sup π 6θ6θ2 |κ2 (reiθ )| satisfies the condition (q, ρ2 ). 2 So in the whole half-plane C+ , in view of (191 ), (192 ) and the conditions of Theorem 2, the function u(z) can be represented in the following form: 2 X π π r ρj 1j π iθ u(re ) = cos ρj θ − − π − cos ρj θ − − sin πρj 2 2 j =1 −
δj+ πρj r ρj sin πρj
sin ρj (θ − π ) +
δj− πρj r ρj sin πρj
sin ρj θ + κ(re ˜ iθ ), (190 )
where for q ∈ (1, ∞) the functions Z π iθ ˜ )| and κ(re ˜ iθ ) sin θ dθ sup |κ(re η6θ6π−η
0
satisfy the condition (q, ρ2 ). Using Theorem 3, we now conclude that τu (t) = 11 t ρ1 + 12 t ρ2 + 8(t)
(20)
and the function 8(t) satisfies the condition (q, ρ2 ). Let us now consider in the half-plane C+ the subharmonic function u1 (z) of the class Uρ1 , which has the Riesz masses the same as u(z) and with the distribution of the boundary measure τu±1 (t) ≡ 0. Then in view of Theorem 1 and (20), we have 2 X π r ρj 1j π π u1 (re ) = cos ρj θ − − π − cos ρj θ − + sin πρj 2 2 j =1 iθ
+ w1 (reiθ ), where the function sup06θ6π |w1 (reiθ )| satisfies the condition (q, ρ2 ). From (190 ), it is easy to see, in C+ u2 (reiθ ) := u(reiθ ) − u1 (reiθ ) 2 X δj+ πρj r ρj δj− πρj r ρj − sin ρj (θ − π ) + sin ρj θ + = sin πρj sin πρj j =1 + w2 (reiθ ), where the function supη6θ6π−η |w2 (reiθ )|, η > 0, satisfies the condition (q, ρ2 ). From this we obtain that πρ1 u2 (reiθ ) = [δ + sin ρ1 (π − θ1 ) + δ1− sin ρ1 θ1 ] ρ r→∞ r 1 sin πρ1 j lim
132
P. AGRANOVICH
and π
πρ1 sin ρ1 π2 + u2 (rei 2 ) lim = [δ1 + δ1− ]. r→∞ r ρ1 sin πρ1 So we can calculate δ1+ and δ1− . Now let us consider the function (reiθ ∈ C+ ) δ1+ πρ1 r ρ1 δ1− πρ1 r ρ1 sin ρ1 (θ − π ) − sin ρ1 θ. u3 (re ) := u2 (re ) + sin πρ1 sin πρ1 iθ
iθ
Evidently, the condition (q, ρ2 ) for the function supη6θ6π−η |w2 (reiθ )| guarantees the existence of a sequence {rj } such that limj →∞ rj = ∞ and π
w2 (rj ei 2 ) w2 (rj eiθ1 ) lim = lim = 0. ρ ρ j →∞ j →∞ rj 2 rj 2 Going over to the limit with respect to the sequence {rj }, we have π
πρ2 u3 (rj ei 2 ) π = [δ2+ + δ2− ] sin ρ2 ρ2 j →∞ sin πρ2 2 rj lim
and u3 (rj eiθ1 ) πρ2 = [δ + sin ρ2 (π − θ1 ) + δ2− sin ρ2 θ1 ]. ρ2 j →∞ sin πρ2 2 rj lim
So we can calculate δ2+ and δ2− , too. Thus, taking into account (20), we obtain that τu (r) = 11 t ρ1 + 12 t ρ2 + 8(t), τu+ (r) = δ1+ r ρ1 + δ2+ r ρ2 + γ + (r), τu− (r) = δ1− r ρ1 + δ1− r ρ2 + γ − (r), where the function 8(t) satisfies the condition (q, ρ2 ). It remains to estimate the terms γ +(r) and γ − (r). First we assume that sup u(z) < 0.
(21)
Im z>0
By virtue of the generalized Carleman formula [5, 9] we have the equality Z r Z π 1 1 1 dσ (x) = − u(reiθ ) sin θ dθ + A(u, r), 2 2 x r π r 1 0 where
Z
σ (r) = 1
r
Z t dτu (t) −
16|t |6r
t dτ˜u (t);
(22)
POLYNOMIAL ASYMPTOTIC REPRESENTATION OF SUBHARMONIC FUNCTIONS
133
A(u, r) = C1 r12 + C2 with the constants C1 and C2 independent of r. Integrating (22) by parts and taking into account that σ (1) = 0, we conclude that Z π Z r σ (x) 1 1 dx = u(reiθ ) sin θ dθ + A(u, r). 3 x 2π r 2 1 0 If we replace r by kr, 0 < k < 1, and subtract the second equation from the first, we obtain Z π Z π Z r 1 1 σ (x) iθ dx = u(re ) sin θ dθ − u(kreiθ ) sin θ dθ + 3 2π r 0 2π kr 0 kr x d(1 − k 2 ) + . (23) 4π r 2 k 2 By virtue of the condition (21), the measure t dτ˜u (t) is nonpositive on R, and that implies monotonic growth of the function σ (x). Hence, Z r 2 2 σ (r) −2 1 − k −2 1 − k 6 . dx 6 σ (r)(rk) σ (kr)(kr) 3 2 2 kr x From this and (23) we have Z π Z π kr iθ k σ (kr) 6 u(re ) sin θ dθ − u(kreiθ ) sin θ dθ + (1 − k 2 )π 0 0 d(1 − k 2 ) 6 σ (r). + 2kr Rπ Let us calculate 0 u(reiθ ) sin dθ. According to (190 ), Z
π
u(re ) sin θ dθ = iθ
0
2 X πρj r ρj j =1
ρj2 − 1
(1j −
δj+
−
δj− )
Z +
π
κ(re ˜ iθ ) sin θ dθ.
0
Therefore X 2 πρj r ρj kr k (1j − δj+ − δj− ) + σ (kr) 6 (1 − k 2 )π j =1 ρj2 − 1 Z
π
2 X πρj r ρj
(1 − δj+ − δj− ) − 2 ρ − 1 j =1 j Z π d(1 − k 2 ) iθ 6 σ (r). κ(kre ˜ ) sin θ dθ + − 2kr 0 +k
κ(re ˜ ) sin θ dθ − iθ
0
It is easy to show that τu (r) − τu+ (r) − τu− (r) =
σ (r) + r
Z 1
r
σ (x) dx. x2
(24)
134
P. AGRANOVICH
From this and (24), we obtain τu (r) − τu+ (r) − τu− (r) " 2 X πρj r ρj k > (1j − δj+ − δj− )(k − k ρj )+ (1 − k 2 )π j =1 ρj2 − 1 Z π Z π d(1 − k 2 ) iθ iθ +k κ(re ˜ ) sin θ dθ − κ(kre ˜ ) sin θ dθ + + 2kr 0 0 2 X π(r ρj − 1) + (1j − δj+ − δj− )(k − k ρj )+ 2 ρ − 1 j j =1 Z Z Z r Z r dt π dt π iθ iθ κ(te ˜ ) sin θ dθ − κ(kte ˜ ) sin θ dθ+ +k t t 1 0 1 0 # d(1 − k 2 ) 1 + 1− . 2k r
(25)
On the other hand, from (24) it follows that (t = kr) τu (t) − τu+ (t) − τu− (t) X 2 1 πρj t ρj 6 (1j − δj+ − δj− )(k 1−ρj − 1)+ (1 − k 2 )π j =1 ρj2 − 1 Z π Z π d(1 − k 2 ) t iθ sin θ dθ − e + κ˜ κ(te ˜ iθ ) sin θ dθ + + k 2t 0 0 2 X
π (1j − δj+ − δj− )(k 1−ρj − 1)(t ρj − 1)+ −1 Z dx π x iθ +k κ˜ sin θ dθ− e k 1 x 0 # Z t Z 1 d(1 − k 2 ) dx π iθ 1− . − κ(xe ˜ ) sin θ dθ + 2 t 1 x 0
+
ρ2 j =1 j Z t
Since k 1−ρj − 1 k(k − k ρj ) 1 − ρj = lim =− , 2 2 k→1 1 − k k→1 1−k 2 from (25) and (26) we have that lim
τu (r) − τu+ (r) − τu− (r) − Z > lim sup k k→1
0
2 X (1j − δj+ − δj− )r ρ j =1
π
Z
κ(re ˜ ) sin θ dθ − iθ
0
π
iθ κ(kre ˜ ) sin θ dθ+
(26)
135
POLYNOMIAL ASYMPTOTIC REPRESENTATION OF SUBHARMONIC FUNCTIONS
Z +k
r
1
dt t
Z
π
Z κ(te ˜ iθ ) sin θ dθ −
0
1
r
dt t
Z
π
iθ κ(kte ˜ ) sin θ dθ
0
k , 1 − k2
and τu (r) − τu+ (r) − τu− (r) − Z
2 X (1j − δj+ − δj− )r ρj j =1
r iθ sin θ dθ− 6 lim inf κ˜ e k→1 k 0 Z Z π Z r dt π t iθ iθ − κ(re ˜ ) sin θ dθ + k κ˜ sin θ dθ− e k 0 1 t 0 Z Z r dt π 1 − κ(te ˜ iθ ) sin θ dθ . 1 − k2 1 t 0 Rπ By virtue of the condition (q, ρ2 ) for the function 0 κ(te ˜ iθ ) sin θ dθ, it follows that q Z 2T Z r Z dt π iθ κ(te ˜ ) sin θ dθ dr 6 o(T ρ2 q+1 ), T → ∞, T 1 t 0 π
hence τu (r) − τu+ (r) − τu− (r) −
2 X (1j − δj+ − δj− )r ρj = γ (r), j =1
where the function γ (r) satisfies the condition (q, ρ2 ). So Theorem 2 is proved for this special case. For the complete proof of Theorem 2, we have to remove restriction (21). To this end, it suffices to consider the function U (z) = u(z) − KRe zρ1 e−iρ1 2 − ln M, π
where K and M are sufficiently large constants. Theorem 2 is proved.
2
4. General Case The statements of Theorems 1 and 2 can be extended to the case when the Riesz measure of a subharmonic function is concentrated on a finite system of rays. Let us formulate the corresponding results. / THEOREM 4. Let u(z), z ∈ C+ , be a subharmonic function of the class Uρ1 , ρ1 ∈ Z, and let the support of its Riesz measure µu be contained on a finite system of rays {arg z = θj },
j = 1, . . . , m, 0 = θ0 < θ1 < · · · < θm < θm+1 = π.
136
P. AGRANOVICH
Suppose that τu,j (r) =
n X
(j )
1k r ρk + 8j (r),
j = 1, . . . , m,
k=1
where τu,j (r) = µu ({teiθj : 0 < t 6 r}), and τu+ (r) = τu− (r) =
n X k=1 n X
δk r ρk + γ + (r), δk− r ρk + γ − (r).
k=1
Here [ρ1 ] < ρn < ρn−1 < · · · < ρ1 and the functions 8j (r), j = 1, . . . , m, γ ± (r) satisfy the condition (q, ρn ), q > 1. Then m n X π r ρk X (j ) 1 [cos ρk (|θ − θj | − π ) − cos ρk (θ − θj )] − u(z) = sin πρk j =1 k k=1 δk− πρk r ρk δk+ πρk r ρk sin ρk (θ − π ) + sin ρk θ + κ(reiθ ), − sin πρk sin πρk where the function κ(reiθ ) = o(r ρn ) when r → ∞ uniformly for θ ∈ [η, π − η], 0 < η < π/2, if the point z = reiθ does not belong to any C0,1 -set. If q > 1, then the functions π sup |κ(reiθ )|, 0 < η < , 2 η6θ6π−η and
Z
π
κ(reiθ ) sin θ dθ 0
satisfy the condition (q, ρn ). / THEOREM 5. Let u(z), z ∈ C+ , be a subharmonic function of the class Uρ1 , ρ1 ∈ Z, let its Riesz measure be concentrated on a finite system of rays {arg z = θj }, j = 1, . . . , m, 0 = θ0 < · · · < θm < θm+1 = π . Suppose, further, that m n X π r ρk X (j ) iθ 1 [cos ρk (|θ − θj | − π ) − cos ρk (θ − θj )] − u(re ) = sin πρk j =1 k k=1 δk− πρk r ρn δk+ πρk r ρn ρk (θ − π ) + sin ρk θ + κ(reiθ ), − sin πρk sin πρk
POLYNOMIAL ASYMPTOTIC REPRESENTATION OF SUBHARMONIC FUNCTIONS
θ ∈ 0 = (0, α1 ) ∪ (α2 , π ) ∪
[ m
137
{θj } ,
j =1
where ρn /Z (θj +1 − θj ) ∈ π
[ρ1 ] < ρn < · · · < ρ1 , for any
j, j = 1, 2, . . . , m − 1, 0 < α1 < θ1 , θm < α2 < π and the functions n [ o sup |κ(reiθ )| : {η 6 θ 6 α1 } ∪ {α2 6 θ 6 π − η} ∪ θj , η > 0 and
Z
α1
Z κ(re ) sin θ dθ +
π
iθ
κ(reiθ ) sin θ dθ α2
0
satisfy the condition (q, ρn ). Then τu,j (r) =
n X
(j )
1k r ρk + 8j (r),
k=1
τu+ (r) = τu− (r) =
n X k=1 n X
δk r ρk + γ + (r), δk− r ρk + γ − (r),
k=1
where the functions 8(r), j = 1, . . . , m, and γ ± satisfy the condition (q, ρn ).
Acknowledgements With deep sorrow I say my last thankful words to Professor L. Ronkin for some very useful conversations related to this problem. I am grateful to the referee for constructive criticism directed at improving the quality of my exposition.
References 1.
Agranovich, P. Z.: Polynomial asymptotic representations of subharmonic functions with masses on the finite system of rays, MAG 3(3/4) (1996), 219–230 (Russian).
138 2.
3. 4. 5. 6. 7. 8. 9. 10.
P. AGRANOVICH
Agranovich, P. Z. and Logvinenko, V. N.: The analogue of the Valiron–Titchmarsh theorem for two-term asymptotics of the subharmonic function with masses on a finite set of rays, Sibirsk. Mat. Z. 24(5) (1985), 3–19 (Russian). Agranovich, P. Z. and Logvinenko, V. N.: Polynomial asymptotic representation of subharmonic function in the plane, Sibirsk. Math. Z. 32(1) (1991), 3–21 (Russian). Anderson, J. M.: Integral functions and Tauberian theorems, Duke Math. J. 32(4) (1965), 145– 163. Govorov, N. V.: Riemann Boundary Value Problems with Infinite Index, Nauka, Moscow, 1986 (Russian). Logvinenko, V. N.: About entire functions with zeros on a half-line. I, Theory of functions, Funct. Anal. Appl. 16 (1972), 154–158. Logvinenko, V. N.: About entire functions with zeros on a half-line. II, Theory of functions, Funct. Anal. Appl. 17 (1973), 84–99. Nevanlinna, R.: Uber die Eigenschaften Meromorpher Funktionen in einem Winkebraum, Acta Soc. Sci. Fenn. 50(12) (1925), 1–45. Ronkin, L. I: Functions of Completely Regular Growth of Several Variables, Kluwer Acad. Publ., Dordrecht, 1992. Zygmund, A.: Integrales Singulieres, Lecture Notes in Math. 204, Springer-Verlag, New York, 1971.
Mathematical Physics, Analysis and Geometry 3: 139–177, 2000. © 2000 Kluwer Academic Publishers. Printed in the Netherlands.
139
Differential Equations Compatible with KZ Equations G. FELDER1, Y. MARKOV2, V. TARASOV3 and A. VARCHENKO2?
1 Departement Mathematik, ETH-Zentrum, 8092 Zürich, Switzerland. e-mail:
[email protected] 2 Department of Mathematics, University of North Carolina, Chapel Hill, NC 27599 – 3250, U.S.A.
e-mail: {markov, av}@math.unc.edu 3 St. Petersburg Branch of Steklov Mathematical Institute, Fontanka 27, St. Petersburg, 191011, Russia. e-mail:
[email protected] (Received: 3 April 2000) Abstract. We define a system of ‘dynamical’ differential equations compatible with the KZ differential equations. The KZ differential equations are associated to a complex simple Lie algebra g. These are equations on a function of n complex variables zi taking values in the tensor product of n finite dimensional g-modules. The KZ equations depend on the ‘dual’ variable in the Cartan subalgebra of g. The dynamical differential equations are differential equations with respect to the dual variable. We prove that the standard hypergeometric solutions of the KZ equations also satisfy the dynamical equations. As an application we give a new determinant formula for the coordinates of a basis of hypergeometric solutions. Mathematics Subject Classifications (2000): Primary 35Q40; secondary 17B10. Key words: hypergeometric solutions, Kac–Moody Lie algebras, KZ equations.
1. Introduction In the theory of the bispectral problem [5, 12], one considers a commutative algebra A of differential operators L(z, ∂/∂z) acting on functions of one complex variable z. Such an algebra is called bispectral if there exists a non-trivial family u(z, µ) of common eigenfunctions depending on a spectral parameter µ Lu(z, µ) = fL (µ)u(z, µ),
L ∈ A,
(1)
which is also a family of common eigenfunctions of a commutative algebra B of differential operators 3(µ, ∂/∂µ) with respect to µ: 3u(z, µ) = θ3 (z)u(z, µ),
3 ∈ B.
(2)
J. Duistermaat and A. Grünbaum [5] studied the case where A is the algebra of differential operators that commute with a Schrödinger operator (d2 /dz2 ) − V (z) with meromorphic potential V (z). They give a complete classification of bispectral ? The last author is supported in part by NSF grant DMS-9801582.
140
G. FELDER ET AL.
algebras arising in this way. In particular they show that A is bispectral if V (z) is a rational KdV potential (a rational function which stays rational under the flow of the Korteweg–de Vries equation). G. Wilson [12] classified bispectral algebras of rank one, i.e., such that the greatest common divisor of the orders of the differential operators in A is one. He showed that the maximal bispectral algebras of rank one are in one to one correspondence with conjugacy classes of pairs (Z, M) of square matrices so that ZM − MZ + I has rank one. The bispectrality then follows from the existence of the involution (Z, M) 7→ (M T , Z T ), which corresponds to exchanging z and µ. The higher-dimensional version of the bispectral problem, in which A consists of partial differential operators in z ∈ Cn is open. However, O. Chalykh, M. Feigin and A. Veselov [3, 2] constructed examples of algebras in higher dimensions which have the bispectral property (see Veselov’s contribution to [7]). In these examples, A consists of differential operators commuting with an n-particle Schrödinger operator with certain special rational potentials, including those of Calogero–Moser. These potentials are in many respects the natural generalization of rational KdV potential associated to rank one algebras. In these examples, the Baker–Akhiezer function u(z, µ) is symmetric in the two arguments, thus B = A. A good source of material on the bispectral problem is the volume [7]. In this paper we study a class of examples of commutative algebras of partial differential operators acting on vector-valued functions with the bispectral property. This means that in (1), (2), u takes values in a vector space and fL (µ), θ3 (z) are endomorphisms of the vector space. In our class of examples, the algebra A is generated by Knizhnik–Zamolodchikov differential operators. They are commuting first-order differential operators associated to a complex simple Lie algebra g with a fixed non-degenerate invariant bilinear form and a non-zero complex parameter κ. They act on functions of n complex variables zi taking values in the tensor product of n finite-dimensional g-modules. The ‘dual’ variable µ is in a Cartan subalgebra of g. The first set of Equations (1) is then the set of generalized Knizhnik–Zamolodchikov equations X (ij ) ∂ − u(z, µ) = µ(i) u(z, µ), i = 1, . . . , n. κ ∂zi j :j 6=i zi − zj Here ∈ g ⊗ g is dual to the invariant bilinear form and (ij ) acts as on the ith and j th factors of the tensor product and as the identity on the other factors. Similarly µ(i) is µ acting on the ith factor. It is well-known that these equations form a compatible system, i.e., they are the equations defining horizontal sections for a flat connection. For µ = 0 they reduce to the classical Knizhnik–Zamolodchikov equations. The algebra B is generated by rank(g) first order partial differential operators in µ with rational coefficients. We call the corresponding Equations (2) dynamical differential equations, and show that they form, together with the generalized Knizhnik–Zamolodchikov equations, a compatible system. We also give simultaneous solutions of both systems of equations in terms of hypergeometric
DIFFERENTIAL EQUATIONS COMPATIBLE WITH KZ EQUATIONS
141
integrals for a more general class of Lie algebras, which includes in particular all Kac–Moody Lie algebras. In the case of g = sl2 , the algebra B is generated by one ordinary differential operator. In this case, the corresponding equations where first written and solved by H. Babujian and A. Kitaev [1], who also related the equations to the Maxwell– Bloch system. Our paper is organized as follows. In Section 2 we introduce the systems of Knizhnik–Zamolodchikov and dynamical differential equations for arbitrary simple Lie algebra and prove their compatibility. We then give formulae for hypergeometric solutions in Section 3 and give as an application a determinant formula. The fact that the hypergeometric integrals provide solutions is a consequence of a general theorem valid for a class of Lie algebras with generic Cartan matrix, introduced in [10]. We introduce in Section 4 the Knizhnik–Zamolodchikov and dynamical differential equations in this more general context and explain in the next Section the results on complexes of hypergeometric differential forms from [10]. In Section 6 we prove that the hypergeometric integrals for generic Lie algebras satisfies the dynamical differential equations. Finally in Section 7 we prove that our hypergeometric integrals are solutions of both systems of equations for any Kac–Moody Lie algebra. We also find a determinant formula, which implies a completeness result for solutions in the case of generic parameters. 2. Dynamical Differential Equations 2.1. Let g be a simple complex Lie algebraL with an invariant bilinear form ( , ) and a root space decomposition g = h ⊕ ( α∈1 Ceα ). The root vectors eα are = 1. Then the quadratic Casimir element of g ⊗ g has normalized so P that (eα , e−α ) P the form = s hs ⊗ hs + α∈1 eα ⊗ e−α , for any orthonormal basis (hs ) of the Cartan subalgebra h. We also fix a system of simple roots α1 , . . . , αr . Consider the Knizhnik–Zamolodchikov (KZ) equations with an additional parameter µ ∈ h, for a function u on n variables taking values in a tensor product V = V1 ⊗ · · · ⊗ Vn of highest weight modules of g with corresponding highest weights 31 , . . . , 3n , κ
X (ij ) ∂u = µ(i) u + u, ∂zi z − z i j i6=j
i = 1, . . . , n,
(3)
where κ is a complex parameter. We are interested in a differential equation for u with respect to µ which are compatible with KZ equations. If µ0 ∈ h, denote by ∂µ0 the partial derivative with respect to µ in the direction of µ0 . THEOREM 2.1. The equations κ∂µ0 u =
n X i=1
zi (µ0 )(i) u +
X hα, µ0 i α>0
hα, µi
e−α eα u,
µ0 ∈ h,
(4)
142
G. FELDER ET AL.
form together with the KZ equations (3), a compatible system of equations for a function u(z, µ) taking values in V = V1 ⊗ · · · ⊗ Vn . Equations (4) will be called dynamical differential equations. EXAMPLE. Let g = slN = glN /C. View slN -modules as glN -modules by letting the center of glN act trivially. Denote by Ea,b ∈ glN the matrix whose entries are zero except for a one at the intersection of the bth column. The P ath row with the P fundamental coweights $a = (1 − a/N) b6a Eb,b − (a/N) b>a Eb,b , a = 1, . . . , N − 1 form a basis of the standard Cartan subalgebra of slN . Write µ = P N−1 a=1 µa $a . Then our equations may be written as
2.2.
κ
(i) (j ) N−1 X X Ea,b X Eb,a ∂u = µa $a(i) u + u, ∂zi z − z i j a=1 j :j 6=i a,b
κ
(i) (j ) n X X X Eb,c Ec,b ∂u = zi $a(i) u + u. ∂µa µb + µb+1 + · · · + µc−1 i=1 b,c:b6a
PROOF OF THEOREM
2.1
It is rather easy to verify that most terms of the compatibility equations vanish. P The only non-trivial thing to check is that the operators α>0 (hα, λi/hα, µi)e−α eα commute for different values of λ. The operators obtained by extending the sum to all roots differ from the sum over positive roots by an element of the Cartan subalgebra. Since the operators commute with the Cartan subalgebra, it is sufficient to prove the following proposition. PROPOSITION 2.2. Let for λ, µ ∈ h, T (λ, µ) = Then for all λ, µ ∈ h,
P
α∈1 (hα, λi/hα, µi)e−α eα .
T (λ, µ)T (ν, µ) = T (ν, µ)T (λ, µ). The proof is based on the following fact. LEMMA 2.3. Let α, β ∈ 1 with α 6= ±β,P and let S = S(α, β) be the set of integers j such that β + j α ∈ 1. Then j ∈S [eα , eβ+j α e−β−j α ] = 0 and P [e , e e ] = 0. j ∈S −α β+j α −β−j α Proof. For roots γ , δ such that γ +δ is a root, let Nγ ,δ = ([eγ , eδ ], e−γ −δ ), so that [eγ , eδ ] = Nγ ,δ eγ +δ . By considering the adjoint action on g of the sl2 sub-algebra generated by e±α , we see that for β 6= ±α, S is a finite sequence of subsequent integers. We may thus assume that S = {0, . . . , k} by replacing β by β − j α for some j if necessary.
DIFFERENTIAL EQUATIONS COMPATIBLE WITH KZ EQUATIONS
143
We then have k X [eα , eβ+j α e−β−j α ] j =0
=
k−1 X
Nα,β+j α eβ+(j +1)α e−β−j α +
j =0
=
k X
Nα,−β−j α eβ+j α e−β−(j −1)α
j =1
k−1 X (Nα,β+j α + Nα,−β−(j +1)α )eβ+(j +1)α e−β−j α . j =0
By the invariance of the bilinear form, Nα,β+j α = ([eα , eβ+j α ], e−β−(j +1)α ) = −(eβ+j α , [eα , e−β−(j +1)α ]) = −Nα,−β−(j +1)α . P Therefore j ∈S [eα , eβ+j α e−β−j α ] vanishes. The other statement is proved by replacing α by −α and noticing that S(−α, β) = S(α, β). 2 Proof of Proposition 2.2. Consider X hα, λihβ, νi T (λ, µ)T (ν, µ) = [eα e−α , eβ e−β ]. hα, µihα, µi α,β
(5)
Let us show that this expression is a regular function of µ ∈ h. Since for nontrivial λ, µ it converges to zero at infinity, it then vanishes identically. We compute the residue of (5) at hα, µi = 0: X hα, λihγ , νi − hγ , λihα, νi (6) [eα e−α , eγ e−γ ], hγ , µi γ :γ 6=±α a function on the hyperplane hα, µi = 0. The sum over γ of the form β + j α, j ∈ S(α, β) gives for hα, µi = 0, X hα, λihβ, νi − hβ, λihα, νi [eα e−α , eβ+j α e−β−j α ] = 0, (7) hβ, µi j ∈S by the previous lemma. Since the sum over γ in (6) can be written as a sum of such terms, it vanishes.
3. Hypergeometric Solutions Let g be a simple complex Lie algebra. Choose a set f1 , . . . , fr , e1 , . . . , er of Chevalley generators of the Lie algebra g associated with simple roots α1 , . . . , αr .
144
G. FELDER ET AL.
P Let λ = (m1 , . . . , mr ) ∈ Nr . Let Q+ = P Nαi be the positive root lattice for mi αi . Let V be a tensor product g. Define a map α: Nr → Q+ by α(λ) = of highest weight modules V of g with respective highest weights 3j , where Pn j j = 1, . . . , n. Set 3 = j =1 3j . Denote Vλ the weight space of V with weight 3 − α(λ). The hypergeometric solutions of the KZ equations in Vλ , see [10], have R the form u(z) = γ (z) 8(z, t)1/k ω(z, t). We will describe the explicit construction. P The number of integration variables (tk )m mi . Let c be the unique k=1 is m = non-decreasing function from {1, . . . , m} to {1, . . . , r} (i = 1, . . . , r), such that #c−1 ({i}) = mi . Define Y Y Y 8(z, t) = (zi − zj )(3i ,3j ) (tk − zj )−(αc(k) ,3j ) (tk − tl )(αc(k) ,αc(l) ) . i<j
k,j
k
The m-form ω(z, t) is a closed logarithmic differential form on Cn × Cm with values in Vλ . It has the following combinatorial description. Let P (λ, n) be the set of sequences I = (i11 , . . . , is11 ; . . . ; i1n , . . . , isnn ) of integers in {1, . . . , r} with sj > 0, j = 1, . . . , n and such that, for all 1 6 j 6 r, j appears precisely |c−1 (j )| times in I . For I ∈ P (λ, n), and a permutation σ ∈ 6m , set σ1 (l) = σ (l) and σj (l) = σ (s1 + · · · + sj −1 + l), j = 2, . . . , n, 1 6 l 6 sj . Define 6(I ) = {σ ∈ 6m | c(σj (l)) = ijl for all j and l}. Fix a highest-weight vector vj for each representation Vj , j = 1, . . . , n. To every I ∈ P (λ, n) we associate a vector fI v = fi11 · · · fis1 v1 ⊗ · · · ⊗ fi1n · · · fisnn vn 1 in Vλ , and meromorphic differential m-forms ωI,σ = ωσ1 (1),...,σ1 (s1 ) (z1 ) ∧ · · · ∧ ωσn (1),...,σn (sn ) (zn ), labeled by σ ∈ 6(I ), where ωi1 ,...,is (z) = d log(ti1 − ti2 ) ∧ · · · ∧ d log(tis−1 − tis ) ∧ d log(tis − z) is a meromorphic one form on C × Cs . Finally X X (−1)|σ | ωI,σ fI v. ω(z, t) = I ∈P (λ,n) σ ∈6(I )
P dz −dz It obeys the equation 8−1 d8 ∧ ω = K ∧ ω, where K = i<j (ij ) zii −zj j . As a consequence, for each horizontal family of twisted cycles γ (z) in {z}×Cm, u obeys the KZ equations (3) with µ = 0 (see [10]). This construction can be modified to give solutions for general µ: THEOREM 3.1. The integrals ! Z m n X X 1 8µk ω, 8µ = exp − hαc(i) , µiti + h3j , µizj 8, γ (z)
i=1
are solutions of the KZ equations (3). Proof. The proof follows from the identity 8−1 µ d8µ ∧ ω =
n X i=1
µ(i) dzi ∧ ω + K ∧ ω.
j =1
(8)
145
DIFFERENTIAL EQUATIONS COMPATIBLE WITH KZ EQUATIONS
To prove this identity notice that ω is a sum of several terms ωs , which can be grouped according to how the integration variables t1 , . . . , tm are distributed among the points z1 , . . . , zn . If the variable tk is associated to the point zi in one term ωs , then (dtk − dzi ) ∧ ωs = 0. Moreover, if the term ωs obeys h(i) ωs = (h3i , hi − P m0j (i)hαj , hi)ωs , h ∈ h, where 3i is the highest weight of the ith representation, then the number of variables tk associated to zi such that c(k) = j is m0j (i) (j = 1, . . . , r). Thus we have n X X −1 0 h3i , µi − mj (i)hαj , µi dzi ∧ ω + K ∧ ω. 8µ d8µ ∧ ω = i=1
j
2
The proof is complete.
THEOREM 3.2. The hypergeometric integrals of Theorem 3.1 obey the Dynamical differential equations (4). The proof of the Theorem is given in Section 7. An application of the above two theorems is a determinant formula. COROLLARY Pd3.3. Fix a basis v1 , . . . , vd of a weight space Vλ . Suppose that ui (µ, z) = j =1 ui,j vj , i = 1, . . . , d is a basis of the space of solutions in a neighbourhood of a generic point (µ, z) ∈ h × Cn . Let δα = trVλ (e−α eα ) (α ∈ 1, α > 0), ij = trVλ (ij ). Then there is a constant C = C(V1 , . . . , Vn , λ, κ) 6= 0 such that ! n X Y Y zi det(uij ) = C exp hα, µiδα /κ (zi − zj )ij /κ . trVλ (µ(i) ) κ i=1 α>0 i<j
4. Free Lie Algebras. Dynamical Differential Equations 4.1.
THE DEFINITION OF KZ AND DYNAMICAL DIFFERENTIAL EQUATIONS
Following [10] let us fix the following data: (1) A finite-dimensional complex vector space h; (2) A non-degenerate symmetric bilinear form ( , ) on h; (3) Linearly independent covectors (‘simple roots’) α1 , . . . , αr ∈ h∗ . We denote by b: h → h∗ the isomorphism induced by ( , ), and we transfer the form ( , ) to h∗ via b. Set bij = (αi , αj ), hi = b−1 (αi ) ∈ h. Denote by g the Lie algebra generated by ei , fi for i = 1, . . . , r and h subject to the relations: [h, ei ] = hαi , hiei ,
[h, fi ] = −hαi , hifi ,
[ei , fj ] = δij hi ,
[h, h0 ] = 0,
for all i, j = 1, . . . , r and h, h0 ∈ h. Thus we have constructed a Kac–Moody Lie algebra without Serre’s relations. We denote by n− (resp. by n+ ) the subalgebra
146
G. FELDER ET AL.
of g generated by fi (resp. ei ) for i = 1, . . . , r. We have g = n− ⊕ h ⊕ n+ . Set b± = n± ⊕ h. These are subalgebras of g. Let 3 ∈ h∗ . Denote by M(3) the Verma module over g generated by a vector v subject to the relations n+ v = 0 and hv = h3, hiv for all h ∈ h. Let us fix weights 31 , . . . , 3n ∈ h∗ , and let M = M(31 ) ⊗ · · · ⊗ M(3n ). For λ = (m1 , . . . , mr ) ∈ Nr , set D X E (n± )λ = x ∈ n± | [h, x] = ± mi αi , h x, for all h ∈ h , D E X mi αi , h x, for all h ∈ h , M(3)λ = x ∈ M(3) | hx = 3 − DX E X Mλ = x ∈ M | hx = 3j − mi αi , h x, for all h ∈ h . L L L We λ (n± )λ , b± = h ⊕ ( λ (n± )λ ), M(3) = λ M(3)λ , M = L have n± = λ Mλ . Let τ : g → g be the Lie algebra automorphism such that τ (eL i ) = −fi , τ (fi ) = L −ei , τ (h) = −h, for h ∈ h. Set n∗± = λ (n± )∗λ . Set M(3)∗ = λ M(3)∗λ . Define a structure of a g-module on M(3)∗ by the rule hgφ, xi = hφ, −τ (g)xi
for φ ∈ M(3)∗ , g ∈ g, x ∈ M(3).
(9)
There is a unique bilinear form K( , ) on g such that: K coincides with ( , ) on h; K is zero on n+ and n− ; h and n− ⊕ n+ are orthogonal; K(fi , ej ) = K(ej , fi ) = δij for i, j = 1, . . . , r; K is g-invariant, that is K([x, y], z) = K(x, [y, z]) for all x, y, z ∈ g. A bilinear form S on g is defined by the rule S(x, y) = −K(τ (x), y). The form S is symmetric, τ -invariant, and S([x, y], z) = S(x, [τ (y), z]). The subspaces n+ , h, n− are pairwise orthogonal with respect to S. For a Verma module M(3) with highest weight 3 and generating vector v there is a unique bilinear form S on M(3) such that: S(v, v) = 1; S(ei x, y) = S(x, fi y); S(fi x, y) = S(x, ei y), for all x, y ∈ M(3) and i = 1, . . . , r. S is symmetric. The subspaces M(3)λ are pairwise orthogonal with respect to S. The form S induces a homomorphism of g-modules S: M(3) → M(3)∗ . The module M(3)/ ker S V is the irreducible g-module with highest weight 3. More generally, on the space p n− ⊗ M(31 ) ⊗ · · · ⊗ M(3n ) for 3j ∈ h∗ , a bilinear form S is defined by the rule S(g1 ∧ · · · ∧ gp ⊗ x1 ⊗ · · · ⊗ xn , g10 ∧ · · · ∧ gp0 ⊗ x10 ⊗ · · · ⊗ xn0 ) = det S(gi , gj0 ) ·
n Y
S(xi , xi0 ).
i=1
This form induces a map ^p n− ⊗ M(31 ) ⊗ · · · ⊗ M(3n ) MS:
DIFFERENTIAL EQUATIONS COMPATIBLE WITH KZ EQUATIONS
→
^p
∗ n− ⊗ M(31 ) ⊗ · · · ⊗ M(3n ) ,
147 (10)
S is called the contravariant form. The linear map (10) depends analytically on (bij ) and 31 , . . . , 3n ∈ h∗ for a fixed r. It is non-degenerate for general values of the parameters, see Theorem 3.7 in [10]. A Lie Bialgebra Structure on b± [10]. A Lie bialgebra is a vector space g with a Lie algebra structure and a Lie coalgebra structure, such that the cocommutator map ν: g → g ∧ g is a one-cocycle: xν(y) − yν(x) = ν([x, y]), for all x, y ∈ g. Here the action of g on g ∧ g is the adjoint one: a(b ∧ c) = [a, b] ∧ c + b ∧ [a, c]. The dual map to the cocommutator map, ν ∗ : (g ∧ g)∗ → g∗ , defines a Lie algebra structure on g∗ . Let g be a Lie bialgebra. The double of g is the Lie algebra equal to g ⊕ g∗ as a vector space with the bracket on g and g∗ defined by the Lie algebra structure on g and g∗ , and for x ∈ g and l ∈ g∗ , [l, x] = l¯ + x, ¯ where x¯ ∈ g and l¯ ∈ g∗ are defined ¯ by the rules l(y) = l([x, y]), m(x) ¯ = [m, l](x), for all y ∈ g, m ∈ g∗ . The double is denoted by D(g). Let g be the Kac–Moody algebra we defined at the beginning of the section. g is a Lie bialgebra with respect to the following cobracket. There exists a unique map ν: g → g ∧ g such that xν(y) − yν(x) = ν([x, y]), and ν(h) = 0, and ν(fi ) = 1 f ∧ hi , and ν(ei ) = 12 ei ∧ hi . In the previous four equalities, h ∈ h, i = 1, . . . , r, 2 i the action of g on g ∧ g is the adjoint one, see [4], Example 3.2 and [10]. b− and b+ are subbialgebras. The map ν has the property τ ν + ντ = 0. Thus, if ρ: b− ∗ → End(V ) is a representation of the Lie algebra (ν|b− )∗ : 32 b− ∗ → b− ∗ , then −ρ ◦ τ : b+ ∗ → End(V ) is a representation of the Lie algebra (ν|b+ )∗ : 32 b+ ∗ → b+ ∗ . Note that the coalgebra map ν defines a Lie algebra structure on b∗± . Comultiplication. Let M = M(31 ) ⊗ · · · ⊗ M(3P n ). Let v = v1 ⊗ · · · ⊗ vn ∈ M be the product of the generating vectors. Set 3 = 3j . Let b− act on b− ⊗ M by the rule a(b ⊗ m) = [a, b] ⊗ m + b ⊗ am. For 1 6 i 6 n, and a, b ∈ g, m = x1 ⊗ · · · ⊗ xn ∈ M, set a (i) m = x1 ⊗ · · · ⊗ xi−1 ⊗ axi ⊗ xi+1 ⊗ · · · ⊗ xn and a (i) (b ⊗ m) = [a, b] ⊗ m + b ⊗ a (i) m. There is a unique linear map νM : M → b− ⊗ M such that νM (h · x) = h · νM (x) for any h ∈ h and x ∈ M; νM (x) = 12 (b−1 (3 − α(λ))) ⊗ x + νM− (x) for x ∈ Mλ . P P Recall that b−1 (α(λ)) = mi b−1 (αi ) = mi hi , and b−1 : h∗ → h is defined n− ⊗ M is defined via an at the beginning of Section 4.1. The map νM− P:n M → (k) inductive definition. νM− (v) = 0, νM− (x) = k=1 νM− (x) where (j )
(j )
(k) (k) (fi x) = fi νM− (x) νM− (k) (fi(k) x) νM−
= fi ⊗
h(k) i x
for k 6= j, and
(k) + fi(k) νM− (x).
148
G. FELDER ET AL.
In all formulae we have 1 6 i 6 r, and 1 6 j, k 6 n. Remark. The corresponding definition of νM− (x) in [10] should be corrected as above. Note that the two definitions coincide if we have one tensor factor, i.e. n = 1. We have the following lemma (cf. Lemma 6.15.2 in [10]). LEMMA 4.1. For any x, y ∈ M, a ∈ b− , 1 S(x, ay) if a ∈ h; S(νM (x), a ⊗ y) = 2 S(x, ay) if a ∈ n− .
(11)
Here S is defined on b− ⊗ M by the rule S(a ⊗ x, b ⊗ y) = S(a, b)S(x, y), cf. (10). The proof of Lemma 4.1 is given in Section 4.5. Note that the above lemma renders the following diagram commutative. n− ⊗ M
standard
S
n− ∗ ⊗ M ∗
M S
∗ νM−
M∗
S is an isomorphism for general values of parameters (bij ), (3k )nk=1 , see [10], ∗ : b− ∗ ⊗ M ∗ → M ∗ is a b− ∗ -module structure sections (3.7) and (6.6). Hence, νM with respect to the Lie algebra structure ν ∗ : 32 b− ∗ → b− ∗ for any values of the above parameters. COROLLARY 4.2. If y ∈ ker S: n− → n− then yM ⊂ ker S: M → M.
ACTIONS OF THE DOUBLES OF
b±
ON
m AND m∗
RESPECTIVELY
Consider the standard action of b− on M ∗ , i.e. ∀a ∈ b− ∀φ ∈ M ∗ , ha · φ, · i = hφ, −a · i, where the action on the right-hand side is the standard action of g on M. ∗ defines an action of b− ⊕ b− ∗ on M ∗ . Lemma 6.17.1 This map together with νM− ∗ [10] asserts that M is a D(b−)-module under this action, where D(b−) is the double of b−. ∗ ∗ defines a map νM (a, ·): M ∗ → M ∗ . Set For any a ∈ b− ∗ , the action of νM ∗ ∗ ∗ ρ(a) = −(νM (a, ·)) : M → M. The map ρ: b− → End(M) gives an action of b− ∗ on M. An action of b+ ∗ on M is defined by ω = −ρ ◦ τ : b+ ∗ → End(M). The rule a ⊗ x → τ (a)x, for a ∈ b+ , x ∈ M, defines an action of b+ on M. This action and the map ω define an action of b+ ⊗ b+∗ on M. Lemma 11.3.28 [11] implies that M is a D(b+) module. Note that for any Kac–Moody Lie algebra g without Serre’s relations, and a g-module V we can define a g-module structure on V ∗ by the rule hgφ, · i =
149
DIFFERENTIAL EQUATIONS COMPATIBLE WITH KZ EQUATIONS
hφ, −g · i for all g ∈ g and φ ∈ M ∗ . With respect to this module structure, M becomes a D(b−)-module and M ∗ becomes a D(b+)-module. KZ Equations and Dynamical Differential Equations in M and M ∗ . For a vector space V denote by (V ) ∈ V ⊗V ∗ the canonical element. For λ ∈ Nr , set − λ,± := + ∗ ∗ ∗ ((n± )λ ) ∈ (n± )λ ⊗ (n± )λ , set λ,± := ((n± )λ ) ∈ (n± )λ ⊗ (n± )λ , 0 := ((h) + (h∗ )) ∈ h ⊗ h∗ + h∗ ⊗ h. Set X X 0 − + ± = λ,± + + λ,± ∈ D(b± ) ⊗ D(b± ). λ
λ
Let +,ij be the operator on M (or M ∗ ) acting as + on M(3i )⊗M(3j ), (M(3i )∗ ⊗M(3j )∗ respectively) and as the identity on the other factors. The action of D(b±) on M(3j ) is the action decribed in the previous paragraph when n = 1. The KZ equations with additional parameter µ ∈ h, for a function u(µ, z) on n variables z = (z1 , . . . , zn ) taking values in M (or M ∗ ) are κ
X +,ij ∂u = µ(i) u + u, ∂zi z − zj i6=j i
i = 1, . . . , n, κ ∈ C.
(12)
Let α be a positive root for g and (yi(α) ) a basis of (n+ )α . Set xi(α) = τ (yi(α) ). Then ((yi(α) )∗ ), (xi(α) ) and ((xi(α) )∗ ) are bases of (n+∗ )α , (n− )α , (n− ∗ )α , respectively. P (α) (α) ∗ Define operators 1±,α on M (or M ∗ ) via the formulae 1+,α = i yi (yi ) , P (α) ∗ (α) 1−,α = i (xi ) xi . The dynamical differential equations for the function u(µ, z) with values in M (M ∗ respectively) are κ∂µ0 u =
n X
zi (µ0 )(i) u +
i=1
4.2.
X hα, µ0 i α>0
hα, µi
PROPERTIES OF THE OPERATORS
1+,α u,
µ0 ∈ h, κ ∈ C.
(13)
1+,α
The properties of the operator +,ij are thoroughly described in [11]. Now we are going to study the operators 1+,α . LEMMA 4.3. The following diagram is commutative: M
1+,α
S
M∗
M S
−1−,α
M∗
In particular, the operators 1+,α preserve the kernel of the map S: M → M ∗ .
150
G. FELDER ET AL.
Proof. We fix α throughout this proof and will drop it from the notation of the bases. Fix a basis (uk ) of ker S: (n+ )α → (n+ ∗ )α . Complete it to a basis of (n+ )α by vectors (vl ). Let (u∗k ), (vl∗ ) be the dual basis of (n+ ∗ )α . Moreover, (vl∗ ) =
X (A−1 )lp S(vp , ·)
and
(τ (vl )∗ ) =
X (A−1 )lp S(τ (vp ), ·),
p
(14)
p
where A = (alp ) is a nondegenerate matrix with entries alp = S(vl , vp ). For y ∈ (n+ )α , consider the map y: M → M via the action of D(b+). Let y · p denote the D(b+) action, and yp denote the standard action of b+ . Then we have S(y · p, q) = S(τ (y)p, q) = S(p, −yq)
for any p, q ∈ M.
(15)
Consider the map (vl )∗ : M → M via the action of D(b+). For any p, q ∈ M we have ∗ (τ (vl )∗ , ·))∗ p, S(q, · )i S((vl )∗ p, q) = h(vl )∗ p, S(q, · )i = h(νM ∗ = hp, νM (τ (vl )∗ , S(q, · ))i X ∗ −1 (A )lj S(τ (vj ), · , S(q, · )) = p, νM j
X (A−1 )lj hp, S(τ (vj )q, ·)i = j
X = (A−1 )lj S(p, τ (vj )q).
(16)
j
The first three equalities come from the definition of the action of D(b+) on M, the last two from formula (14). We combine (15), (16), and Corollary 4.2 to obtain: X X ∗ ∗ uk (uk ) + vl (vl ) p, q S(1+,α p, q) = S k
= S
X
vl (vl∗ )p, q
l
= S((vl∗ )p, (−vl )q)
l
X (A−1 )lj S(p, τ (vj )vl q). = −
(17)
j,l
Now we trace the arrows in the alternative direction. For x ∈ (n− )α , consider the map x: M ∗ → M ∗ via the action of D(b−). Denote this action by ‘·’. Let xq denote the standard action of b− on M. We have hx · S(p, ·), qi = hS(p, −τ (τ (x)) · ), qi = S(p, −xq) for all x ∈ b− ; p, q ∈ M.
(18)
DIFFERENTIAL EQUATIONS COMPATIBLE WITH KZ EQUATIONS
151
Consider the map (τ (vl ))∗ : M ∗ → M ∗ via the action of D(b− ). h(τ (vl ))∗ S(p, · ), qi =
X (A−1 )lj S(τ (vj ), ·) S(p, · ), q j
X (A−1 )lj hS(τ (vj )p, · ), qi = j
X = (A−1 )lj S(τ (vj )p, q).
(19)
j
Finally combine (18), (19), and Corollary 4.2 with uk ∈ ker S to get h−1−,α S(p, · ), qi X X ∗ ∗ τ (uk ) τ (uk ) + τ (vl ) τ (vl ) S(p, · ), q = − k
l
X X ∗ ∗ τ (uk ) S(p, −τ (uk ) · ), q − τ (vl ) S(p, −τ (vl ) · ), q =− k
= −
X
k
+ −
l
τ (uk )∗ S(−τ (−τ (uk ))p, · ), q +
X
∗
τ (vl ) S(−τ (−τ (vl ))p, · ), q
l
X −1 (A )lj S(τ (vj )(vl )p, · ), q =0− j,l
X (A−1 )lj S(τ (vj )(vl )p, q) =− j,l
X =− (A−1 )lj S(p, τ (vl )vj q).
(20)
j,l
Since the matrix A is symmetric (17) and (20) prove that the diagram is commutative. 2 As a corollary of the lemma we have that 1+,α naturally acts on L = M/ ker(S: M → M ∗ ). We describe this action. Consider the Kac–Moody algebra g¯ = g/ ker(S: g → g∗ ). Let x 7→ x¯ denote the canonical projections M → L, g → g¯ . ker S is an ideal and the form S induces a non-degenerate Killing form on g¯ via the formula K(x, y) = −S(τ (x), y), see [11]. K induces a non-degenerate pairing between root spaces g¯ α and g¯ −α . Let (el(α) ) be a basis of g¯ α , and let (fl(α) ) be the P ¯ α = l fl(α)el(α) . dual basis of g¯ −α with respect to K. Let 1
152
G. FELDER ET AL.
COROLLARY 4.4. The following diagram is commutative M
L
1+,α D(b+ )-act ion ¯α 1 st andard
M
L.
∗ x¯ 7→ S(x, · ). We keep the notation Proof. L(3) ∼ = Im{S: M → M P } via from the lemma above. Set wl = − j (A−1 )lj τ (vj ). From the computation in the lemma we have X X −1 − 1+,α p = S(1+,α p, · ) = S (A )lj τ (vj ) vl p, ·
= S
X l
l
wl vl p, ·
=
j
X l
wl vl p =
X
w¯ l v¯l p. ¯
l
Finally notice that the set (v¯l ) forms a basis of g¯ α , and the set (w¯ l ) forms the dual 2 basis of g¯ −α with respect to K. COROLLARY 4.5. Fix λ ∈ Nr . Let m ∈ Mλ , and let (mj ) be a basis of Mλ , and let (m∗j ) be the dual basis of Mλ∗ . Then the following decomposition holds X h−1−,α m∗j , mimj . 1+,α m = j
Proof. Let y ∈ b+ and x = τ (y) ∈ b− . Let p ∈ M, and φ ∈ M ∗ . As in the proof of Lemma 4.3 hy ∗ p, φi = hp, x ∗ φi, where D(b+) acts on M and D(b−) acts on M ∗ . Moreover hy · p, φi = hp, −xP · φi, where D(b+) acts on P M and D(b−) acts on M ∗ . Finally noting that 1+,α = i yi(α) (yi(α) )∗ and 1−,α = i τ (yi(α) )∗ τ (yi(α) ) we have X X hm∗j , 1+,α mimj = h−1−,α m∗j , mimj . 2 1+,α m = j
4.3.
j
AN INTEGRAL FORM OF THE DYNAMICAL DIFFERENTIAL EQUATIONS
Our aim now is to rewrite the Dynamical equations in a form related to the hypergeometric solutions. Fix λ = (m1 , . . . , mr ) ∈ Nr . Let M be a tensor product of highest weight P modules of the Kac–Moody Lie algebra without Serre’s relations g. Set 3 = 3j , the sum of the respective highest weights. P Consider the weight mi αi . Fix a highestspace Mλ of M with weight 3 − α(λ), where α(λ) = weight vector vj for each module M(3j ), j = 1, . . . , n. To every I ∈ P (λ, n) we associate a vector fI v = fi11 · · · fis1 v1 ⊗ · · · ⊗ fi1n · · · fisnn vn in Mλ , cf. Section 3. 1 Note that the vectors (fI v)I ∈P (λ,n) form a basis of Mλ .
DIFFERENTIAL EQUATIONS COMPATIBLE WITH KZ EQUATIONS
153
n− acts on Mλ∗ via the D(b−) action. Therefore U (n−) acts on Mλ∗ . Explicitly, x · φ( · ) = φ(−x · ) for x ∈ n−, φ ∈ M ∗ (cf. Section 4.1), where the action on the left-hand side is the D(b−) one and the action on the right-hand side is the standard one. Let V be a vector space freely generated by f1 , . . . , fr . Therefore we have an inclusion of tensor algebras T (V ) ⊂ T (n− ). Moreover T (V ) is an associative enveloping algebra of the Lie algebra n− . Since T (V ) is a free associative algebra, T (V ) is isomorphic to the universal enveloping algebra U (n−). From now on we will refer to the monomial basis of T (V ) as to the monomial basis of U (n− ), and to the dual of the monomial basis of T (V ) as to the monomial basis of UP (n−)∗ . aj xj Rewrite a commutator x ∈ n− as an element of U (n−) in the form x = where aj ∈ Z, and xj ’s are elements of the monomial basis of U (n−). Thus E D X (21) hx · φ, · i = φ, − aj xj · . Denote i: n− → U (n− ) the inclusion monomorphism. Let σj ∈ {1, . . . , r} P for j = 1, . . . , k. Let the positive root of g, nj=1 ασj , correspond to the r-tuple P λ0 ∈ Nr , i.e. α(λ0 ) = nj=1 ασj . Define an element 1σ1 ,...,σk of (n− )∗λ0 via the rule h1σ1 ,...,σk , xi = h(fσ1 · · · fσk )∗ , i(x)i, (fσ1 · · · fσk )∗ ∈ U (n−)∗ .
where x ∈ n− , and (22)
Thus h1σ1 ,...,σk , xi is the coefficient of fσ1 · · · fσk in the decomposition of i(x) into a sum of monomials. P LEMMA 4.6. Let α = rk=1 m0i αi be a positive root of g. Let I ∈ P (λ, n). Set λ0 = (m01 , . . . , m0r ) ∈ Nr . Then we have X ∗ 1i1 ,...,im0 (eim0 · · · ei1 ) (fI v)∗ , (23) −1−,α (fI v) = (i1 ,...,im0 )∈P (λ0 ,1)
where 1i1 ,...,im0 acts according to the D(b−) action on Mλ∗ , the product of e’s acts P (j ) (j ) on one tensor factor at a time (eim0 · · · ei1 ) = nj=1 eim0 · · · ei1 , and ej acts via the standard action (9) on each tensor factor. Proof. Note that α = α(λ0 ). Let x ∈ (n− )λ0 . If there exists i such that m0i > mi , then formula (21) implies x · (fI v)∗ = 0 for every I ∈ P (λ, n) because each than fI v. monomial xj in the U (n−) expansion of x has more fi ’s P m0i . First consider the Now let m0i 6 mi for any 1 6 i 6 r. Set m0 = ∗ case n = 1, I = (i1 , . . . , im ) ∈ P (λ, 1). Let (xj ) be a basis of (n− )∗λ0 such that x1∗ = 1i1 ,...,im0 , and let (xj ) be the dual basis of (n−)λ0 . Formula (22) implies that the coefficient of the monomial fi1 · · · fim0 in the U (n−) expansion of x1 is 1, and the coefficient of fi1 · · · fim0 in the U (n−) expansion of xj is 0 for j > 1. Now use (21) to obtain −x1 · (fI v)∗ = (fim0 +1 · · · fim v)∗ = eim0 · · · ei1 (fI v)∗ , −xj · (fI v)∗ = 0 for j > 1.
(24)
154
G. FELDER ET AL.
For any element I 0 = (i10 , . . . , im0 0 ) ∈ P (λ0 , 1), such that I 0 6= (i1 , . . . , im0 ) we have eim0 0 · · · ei10 (fI v)∗ = 0. The proof for n = 1 is finished. Let n be arbitrary natural number. (fI v)∗ = (fi11 · · · fis1 v1 )∗ ⊗ · · · ⊗ (fi1n · · · fisnn vn )∗ . 1
Formula (21) implies h−x · (fI v)∗ , · i = h(fI v)∗ , x · i = h(fI v)∗ , This and the computation for n = 1 give X
∗
−1−,α (fI v) =
n X
(i1 ,...,im0 )∈P (λ0 ,1) j =1
Pn
j =1
x (j ) · i.
1i1 ,...,im0 eim0 · · · ei1 (fI v)∗ . (j )
(j )
2
P α For every positive root α = mi αi of g, set λα = (mα1 , . . . , mαr ) ∈ Nr . Now we combine Corollary 4.5 and Lemma 4.6 to obtain the following form of the Dynamical KZ equation. P 0 LEMMA 4.7. Let u(µ, z) = I ∈P (λ,n) uI fI v, and let µ ∈ h be a direction of differentiation. The Dynamical differential equation (13) is equivalent to the equation κ∂µ0 u =
n X
zi (µ0 )(i) u +
i=1
×
X J ∈P (λ,n)
4.4.
X
X X hα, µ0 i × hα, µi I ∈P (λ,n) α>0 1i1 ,...,im0 (eim0
∗ · · · ei1 ) (fJ v) , fI v uI fJ v. 2
(25)
(i1 ,...,im0 )∈P (λα ,1)
A SYMMETRIZATION PROCEDURE
The definition of the hypergeometric differential form involves a symmetrization procedure, see Section 3. Now we will study the behavior of the operator 1i1 ,...,im0 eim0 · · · ei1 for (ii , . . . , im0 ) ∈ P (λ0 , 1), where α = α(λ0 ), under the same type of symmetrization procedure. Complexes [10]. For a Lie algebra g and a g-module M, denote by C• (g, M) the standard chain complex of g with coefficients in M. Cp (g, M) = 3p g ⊗ M and d: gp ∧ · · · ∧ g1 ⊗ x p X = (−1)i−1 gp ∧ · · · ∧ gbi ∧ · · · ∧ g1 ⊗ gi x+ i=1
+
X
16i<j 6p
(−1)i+j gp ∧ · · · ∧ gbj ∧ · · · ∧ gbi ∧ · · · ∧ g1 ∧ [gj , gi ] ⊗ x.
155
DIFFERENTIAL EQUATIONS COMPATIBLE WITH KZ EQUATIONS
the complex Let 31 , . . . , 3n ∈ h∗ . Set M = M(31 ) ⊗ · · · ⊗ M(3n ). Consider L C• (n− , M). We have the weight decomposition C• (n− , M) = λ∈Nr C• (n− , M)λ . In Section 4.1 we recalled a Lie algebra structure on n− ∗ and a n− ∗ -module Let C• (n− ∗ , M ∗ ) be the corresponding standard chain complex: structure on M ∗ . L ∗ ∗ ∗ ∗ C• (n− , M ) = λ∈Nr C• (n− , M )λ . The covariant form induces a graded homomorphism of complexes, S: C• (n− , M) →PC• (n−∗ , M ∗ ), see [10]. mi . Define a subgroup 6λ of the Let λ = (m1 , . . . , mr ) ∈ Nr , and m = symmetric group 6m via the direct product 6λ = 6m1 × · · · × 6mr , where 6mj Pj −1 Pj −1 permutes the set of indices { p=1 mp + 1, . . . , p=1 mp + mj }. Introduce a free f˜1 , . . . , f˜m . Define a map of Lie algebras n− → n f Lie algebra n f − on − Pgenerators mi ˜ by setting fi 7→ j =1 fm(i)+j , where m(i) = m1 + · · · + mi−1 . It induces a map of n− , U (f n− )⊗n )6λ . Set e λ = (1, 1, . . . , 1). Let complexes C• (n− , U (n− )⊗n ) → C• (f | {z } m ⊗n
s: C• (n−, U (n− ) )λ →
C• (f n− , U (f n− )⊗n )eλ6λ
(26)
be the previous map composed with the projection on the e λ-component. e On the other hand there is a map of Lie algebras n f − → n− defined by fj 7→ ⊗n n− , U (f n− ) )eλ → fi , for m(i) < j 6 m(i + 1). It induces the map πλ : C• (f C• (n− , U (n− )⊗n )λ . Note that s(y) equals the sum over the preimages of y under πλ , for any y ∈ C• (n− , U (n−)⊗n )λ . Each such preimage is uniquely described by an element σ ∈ 6λ . . . . , im ) ∈ P (λ, 1). Consider fI as an element EXAMPLE. Let n = 1 and I = (i1 ,P of C0 (n− , U (n−))λ . Then s(fI ) = σ feσ1 . . . feσm , where the sum is over the set {σ ∈ 6(I )} ∼ = 6λ . LEMMA 4.8. Let n = 1, I ∈ P (λ, 1), and m0 6 m. Then the map s ∗ : n− , U (f n−))eλ6λ )∗ → (C1 (n− , U (n−))λ )∗ has the following property (C1 (f 1 X ∗ ∗ e e e s 1σ ,...,σ 0 ⊗ (fσm0 +1 . . . fσm ) |6λ | σ ∈6(I ) 1 m = 1i1 ,...,im0 ⊗ (fim0 +1 . . . fim )∗ .
(27)
Proof. Choose a basis (xj∗ ) of n− ∗ such that x1∗ = 1i1 ,...,im0 and let (xj ) be the dual basis of n− . Take a tensor product of the basis (xj ) with the monomial basis in U (n− ) to get a basis in n− ⊗ U (n− ). Let us compare the two sides of (27) on that basis. 1i1 ,...,im0 ⊗ (fim0 +1 . . . fim )∗ (x1 ⊗ fim0 +1 . . . fim ) = 1. The right-hand side is zero on any other element of the basis. i: n f n− ) → U (f n−)⊗U (f n− ) Let i: n− ⊗U (n−) → U (n− )⊗U (n−), ande − ⊗U (f e e be the natural inclusion maps. Clearly i ◦ s = s ◦ i. Moreover, h1σi1 ,...,σi 0 , s(x)i = m i ◦ s(x))i by definition for x ∈ n− . For a fixed σ ∈ 6λ we have h(feσi1 · · · feσi 0 )∗ , (e m
eσ1 ,...,σ 0 ⊗ (feσ 0 . . . feσm )∗ ), x1 ⊗ fi 0 . . . fim i hs (1 m m +1 m +1 ∗
156
G. FELDER ET AL.
eσi ,...,σi ⊗ (feσi = h1 0 1
m0 +1
m
. . . feσim )∗ , s(x1 ⊗ fim0 +1 . . . fim )i
= h(feσ1 . . . feσm0 )∗ ◦e i ⊗ (feσm0 +1 . . . feσm )∗ , s(x1 ⊗ feim0 +1 . . . feim )i = h(feσ1 . . . feσ 0 )∗ ⊗ (feσ 0 . . . feσm )∗ , s(i(x1 ) ⊗ fei 0 . . . feim )i. m
m +1
m +1
The duality of the bases implies i(x1 ⊗ fim0 +1 . . . fim ) = (fi1 . . . fim0 + other monomials) ⊗ fim0 +1 · · · fim , and s ◦ i(x1 ⊗ fim0 +1 . . . fim ) = (feσ1 . . . feσm0 + other eσ1 ,...,σ 0 ⊗ (feσ 0 . . . feσm )∗ ), x1 ⊗ monomials) ⊗ feσm0 +1 · · · feσm . Therefore, hs ∗ (1 m m +1 fim0 +1 . . . fim i = 1. The same way we check that the left-hand side is zero on the other basis elements. 2 COROLLARY 4.9. Let π be the restriction of the projection πλ to the subspace n− , U (f n− )⊗n )eλ6λ of C. (f n−, U (f n− )⊗n )eλ . Let J = (j1 , . . . , jm ) ∈ P (λ, n) and C. (f I = (i1 , . . . , im0 ) ∈ P (λα , 1) for a positive root α. Then we have X X ∗ e e eτ1 ,...,τ 0 (e e . . .e e ) ( f . . . f ) 1 τm0 τ1 σ1 σm m τ ∈6(I ) ∗
σ ∈6(J )
= π ((1i1 ,...,im0 (eim0 . . . ei1 ))(fj1 · · · fjm )∗ ).
(28)
Proof. Assume n = 1. The general case follows from this one because the eτ1 ) and (eim0 . . . ei1 ) act on one tensor factor at a time. Since operators (e eτm0 . . .e s ◦ π = |6λ |id we have π ∗ ◦ s ∗ = |6λ |id. s ∗ and π ∗ are maps of complexes, i.e. they commute with the corresponding differentials. Thus, Lemma 4.8 implies after applying differentials and taking π ∗ from both sides X eσ1 ,...,σ 0 (feσ 0 · · · feσm )∗ . π ∗ (1j1 ,...,jm0 (fjm0 +1 · · · fjm )∗ ) = 1 m m +1 σ ∈6(J )
The right-hand side of our formula is non-zero if and only if m0 6 m and ik = jk for 1 6 k 6 m0 . Thus we compute π ∗ (1i1 ,...,im0 eim0 . . . ei1 (fj1 . . . fjm )∗ ) = π ∗ (1j1 ,...,jm0 (fjm0 +1 . . . fjm )∗ ) X eσ1 ,...,σ 0 (feσ 0 . . . feσm )∗ = 1 m m +1 σ ∈6(J )
=
X
eσ1 ,...,σ 0e e . . .e eσ1 (feσ1 . . . feσm )∗ . 1 m σm0
σ ∈6(J )
Note that (σ1 , . . . , σm0 ) ∈ 6(I ). For any other τ ∈ 6(I ), we have e eτm0 . . .e eτ1 (feσ1 . . . feσm )∗ = 0. Therefore we rewrite the last equality in the form (28).
2
DIFFERENTIAL EQUATIONS COMPATIBLE WITH KZ EQUATIONS
157
Our last step is to show that the Dynamical equations for any Lie algebra in any weight space follow from the Dynamical equations for any Lie algebra in a weight space with weight e λ = (1, 1, . . . , 1). Fix a finite-dimensional complex vector space h, a non-degenerate symmetric bilinear form ( · , · ) on h, and a set of linearly independent ‘simple roots’ α1 , . . . , αr ∈ h∗ . Consider the corresponding Kac–Moody Lie algebra without Serre’s relations, g, defined at the beginning of Section 4.1. Recall that b: h → h∗ denotes the isomorphism induced by the bilinear form, hi = b−1 (αi ), and and the form ( · , · ) is transfered to h∗ via the map b. r Ppα(λ) = PrFix λ = (m1 , . . . , mr ) ∈ N . Consider the corresponding positive root m α of g. Up to reordering of the α’s we can assume that α(λ) = i=1 mi αi i=1 i i where mi > 0 for 1 6 i 6 P p and p 6 r is fixed. The corresponding coloring mi } → {1, . . . , p}. We use the following linear function is cλ : {1, . . . , m = algebraic fact when symmetrizing. PROPOSITION 4.10. Let h be a finite-dimensional vector space with a non-degenerate symmetric bilinear form ( · , · ), and a set of linearly independent vectors h with a non(hi )ri=1 ⊂ h. Then there exists a finite-dimensional vector space e degenerate symmetric bilinear form ( · , · )1 , a set of linearly independent vectors e e (e hj )m j =1 ⊂ h, and a monomorphism sh : h → h such that P i e (a) sh (hi ) = m1i m j =1 hm(i)+j , where m(i) = m1 + · · · + mi−1 , i = 1, . . . , p; 0 (b) (e hj , sh (h ))1 = (hc(j ) , h0 ) and (h0 , h00 ) = (sh (h0 ), sh (h00 ))1 for any h0 , h00 ∈ h, j = 1, . . . , m. Proof. Let q = dim h. Complete the set h1 , . . . , hr to a basis h1 , . . . , hr , hr+1 , h0 = C{e h1 , . . . , e hm , e hm+1 , . . . , hq of h. Consider a complex linear space e . . . ,e hm+q−p }. Extend the coloring function c: {1, . . . , m + q − p} → {1, . . . , q} setting c(m + j ) = r + j for j = 1, . . . , q − p. Define a symmetric degenerate hj , e hk )1 = (hc(j ) , hc(k) ) for 1 6 j, k 6 m + q − p. bilinear form on e h0 by the rules (e The rank of the form is q and the dimension of its kernel is m − p. There exists an extension e h of the vector space e h0 and an extension of ( · , · )1 to a non-degenerate symmetric bilinear form on e h. P i e Define a monomorphism sh by sh (hi ) = m1i m j =1 hm(i)+j , where m(i) = m1 + hm+j for j = 1, . . . , q − p. · · · + mi−1 , i = 1, . . . , q, and note that sh (hp+j ) = e Now checking (b) on a basis is straightforward. 2 hj , · )1 ∈ e h∗ for j = 1, . . . , m. Consider e g, a Kac–Moody Lie algeSet e αj = (e αj )m bra without Serre’s relations corresponding to the data e h, ( · , · )1 , and (e j =1 . Note 0 0 0 e that 1 6 j 6 m implies he αj , sh (h )i = (hj , sh (h ))1 = (hc(j ) , h ) = hαc(j ) , h0 i for 0 any h ∈ h. Let M = M(31 )⊗· · ·⊗M(3n ) be a tensor product of Verma modules for g with corresponding highest weights 31 , . . . , 3n ∈ h∗ . Since sh is a monomorphism, e1 , . . . , 3 en ∈ e h∗ → h∗ is a linear epimorphism. Choose highest weights 3 h∗ sh∗ : e
158
G. FELDER ET AL.
ej ) = 3j for 1 6 j 6 n, and consider the corresponding tensor such that sh∗ (3 e 3 e = M( e 3 en ). e1 ) ⊗ · · · ⊗ M( product of Verma modules for e g, M P LEMMA 4.11. Let e λ = (1, 1, . . . , 1). Let e u(e µ, z) = K∈P (eλ,n) e uK feK be a hyper| {z } m
geometric solution of the Dynamical equations with values in the e λ weight space e∼ u)(sh (µ), z) is a hypergeometof a e g-module M n−)⊗n . Then u(µ, z) = π(e = U (f ric solution of the Dynamical equations with values in the λ weight space of a P ⊗n ∼ g-module M = U (n−) , i.e. u = I ∈P (λ,n) uI fI . S Proof. Note that by definition P (e λ, n) = I ∈P (λ,n) {K ∈ 6(I )}. From the definition P of the hypergeometric differential form, see Section 3, it follows that uK . Therefore uI = k∈6(I ) e X X e uk feK π(e u) = π I ∈P (λ,n) K∈6(I )
X X
=
I ∈P (λ,n)
e uk fI =
X
uI fI .
(29)
I ∈P (λ,n)
K∈6(I )
µ = sh (µ) and Fix a point µ ∈ h and a direction of differentiation µ0 ∈ h. Denote e e µ0 = sh (µ0 ). Since ! m n X X hαc(j ) , µitj + h3l , µi ∂µ0 exp − j =1
=
−
l=1
m X
0
hαc(j ) , µ itj +
j =1
=
−
0
he αj , e µ itj +
= ∂e µ0 exp
! m n X X − hαc(j ) , µitj + h3l , µi
0
h3l , µ i exp
l=1
m X j =1
!
n X
j =1
!
n X
el , e h3 µ i exp 0
−
l=1
! m n X X el , e − he αj , e µitj + h3 µi , j =1
l=1
m X
n X
j =1
l=1
he αj , e µitj +
!
el , e h3 µi
(30)
l=1
we have π(∂e u(e µ, z)) = ∂µ0 π(e u)(µ, z). If I = (I1 , . . . , In ) ∈ P (λ, n) and K = µ0e j j j j (K1 , . . . , Kn ) ∈ 6(I ), where Kj = (k1 , . . . , ksj ) and Ij = (i1 , . . . , isj ), then + + * * sj sj X X 0(j ) e 0 0 ej − e α j,e µ feK = 3j − α j , µ feK , e µ fK = 3 kl
il
l=1
l=1
+ * sj X X X 0(j ) 0 e e π e µ e uK fK = 3j − αi j , µ e uK π(fK ) l
K∈6(I )
* = 3j −
l=1 sj
X
0
αi j , µ
+
K∈6(I )
X
l
l=1
K∈6(I )
e uK fI
DIFFERENTIAL EQUATIONS COMPATIBLE WITH KZ EQUATIONS
* = 3j −
sj X
159
+ 0
αi j , µ uI fI = µ0(j ) fI .
(31)
l
l=1
Combine formulae (29) and (31) to obtain ! n n X X 0(j ) zj e µ e u = zj µ0(j ) u. π j =1
(32)
j =1
P Let α = ri=1 m0i αj be a positive root for g. Lemma 4.6 gives a necessary conMλ∗ . Namely m0i 6 mi for all i = 1, . . . , r. dition for a non-zero action of 1−,α on P αj of e g a necessary condition for Analogously, for a positive root e α = m j =1 pj e ee∗ is pj = 0, 1 for j = 1, . . . , m. Call all such a non-zero action of 1−,eα on M λ α’s (e α ’s) λ-admissible (e λ-admissible). Since sh∗ (e αj ) = αc(j ) , sh∗ maps the set of e λ-admissibleProots of e g onto the set of λ-admissible roots ofP g. 0 0 m0i . For any e α , such Let α = mi αi be a λ-admissible root for g and m = ∗ 0 0 α ) = α, we have he α, e µ i/he α, e µi = hα, µ i/hα, µi. Consider that sh (e X he α, e µ0 i 1+,eαe π u , he α, e µi ∗ e α , sh (e α)=α
where the sum is over e λ-admissible roots. Corollary 4.5 applied to the basis (feK )K∈P (eλ,n) of Mλ gives X hα, µi he α, e µ0 i u π 1+,eαe hα, µ0 i he α, e µi e α, sh∗ (e α)=α X X ∗ − =π 1−,eα (feK ) ,e u feK e α , sh∗ (e α)=α
K∈P (e λ,n)
=
X I ∈P (λ,n)
X
−
X u fI . 1−,eα (feK )∗ ,e
e α, sh∗ (e α)=α
Lemma 4.6 asserts that X − 1−,eα =
X
e α , sh∗ (e α)=α
e α , sh∗ (e α)=α
(33)
K∈6(I )
X
e elm0 . . .e el1 ) . 1l1 ,...,lm0 (e
(l1 ,...,lm0 )∈P (e λe α ,1)
Rearrange the summation using that sum over (l1 , . . . , lm0 ) ∈ P (e λeα , 1) such that α ) = α equals the sum over (p1 , . . . , pm0 ) ∈ 6(J ) such that J = (j1 , . . . , lm0 ) ∈ sh∗ (e P (λα , 1). Combine such rearrangement with Lemma 4.6 and Corollary 4.9 to simplify formula (33). X he α, e µ0 i hα, µi π 1 e u +,e α hα, µ0 i he α, e µi ∗ e α, sh (e α)=α
160
G. FELDER ET AL.
=
X I ∈P (λ,n)
J ∈P (λα ,1)
× =
X I ∈P (λ,n)
=
X
I ∈P (λ,n)
=
X
X
X
ep1 ,...,p 0 (e e . . .e e ) × 1 pm0 p1 m
X (p1 ,...,pm0 )∈6(J )
∗ e u fI (fK ) ,e
K∈6(I )
π ∗ (1j1 ,...,jm0 (ejm0 . . . ej1 )(fI )∗ ),e u fI
X J ∈P (λα ,1)
X
. . . ej1 )(fI ) , π(e u) fI ∗
1j1 ,...,jm0 (ejm0
J ∈P (λα ,1)
h−1−,α (fI )∗ , uifI = 1+,α u.
(34)
I ∈P (λ,n)
Finally (32) and (34) imply n X
∂µ0 u = π(∂e u) = π µ0e
0(j )
zj e µ
+
j =1
=
n X
zj µ0(j ) +
j =1
4.5.
X he α, e µ0 i e α >0
he α, e µi
X hα, µ0 i α>0
hα, µi
! ! 1+,eα e u !
1+,α u.
2
(35)
THE PROOF OF LEMMA 4.1
Recall that the linear map νM : M → b− ⊗ M has the following property νM (x) = 1 −1 (b (3−α(λ)))⊗x +νM− (x), where x ∈ Mλ , νM− (x) ∈ n− ⊗M, and b−1 : h∗ → 2 h is defined at the beginning of Section 4.1. Let a ∈ h, x ∈ Mλ . Since S(b ⊗ x, a ⊗ y) = S(a, b)S(x, y) for any b ∈ g, y ∈ M, and h is orthogonal to n− with respect to S, and ( · , · ) coincides with S on h we have (36) S(νM− (x), a ⊗ y) = S 12 (b−1 (3 − α(λ))), a S(x, y) = 12 S(x, ay). This proves the first equality in the lemma. To prove the second part for a monomial x, we use double induction by the number of tensor factors and the number of f 0 s in x. We use νM− (x) instead of νM (x) because of the orthogonality mentioned above. Let M = M(31 ) be a highest weight module of g with a highest vector v, and a ∈ n− , y ∈ M. Since S(νM− (v)) = 0, we have S(νM− (v), a ⊗y) = 0 = S(v, ay). The inductive step is as follows. Assume S(νM− (x), a ⊗ y) = S(x, ay). Then S(νM− (fi x), a ⊗ y) = S(fi ⊗ hi x, a ⊗ y) + S(fi νM− (x), a ⊗ y) = S(fi , a)S(hi x, y) + S(νM− (x), ei (a ⊗ y))
DIFFERENTIAL EQUATIONS COMPATIBLE WITH KZ EQUATIONS
161
= S(fi , a)S(hi x, y) + S(νM− (x), [ei , a] ⊗ y) + S(νM− (x), a ⊗ ei y) = S(fi , a)S(hi x, y) + S(νM− (x), [ei , a] ⊗ y) + S(x, aei y) = S(fi , a)S(hi x, y) + S(νM− (x), [ei , a] ⊗ y)+ + S(x, ei ay) − S(x, [ei , a]y) = S(fi , a)S(hi x, y) + S(νM− (x), [ei , a] ⊗ y)+ + S(fi x, ay) − S(x, [ei , a]y). (37) If a = fi , then (37) and the properties S(x, hy) = S(hx, y), S(νM− (x), h ⊗ y) = 0 for h ∈ h imply S(νM− (fi x), a ⊗ y) = S(hi x, y) + S(νM− (x), hi ⊗ y) + S(fi x, ay) − S(x, hi y) = S(fi x, y). If a is orthogonal to fi with respect to S, then (37) and the inductive hypothesis give S(νM− (fi x), a ⊗ y) = 0 + S(x, [ei , a]y) + S(fi x, ay) − S(x, [ei , a]y) = S(fi x, a). Thus the statement is proved for one tensor factor. Assume that S(νM− (fi x), a ⊗ y) = S(x, ay) for a module M, which is a tensor product of up to n − 1 tensor factors (n > 2). Pn (k) Let M = M(31 ) ⊗ · · · ⊗ M(3n ). Recall that νM− (x) = k=1 νM (x)− , (j ) (j ) (k) (k) (k) (fi x)− = fi νM (x)− for k 6= j , and νM (fi(k) x)− = fi ⊗ h(k) where νM i x + (j ) (k) (k) (k) fi νM (x)− . The following commutation relations will be useful. S(fi νM− (x), (j ) (j ) (j ) (j ) (k) (x), a ⊗ei y), for j 6= k, and S(fi νM (x)− , a ⊗y) = S(νM− (x), a ⊗y) = S(νM− (j ) (j ) [ei , a]⊗y)+S(νM− (x), a⊗ei y). Both equalities are corollaries of the Lemma 4.1 for one tensor factor, and the definition of S, e.g. S(x1 ⊗ · · · ⊗ xj −1 Q⊗ νM− (xj ) ⊗ xj +1 ⊗· · · ⊗xn , a ⊗y1 ⊗· · · ⊗yj ⊗· · ·⊗yn ) = S(νM− (xj ), a ⊗yj ) k6=j S(xk , yk ). In all formulae 1 6 i 6 r, and 1 6 j, k 6 n, and the upper script indicates the tensor factor where the action is applied. Let a ∈ n− . The base for the induction is exactly as for n = 1. The inductive step is as follows. (j )
S(νM− (fi x), a ⊗ y) X (j ) (j ) (j ) (k) = S(νM− (fi x), a ⊗ y) + S(νM− (fi x), a ⊗ y) k6=j
=
X
(j )
(k) S(fi νM− (x), a ⊗ y)+
k6=j (j )
(j ) (j )
+ S(fi ⊗ hi x, a ⊗ y) + S(fi νM (x)− , a ⊗ y) (j )
(j )
= S(fi , a)S(hi x, y) + S(νM− (x), [ei , a] ⊗ y)+ n X (j ) (k) + S(νM− (x), a ⊗ ei y). k=1
(38)
162
G. FELDER ET AL.
The result for one tensor factor gives X X (j ) (j ) (k) S(νM− (x), a ⊗ ei y) = S(x, a (k) ei y) k
k
=
X
(j )
S(x, ei a (k) y) − S(x, [ei , a](j ) y)
k (j )
= S(fi x, ay) − S(x, [ei , a](j ) y).
(39)
(k) If a = fi , then (38), (39) and the properties S(x, h(k) y) = S(h(k) x, y), S(νM− (x), h ⊗ y) = 0 for h ∈ h, k = 1, . . . , n imply (j )
(j )
(j )
(j )
S(νM− (fi x), a ⊗ y) = S(hi x, y) + 0 + S(fi x, ay) − S(x, hi y) (j )
= S(fi x, ay). If a is orthogonal to fi with respect to S, then (38), (39) and the result for one tensor factor give (j )
S(νM− (fi x), a ⊗ y) (j )
= 0 + S(x, [ei , a](j ) y) + S(fi x, ay) − S(x, [ei , a](j ) y) (j )
= S(fi x, ay). This finishes the inductive argument. The lemma is proved.
2
5. Flags, Orlik–Solomon Algebra, Hypergeomertic Differential Forms In this section we will formulate results from [10] which define a map between the complex of hypergeometric differential forms and the complex C• (n− ∗ , M ∗ ) for a suitable Lie algebra n− and a n− -module M. 5.1.
COMPLEXES
Let W be an affine complex m-dimensional space and let C be a configuration of hyperplanes in W . Define Abelian groups Ak (C, Z), 0 6 k 6 m, as follows. A0 (C, Z) = Z. For k > 1, Ak (C) is generated by k-tuples (H1 , . . . , Hk ), Hi ∈ C, subject to the relations: (H1 , . . . , Hk ) = 0 if H1 , . . . , Hk are not in general position (codim H1 ∩ · · · ∩ Hk 6= k); (Hσ (1), . . . , Hσ (k) ) = (−1)|σ | (H1 , . . . , Hk ) for any permutation σ ∈ 6k ; Pk+1 i ˆ i=1 (−1) (H1 , . . . , Hi , . . . , Hk+1 ) = 0 for any (k + 1)-tuple H1 , . . . , Hk+1 which is not in general position L and such that H1 ∩ · · · ∩ Hk 6= 0. k The direct sum A• (C, Z) = m k=0 A (C, Z) is a graded skew commutative algebra with respect to the multiplication (H1 , . . . , Hk ) · (H10 , . . . , Hl0 ) =
DIFFERENTIAL EQUATIONS COMPATIBLE WITH KZ EQUATIONS
163
(H1 , . . . , Hk , H10 , . . . , Hl0 ). A• (C, Z) is called the Orlik–Solomon algebra of the configuration C. Flags. For 0 6 k 6 m, denote by Flagk (C) the set of all flags L0 ⊃ L1 ⊃ k · · · ⊃ Lk , where Li is an edge of C of codimension i. Denote by Flag (C) the k free Abelian group on Flagk (C) and by F l k (C, Z) the quotient of Flag (C) by the following relations. For every i, 0 < i < k, and a flag with a gap, Fˆ = P (L0 ⊃ · · · ⊃ Li−1 ⊃ Li+1 ⊃ k j L ), where L is an edge of codimension j , we set F ⊃Fˆ F = 0 in F l k (C, Z), where the summation is over all flags F = (L˜ 0 ⊃ L˜ k ) ∈ Flagk (C) such that L˜ j = Lj for all j 6= i. To define the relation between Ak (C, Z) and F l k (C, Z) we define the following map. For (H1 , . . . , Hk ) in the general position, Hi ∈ C, define F (H1 , . . . , Hk ) = (H1 ⊃ H12 ⊃ · · · ⊃ H12...k ) ∈ Flagk (C), where H12...i = H1 ∩ H2 ∩ · · · ∩ Hi . For a flag F ∈ Flagk (C), define a functional δF ∈ F l k (C, Z)∗ as δF (F 0 ) = 1 if F 0 = F and δF (F 0 ) = 0 otherwise. For (H1 , . . . , Hk ) in general position, define a map X (−1)|σ | δF (Hσ1 ,...,Hσk ) . (40) ϕ k (H1 , . . . , Hk ) = σ ∈6k
Thus we have a homomorphism ϕ k : Ap (C, Z) → F l p (C, Z)∗ . The following statements are from [10]. All groups F l p (C, Z) are free over Z. Ap (C, Z) and F l p (C, Z) are dual and the map ϕ k is an isomorphism. Set Ak (C) = Ak (C, Z) ⊗Z C and F l k (C) = F l k (C, Z) ⊗Z C for all k. From now on we assume that the configuration C is weighted, that is, to any hyperplane H ∈ C its weight, a number a(H ) ∈ C, is assigned. Define the quasiclassical weight of any edge L of C as the sum of the weights of all hyperplanes that contain the edge. Say that a k-tuple H¯ = (H1 , . . . , Hk ), Hi ∈ C, is adjacent to a flag F if there exists σ ∈ 6k such that F = F (Hσ1 , . . . , Hσk ). This permutation σ is unique. Denote it by σ (H¯ , F ). Define a symmetric bilinear form S k on F l k (C). For F, F 0 ∈ Flagk (C), set S k (F, F 0 ) =
1X ¯ ¯ 0 (−1)σ (H ,F )σ (H ,F ) a(H1 ) . . . a(Hk ), k
(41)
where the summation is over all H¯ = (H1 , . . . , Hk ) adjacent to both F and F 0 . The form S k is called the quasiclassical contravariant form of the configuration C. It defines a bilinear symmetric form on F l k (C). See [10]. Flag Complex. Define a differential d: F l k → F l k+1 by d(L0 ⊃ · · · ⊃ Lk ) = P 0 k k+1 ), where the sum is taken over all edges Lk+1 of Lk+1 (L ⊃ · · · ⊃ L ⊃ L k codimension k + 1 such that L ⊃ Lk+1 . From the definition of the groups F l k it follows that d2 = 0.
164
G. FELDER ET AL.
P A Complex (A• , d(a)). Set ω = ω(a) = H ∈C a(H )H, ω(a) ∈ A1 . Define a differential d = d(a): Ak → Ak+1 by the rule dx = ω(a) · x. It is clear that d2 = 0. For any k, the quasiclassical bilinear form on C defines a homomorphism S k : F l k → (F l k )∗ ' Ak ,
(42)
where S k (F ) = (−1)k(k−1)/2S(F, ·). LEMMA 5.1. S • defines a map of complexes S • = S • (a): (F l • (C), d) → (A• (C), d(a)). Note. There is a misprint in [10] in the definition of S k where the factor (−1)k(k−1)/2 is missing. P 1 Proof. For any edge L, set S(L) = H ∈C, L⊂H a(H )H , S(L) ∈ A . It is k k 0 easy to see that the homomorphism S is defined by S (L ⊃ · · · ⊃ Lk ) = (−1)k(k−1)/2S(L1 ) · S(L2 ) · · · S(Lk ). In other words X a(H1 ) . . . a(Hk )(H1 , . . . , Hk ), S k (L0 ⊃ · · · ⊃ Lk ) = (−1)k(k−1)/2 where the sum is over all k-tuples (H1 , . . . , Hk ) such that Hi ⊃ Li for all i. Therefore, we have S k+1 d(L0 ⊃ · · · ⊃ Lk ) X = S k+1 (L0 ⊃ · · · ⊃ Lk ⊃ Lk+1 ) Lk+1 ,Lk+1 ⊂Lk
(k+1)k X = (−1) 2 a(H1 ) . . . a(Hk )a(Hk+1 )(H1 , . . . , Hk , Hk+1 ) X k(k−1) X = (−1) 2 a(H1 ) . . . a(Hk )(H1 , . . . , Hk ) · (−1)k a(Hk+1 )Hk+1 k(k−1) X = (−1) 2 a(H1 ) . . . a(Hk )(H1 , . . . , Hk ) · (−1)k ω(a)
= S k (L0 ⊃ · · · ⊃ Lk ) · (−1)k ω(a) = ω(a) · S k (L0 ⊃ · · · ⊃ Lk ) = d(a)S k (L0 ⊃ · · · ⊃ Lk ). The second, third and fourth sum are over all Hi , such that Hi ⊃ Li , for 1 6 i 6 k and Hk+1 ∩ Lk 6= 0. Note that Hk+1 ∩ Lk = 0 implies (H1 , . . . , Hk , Hk+1 ) = 0, and thus the fourth equality is justified. The sixth one comes from the skew symmetry 2 in A• (C). Recall that we have a weighted configuration of hyperplanes in a complex mdimensional space W , and a = {a(H ) | H ∈ C} are the weights. Fix an affine S equation lH = 0 for each hyperplane H ∈ C. Set Y = W − H ∈C H . Consider the trivial line bundle L(a) over Y with an integrable connection d(a): O → 1
DIFFERENTIAL EQUATIONS COMPATIBLE WITH KZ EQUATIONS
165
P given by d + (a) = d + H ∈C a(H )d log lH , where d is the de Rham differential. Denote by • (L(a)) the complex of Y -sections of the homomorphic de Rham complex of L(a). To any H ∈ C assign the one-form i(H ) = d log lH ∈ 1 (L(a)). This construction defines a monomorphism i(a): (A• (C), d(a)) → (• (L(a)), d(a)). The image of this monomorphism is called the complex of the hypergeometric differential forms of weight a. It is denoted by (A• (C, a), d(a)). The image of the homomorphism i(a)S: (F l • (C), d) → (• (L(a)), d(a)) is called the complex of the flag hypergeometric differential forms of weight a. It is denoted by (F l • (C, a), d(a)). For further details see [10]. 5.2.
DISCRIMINANTAL CONFIGURATIONS
Let W be an affine complex space of dimension m. Let z1 , . . . , zn be pairwise distinct complex numbers. Denote by Cm a configuration in W consisting of hyperplanes Hkl : tk − tl = 0; 1 6 k < l 6 m. So C1 = ∅, and Y (Cm ) is the space of m-tuples of ordered distinct points in C. Denote by Cn;m (z) a configuration in j W consisting of hyperplanes Hk : tk − zj = 0, 1 6 k 6 m, 1 6 j 6 n, and Hkl , 1 6 k < l 6 m. Thus, Y (Cn;m (z)) = p −1 (z) where p: Y (Cn+m ) → Y (Cn ) is the projection on the first n coordinates. Define C0;m = Cm . Edges and Flags of Cn,m . For every non-empty subset J = {j1 , . . . , jk } ⊂ [m] k−1 . L is an edge of codimension k −1. In set LJ = Hj1 j2 ∩Hj2 j3 ∩· · · ∩Hjk−1 jk ∈ Cn;m particular set LJ = W , for k = 1. For i ∈ [n] define LiJ = Hji1 ∩ Hji2 ∩ · · · ∩ Hjik ∈ k Cn;m . LiJ is an edge of codimension k. Set Li∅ = W . Given non-intersecting subsets T T J1 , . . . , Jk ; I1 , . . . , In ⊂ [m], define LJ1 ,...,Jk ;I1 ,...,In = ( kj =1 LJj ) ∩ ( ni=1 LiIi ). Multiplication of Flags. Given two subsets J ⊂ [m] and I ⊂ [n], denote by CJ ;I ⊂ Cm;n the subset consisting of all hyperplanes Hj1 j2 with j1 , j2 ∈ J and Hji with j ∈ J, i ∈ I . Given subsets J, J 0 ⊂ [m]; I, I 0 ⊂ [n] such that J ∩ J 0 = ∅; I ∩ I 0 = ∅, define maps ◦: Flagk (CI ;J ) × Flagl (CI 0 ;J 0 ) → Flagk+l (CI ∪I 0 ,J ∪J 0 ) as follows. For F = F (H1 , . . . , Hk ) ∈ Flagk (CI ;J ), F = F (H10 , . . . , Hl0 ) ∈ Flagl (CI 0 ;J 0 ), set F ◦ F 0 = (H1 , . . . , Hk , H10 , . . . , Hl0 ). The following lemma, [10], Lemma 5.7.2, takes place. LEMMA 5.2. The above map correctly defines the map F l k (CI ;J )⊗F l l (CI 0 ;J 0 ) → F l k+l (CI ∪I 0 ,J ∪J 0 ). Moreover, for all x ∈ F l k (CI ;J ), y ∈ F l l (CI 0 ;J 0 ) we have x◦y = (−1)kl y ◦ x.
166 5.3.
G. FELDER ET AL.
TWO MAPS OF COMPLEXES
Let g be a Kac–Moody Lie algebra without Serre’s relations. Let M = M(31 ) ⊗ · · ·⊗M(3n ) be a tensor product of Verma modules with weights 31 , . . . , 3n ∈ h∗ . Set λ = (1, 1, . . . , 1). In this case the number of generators (fj )r1 of n− equals m, | {z } m
i.e. r = m. Two maps of complexes ψ• and η• are described in [10]: ψp : Cp (n− , M)λ → F l m−p , ηp = ϕ −1 ◦ (ψp∗ )−1 : Cp (n− ∗ , M ∗ )λ → Am−p ,
(43)
where ϕ is the map (40). Note. The maps ψp define isomorphism of complexes. Theorem 6.6 [10] implies that the maps (−1)p (−1)(m−p)(m−p−1)/2ηp define isomorphism of complexes. The sign is due to Lemma 5.1 and the fact that the contravariant form S in this paper is minus the contravariant form in [10], see formula (10) and [10], formula (6.2.3). We will recall the explicit description of ψ• under the above assumption on λ. Let g ∈ n− . A length l = l(g) of a commutator g is given via an inductive definition. Set l(fj ) = 1 for j = 1, . . . , m. If g = [g1 , g2 ] and l1 = l(g1 ), l2 = l(g2 ), then set l(g) = l1 + l2 . So l(g) = ‘the number of f ’s in g’. To every commutator g assign a bracket sign b(g) ∈ Z/2Z as follows. Set b(fj ) = 0; b([g1 , g2 ]) = b(g1 ) + b(g2 ) + l(g1 ) mod 2. To every commutator g assign a flag F l(g) ∈ F l l(g)−1(C0;|g|) as follows. Set F l(fj ) = . If g = [g2 , g1 ], set F l(g) equal to F l(g1 ) ◦ F l(g2 ) completed by the edge L|g| . Finally, for a commutator g set F (g) = (−1)b(g) F l(g) ∈ F l l(g)−1(C0;|g| ). For I = (i1 , . . . , il ) ⊂ {1, . . . , m} and 1 6 i 6 n, set fI = fil . . . fi1 ∈ U (n− ) and F i (fI ) = F (Hii1 , . . . , Hiil ) ∈ F l l (C{i};I ). Let z ∈ Cp (n− , U (n− )⊗n )λ and z = gp ∧ gp−1 ∧ · · · g1 ⊗ fIn ⊗ fIn−1 ⊗ · · · ⊗ fI1 , where all gi are commutators, li = l(gi ). Let {fi1 , . . . , fim } be the list of fi ’s in z read from right to left. Define σ (z) ∈ 6m by σ (z)(j ) = ij . Set ψp (z)
Pp
= (−1)|σ (z)|+
i=1 (i−1)(li −1)
F 1 (fI1 ) ◦ · · · ◦ F n (fIn ) ◦ F (g1 ) ◦ · · · ◦ F (gp ).
(44)
Note. There is a correction of the sign in the definition of ψ compared with [10]. EXAMPLES. Let n = 1. ψ(fm . . . f1 ) = F (H11 , . . . , Hm1 ), and η((fσ1 . . . fσm )∗ ) = (−1)|σ | Hσ1 ,σ2 ◦ · · · ◦ Hσm−1 ,σm ◦ Hσ1m . Compose the inclusion map i(a): (A• (C), d(a)) → (• (L(a)), d(a)) (see Section 5.1) with the map η to get i(a) ◦ η((fσ1 . . . fσm )∗ ) = (−1)|σ | d ln(tσ1 − tσ2 ) ∧ · · · ∧ d ln(tσm−1 − tσm ) ∧ d ln(tσm − z1 ).
167
DIFFERENTIAL EQUATIONS COMPATIBLE WITH KZ EQUATIONS
Let I ∈ P (λ, n) and I = (i11 , . . . , is11 , . . . , i1n , . . . , isnn ). Since λ = (1, 1, . . . , 1), j j ∗ I ∈ 6m . Let Ij = (isj , . . . , i1 ) for 1 6 j 6 P n. We have∗ i(a) ◦ η((fIn ) ⊗ ∗ |I | · · · ⊗ (fI1 ) ) = (−1) ωI . Therefore i(a) ◦ η( I ∈P (λ,n)(fI ) fI ) = ω(z, t), see Section 3. Let I = (i11 , . . . , is11 , . . . , i1n , . . . , isnn ) ∈ P (λ, n), and 1 6 k 6 sj . Define fI ;i j = fIn ⊗ · · · ⊗ fIj+1 ⊗ fi j . . . fisj ⊗ fIj−1 ⊗ · · · ⊗ fI1 , k
k+1
j
θI ;i j = ωi 1 ,...,is1 ∧ · · · ∧ ωi j−1 ,...,isj−1 ∧ ωi j 1
k
∧ω
1
1
i1n ,...,isnn
1
j −1
n
◦Hj
ik+1 ,...,i
j k−1 ,ik
1 2
p
j+1
− ti j )],
j sj
(45)
k
j +1
j
j−1
,...,isj−1
◦Hinn ,...,isn ◦ [Hi j ,i j ◦ · · · ◦ Hi j 1
1
k−1
2
i1
s1
1
∧ ωi j+1 ,...,isj+1 ∧ · · ·
∧ [d ln(ti j − ti j ) ∧ · · · ∧ d ln(ti j
HI ;i j = Hi11 ,...,i 1 ◦ · · · ◦ H j−1 k
k+1 ,...,is j j
j−1
◦ H j+1 i1
j+1
,...,isj+1
◦ ···
],
p
where Hi1 ,...,il = Hi1 ,i2 ◦ · · · ◦ Hil−1 ,il ◦ Hil . j
LEMMA 5.3. Let I ∈ P (λ, n). Let k = k((sj − k) + sj +1 + · · · + sn ). Then j
i(a) ◦ η(1σ j ,...,σ j ⊗ (fI ;i j )∗ ) = (−1)|I |+k θI ;i j . i1
ik
k
(46)
k
Proof. The statement of the lemma is equivalent to the equation: j
1σ j ,...,σ j ⊗ (fI ;i j )∗ = (−1)|I |+k η−1 (HI,i j ). i1
ik
k
(47)
k
It is sufficient to compute the two sides on elements of type g ⊗ fI,i j where g is a k commutator of length k on fi j , . . . , fi j . Let σ ∈ 6k and fiσj , . . . , fiσj be the list k 1 1 k of fi j ’s entering g from right to left. The left hand side and the right-hand side of (47) evaluated on g ⊗ fI,i j give k
∗
1σ j ,...,σ j ⊗ (fI ;i j ) (g ⊗ fI ;i j ) = 1σ j ,...,σ j (g), i1
ik j |I |+k
(−1)
k
i1
k
ik
j
η−1 (HI,i j )(g ⊗ fI ;i j ) = (−1)|I |+k ϕ(HI,i j )(ψ(g ⊗ fI ;i j )), (48) k
k
k
k
respectively. See formula (43). Use the definition of ψ to obtain ψ(g ⊗ fI ;i j ) = (−1)|τ | F (Hi11 , . . . , Hi11 ) ◦ · · · ◦ F (Hi 1 , . . . , H j ) ◦ · · · s1
k
1
j
j
sj
ik+1
◦ F (Hinsn , . . . , Hinn ) ◦ F (g), n
1
1 . . . . . . . . . . . j j is11 · · · i11 · · · isj · · · ik+1 · · · isnn · · · i1n iσj1 I as an element of 6m has the form 1 . . . . . . . . . m . I= 1 j j i1 · · · is11 · · · i1 · · · isj · · · i1n · · · isnn where τ =
. ···
!
m . iσjk
168
G. FELDER ET AL.
Thus j
|I | = (|τ | + Sk + k((sj − k) + sj +1 + · · · + sn ) + |σ |) mod 2,
(49)
where j
Sk =
n X sl (sl − 1) (sj − k)(sj − k − 1) + . 2 2 l=1, l6=j p
p
Note that Hi1 ,...,il = (−1)l(l−1)/2Hil ◦ Hil−1 ,il ◦ · · · ◦ Hi1 ,i2 . Use the definition of ϕ, (40), to compute p
p
ϕ(Hi1 ,...,il ) = (−1)l(l−1)/2ϕ(Hil ◦ Hil−1 ,il ◦ · · · ◦ Hi1 ,i2 ) = (−1)l(l−1)/2δF (Hip ,...,Hip ) + other δ-summands. l
j
ϕ(HI,i j ) = (−1)Sk ϕ(Hi11 ◦ Hi 1
1 s1 −1 ,is1
s1
k
j
◦ H j ◦ Hi j
j sj −1 ,isj
isj
◦ · · · ◦ Hi 1 ,i 1 ◦ · · · 1 2
◦ · · · ◦ Hi j
j k+1 ,ik+2
◦ Hi1n ,i2n ◦ [Hi j ,i j ◦ · · · ◦ Hi j
j k−1 ,ik
1 2
(50)
1
◦ · · · ◦ Hinsn ◦ Hisnn −1 ,isnn ◦ · · · n
]).
(51)
We use formulae (49), (50), (51) to simplify (48). j
(−1)|I |+k η−1 (HI,i j )(g ⊗ fI ;i j ) k
k
= (−1)|σ | ϕ(Hi j ,i j ◦ · · · ◦ Hi j 1 2
j k−1 ,ik
)(F (g)).
The proof of Lemma 5.3 is finished modulo the following result.
(52) 2
LEMMA 5.4. Let ηI , ψI be the combinatorial maps (43), (44) defined on the set of distinct indices I, I = {i1 , . . . , ik } ⊂ {1, . . . , m}. Then ηI (1i1 ,...,ik ) = Hi1 ,i2 ◦ · · · ◦ Hik−1 ,ik , for any k = 2, . . . , m. Proof. Induction by k. For k = 2, g = [fi1 , fi2 ] forms a base of the commutators of length 2 on fi1 and fi2 . 1i1 ,i2 (g) = 1. Since b(g) = 1 and F (g) = (−1)b(g)F (Hi1 ,i2 ) and σ = 12 21 , we have η−1 (Hi1 ,i2 )([fi1 , fi2 ]) = ϕ(Hi1 ,i2 )(ψ(g)) = δF (Hi1 ,i2 ) ((−1)|σ |+b(g)F (Hi1 ,i2 )) = 1. Let 2 < k 6 m. Assume that for any j , 2 6 j < k, and 1 6 s1 < · · · < sj 6 k we have η(1is1 ,...,isj ) = His1 ,is2 ◦ · · · ◦ Hisj−1 ,isj . Let g be a commutator of length k on fi1 , . . . , fik . Then g = [g1 , g2 ] with l(g1 ) = l1 , and l(g2 ) = l2 , and l1 + l2 = k. Let σ ∈ 6k be such that fiσ1 , . . . , fiσk is the list of fi ’s in g read from right to left. In order to evaluate η−1 (Hi1 ,i2 ◦ · · · ◦ Hik−1 ,ik )(g) = ϕ(Hi1 ,i2 ◦ · · · ◦ Hik−1 ,ik )(ψ(g)) remark that ψ(g) = (−1)|σ |+b(g) (F l(g2 ) ◦ F l(g1 ), L|g| )
169
DIFFERENTIAL EQUATIONS COMPATIBLE WITH KZ EQUATIONS
= (−1)|σ |+l(g1 ) (F (g2 ) ◦ F (g1 ), Li1 ,...,ik ), η−1 (Hi1 ,i2 ◦ · · · ◦ Hik−1 ,ik )(ψ(g)) X = (−1)|τ | δF (Hiτ ,iτ +1 ,...,Hiτ ,iτ +1 ) (ψ(g)). 1
τ ∈6k−1
k−1
1
(53)
k−1
Since a link corresponding to a hyperplane Hj,j +1 connects only neighbouring indices in a flag of a type F (Hiτ1 ,iτ1 +1 , . . . , Hiτk−1 ,iτk−1+1 ) we have F (Hiτ1 ,iτ1 +1 , . . . , Hiτk−1 ,iτk−1 +1 ) = (, . . . , (ti1 = · · · = tiτk−1 ; tiτk−1 +1 = · · · = tik ), Li1 ,...,ik = (ti1 = · · · = tik )).
(54)
|σ |+b(g)
Let δF (Hiτ ,iτ +1 ,...,Hiτ ,iτ +1 ) (ψ(g)) 6= 0. Since (−1) ψ(g) = (F l(g2 ) ◦ 1 1 k− k−1 F l(g1 ), L|g| ) = (, . . . , L|g2 | ∩ L|g1 | , L|g| ) formula (53) implies either L|g1 | = (ti1 = · · · = tiτk−1 ); L|g2 | = (tiτk−1 +1 = · · · = tik ), or L|g2 | = (ti1 = · · · = tiτk−1 ); L|g1 | = (tiτk−1 +1 = · · · = tik ). Without loss of generality we will assume that the second case takes place, i.e. L|g2 | = (ti1 = ti2 = · · · = tiτk−1 ); L|g2 | = (tiτk−1 +1 = · · · = tik−1 = tik ). Compare the lengths of the flags to conclude that τk−1 = l2 . In order to have non-zero multiples in the product δF (Hiτ
1 ,iτ1 +1
,...,Hiτ
k−1 ,iτk−1 +1
= δF (Hiτ
1 ,iτ1 +1
,...,Hiτ
◦ F l(g1 ), L|g| )
) (F l(g2 )
l2 −1 ,iτl2 −1 +1
) (F l(g2 ))δF (Hiτ
l2 ,iτl2 +1
,...,Hiτ
k−2 ,iτk−2 +1
) (F l(g1 ))
(55)
we need (τ1 , . . . , τl2 −1 ) to be a permutation of the set (1, . . . , l2 − 1) and (τl2 , . . ., τk−2 ) to be a permutation of the set (l2 + 1, . . . , k − 1). Set τ 0 = τ1 ·· ·· ·· lτ2 − 1
and τ 00 = τ 1− l ·· ·· ·· τ l1 −−1l . Then (−1)|τ | = (−1)|τ |+|τ l2 2 k−2 2 b(g1 ) + b(g2 ) mod 2, and 0
00 |+l
η−1 (Hi1 ,i2 ◦ · · · ◦ Hik−1 ,ik )(ψ(g)) X 0 (−1)|τ | δF (Hi = (−1)|σ |+b(g)+l1 −1 ×
X τ 00 ∈6l1 −1
1 −1
,...,Hi 0 ,i 0 ) ,i τ10 τ10 +1 τl −1 τl −1 +1 2 2
τ 0 ∈6l2 −1 00
(−1)|τ | δF (Hi
,...,Hi 00 ) ,i ,i τ100 +l2 τ100 +l2 +1 τl −1 +l2 τl00 −1 +l2 +1 1 1
1
l2 −1
, b(g) + l1 =
(F l(g2 ))×
(F l(g1 ))
= (−1)|σ |−1 η−1 (Hi1 ,i2 ◦ · · · ◦ Hil2 −1 ,il2 )(F (g2 ))× ×η−1 (Hil2 +1 ,il2 +2 ◦ · · · ◦ Hik−1 ,ik )(F (g1 )) = (−1)η−1 (Hi1 ,i2 ◦ · · · ◦ Hil2 −1 ,il2 )(ψ(g2 ))× ×η−1 (Hil2 +1 ,il2 +2 ◦ · · · ◦ Hik−1 ,ik )(ψ(g1 )). 1 σ1
··· k · · · σk
The last equality holds because σ = Using the inductive hypothesis rewrite (56) as
=
(56) 1 σ1
· · · l2 l2 + 1 · · · lk · · · σl2 × σl2 +1 · · · σlk .
η−1 (Hi1 ,i2 ◦ · · · ◦ Hik−1 ,ik )(ψ(g)) = (−1)1i1 ,...,il2 (g2 )1il2 +1 ,...,ik (g1 ) 2 = 1i1 ,...,ik ([g1 , g2 ]).
(57)
170
G. FELDER ET AL.
6. Derivation of the Dynamical Differential Equation In this section g will be a Kac–Mody Lie algebra without Serre’s relations, λ = (1, 1, . . . , 1), r = m. We will work in a weight space Mλ of the module M = M(31 ) ⊗ · · · ⊗ M(3n ). We will differentiate the hypergeometric form ω(z, t), express the result in terms of the complex C• (n− ∗ , M ∗ ), and derive the Dynamical differential equation in the form (25). The integrand of a hypergeometric solution have the following form, see Section 3. !! m n X X 1 − 81/κ , hαc(i) , µiti + h3j , µizj 81/κ µ ω = exp κ i=1 j =1 where 8(z, t) =
Y
(zi − zj )(3i ,3j )
Y Y (tk − zj )−(αc(k) ,3j ) (tk − tl )(αc(k) ,αc(l) ) .
i<j
k,j
k
Fix µ0 ∈ h and let ∂µ0 be the partial derivative with respect to the parameter µ in the direction of µ0 . Then κ∂µ0 (81/κ µ ω) =
! m n X X X hαc(i) , µ0 iti + h3j , µ0 izj 81/κ (−1)|I | ωI fI v. (58) − µ j =1
i=1
I ∈P (λ,n)
Let I = (i11 , . . . , is11 ; . . . ; i1n , . . . , isnn ) ∈ P (λ, n). Moreover I ∈ 6m because of the form of λ. Since r = m we have c(i) = i. Set ti j = zj and αik;j = αi j + αi j + · · · + αi j . Rearrange the following expression:
sj +1
k
k−1
1
! m n X X 0 0 − hαi , µ iti + h3j , µ izj ωI j =1
i=1
=
−
! n X 0 hαi j , µ iti j + h3j , µ izj ωI
sj n X X
k
j =1 k=1
=
−
0
k
j =1
sj n X X
hαi j , µ0 i(ti j − ti j
j =1 k=1
k
k
k+1
+ ti j
k+1
− ti j
!
n X h3j , µ0 izj ωI + j =1
=
! sj n X X − hαik;j , µ0 i(ti j − ti j ) ωI + j =1 k=1
k
k+1
k+2
+ · · · + tisj − zj + zj )+ j
171
DIFFERENTIAL EQUATIONS COMPATIBLE WITH KZ EQUATIONS
!
n X
+
0
zj h3j − αi j − · · · − αisj , µ i ωI . 1
j =1
LEMMA 6.1. Let u(µ, z) = κ∂µ0 u −
n X
=
R
1/κ γ (z) 8µ ω.
Then
zj µ0 u (j )
j =1
X
(59)
j
(−1)|I |
! sj n X X − hαik;j , µ0 i(ti j − ti j ) ωI fI v. (60)
Z 81/κ µ γ (z)
I ∈P (λ,n)
k
j =1 k=1
k+1
Proof. Combine formulae (58), (59) with the fact µ0 (j ) fI v = h3j − αi j − · · · − 1 2 αisj , µ0 ifI v to obtain the result. j
LEMMA 6.2. Define an operator L by Lu = κ∂µ0 u − X hα, µ0 i Lu = 1+,α u. hα, µi α>0
Pn
0 (j ) u. j =1 zj µ
Then
Proof. Let I ∈ P (λ, n) and 1 6 k 6 sj . (ti j − ti j )ωI k
k+1
= ωi 1 ,...,is1 ∧ · · · ∧ (ti j − ti j )ωi j ,...,isj ∧ · · · ∧ ωi1n ,...,isnn 1
k
1
k