Abstract
We study k-point correlators of characteristic polynomials in non-Hermitian ensembles of random matrices, focusing on the Ginibre and truncated unitary random matrices. Our approach is based on the technique of character expansions, which expresses the correlator as a sum over partitions involving Schur functions. We show how to sum the expansions in terms of representations which interchange the role of k with the matrix size N. We also provide a probabilistic interpretation of the character expansion analogous to the Schur measure, linking the correlators to the distribution of the top row in certain Young diagrams. In more specific examples, we evaluate these expressions in terms of \(k \times k\) determinants or Pfaffians.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction and Main Results
Characteristic polynomials of random matrices play an important role in both mathematical and physical applications of random matrix theory. For example, they have been successfully used to model statistical properties of the Riemann zeta-function or other L-functions [42, 43]. This motivated several works on the problem of computing multi-point correlators of characteristic polynomials [12, 13, 17], defined as averaging a product of characteristic polynomials with respect to the distribution of the underlying random matrix. Bump and Gamburd [14] performed such calculations using a character expansion technique in the setting of classical compact groups, building on earlier work of Diaconis and Shashahani [19] who used it to study power traces in the same setting. The approach has been successfully applied to computing other integrals over the unitary group [2, 8, 55]. The evaluation of group integrals and the character expansion technique in general has a wide applicability in several topics including quantum field theory, quantum chromodynamics, and string theory, see [8, 60] and references therein.
The purpose of this paper is to develop a character expansion technique in the context of multi-point correlators of non-Hermitian ensembles. We work with the three classical Ginibre ensembles, though we also apply the method to some non-Gaussian models such as truncations of random unitary matrices. The Ginibre ensembles are defined in terms of \(N \times N\) matrices G whose entries \(G_{ij}\) are i.i.d. standard Gaussian random variables over the real, complex, or quaternion number systems. In the complex case, this is equivalent to a probability measure on the set \({\mathbb {C}}^{N \times N}\) of all complex \(N \times N\) matrices of the form
where dG is the Lebesgue measure on \({\mathbb {C}}^{N \times N}\). Such ensembles were introduced by Ginibre [31] as non-Hermitian counterparts of the ensembles such as the Gaussian Unitary Ensemble (GUE) that were previously introduced by Wigner in 1955. Hence, (1.1) is sometimes also known as the Ginibre Unitary Ensemble (GinUE).
The characteristic polynomial of the complex Ginibre ensemble appeared in the work of Rider and Virag [57] who showed how it is closely related to the Gaussian Free Field—see the survey [7] on the recent activity surrounding such connections. In the case of the real Ginibre ensemble, related averages have appeared in the analysis of spin glasses [47] and neural networks [67]. See [15, 16] for two recent review articles regarding progress on the Ginibre ensembles.
1.1 Complex Ensembles
Before stating our results, we fix some notation that will be used throughout the paper. Let \(\varvec{z} = (z_1,\ldots ,z_k)\) and \(\varvec{w} = (w_1,\ldots ,w_k)\) be vectors in \({\mathbb {C}}^{k}\) and define the following correlator of characteristic polynomials
where the expectation is taken with respect to (1.1). In (1.2), \(\Omega \) and \(\Sigma \) are \(N \times N\) deterministic complex matrices, and throughout the paper \(I_{N}\) is the \(N \times N\) identity matrix. The matrix \(\Omega G\) can be viewed as a multiplicative perturbation which deforms the original matrix G, see, e.g., [26, 28, 30]. If \(\Omega = \Sigma = I_{N}\), we refer to (1.2) as \(R^{\text {GinUE}}_{N}(\varvec{z},\varvec{w})\). Throughout the paper, we reserve the notation
for the Vandermonde determinant in the variables \(z_{1},\ldots ,z_{k}\), whose dimension may vary depending on the context. Finally, for any two matrices A and B, \(A \otimes B\) is their Kronecker product.
Theorem 1.1
Let G be a complex Ginibre matrix of size \(N \times N\) and consider the multi-point correlator (1.2), setting \(Z = \text {diag}(\varvec{z})\) and \(W = \text {diag}(\varvec{w})\). Then,
where \({\mathbb {E}}_{X}\) denotes the expectation with respect to a complex Ginibre matrix X of size \(k \times k\). If \(\Omega = \Sigma = I_N\), we have
Proof
See Sect. 2.2. \(\square \)
Formula (1.4) was obtained by Nishigaki and Kamenev [53] by the use of Grassmann variables, in special cases where the vectors \(\varvec{z}\) and \(\varvec{w}\) contain several coinciding points. The determinantal expression (1.5) was obtained by Akemann and Vernizzi [6] by using the formula for the joint probability density function of eigenvalues of G, see also [9]. Formula (1.3) generalizes these two results, and our approach does not rely on the joint distribution of eigenvalues of G. These expressions replace the initial average over \(N \times N\) random matrices with an average over a smaller ensemble of \(k \times k\) matrices, or a \(k \times k\) determinant, a phenomenon sometimes referred to as duality in the literature. It is particularly convenient for asymptotic analysis as \(N \rightarrow \infty \) with fixed k [18].
For the Gaussian measure considered in Theorem 1.1, an approach based on the diffusion equation has been developed, see [33] and recent work of Liu and Zhang [48] who obtained formulae (1.3) and (1.4) with the method. However, this approach does not seem to be obviously applicable to the non-Gaussian measures we will discuss below, such as truncations of random unitary matrices.
Our approach is quite different to [6, 48, 53] mentioned above. It is based on expanding (1.2) as a sum over partitions involving Schur functions. The utility of a Schur function approach is somewhat less appreciated in the non-Hermitian context, though it has been successfully applied to computing moments of characteristic polynomials in [26, 29, 61] and was applied to circular ensembles in [21]. It was applied to Christoffel–Darboux type random matrix ensembles in [59]. Building on these works, we show how to evaluate the general quantity (1.2) and resum the resulting expansions as determinants or as dual matrix integrals, revealing directly the connection between (1.4) and (1.5).
We also provide a probabilistic interpretation behind these expansions. Consider the set \({\mathcal {P}}_{k}\) of all partitions \(\eta = (\eta _1,\ldots ,\eta _k)\) of length \(l(\eta ) \le k\), where \(\eta _{1} \ge \eta _{2} \ge \ldots \ge \eta _{k}\) are weakly decreasing non-negative integers. To each \(\eta \in {\mathcal {P}}_{k}\), we introduce the probability
where \(s_{\eta }(\varvec{z})\) is the Schur function, and \({\mathcal {Z}}_{k}\) is a normalization constant. See Sect. 2 for further background on partitions and Schur functions, in particular Sect. 2.4 where a determinant formula is given for \({\mathcal {Z}}_{k}\).
Theorem 1.2
When \(\Omega = \Sigma = I_{N}\), we have the following equivalent representation for the quantity (1.2)
If the entries of \(\varvec{z}\) and \(\varvec{w}\) are positive, we have
i.e., the distribution of the largest part with respect to (1.6).
Proof
See Sect. 2.4. \(\square \)
The distribution (1.6) is reminiscent of the Schur measure which plays an important role in problems related to KPZ universality, see [37, 54]. Character sums of this type also appeared in [25] related to an inhomogeneous model of last passage percolation with geometric weights.
Our method is well adapted to obtaining analogous results for another non-Hermitian ensemble known as truncations. Let \(\text {O}(N)\), \(\text {U}(N)\) and \(\text {Sp}(2N)\) denote the three classical compact groups of N \(\times \) N orthogonal, unitary, or symplectic matrices. These matrix groups each come with a unique translation invariant measure known as Haar measure, see [50] for details. Choosing an element U from one of \(\text {O}(N)\), \(\text {U}(N)\) or \(\text {Sp}(2N)\) with respect to the Haar measure, we consider the sub-block decomposition
where the principal sub-matrix T is of size \(M \times M\) with \(M < N\). Then, the random matrix T is said to belong to the truncated orthogonal, unitary, or symplectic ensemble (denoted TOE, TUE, or TSE, respectively). These ensembles were introduced and studied in the works [44, 45, 69]. When M is fixed and \(N \rightarrow \infty \), after suitable rescaling they recover the Ginibre ensembles discussed previously. On the other hand, if \(N-M\) stays of finite order, their eigenvalue statistics resemble those from the corresponding classical compact group. In particular, we point out that our results below continue to apply for averages over the whole group, i.e., when \(N=M\). In particular, if \(\Sigma = \Omega = I_{N}\), as a special case our results reduce to averages of characteristic polynomials over the CUE (Circular Unitary Ensemble).
For definiteness, consider the complex case. Let T denote a truncated unitary random matrix and define
where \(\Omega \) and \(\Sigma \) are again deterministic source matrices. Here the expectation is taken with respect to truncated matrices T introduced above. We mention that multiplicative perturbations of the form \(\Omega T\) appeared recently in the context of extreme eigenvalue statistics of rank-one perturbations [30].
In order to state our results, we need to define the appropriate analogue of the dual averaging \({\mathbb {E}}_{X}\) in Theorem 1.1. We define the following probability measure on \({\mathbb {C}}^{k \times k}\)
where \(S^{(2)}_{k}\) is a normalization constant and dX is the Lebesgue measure on \({\mathbb {C}}^{k \times k}\). This dual measure also played an important role in the work [29].
Theorem 1.3
Let T belong to the truncated unitary ensemble (1.7) and consider multi-point correlator (1.8), setting \(Z = {{\,\textrm{diag}\,}}(\varvec{z})\) and \(W = {{\,\textrm{diag}\,}}(\varvec{w})\). Then we have
where \({\mathbb {E}}_{X}\) denotes expectation with respect to (1.9). If \(\Omega = \Sigma = I_M\), we have
where
and
Proof
See Sect. 2.3. \(\square \)
1.2 Real Ensembles
An emphasis of the present work is the question of whether the above results have analogues for the real Ginibre ensembles. In Sect. 3.5, we also present results for quaternionic ensembles, but to keep the introduction concise we restrict ourselves here to real matrices. The real Ginibre ensemble is defined by the following probability measure on the set \({\mathbb {R}}^{N \times N}\) of real \(N \times N\) matrices of the form
where dG is the Lebesgue measure on \({\mathbb {R}}^{N \times N}\). This ensemble is sometimes also known as the Ginibre Orthogonal Ensemble (GinOE). For real ensembles, we assume throughout for simplicity that N is even. The corresponding multi-point correlator is
where \(\varvec{z} = (z_1,\ldots ,z_{2k}) \in {\mathbb {C}}^{2k}\) and \(\Omega \) is again an \(N \times N\) deterministic source matrix. As before, we abbreviate \(R^{\text {GinOE}}_{N}(\varvec{z}) := R^{\text {GinOE}}_{N}(\varvec{z};I_{N})\).
We denote the space of \(2k \times 2k\) complex anti-symmetric matrices by \({\mathcal {A}}_{2k}({\mathbb {C}})\). The Pfaffian of an anti-symmetric matrix \(A \in {\mathcal {A}}_{2k}({\mathbb {C}})\) is a polynomial in the entries of A defined by the formula
where \(S_{2k}\) is the group of permutations on 2k symbols. The Pfaffian satisfies many properties analogous to the determinant, to which it is related by the formula \(\det (A) = \text {Pf}(A)^{2}\). Note that we only define the Pfaffian for anti-symmetric matrices, so we will frequently use a notation that only indexes the upper triangular part of A, i.e., \(\text {Pf}(A) = \text {Pf}(A_{i,j})_{1 \le i < j \le 2k}\). We introduce the following dual probability measure on \({\mathcal {A}}_{2k}({\mathbb {C}})\),
where dX denotes the product of Lebesgue measures over the upper triangular entries of X.
Theorem 1.4
Let G be a real Ginibre matrix of size \(N \times N\) and consider the multi-point correlator (1.16), setting \(Z = \text {diag}(\varvec{z})\). Then
where \({\mathbb {E}}_{X}\) denotes expectation with respect to \(2k \times 2k\) anti-symmetric matrices distributed according to (1.17). If \(\Omega = I_N\), we have
Proof
See Sects. 3.1 and 3.4. \(\square \)
Expression (1.19) recovers a result due to Tribe and Zaboronski [65, 66] derived using Grassmann calculus, including an odd number of products in (1.16). See again [48] where the diffusion method remains applicable to the real Gaussian measure. However, the exact Pfaffian formula (1.20) does not appear in [48, 65, 66], except in the limiting case \(N \rightarrow \infty \). Derivation of (1.20) is possible following along the same lines as the complex case in [6], by using the formula for the joint probability density function of eigenvalues, see [4, 5, 46, 47]. Analogously to the complex case, our approach does not make use of the joint probability density function of eigenvalues of G and again reveals the connection between the dual integral (1.19) and its Pfaffian evaluation (1.20). Asymptotic expansion as \(N \rightarrow \infty \) for the real Ginibre averages considered here has recently attracted attention, for example, in the works [1, 47, 65,66,67]. In Sect. 4, we discuss the known asymptotic results for the real Ginibre ensemble and show how Theorem 1.4 allows one to obtain new asymptotic expansions.
To conclude our statement of results, we consider analogous results for truncations. Let T be an \(M \times M\) truncation of a Haar distributed orthogonal matrix from \(\text {O}(N)\), and define
where \(\varvec{z} = (z_1,\ldots ,z_{2k}) \in {\mathbb {C}}^{2k}\). Now introduce the following dual probability measure on \({\mathcal {A}}_{2k}({\mathbb {C}})\),
where \(S^{(1)}_{k}\) is a normalization constant and dX as in (1.17).
Theorem 1.5
Let T belong to the truncated orthogonal ensemble and consider the multi-point correlator (1.21), setting \(Z = \textrm{diag}(\varvec{z})\). Then
where \({\mathbb {E}}_{X}\) denotes expectation with respect to \(2k \times 2k\) complex anti-symmetric matrices X distributed according to (1.22). If \(\Omega = I_M\), we have
where
and
Proof
See Sects. 3.2 and 3.4. \(\square \)
We remark that the results given for the Ginibre ensembles can be viewed as limiting cases of the results for truncations, which are more general in the following sense. Rescaling \(X \rightarrow X/\sqrt{N}\) and rescaling the characteristic variables \(Z \rightarrow Z/\sqrt{N}\), \(W \rightarrow W/\sqrt{N}\), taking the limit \(N \rightarrow \infty \) with fixed M recovers Theorems 1.1 and 1.4 as limiting cases of Theorems 1.3 and 1.5, respectively.
The structure of this paper is the following. We begin in Sect. 2 with the complex ensembles, giving the proofs of Theorems 1.1, 1.2, and 1.3. This allows us to recall several key notions of character expansion techniques. In Sects. 3.1 and 3.2, we give the proofs for the Pfaffian expressions of the real correlators, arriving at equations (1.20) and (1.24). Then in Sect. 3.4 we complete the full proof of Theorems 1.4 and 1.5. In Sect. 3.5, we present analogous results for quaternionic ensembles. Finally, on the basis of these exact results, in Sect. 4 we study the asymptotics as \(N \rightarrow \infty \) for the real Ginibre ensemble.
2 Character Expansion Technique for Complex Ensembles
The goal of this section is to evaluate multi-point correlators in the complex Ginibre ensemble and truncated unitary ensemble. We begin by recalling some standard facts about partitions. The book by Macdonald [49] provides relevant background.
2.1 Partitions and Schur Function Expansions
A partition \(\lambda \) is a weakly decreasing sequence \(\lambda = ( \lambda _1, \lambda _2,\ldots , \lambda _N,0,0,\ldots )\) with finitely many nonzero positive integer terms \(\lambda _1 \ge \lambda _{2} \ge \ldots \), where each term in the sequence is referred to as a part. We only work with the nonzero parts, so one may write \(\lambda = ( \lambda _1, \lambda _2,\ldots , \lambda _N)\). The number of nonzero terms in such a sequence is called the length of the partition and is denoted as \(l(\lambda ) = N\) if \(\lambda _N > 0\). We define the weight associated with the partition \(\lambda \) as the sum of its parts, i.e., \(|\lambda | = \sum _{i=1}^{N} \lambda _i\).
Partitions are represented by their Young diagram. We adopt the convention of drawing boxes top-down and placing \(\lambda _j\) boxes left to right for each \(j=1,\ldots , N\). The conjugate of \(\lambda \), denoted \(\lambda '\), is the partition obtained by transposing the Young diagram: row j of \(\lambda '\) is built from column j of \(\lambda \). For example, let \(\lambda = ( 4, 2,1 )\). Then the corresponding Young diagram is shown in Fig. 1 with its conjugate in Fig. 2.
It will be useful to introduce the following coefficient
where a(i, j) is the number of boxes to the right of (i, j), known as arm length, while l(i, j) is the number of boxes below (i, j), known as leg length. The term \(a(i,j)+l(i,j)+1\) in (2.1) is the hook length associated with box (i, j), and hence, \(d'_{\lambda }\) is the product of all hook lengths in the diagram. Note that conjugating a partition has the effect of interchanging the arm and leg lengths which leaves (2.1) unchanged, implying that
A function f of N variables is said to be symmetric if for each \(\sigma \in S_N\), the group of permutations on N symbols, we have \(f(x_1,\ldots ,x_N) = f(x_{\sigma (1)}, \ldots ,x_{\sigma (N)})\). In general, if X is a matrix with eigenvalues \(x_{1},\ldots ,x_{N}\), we define \(f(X) = f(x_1,\ldots ,x_N)\). An important basis in the space of symmetric polynomials is given by the Schur functions
If \(l(\lambda ) \le N\), the definition above continues to hold provided the partition is padded out with an appropriate number of 0s. The Schur functions are characters of irreducible representations of the unitary group, see, e.g., [8, 50]. In that context, (2.3) is a particular case of the Weyl character formula and satisfies various orthogonality relations with respect to integration over the unitary group. In particular, the coefficient \(s_{\lambda }(I_N)\) appears in these formulas and is closely related to the dimension of an irreducible representation with signature \(\lambda \). This quantity is related to (2.1) via
where the hypergeometric coefficient is defined as
Furthermore, Schur functions are a convenient basis in which to expand the correlators (1.16) and (1.2), due to the dual Cauchy identity
Note that both sides of the above are polynomial expressions in the variables \(\{y_{j}\}_{j=1}^{N}\) and \(\{x_{j}\}_{j=1}^{k}\), as expected. This is because the summation in (2.5) only extends over partitions satisfying \(l(\eta ) \le k\) and \(l(\eta ') \le N\), i.e., only the partitions \(\eta \) whose Young diagrams belong to the rectangle \(R = \{1,\ldots ,k\} \times \{1,\ldots ,N\}\) are relevant. Given a partition \(\eta \) in R, we define the complement partition \(\tilde{\eta }\) to be the complement of \(\eta \) in R, with its parts reversed so that it is weakly decreasing, see Fig. 3. Then \(\tilde{\eta }\) has parts
The complement partition arises due to the following property of Schur functions:
In order to sum the expansions that occur in practice, we use the following.
Lemma 2.1
(Cauchy–Binet identity) Let A and B be matrices of size \(M \times N\) and \(N \times M\), respectively. Then
where \(l_i \in \{ 1, \ldots , N \}\) and \(\textrm{cols}_{l_1,\ldots ,l_M} (A)\) is the matrix of columns of A labeled by \(l_1,\ldots ,l_M\).
2.2 Complex Ginibre Ensemble
The goal of this subsection is to prove Theorem 1.1. The starting point will be to expand the determinants in (1.2) using the dual Cauchy identity (2.5). The subsequent averaging will then be done with the help of the orthogonality relation
This is proved in [26] by applying the left and right invariance of G under unitary transformations, allowing one to exploit the better known orthogonality of Schur functions over U(N). One of our goals will be to write (1.2) as a \(k \times k\) matrix integral. For this purpose, let X be a standard complex Ginibre random matrix of size \(k \times k\). The following relation will be used
where A and B are deterministic \(k \times k\) complex matrices. Identities of this type were discussed in [34], see Lemma B.6 for a proof.
Proof of Theorem 1.1
By the dual Cauchy identity (2.5) we expand the multi-point average in (1.2) as
where we applied the orthogonality relation (2.8). Next we make use of \(d'_{\eta '} = d'_{\eta }\) and apply (2.9), so that the sum in expansion (2.10) becomes
Now the dual integral (1.3) follows directly from evaluating the sum in (2.11) again using the dual Cauchy identity (2.5). To obtain representation (1.4), let \(\Omega = \Sigma = I_N\). Then (1.3) can be simplified using the formula for the determinant of a block matrix. This gives
which completes the proof of (1.4).
Finally, we prove the determinantal formula (1.5). Starting from expression (2.10) with \(\Omega =\Sigma =I_{N}\), from (2.4), we have \(d'_{\eta '}s_{\eta '}(I_{N}) = [N]_{\eta '}\). Then we use Lemma A.1, equation (A.3) with \(u=N\) and \(\alpha =1\) to write
where \(\tilde{\eta }\) is the complement partition of \(\eta \), recall definition (2.6). Then (2.10) takes the form
where we used (2.7). In (2.14), we can replace the complement partition \(\tilde{\eta }\) with \(\eta \) as this amounts to rearranging the terms of the sum. To compute it, introduce indices
Then the weakly decreasing property of \(\eta \) holds if and only if \(l_{j+1} < l_{j}\) for all \(j=1,\ldots ,k-1\), and furthermore, \(\eta _{1} \le N\) if and only if \(l_{1} \le N+k-1\). Using definition (2.3) of Schur functions, the sum in (2.14) becomes
which is a consequence of the Cauchy–Binet identity given in Lemma 2.1. Inserting this into (2.14) establishes the determinantal form of the correlator in (1.5) and completes the proof. \(\square \)
Corollary 2.2
Let A, B be invertible \(k \times k\) matrices with eigenvalues \(\{a_j\}_{j=1}^k\) and \(\{b_j\}_{j=1}^k\), respectively. Then
Proof
Identifying \(Z=A\), \(W=B\), we reverse the steps starting from the left-hand side of (2.12), so that the left-hand side of (2.16) is written in the form of the character expansion in (2.10). Then apply the steps (2.13)–(2.15) to sum the expansion and obtain the right hand of (2.16). \(\square \)
2.3 Truncated Unitary Ensemble
Let U be an \(N \times N\) random unitary matrix sampled uniformly with respect to Haar measure from U(N). We let T denote the upper-left \(M \times M\) sub-block of U. Our goal here is the computation of the associated multi-point correlator
where the expectation is over the \(M \times M\) matrix T and where \(\Omega \) and \(\Sigma \) are arbitrary deterministic complex matrices.
It turns out that the two important ingredients needed in the Ginibre setting generalize to truncations. As shown in [61], we have the orthogonality relation
In order to describe the appropriate analogue of (2.9) in this context, we introduce a \(k \times k\) matrix X with complex entries sampled with respect to the joint probability density function proportional to
Then the following identity holds
where \({\mathbb {E}}_{X}\) is expectation with respect to (2.19), for the proof see Lemma B.5. With (2.18) and (2.20) in hand, we can proceed to proving our main result for truncated unitary matrices.
Proof of Theorem 1.3
The proof follows a similar pattern to the proof of Theorem 1.1 so we just emphasize the main points. Proceeding similarly but now using orthogonality (2.18), we expand (2.17) as
where we made use of (2.2) and (3.4). Now applying integral representation (2.20) in (2.21) results in a sum involving only two Schur functions, which is amenable to the dual Cauchy identity (2.5). This leads directly to expression (1.10). Setting \(\Omega = \Sigma = I_{M}\) in (1.10) and using the formula for the determinant of a block matrix immediately yields (1.11). The computation of the normalization constant in (1.11) is given in Lemma B.11.
To obtain the determinantal representation (1.12), we use \(d'_{\eta '}s_{\eta '}(I_M) = [M]_{\eta '}\), so that the right-hand side of (2.21) becomes
Using (A.3), we write expression (2.22) as
Proceeding as in the proof of Theorem 1.1, we can express the sum in (2.23) as
where
The closed form evaluation of sum (2.24) again follows from the Cauchy–Binet identity (Lemma 2.1) and leads directly to the determinantal form (1.12). \(\square \)
We have an analogous result to Corollary 2.2. The proof follows the same pattern so we omit the details.
Corollary 2.3
Let A, B be invertible \(k \times k\) matrices with eigenvalues \(\{a_j\}_{j=1}^k\) and \(\{b_j\}_{j=1}^k\), respectively. Then
where \(S^{(2)}_{k}\) is given by (1.14) and \(D^{(2)}_{k}\) is given by (1.13).
2.4 Probabilistic Interpretation
In this subsection, we prove Theorem 1.2 and give an analogous result for truncated unitary matrices.
Proof of Theorem 1.2
In the proof of Theorem 1.1, expression (2.14), it is shown that for \(\Omega = \Sigma = I_{N}\), we have
Consider the set \({\mathcal {P}}_{k}\) of all partitions \(\eta \) of length \(l(\eta ) \le k\) and recall the probability distribution introduced in (1.6) which assigns to each \(\eta \in {\mathcal {P}}_{k}\) the probability
where \({\mathcal {Z}}_{k}\) is a normalization factor. We assume that the entries of \(\varvec{z}\) and \(\varvec{w}\) are positive. With this assumption, by Schur positivity, the factors \(s_{\eta }(\varvec{z})\), \(s_{\eta }(\varvec{w})\) are positive and (2.25) is a well-defined probability distribution. The normalization in (2.25) is obtained by summing over all partitions of length less than or equal to k. Following the steps below (2.14) without the restriction on \(\eta _{1}\) leads to
where the last equality is the Harish Chandra Itzykson–Zuber formula [8] and dU denotes the normalized Haar measure on U(k). By definition of the probability distribution (2.25), we arrive at the following probabilistic interpretation for the multi-point correlator in (1.2):
which completes the proof. \(\square \)
Remark 2.4
Like the Schur measure [37, 54], it is reasonable to expect that (2.25) gives rise to a determinantal point process, and that the distribution of the top row in (2.27) can be expressed as a Fredholm determinant.
In the case of truncations, an analogous description is possible, replacing the factorials in (2.25) with the ratio appearing in (2.23). We define the probability distribution
where the normalization constant is
where for this interpretation we assume the parameters \(z_{i}, w_{i}\) belong to (0, 1) for all \(i=1,\ldots ,k\). The final equivalence with a group integral in (2.29), generalizing identity (2.26), is due to Orlov [55, Sec 3]. The proof of the corollary is identical to the proof of Theorem 1.2, and we omit the details.
Corollary 2.5
When \(\Omega = \Sigma = I_{N}\), we have the following probabilistic interpretation for multi-point correlator in (1.8):
where the probability on the right-hand side is defined with respect to distribution (2.28), \({\mathcal {Z}}_{k}\) is given by (2.29), and \(D^{(2)}_{k}\) is given by (1.13).
3 Real and Quaternionic Ensembles: Pfaffians and Duality
The goal of this section is to establish exact results for multi-point correlators of real and quaternionic ensembles, focusing mainly on the real case. We begin without the multiplicative perturbations in (1.16). The reason is mainly pedagogical; for \(\Omega = I_{N}\) we can derive Pfaffian closed form expressions for (1.16) with only a few modifications of the approach of Sect. 2, while the case of general \(\Omega \) will require the use of zonal spherical functions which are discussed separately in Sect. 3.4.
Let \(\eta \) be a partition of length k. We say that \(\lambda = 2\eta = (2\eta _1, 2\eta _2,\ldots ,2\eta _{k})\) is a doubled partition derived from the partition \(\eta \) if it is constructed by doubling each part of the original partition. In what follows, we will say that a partition \(\lambda \) is an even partition if each of its parts is even, and consequently, there is a partition \(\eta \) such that \(\lambda = 2\eta \). If we repeat each part of \(\lambda \) twice, we then say it is a repeated partition, denoted \(\lambda ^2 = (\lambda _1, \lambda _1, \lambda _2, \lambda _2,\ldots )\). One may observe that \((2\lambda )' = (\lambda ')^2\). See Figs. 4 and 5.
3.1 Real Ginibre Ensemble: \(\Omega =I_{N}\)
We start by computing the multi-point correlator
After the same expansion of the determinants using the dual Cauchy identity, the relevant Schur function average can be computed with a result of Sommers and Khoruzhenko [62] and Forrester and Rains [26]:
where the generalized hypergeometric coefficient is
Note the following property of generalized hypergeometric coefficients
see Lemma A.1. These results allow one to proceed as in the complex case, starting with an expansion of (3.1) in terms of Schur functions via the dual Cauchy identity (2.5) and then applying (3.2). This step was essentially pointed out in [62]. The problem is then to evaluate the resulting sum over partitions which to our knowledge was never achieved. We show that this can be done with the following Pfaffian version of the Cauchy–Binet identity.
Lemma 3.1
[Ishikawa–Wakayama Pfaffian identity [35, 36]]. Let A be an \(N \times N\) anti-symmetric matrix and B be an \(M \times N\) matrix such that \(M \le N\) with M even. Then
This leads to a general summation identity over repeated partitions which results in a Pfaffian instead of a determinant. We were motivated to prove the following lemma by work of Betea and Wheeler [10] who proved a particular case in relation to so-called refined Cauchy–Littlewood identities.
Lemma 3.2
Let \(\varvec{x} \in {\mathbb {C}}^{2k}\) and let \(f:{\mathbb {N}} \rightarrow {\mathbb {C}}\). Then
Proof
Inserting definition (2.3) of the Schur functions, the left-hand side of (3.5) is
We introduce indices \(l_{j} = \mu _{j}+2k-j\) for \(j=1,\ldots ,2k\). Then \(\{\mu _{j}\}_{j=1}^{2k}\) satisfies the weakly decreasing property with \(\mu _{2k} \ge 0\) and \(\mu _{1} \le N\) if and only if we have the constraints
Furthermore, \(\mu _{2j}=\mu _{2j-1}\) if and only if \(l_{2j}+1=l_{2j-1}\) for each \(j=1,\ldots ,k\). Then we can write the sum in (3.6) as
Now we express the product over j as the Pfaffian of an anti-symmetric tridiagonal matrix, using the fact that for a generic sequence \(\{ a_j \}_{j=1}^{2k-1}\), we have
This identity follows inductively from a Laplace expansion along the first row. Hence, we have
where \(\delta _{i,j-1}\delta _{l_{i},l_{i+1}+1} = \delta _{l_{i},l_{j}+1}\) follows from the strictly decreasing property of \(\{l_{j}\}_{j=1}^{2k}\). Hence, after reordering indices, we can write (3.7) as
to which Lemma 3.1 is applicable. For this, we identify the matrix elements in the lemma as
so that (3.8) becomes
which completes the proof. \(\square \)
Theorem 3.3
We have
where
Proof
By the dual Cauchy identity (2.5) and average (3.2), we can write the correlator (3.1) as
where \(\eta ' = \mu '/2\) and we used (2.7). Using Lemma A.1, equation (A.3) with \(u=N\) and \(\alpha =1/2\), we can express the coefficient in (3.9) as
where \(\tilde{\eta }\) is the complement partition of \(\eta \), recall definition (2.6). Since \(\mu '\) is even, we deduce that \(\mu \) is repeated and \(\tilde{\eta }_{j} = \tilde{\mu }_{2j}\). Then replacing \(\tilde{\mu }\) with \(\mu \), the sum in (3.9) becomes
Applying Lemma 3.2 with \(f(l) = \frac{1}{l!}\) completes the proof. \(\square \)
3.2 Truncated Orthogonal Ensemble: \(\Omega =I_{M}\)
We now prove an analogous Pfaffian formula for truncations of random orthogonal matrices.
Theorem 3.4
Let \(\varvec{z} \in {\mathbb {C}}^{2k}\) and T be a \(M \times M\) truncation of a \(N \times N\) Haar distributed orthogonal matrix. Then
where \(D^{(1)}_{k}\) is given in (1.25).
Proof
We proceed as in the real Ginibre case and expand the characteristic polynomial product in terms of Schur functions. The relevant average is now the one computed in [61] as
We then have
Rewriting the hypergeometric coefficients using Lemma A.1 with \(\alpha =1/2\), we obtain
Finally, we employ Lemma 3.2 with the choice
to obtain the result. \(\square \)
3.3 Probabilistic Interpretation: Real Ensembles
In the spirit of Sect. 2.4, we introduce the following interpretation of the real ensemble multi-point correlators.
From the proof of Theorem 3.3, we recall the representation
Consider the set \({\mathcal {P}}_{k}\) of all partitions \(\eta \) of length \(l(\eta ) \le k\) and assign each \(\eta \in {\mathcal {P}}_{k}\) the probability
where \({\mathcal {Z}}_{k}\) is a normalization factor. To have a probabilistic interpretation, we require \(\varvec{z} \in {\mathbb {R}}^{2k}_{+}\) so that (3.10) is positive by Schur positivity. The normalization constant in (3.10) is obtained by following the steps in the proof of Theorem 3.3 with no restriction on \(\eta _{1}\). This gives
where CSE(k) denotes the Circular Symplectic Ensemble of \(2k \times 2k\) self-dual unitary matrices U, i.e., unitary matrices satisfying \(U^D = J^{-1} U^T J = U\), where
The measure dU in (3.11) is the restriction of the Haar measure on U(2k) to self-dual unitary matrices satisfying \(U^{\textrm{D}}=U\) and normalized as a probability measure. See, e.g., [24] for further details about the CSE. The integration formula (3.11) in terms of a Pfaffian has been the subject of some attention recently, see the works [39, 65, 66].
Corollary 3.5
When \(\Omega = I_{N}\), we have the following probabilistic interpretation for the multi-point correlator in (1.16):
where the probability on the right-hand side is defined with respect to distribution (3.10) and \({\mathcal {Z}}_{k}\) is given by (3.11).
In a similar manner, we provide a description for the orthogonal truncations, assuming \(\varvec{z} \in (0,1)^{2k}\). We define the probability distribution as
where the normalization constant is
The above integration identity can be deduced by application of Lemma 3.2; we postpone the details to a separate publication.
Corollary 3.6
When \(\Omega = I_{N}\) and \(\varvec{z} \in (0,1)^{2k}\), we have the following probabilistic interpretation for the multi-point correlator in (1.21):
where the probability on the right-hand side is defined with respect to distribution (3.12), \({\mathcal {Z}}_{k}\) is given by (3.13), and \(D^{(1)}_{k}\) is given by (1.25).
Similar character expansions were obtained in [25] for averages of O(N) characteristic polynomials in the context of last passage percolation models.
3.4 Duality Formulas in Real Ensembles: General \(\Omega \)
In this subsection, we derive dual integral representations of the multi-point correlators of real and quaternionic ensembles. Compared to the previous subsection, we now include the multiplicative perturbations \(\Omega \). Starting with the real Ginibre ensemble, our goal is to evaluate
Before we proceed, let us introduce a few concepts we shall require. The generalized hypergeometric coefficients (3.3) obey the following relations, see [24],
Further, we introduce \(\alpha \)-deformations of (2.1)
In what follows, we denote \(d'_\eta (1) = d'_\eta \). Note that \(d'_\eta = h_\eta (1)\). These coefficients with different values of \(\alpha \) can be related as follows, see, e.g., [26]
Let us introduce Jack polynomials \(P_\eta ^{(\alpha )}(X)\) that can be defined as eigenfunctions of the differential operator
with the additional structure
where \(m_{\mu }\) are the monomial symmetric functions, \(a^{(\alpha )}_{\mu \eta }\) are coefficients that do not depend on X, and \(\mu < \eta \) is the dominance ordering on partitions, see [24, 52, 63] for further details. In the Schur polynomial case \(\alpha =1\), these coefficients are better known as Kostka numbers. Jack polynomials are evaluated at the identity, see, e.g., [26], as
Remark 3.7
In the case \(\alpha = \frac{1}{2}\), Jack polynomials are frequently evaluated on self-dual matrices that have doubly degenerate eigenvalues. It is convenient for the rest of the paper to define the Jack polynomial for \(\alpha = \frac{1}{2}\) to evaluate only the distinct eigenvalues, see [21]. Hence, if X has eigenvalues \(\varvec{x} = (x_1,x_1,x_2,x_2,\ldots ,x_k,x_k)\), we define \(P^{(\frac{1}{2})}_{\eta }(X) := P^{(\frac{1}{2})}_{\eta }(x_1,x_2,\ldots ,x_k)\). If \(\varvec{y} \in {\mathbb {C}}^{N}\), the generalized Cauchy identity gives
For more details, see [21].
We begin with the duality formula for the real Ginibre ensemble, completing the proof of Theorem 1.4.
Proof of Theorem 1.4
Proceeding as in the unperturbed case, we employ the dual Cauchy identity, which yields
Now, we recognize the Schur function average as that computed by Forrester and Rains in [26], namely
Substituting this average into the sum over partitions, we find
where to obtain the last line we have used that \((2\eta ')'=\eta ^2\). Next we recognize the summand as the integral of Lemma B.9, that is,
where \(C_{A} = \int _{{\mathcal {A}}_{2k}({\mathbb {C}}) } dX e^{-\frac{1}{2} {{\,\textrm{Tr}\,}}XX^\dag }\). Substituting this expression in (3.19) and interchanging it with the summation, we obtain
Computing the sum using the Cauchy identity (3.17) proves (1.18). Setting \(\Omega = I_{N}\) in (1.18), similar to the complex case, we recognize the Pfaffian of a block matrix leading directly to (1.19). Together with the Pfaffian evaluation derived in Sect. 3.1, this completes the proof of Theorem 1.4. \(\square \)
We now conclude the proof of Theorem 1.5 regarding truncated orthogonal random matrices.
Proof of Theorem 1.5
As the proof is similar to the proof of Theorem 1.4, we just give the main ideas. Having expanded the characteristic polynomial product (1.21) in terms of Schur functions, the relevant Schur function average generalizing (3.18) was computed in [61] as
Substituting this average in the sum over partitions, we have
where we used the transposition property (3.4) of \([u]^{(\alpha )}_{\eta '}\). Next we recognize the summand as the integral of Lemma B.10, namely the coefficient \(\frac{s_{\eta ^2}(\varvec{z}^{-1}) h_{\eta '}(2)}{ (-1)^{|\eta |} \left[ -N \right] ^{(1/2)}_{\eta }}\) is equal to
Substituting this integral in (3.20) and computing the sum over partitions with the help of the generalized Cauchy identity yield the result. The computation of the normalization constant \(S^{(1)}_{k}\) is given in Lemma B.11. \(\square \)
3.5 Quaternionic Ensembles
Consider the multi-point correlator
where the average is over a matrix G from the quaternionic Ginibre ensemble. As in previous cases, we abbreviate \(R^{\textrm{GinSE}}_{N}(\varvec{z}) := R^{\textrm{GinSE}}_{N}(\varvec{z};I_{2N})\). For background on the quaternionic Ginibre ensemble, we refer the reader to the review article [16]. As this ensemble has a natural symplectic symmetry, it is also referred to as Ginibre Symplectic Ensemble (GinSE). Here, we will adopt the same underlying Gaussian measure as in [26]. The character expansion approach here goes through similarly to the real case so we just present the main differences. The required Schur function average was obtained in [26] as
Since the above will be applied to the conjugate partition \(\mu '\), we need an analogue of Lemma 3.2 where we sum over even partitions rather than repeated partitions.
Lemma 3.8
Let \(\varvec{x} \in {\mathbb {C}}^{2k}\) and let \(f:{\mathbb {N}} \rightarrow {\mathbb {C}}\). Then
Proof
Following the same steps as for the proof of Lemma 3.2, here we need to evaluate
and we begin with the same substitution \(l_{j} = \mu _{j}+2k-j\) for \(j=1,\ldots ,2k\). Note that the partition \(\mu \) is even if and only if \(l_{2j-1}\) is odd and \(l_{2j}\) is even for \(j=1,\ldots ,k\). We write the product of indicator functions in (3.22) as the Pfaffian
This identity follows from the fact that for generic sequences \(\{a_{j}\}_{j=1}^{2k-1}\), \(\{b_{j}\}_{j=2}^{2k}\), we have
which follows by induction on k and Laplace expansion. Therefore, after shifting indices by 1 and reordering, we see that (3.22) is equal to
Now the proof is completed by applying Lemma 3.1 with matrix elements
\(\square \)
Applying the above lemma, we obtain Pfaffian formulas for multi-point correlators.
Theorem 3.9
For \(\varvec{z} \in {\mathbb {C}}^{2k}\), we have
where the kernel is
Proof
For \(\Omega = I_{2N}\), the result (3.21) is
We apply Lemma A.1, equation (A.3) with \(\alpha =2\), so that
Then we have
The proof is completed by applying Lemma 3.8 with
\(\square \)
The Pfaffian expression above recovers a result of Akemann and Basile [3]. Their approach differs from ours through their use of the joint probability density function of eigenvalues of G. The kernel (3.23) was first discovered in the context of eigenvalue correlations of symplectic ensembles in [41]. Using the representation as a character sum in (3.24), one can write an analogous probabilistic interpretation that we obtained in the real case in Corollary 3.5; we omit the details.
Dualities of the type (1.19) and (1.23) also hold in the quaternionic case, the main difference being that the dual integrals are over complex symmetric matrices instead of anti-symmetric matrices. Let \({\mathcal {S}}_{2k}({\mathbb {C}})\) denote the space of \(2k \times 2k\) symmetric matrices with complex entries. On the space \({\mathcal {S}}_{2k}({\mathbb {C}})\), we put the Gaussian measure
where dX is the Lebesgue measure on the upper triangular and diagonal elements of X.
Theorem 3.10
Let \(\varvec{z} \in {\mathbb {C}}^{2k}\). Then with \({\mathbb {E}}_{X}\) denoting the average with respect to (3.25), we have
and for \(\Omega = I_{2N}\) (3.26) equals
Proof
Applying the dual Cauchy identity, we have
Using integral identity (B.6), we recognize the summand as
Inserting (3.29) into (3.28) and summing using the generalized Cauchy identity (3.17) complete the proof. \(\square \)
For a derivation of (3.27) using Grassmann calculus, see [53]. We now give the corresponding result for truncations of symplectic unitary matrices. The main difference with the Ginibre ensemble is that the duality comes from a different integral representation for the coefficients that arise in the character expansion (see (B.9)). Consider the following dual measure on the space \({\mathcal {S}}_{2k}({\mathbb {C}})\)
where \(S_{k}^{(4)} = \int _{{\mathcal {S}}_{2k}({\mathbb {C}}) } dX \det ( I_{2k} + XX^\dag )^{-N-1-2k}\).
Theorem 3.11
Let T be an \(M \times M\) truncation of a \(N \times N\) Haar distributed symplectic matrix. Then with \({\mathbb {E}}_{X}\) denoting the average with respect to (3.30) over \({\mathcal {S}}_{2k}({\mathbb {C}})\), we have
and for \(\Omega = I_{2M}\)
where the kernel is
Remark 3.12
The kernel (3.31) is known in the context of eigenvalue correlation functions of the truncated symplectic ensemble [44], see also [41]. The constants in Theorem 3.11 are
and
see Lemma B.11.
4 Asymptotic Expansion in the Real Ginibre Ensemble
In this section, we discuss the asymptotics of multi-point correlators in the real Ginibre ensemble as the matrix size \(N \rightarrow \infty \). We focus on the real case because analogous results for the complex Ginibre ensemble were obtained in previous works, for moments in [68] and for correlators in [18]. For simplicity, we restrict ourselves to the setting \(\Omega = I_{N}\) throughout this section.
Let \(\{z_{j}\}_{j=1}^{2k}\) be complex numbers and define a normalized correlator
where \(G_{N} = G/\sqrt{N}\) and G is a matrix from the real Ginibre ensemble distributed as in (1.15). This scaling ensures that the support of the limiting distribution of eigenvalues of \(G_{N}\) is the unit disk in the complex plane. Our results suggest two options for carrying out this asymptotic analysis, starting either from the Pfaffian representation (1.20) or from the dual integral (1.19). We will mainly concentrate on the Pfaffian representation in what follows.
4.1 Multi-Point Asymptotics: Bulk and Edge
It is natural to split the results for multi-point asymptotics into 4 cases, depending on whether one scales the characteristic variables on the real line or complex plane and whether one scales near the edge or in the bulk.
Corollary 4.1
[Real bulk]. Let \(z_j = x + N^{-\frac{1}{2}}\zeta _j\) with real and fixed \(x \in (-1,1)\) for \(j=1,\ldots ,2k\) and let \(\{\zeta _j\}_{j=1}^{2k}\) be fixed and distinct complex numbers. Then as \(N \rightarrow \infty \)
Proof
Starting from formula (1.20), we observe that
where the kernel is
Applying, e.g., Lemma 11 from [11] shows that
while Stirling’s formula gives
Inserting (4.5) and (4.4) into (4.3) completes the proof. \(\square \)
The case of the real bulk asymptotics in Corollary 4.1 has been studied recently. Afanasiev [1] developed a Grassmann integral approach to compute (4.1) for general real random matrices M in place of G, where M has i.i.d. entries with finite moments. Using supersymmetry techniques, Afanasiev obtained the asymptotics (Conjecture 1.2 in [1]) in the form
as \(N \rightarrow \infty \), where \(z_{j}\) are defined such that \(z_{2j-1} = x+\xi _{j}/\sqrt{N}\), \(z_{2j}=x+\overline{\xi _{j}}/\sqrt{N}\) with real \(|x|<1\) and complex \(\{\xi _{j}\}_{j=1}^{k}\) fixed for all \(j=1,\ldots ,k\). Quantity \(C_{k,x}\) is an unspecified constant, independent of the choice of distribution on the entries and \(\kappa _{4}={\mathbb {E}}(M_{11}^{4})-3\) (for the Gaussian case, \(\kappa _{4}=0\)). The \(2 \times 2\) matrix kernel inside the above Pfaffian is
It is easy to see that these asymptotics are consistent with Corollary 4.1. Although this is stated as Conjecture 1.2 in [1], the supersymmetry method is applied to show that the asymptotics in (4.6) are given in the form of a matrix integral that was apparently unknown to the author (the work [65] was not cited). In fact, the matrix integral that appears in [1] is precisely the one we discussed in (3.11) and its evaluation in the works [39, 65, 66] as a ratio of Pfaffian and Vandermonde determinant settles Conjecture 1.2 in [1]. Moreover, Corollary 4.1 shows that \(C_{k,x}=1\).
Tribe and Zaboronski [65] also proved the expansion (4.2) in the Gaussian setting in the particular case \(x=0\), based on a steepest descent analysis of the matrix integral (1.19). This representation has the advantage that there is no Vandermonde factor in the denominator (c.f. (4.3)), allowing for better uniformity in the estimates.
Theorem 4.2
[Tribe and Zaboronski [65]] The asymptotic expansion in the real bulk of Corollary 4.1 is valid uniformly in the microscopic variables \(\varvec{\zeta }\) varying in compact subsets of \({\mathbb {C}}^{2k}\).
Proof
Inspecting the proof given in [65], it is clear that the argument given easily extends to any base point \(x \in (-1+\delta ,1-\delta )\) for any fixed \(\delta >0\) and that furthermore the error bounds are uniform in \(\varvec{\zeta }\) as stated. \(\square \)
Although the result in Corollary 4.1 does not include the uniformity in \(\varvec{\zeta }\) in Theorem 4.2, our proof has the advantage that it only requires some mild asymptotics of the correlation kernel (c.f. (4.4)). Another advantage of our approach is that we can compute the asymptotics when the base point x is away from the real bulk, e.g., near the real edges \(\pm 1\), or in the complex plane. Following a similar strategy to the proof of Corollary 4.1, we derive the following edge asymptotics. For this, we need the complementary error function
Corollary 4.3
(Real edge). Let \(z_j = 1 + N^{-\frac{1}{2}}\zeta _j\) for each \(j=1,\ldots ,2k\) where \(\{\zeta _j\}_{j=1}^{2k}\) are fixed and distinct complex numbers. Then as \(N \rightarrow \infty \)
where
Proof
We apply the strategy outlined in the proof of Corollary 4.1. The main difference is that in place of asymptotics (4.4) we need to use the asymptotics of the kernel near the edge, see, e.g., [11]
\(\square \)
Remark 4.4
If one merges points \(\zeta _{j} \rightarrow \zeta \) for all \(j=1,\ldots ,2k\) in expression (4.7), it can be shown that one formally recovers the asymptotics given in [67] in terms of certain largest eigenvalue distributions, see also [18] for results of this type in the complex Ginibre ensemble.
When the base point is away from the real line but still contained within the bulk of the Ginibre circular law \(|z|<1\), we obtain asymptotics similar to those known for the complex Ginibre ensemble given in [18]. This reflects a similar phenomenon known for complex eigenvalues of real matrices [11].
Corollary 4.5
(Complex bulk). In the correlator (4.1), consider the scaling
with distinct \(\zeta _{1}< \ldots < \zeta _{k}\) and distinct \(\xi _{1}< \ldots < \xi _{k}\), and where z is a fixed complex number satisfying \(z \in {\mathbb {C}}\setminus {\mathbb {R}}\) and \(|z|<1\). Then
Proof
We have that
First observe that
Next note that the leading term in the asymptotics of \(B_{N+2k-1}(\sqrt{N}z_{i},\sqrt{N}z_{j})\) is of the form \(e^{Nz_{i}z_{j}}\). Due to the inequality \(|z|^{2}-\textrm{Re}(z^{2}) > 0\) holding provided z has nonzero imaginary part, we see that \(B_{N}(\sqrt{N}z_{i},\sqrt{N}z_{j})\) is exponentially subleading unless both i and j belong simultaneously to the set \(\{1,\ldots ,k\}\) or to the set \(\{k+1,\ldots ,2k\}\). Then we have that
where A is a \(k \times k\) matrix with entries \(A_{ij} = (\overline{z}-z)e^{\zeta _{i}\xi _{j}} = -2i\textrm{Im}(z)e^{\zeta _{i}\xi _{j}}\). Applying Stirling’s formula as in (4.5) and inserting (4.10) and (4.11) into (4.9) complete the proof. \(\square \)
An analogous derivation allows to extend the above result to the complex edge \(|z|=1\).
Corollary 4.6
(Complex edge). Let \(|z|=1\) with \(z \in {\mathbb {C}} \setminus {\mathbb {R}}\) fixed and consider the scaling
with distinct \(\zeta _{1}< \ldots < \zeta _{k}\) and distinct \(\xi _{1}< \ldots < \xi _{k}\)
where \(B_{\textrm{edge}}(\xi _{i},\zeta _{j})\) is given in (4.8).
4.2 Moments and Correlators of Moments
For various applications, one wishes to compute moments which involve taking powers of the characteristic polynomial, such as
As before and throughout this section, we remind that \(G_{N} = G/\sqrt{N}\) where G is a matrix from the real Ginibre ensemble (1.15). The recent work [47] obtained asymptotic estimates for moments of type (4.12) provided that k is either an integer or a half-integer, obtaining lower and upper bounds. In the following, we obtain a more precise asymptotic expansion up to and including the constant term, but we assume that k is an integer. It will be helpful to recall the matrix integral discussed in (3.11),
Corollary 4.7
(Integer moments in the GinOE). Let \(k \in {\mathbb {N}}\) be fixed. The following asymptotic expansion holds for any fixed \(x \in (-1,1)\)
Proof
Starting from Theorem 4.2 and Corollary 4.1, we exploit the uniformity of the estimates in [65] to merge points \(\zeta _{j} \rightarrow 0\) for each \(j=1,\ldots ,2k\). For this purpose, we replace the ratio of Pfaffian and Vandermonde in (4.2) with the integral representation (4.13). Then setting all \(\zeta _{j}=0\) for \(j=1,\ldots ,2k\), the group integral is equal to 1 by normalization. Combining this once again with (4.5) completes the proof. \(\square \)
This can be compared with the case \(x=0\) which can be computed for general non-integer exponents. In such a non-integer setting, we need to make use of the so-called Barnes G-function G(z). Analogously to the Gamma function, it obeys the functional relation
and has an analytic (entire) extension to \({\mathbb {C}}\). For integers \(k \ge 2\), it satisfies
Theorem 4.8
The following asymptotic expansion holds uniformly for \(\gamma \) varying in a compact subset of \((-\frac{1}{2},\infty )\),
where G(z) is the Barnes G-function. Furthermore, when \(\gamma = k \in {\mathbb {N}}\), expansion (4.16) reduces to the \(x=0\) case of (4.14).
Proof
See Appendix Appendix C. \(\square \)
Based on Theorem 4.8 and Corollary 4.7, it seems reasonable to make the following conjecture.
Conjecture 1
The following asymptotic expansion holds uniformly for \(\gamma \) varying in a fixed compact subset of \((-\frac{1}{2},\infty )\), and x in a fixed compact subset of \((-1,1)\),
In the case of the complex Ginibre ensemble, a result analogous to (4.17) was proved by Webb and Wong [68] using Riemann–Hilbert methods. Any proof of Conjecture 1 would be of interest as we are not aware of any analogous approach, such as in [68], that would apply in the context of the real Ginibre ensemble.
Remark 4.9
As well as integer values of \(\gamma \) in Conjecture 1, half-integer values are also interesting. For example, when \(\gamma =\frac{1}{2}\), the left-hand side of (4.17) is directly related to the mean density of real eigenvalues in the real Ginibre ensemble [20], as discussed recently in [16]. In [47, Proposition 8.1], an upper bound for general half-integer exponents is obtained, which captures the same exponential and power law terms predicted in (4.17), but does not specify the value of the constant O(1) term in the asymptotics. The latter task was recently accomplished in [56] where Conjecture 1 is proved in the special case \(\gamma = \frac{m}{2}\) for all \(m \in {\mathbb {N}}\). The case of general non-integer m remains an open problem.
Now we prove a two point extension of Corollary 4.7. To state the result we recall that the Laguerre Symplectic Ensemble (LSE) of dimension \(k_{2}\) and degrees of freedom \(k_{1}\) has joint probability density of eigenvalues \(\lambda _{1},\ldots ,\lambda _{k_2}\) proportional to
with \(\lambda _{j} > 0\) for each \(j=1,\ldots ,k_2\) and we assume here that \(k_{1} \ge k_{2}\). Given such a density, consider the largest eigenvalue \(\lambda _{\textrm{max}}^{\textrm{LSE}_{k_2,k_1}} := \textrm{max}_{\ell =1,\ldots ,k_2}\lambda _{\ell }\). For more background on the LSE, see [24, Chapter 3].
Theorem 4.10
Let \(k_{1},k_{2} \in {\mathbb {N}}\) with \(k_{1} \ge k_{2}\) and let x and y be scaled according to
where \(x_{0} \in (-1,1)\) is fixed. Then as \(N \rightarrow \infty \) the following asymptotic expansion holds uniformly in compact subsets of \(\zeta , \xi \in {\mathbb {C}}\),
where \(F^{(4)}_{k_1,k_2}(x)\) is the distribution function
Proof
Starting from Corollary 4.1 and Theorem 4.2 with \(k=k_{1}+k_{2}\), exploiting the uniformity of the estimates in [65], we employ the integral representation (4.13), setting \(\zeta _{j}=\zeta \) for \(j=1,\ldots ,2k_{1}\) and \(\zeta _{j} = \xi \) for \(j=2k_{1}+1,\ldots ,2(k_{1}+k_{2})\). We then partition the unitary matrix U as
Making use of the unitarity of U, the trace in (4.13) can be explicitly calculated as
where c is the \(2k_2 \times 2k_1\) sub-block in (4.19). Consequently,
As in [18], the random variable \(\textrm{Tr}(cc^{\dagger })\) is well studied, see [23, 51] and references therein. In particular, the full joint distribution of eigenvalues of \(cc^{\dagger }\) is known. Results in [23] imply that the eigenvalues of \(cc^{\dagger }\) are doubly degenerate, and that modulo this degeneracy they are distributed according to the Jacobi Symplectic Ensemble (JSE). The joint density of distinct eigenvalues, say \(t_{1},\ldots ,t_{k_2}\) of \(cc^{\dagger }\) is proportional to
with \(t_{j} \in [0,1]\) for each \(j=1,\ldots ,k_2\). These results imply that, setting \(s=(\xi -\zeta )^{2}\),
where \(c^{\textrm{JSE}}_{k_1,k_2}\) is a known normalization constant, defined such that (4.21) equals 1 when \(\xi =\zeta \). In the second line of (4.21), we changed variable \(t_{j} \rightarrow t_{j}/s\). Finally, we recognize the last line in (4.21) as being proportional to the cumulative distribution function of \(\lambda ^{\mathrm {LSE_{k_2,k_1}}}_{\textrm{max}}\), the largest eigenvalue in the Laguerre Symplectic Ensemble of dimension \(k_{2}\), with \(k_{1}\) degrees of freedom. Plugging in the relevant normalization constants, we have shown that
Now inserting (4.22) and (4.20) back into (4.18), we get the stated result after inserting the one-point expansions (4.14) in the denominator of (4.18). This completes the proof of Theorem 4.10. \(\square \)
References
Afanasiev, I.: On the correlation functions of the characteristic polynomials of real random matrices with independent entries. J. Math. Phys. Anal. Geom. 16(2), 91–118 (2020)
Akemann, G., Aygün, N., Würfel, T.R.: Generalised unitary group integrals of Ingham-Siegel and Fisher-Hartwig type. J. Math. Phys. 65, 023501 (2024)
Akemann, G., Basile, F.: Massive partition functions and complex eigenvalue correlations in matrix models with symplectic symmetry. Nuclear Phys. B 766, 150–177 (2007)
Akemann, G., Kieburg, M., Phillips, M.J.: Skew-orthogonal Laguerre polynomials for chiral real asymmetric random matrices. J. Phys. A: Math. Theor. 43, 375207 (2010)
Akemann, G., Phillips, M.J., Sommers, H.-J.: Characteristic polynomials in real Ginibre ensembles. J. Phys. A: Math. Theor. 42(1), 012001 (2008)
Akemann, G., Vernizzi, G.: Characteristic polynomials of complex random matrix models. Nuclear Phys. B 660(3), 532–556 (2003)
Bailey, E.C., Keating, J.P.: Maxima of log-correlated fields: some recent developments. J. Phys. A 55, 053001 (2022)
Balantekin, A.B.: Character expansions, Itzykson–Zuber integrals, and the QCD partition function. Phys. Rev. D 62(8), 085017 (2000)
Bergère, M.C.: Biorthogonal polynomials for potentials of two variables and external sources at the denominator. arXiv:hep-th/0404126, (2004)
Betea, D., Wheeler, M.: Refined Cauchy and Littlewood identities, plane partitions and symmetry classes of alternating sign matrices. J. Combin. Theory Ser. A 137, 126–165 (2016)
Borodin, A., Sinclair, C.D.: The Ginibre ensemble of real random matrices and its scaling limits. Commun. Math. Phys. 291(1), 177–224 (2009)
Brézin, E., Hikami, S.: Characteristic polynomials of random matrices. Commun. Math. Phys. 214, 111–135 (2000)
Brézin, E., Hikami, S.: Characteristic polynomials of real symmetric random matrices. Commun. Math. Phys. 223(2), 363–382 (2001)
Bump, D., Gamburd, A.: On the averages of characteristic polynomials from classical groups. Commun. Math. Phys. 265(1), 227–274 (2006)
Byun, S., Forrester, P.J.: Progress on the study of the Ginibre ensembles I: GinUE. arXiv:2211.16223, (2022)
Byun, S., Forrester, P.J.: Progress on the study of the Ginibre ensembles II: GinOE and GinSE. arXiv:2301.05022, (2023)
Conrey, J.B., Farmer, D.W., Keating, J.P., Rubinstein, M.O., Snaith, N.C.: Autocorrelation of random matrix polynomials. Commun. Math. Phys. 237(3), 365–395 (2003)
Deaño, A., Simm, N.: Characteristic polynomials of complex random matrices and Painlevé transcendents. Int. Math. Res. Not. IMRN 2022(1), 210–264 (2022)
Diaconis, P., Shahshahani, M.: On the eigenvalues of random matrices. J. Appl. Probab. 31A, 49–62 (1994)
Edelman, A., Kostlan, E., Shub, M.: How many eigenvalues of a random matrix are real? J. Amer. Math. Soc. 7(1), 247–267 (1994)
Feng, Z.M., Song, J.P.: Integrals over the circular ensembles relating to classical domains. J. Phys. A: Math. Theor. 42, 325204 (2009)
Fischmann, J., Forrester, P.J.: One-component plasma on a spherical annulus and a random matrix ensemble. J. Stat. Mech. 2011(10), P10003 (2011)
Forrester, P.J.: Quantum conductance problems and the Jacobi ensemble. J. Phys. A: Math. Gen. 39(22), 6861–6870 (2006)
Forrester, P.J.: Log-Gases and Random Matrices. London Mathematical Society Monographs. Princeton University Press, Princeton, NJ (2010)
Forrester, P.J., Rains, E.M.: Symmetrized models of last passage percolation and non-intersecting lattice paths. J. Stat. Phys. 129(5–6), 833–855 (2007)
Forrester, P.J., Rains, E.M.: Matrix averages relating to Ginibre ensembles. J. Phys. A: Math. Theor. 42, 385205 (2009)
Forrester, P.J., Warnaar, S.O.: The importance of the Selberg integral. Bull. Amer. Math. Soc. 45, 489–534 (2008)
Fyodorov, Y.V.: Spectra of random matrices close to unitary and scattering theory for discrete-time systems. AIP Conf. Proc. 553(1), 191–196 (2001)
Fyodorov, Y.V., Khoruzhenko, B.A.: On absolute moments of characteristic polynomials of a certain class of complex random matrices. Commun. Math. Phys. 273(3), 561–599 (2007)
Fyodorov, Y.V., Khoruzhenko, B.A., Poplavskyi, M.: Extreme eigenvalues and the emerging outlier in rank-one non-Hermitian deformations of the Gaussian unitary ensemble. Entropy 25(1), 74 (2023)
Ginibre, J.: Statistical ensembles of complex, quaternion, and real matrices. J. Math. Phys. 6, 440–449 (1965)
Goodman, N.R.: The distribution of the determinant of a complex Wishart distributed matrix. Ann. Math. Stat. 34, 178–180 (1963)
Grela, J.: Diffusion method in random matrix theory. J. Phys. A: Math. Theor. 49, 015201 (2016)
Hanlon, P.J., Stanley, R.P., Stembridge, J.R.: Some combinatorial aspects of the spectra of normally distributed random matrices. In Hypergeometric Functions on Domains of Positivity, Jack Polynomials, and Applications, volume 138 of Contemp. Math., pp. 151–174. Amer. Math. Soc., (1992)
Ishikawa, M., Okada, S., Wakayama, M.: Applications of minor-summation formula I. Littlewood’s formulas. J. Algebra 183(1), 193–216 (1996)
Ishikawa, M., Wakayama, M.: Applications of minor-summation formula II Pfaffians and Schur polynomials. J. Comb. Theory Ser. A 88, 136–157 (1999)
Johansson, K.: Random growth and random matrices. In European Congress of Mathematics, volume 201 of Progress in Mathematics, pp. 445–456. Birkhäuser, (2001)
Kadell, K.W.J.: The Selberg-Jack symmetric functions. Adv. Math. 130(1), 33–102 (1997)
Kanazawa, T.: Unitary matrix integral for two-color QCD and the GSE-GUE crossover in random matrix theory. Phys. Lett. B 819, 136416 (2021)
Kaneko, J.: Selberg integrals and hypergeometric functions associated with Jack polynomials. SIAM J. Math. Anal. 24(4), 1086–1110 (1993)
Kanzieper, E.: Eigenvalue correlations in non-Hermitean symplectic random matrices. J. Phys. A 35(31), 6631–6644 (2002)
Keating, J.P., Snaith, N.C.: Random matrix theory and \(L\)-functions at \(s=1/2\). Commun. Math. Phys. 214(1), 91–110 (2000)
Keating, J.P., Snaith, N.C.: Random matrix theory and \(\zeta (1/2+it)\). Commun. Math. Phys. 214(1), 57–89 (2000)
Khoruzhenko, B.A., Lysychkin, S.: Truncations of random symplectic unitary matrices. arXiv:2111.02381 (2021)
Khoruzhenko, B.A., Sommers, H.-J., Życzkowski, K.: Truncations of random orthogonal matrices. Phys. Rev. E. 82, 040106 (2010)
Kieburg, M., Guhr, T.: A new approach to derive Pfaffian structures for random matrix ensembles. J. Phys. A Math. Theor. 43, 135204 (2010)
Kivimae, P.: Concentration of equilibria and relative instability in disordered non-relaxational dynamics. arXiv:2212.11452 (2022)
Liu, D-Z., Zhang, L.: Phase transition of eigenvalues in deformed Ginibre ensembles. arXiv:2204.13171 (2022)
Macdonald, I.G.: Symmetric Functions and Hall Polynomials. Oxford Mathematical Monographs (1979)
Meckes, E.S.: The Random Matrix Theory of the Classical Compact Groups. Cambridge Tracts in Mathematics. Cambridge University Press (2019)
Mezzadri, F., Simm, N.J.: Tau-function theory of chaotic quantum transport with \(\beta =1,2,4\). Commun. Math. Phys. 324(2), 465–513 (2013)
Muirhead, R.J.: Aspects of Multivariate Statistical Theory. Wiley (1982)
Nishigaki, S.M., Kamenev, A.: Replica treatment of non-Hermitian disordered Hamiltonians. J. Phys. A: Math. Gen. 35(21), 4571–4590 (2002)
Okounkov, A.: Infinite wedge and random partitions. Selecta Math. 7(1), 57–81 (2001)
Orlov, Y.A.: New solvable matrix integrals. Int. J. Mod. Phys. A 19, 276–293 (2004)
Osman, M.: Least non-zero singular value and the distribution of eigenvectors of non-Hermitian random matrices. arXiv:2404.01149 (2024)
Rider, B., Virág, B.: The noise in the circular law and the Gaussian free field. Int. Math. Res. Not. IMRN (2007)
Rouault, A.: Asymptotic behavior of random determinants in the Laguerre, Gram and Jacobi ensembles. ALEA Lat. Am. J. Probab. Math. Stat. 3, 181–230 (2007)
Santilli, L., Tierz, M.: Schur expansion of random-matrix reproducing kernels. J. Phys. A 54, 435202 (2021)
Schlittgen, B., Wettig, T.: Generalizations of some integrals over the unitary group. J. Phys. A 36(12), 3195–3201 (2003)
Serebryakov, A., Simm, N., Dubach, G.: Characteristic polynomials of random truncations: moments, duality and asymptotics. Random Matrices: Theory Appl. 12(1), 2250049 (2023)
Sommers, H.-J., Khoruzhenko, B.A.: Schur function averages for the real Ginibre ensemble. J. Phys. A 42(22), 222002 (2009)
Stanley, R.P.: Some combinatorial properties of Jack symmetric functions. Adv. Math. 77(1), 76–115 (1989)
Takemura, A.: Zonal polynomials. IMS (1984)
Tribe, R., Zaboronski, O.: The Ginibre evolution in the large-N limit. J. Math. Phys. 55(6), 063304 (2014)
Tribe, R., Zaboronski, O.: Averages of products of characteristic polynomials and the law of real eigenvalues for the real Ginibre ensemble. arXiv:2308.06841 (2023)
Wang, Q., Wang, Y.: Moments of characteristic polynomials in certain random neural networks. Stat. Probab. Lett. 172, 109044 (2021)
Webb, C., Wong, M.D.: On the moments of the characteristic polynomial of a Ginibre random matrix. Proc. Lond. Math. Soc. 118(5), 1017–1056 (2019)
Życzkowski, K., Sommers, H.-J.: Truncations of random unitary matrices. J. Phys. A Math. Gen. 33(10), 2045–2057 (2000)
Acknowledgements
We thank the referee for their constructive comments and feedback.
Funding
Both authors are grateful for financial support from the Royal Society, grant URF\(\backslash \)R1\(\backslash \)180707.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest:
There is no potential conflict of interest.
Additional information
Communicated by Vadim Gorin.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Identities for Partitions and Related Coefficients
Let \(\eta \) be a partition of length k with \(\eta _{1} \le N\), i.e., the Young diagram of \(\eta \) is contained inside the \(k \times N\) rectangle. Also recall the notion of complement partition \(\tilde{\eta }_{j} = N-\eta _{k-j+1}\) for \(j=1,\ldots ,k\). The generalized hypergeometric coefficients are
and for \(\alpha =1\) we denote \([u]_{\eta } = [u]^{(1)}_{\eta }\).
Lemma A.1
For any \(\alpha >0\), we have the identities
and
Together, these give
Proof
For (A.1), we write
For (A.2), similarly
where we replaced j with \(k-j+1\) in the denominator and used the definition of \(\tilde{\eta }_{j}\). \(\square \)
Matrix Integrals
Throughout this section, we make use of the notation introduced in Sect. 3.4. It will be useful to recall Selberg’s integral
valid for \(\textrm{Re}(a)> 0, \textrm{Re}(b)>0\) and \(\textrm{Re}(\gamma ) > -\textrm{min}(1/k,\textrm{Re}(a)/(k-1),\textrm{Re}(b)/(k-1))\). See [27] for a review.
Lemma B.1
[26]. Let the measure \(d\mu (X)\) of a real, complex or quaternion \(k \times k\) matrix X be invariant under rotations \(X \rightarrow UX, XU\) by matrices \(U \in \textrm{O}(k), \textrm{U}(k)\) or \(\textrm{Sp}(2k)\), for which \(\alpha = \frac{1}{2}, 1\) or 2, respectively. Then
where A and B are deterministic complex \(k \times k\) matrices.
While we refer to [26], one can trace this result back to the monograph of Takemura [64, pp.32]. To evaluate the right-hand side of (B.2) in various cases of interest, we will make use of the following class of integrals due to Kadell and Kaneko.
Lemma B.2
[38, 40]. Let \(a, b \in {\mathbb {C}}\) such that \(\textrm{Re}(a),\textrm{Re}(b) > 0\) and \(\alpha >0\). Then
Similarly, we have
Lemma B.3
[21]. Let \(a, b \in {\mathbb {C}}\) such that \(\textrm{Re}(a),\textrm{Re}(b) > -1\), \(\alpha >0\). Then
We further require the limiting form of Lemma B.2.
Lemma B.4
[26]. Let \(a \in {\mathbb {C}}\) such that \(\textrm{Re}(a) > -1\), \(\alpha >0\). Then
where \(C=\int _{[0,\infty )^k}\prod _{j=1}^{k}dt_{j}\, t_j^a e^{-t_j} |\Delta (\varvec{t})|^{2/\alpha }\).
Lemma B.5
We have
where \({\mathbb {E}}_{X}\) is expectation with respect to the \(k \times k\) random matrix X distributed according to (1.9).
Proof
By Lemma B.1, we have
Now let \(t_1,\ldots , t_k\) denote the eigenvalues of \(XX^{\dagger }\). Using Lemma B.3 with \(a=0\), \(b=N\) and \(\alpha =1\), we write (B.3) as
where \(c=\int _{[0,\infty )^{k}}\prod _{j=1}^{k}dt_{j}\,(1+t_{j})^{-N-2k}|\Delta (\varvec{t})|^{2}\) and we used (2.4) in the last line. \(\square \)
Lemma B.6
Let X be a \(k \times k\) standard complex Ginibre random matrix. Then we have
Proof
By Lemma B.1, we have
Following the proof of Lemma B.5, (B.4) can be computed using Lemma B.4 with \(\alpha =1\), \(a=0\), and finally applying (2.4). \(\square \)
Lemma B.7
Let \(\eta \) be a partition of length k. Then
Proof
This follows from combining (3.14), (3.15) and (3.16). \(\square \)
The next Lemma is from [21, Corollary 3.2 and 3.3]. Recall that \({\mathcal {A}}_{k}({\mathbb {C}})\) denotes the space of \(k \times k\) complex anti-symmetric matrices and \({\mathcal {S}}_{k}({\mathbb {C}})\) denotes the space of \(k \times k\) complex symmetric matrices.
Lemma B.8
Let \(A\in {\mathbb {C}}^{k \times k}\) and let \(\rho \) be an invariant Borel measure, i.e., for any \(U \in U(k)\) we have \(d\rho (X) = d\rho ( U X U^T)\). Then
and
Specializing the above to a Gaussian measure and evaluating the right-hand sides, we show the following.
Lemma B.9
We have
where \(C_{A} = \int _{{\mathcal {A}}_{2k}({\mathbb {C}}) } dX e^{-\frac{1}{2} {{\,\textrm{Tr}\,}}XX^\dag }\) and
where \(C_{S} = \int _{{\mathcal {S}}_{k}({\mathbb {C}}) } dX e^{-\frac{1}{2}{{\,\textrm{Tr}\,}}XX^\dag }\).
Proof
We give the proof of (B.5). The proof of (B.6) follows a similar pattern and we omit the details. Choosing the Gaussian measure in Lemma B.8, we obtain
Next we pass to the doubly degenerate eigenvalues of \(XX^\dag \) using a result from [24, Section 3.3.2] on Jacobians of complex anti-symmetric matrices. The latter section in [24] also contains the required Jacobian for complex symmetric matrices when evaluating (B.6). Denoting the distinct eigenvalues \(t_{1},\ldots ,t_{k}\), the integral on the right-hand side of (B.7) is
where \(c=\int _{[0,\infty )^k}\prod _{j=1}^{k}dt_{j}\,e^{-t_j} \,|\Delta (\varvec{t})|^4\) and we used the integral of Lemma B.4. Application of Lemma B.7 completes the proof. \(\square \)
Lemma B.10
For a \(2k \times 2k\) complex matrix A, we have
where \(S^{(1)}_{k} = \int _{{\mathcal {A}}_{2k}({\mathbb {C}})} dX \det ( I_{2k} + XX^\dag )^{-N/2+1-2k}\) is evaluated in Lemma B.11. Furthermore,
where \(S^{(4)}_{k} = \int _{{\mathcal {S}}_{2k}({\mathbb {C}})}dX\,\det (I_{2k}+XX^{\dagger })^{-N-1-2k}\).
Proof
We present the details for (B.8); the integral (B.9) follows a similar pattern. We proceed as in the analogous Gaussian case by employing Lemma B.8, which shows that the left-hand side of (B.8) is equal to
We recall the result from [24, Section 3.3.2] on Jacobians of complex anti-symmetric matrices and pass to the distinct eigenvalues \(t_{1},\ldots ,t_{k}\) of \(XX^\dag \), which implies that the integral in (B.10) is
where \(c = \int _{[0,\infty )^k}\prod _{j=1}^{k} dt_{j}\,( 1 + t_j )^{-N+2-4k}|\Delta (\varvec{t})|^4\) and we applied Lemma B.3. Again, application of Lemma B.7 completes the proof. \(\square \)
Lemma B.11
We have the following integral identities:
and
Proof
By a change of variables to the eigenvalues \(t_{1},\ldots ,t_{k}\) of \(XX^{\dagger }\) we have
where the proportionality constant \(\tilde{c}\) is known, see, for example [22],
To evaluate (B.14) make the change of variables, for each \(j=1,\ldots ,k\), \(t_{j} = x_{j}/(1-x_{j})\), so that \((1+t_{j})^{-1} = (1-x_{j})\). Then (B.14) is equal to
where we used Selberg’s integral (B.1). For (B.12), note that for \(X \in {\mathcal {A}}_{2k}({\mathbb {C}})\), the eigenvalues of \(XX^{\dagger }\) are non-negative and doubly degenerate. Taking \(t_{1},\ldots ,t_{k}\) for the distinct ones, the eigenvalue-dependent part of the Jacobian is proportional to \(|\Delta (\varvec{t})|^{4}\) see [24, Section 3.3.2]. Then there exists some constant \(\tilde{d}\) such that
To compute \(\tilde{d}\), we adopt a similar strategy to [22], noting that
This is again a particular case of Selberg’s integral and solving for \(\tilde{d}\) gives
To evaluate (B.15) make the change of variables \(t_{j} = x_{j}/(1-x_{j})\) as before. Then another instance of Selberg’s integral (B.1) shows that (B.15) is equal to
We omit the proof of (B.13) as it follows a similar pattern to the evaluation of (B.11) and (B.12). \(\square \)
Asymptotics of Non-integer Moments in the Real Ginibre Ensemble
In this appendix, we give the proof of the asymptotics (4.16). We use the fact that for any real \(\gamma \), one has \(|\det (G)|^{2\gamma } = \det (GG^{\textrm{T}})^{\gamma }\), noting that if G is taken from the real Ginibre ensemble, the symmetric positive definite matrix \(W = GG^{\textrm{T}}\) is well studied and known as a real Wishart matrix or Laguerre Orthogonal Ensemble (LOE) see [24, Chapter 3] for background. For such ensembles, we can follow an approach that was used to prove a central limit theorem for the log determinant of a Wishart matrix in [32] and was applied to several Hermitian ensembles in [58].
Proof of Theorem 4.8
We have that \({\mathbb {E}}(|\det (G_{N})|^{2\gamma }) = N^{-\gamma N}{\mathbb {E}}(\det (W)^{\gamma })\) where \(W=GG^{\textrm{T}}\). The joint probability density function of eigenvalues \(\lambda _{1},\ldots ,\lambda _{N}\) of W is well known, for example, in [24, Chapter 3],
where the normalization constant is given in terms of a limiting form of Selberg’s integral [27]
Now provided \(\gamma > -\frac{1}{2}\), our quantity of interest is
where for simplicity we have assumed that N is even. The Barnes G-function satisfies the following asymptotic expansion [24, Eqn (4.185)],
which is uniform for the parameters a and b varying in compact subsets of \({\mathbb {C}}\). Applying these asymptotics to the last line of (C.1) completes the proof of (4.16). The expansion in (4.16) is fully consistent with the \(x=0\) one of (4.14), for the following reasons. By the functional relation (4.15), we have, for integer \(k \in {\mathbb {N}}\),
Applying this, followed by the duplication formula for the Gamma function we have
Inserting (C.2) into (4.16) for \(\gamma =k\) reduces expansion (4.16) to the particular case \(x=0\) of expansion (4.14). \(\square \)
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Serebryakov, A., Simm, N. Schur Function Expansion in Non-Hermitian Ensembles and Averages of Characteristic Polynomials. Ann. Henri Poincaré (2024). https://doi.org/10.1007/s00023-024-01483-6
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s00023-024-01483-6