Abstract
We study the overlaps between eigenvectors of nonnormal matrices. They quantify the stability of the spectrum, and characterize the joint eigenvalues increments under Dyson-type dynamics. Well known work by Chalker and Mehlig calculated the expectation of these overlaps for complex Ginibre matrices. For the same model, we extend their results by deriving the distribution of diagonal overlaps (the condition numbers), and their correlations. We prove: (i) convergence of condition numbers for bulk eigenvalues to an inverse Gamma distribution; more generally, we decompose the quenched overlap (i.e. conditioned on eigenvalues) as a product of independent random variables; (ii) asymptotic expectation of off-diagonal overlaps, both for microscopic or mesoscopic separation of the corresponding eigenvalues; (iii) decorrelation of condition numbers associated to eigenvalues at mesoscopic distance, at polynomial speed in the dimension; (iv) second moment asymptotics to identify the fluctuations order for off-diagonal overlaps, when the related eigenvalues are separated by any mesoscopic scale; (v) a new formula for the correlation between overlaps for eigenvalues at microscopic distance, both diagonal and off-diagonal. These results imply estimates on the extreme condition numbers, the volume of the pseudospectrum and the diffusive evolution of eigenvalues under Dyson-type dynamics, at equilibrium.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
1.1 The Ginibre ensemble
Throughout this article we will consider a complex Ginibre matrix \(G_N =(G_{ij})_{i,j=1}^N\) where the \(G_{ij}\)’s are independent and identically distributed complex Gaussian random variables, with distribution \(\mu =\mu ^{(N)}\):
where m is the Lebesgue measure on \({\mathbb {C}}\). As proved in [28], the eigenvalues of \(G_N\) have joint distribution
where \(Z_N=N^{-N(N-1)/2}\prod _{j=1}^N j!\). The above measure is written \({\mathbb {P}}_N\), with corresponding expectation \({\mathbb {E}}_N\). The limiting empirical spectral measure converges to the circular law, i.e. \(\frac{1}{N}\sum \delta _{\lambda _i}\rightarrow \frac{1}{\pi }\mathbb {1}_{|\lambda |<1}\mathrm{d}m(\lambda )\).
The statistics of eigenvalues of Ginibre matrices have been studied in great details, and other non-Hermitian matrix models are known to be integrable, see e.g. [23, 37]. Much less is known about the statistical properties of eigenvectors of non-Hermitian ensembles.
1.2 Overlaps
Almost surely, the eigenvalues of a Ginibre matrix are distinct and G can be diagonalized with left eigenvectors denoted \((L_i)_{i=1}^N\), right eigenvectors \((R_i)_{i=1}^N\), defined by \(G R_i=\lambda _iR_i\), \(L_i^{\mathrm{t}}G=\lambda _i L_i^{\mathrm{t}}\) (for a column vector \(\mathbf{{x}}\), we write \(\mathbf{{x}}^{\mathrm{t}}=(x_1,\dots ,x_n)\), \(\mathbf{{x}}^{*}=(\overline{x_1},\dots ,\overline{x_n})\) and \(\Vert \mathbf{{x}}\Vert =(\mathbf{{x}}^*\mathbf{{x}})^{1/2}\)). Right and left eigenvectors are biorthogonal basis sets, normalized by
In other words, defining X with ith column \(R_i\), we have \(G= X \Delta X^{-1}\) with \(\Delta =\mathrm{diag}(\lambda _1,\dots ,\lambda _N)\), and \(L_i^{\mathrm{t}}\) is the ith row of \(Y=X^{-1}\). Because of the normalization (1.3), the first interesting statistics to quantify non-orthogonality of the eigenbasis is
These overlaps are invariant under the rescaling \(R_i\rightarrow c_iR_i\), \(L_i\rightarrow c_i^{-1} L_i\) and the diagonal overlaps \({\mathscr {O}}_{ii}= \Vert R_i \Vert ^2 \Vert L_i \Vert ^2\) directly quantify the stability of the spectrum. Indeed, if we assume all eigenvalues of G are distinct and denote \(\lambda _i(t)\) the eigenvalues of \(G+tE\), standard perturbation theory yields (in this paper \(\Vert M\Vert =\sup _{\Vert \mathbf{{x}}\Vert _2=1}\Vert M \mathbf{{x}}\Vert _2\))
so that the \({\mathscr {O}}_{ii}^{1/2}\)’s are also called condition numbers. They also naturally appear through the formulas \({\mathscr {O}}_{ii}^{1/2}=\Vert R_iL_i^{\mathrm{t}}\Vert \) or \({\mathscr {O}}_{ii}^{1/2}=\limsup _{z\rightarrow \lambda _i} \Vert (z-G)^{-1}\Vert \cdot |z-\lambda _i|\). We refer to [52, Sects. 35 and 52] for further discussion and references about the relevance of condition numbers to the perturbative theory of eigenvalues, and to estimates of the pseudospectrum.
Eigenvector overlaps also play a fundamental role in non perturbative dynamical settings. First, the large off-diagonal \({\mathscr {O}}_{ij}\)’s appear when G is the generator of evolution in real or imaginary time, see [13, Appendix B]. More generally, eigenvector correlations are as relevant as eigenvalue distributions in determining evolution at intermediate times, a well known fact in hydrodynamic stability theory [53]. Second, the overlaps also fully characterize the eigenvalue increments when all matrix entries undergo independent Brownian motions, as shown in “Appendix A”, for any deterministic initial condition. For the Dyson Brownian motion on Hermitian matrices, the eigenvalues evolution is autonomous and coincides with Langevin dynamics for a one-dimensional log-gas. On the contrary, in the nonnormal setting, the Dyson and Langevin dynamics strongly differ. More about the Dyson-type dynamics in the context of the Ginibre ensemble can be found in [11, 30], and the Langevin equation related to (1.2) is studied in [7].
1.3 Overlaps statistics
The statistical study of overlaps started with the seminal work of Chalker and Mehlig [12, 13, 43]. They estimated the large N limit of the expectation of diagonal and off-diagonal overlaps, for the complex Ginibre ensemble: for any \( |z_1|, |z_2|<1\),
with (1.6) uniformly in \(|z_1-z_2|\) from the macroscopic up to the microscopic \(N^{-1/2}\) scale.Footnote 1 In [13], (1.5) and (1.6) were rigorously established for \(z_1=0\), and convincing heuristics extended them anywhere in the bulk of the spectrum. From (1.5), one readily quantifies the instability of the spectrum, an order N greater than for normal matrices in the bulk, and more stable closer to the edge.
An increasing interest in the statistical properties of overlaps for nonnormal matrices followed in theoretical physics [5, 29, 33, 48, 50], often with interest in calculating overlaps averages beyond the Ginibre ensemble. For example, eigenvector overlaps appear to describe resonance shift if one perturbs a scattering system [25,26,27]. This was experimentally verified [31]. Remarkably, the exact statistics (1.6) appeared very recently in an experiment from [15] for microscopic separation of eigenvalues, suggesting some universality of this formula. Unfortunately, many of the models considered in the physics literature are perturbative, and most of the examined statistics are limited to expectations.
In the mathematics community, the overlaps were recently studied in [54]. Walters and Starr extended (1.5) to any \(z_1\) in the bulk, established asymptotics for \(z_1\) at the edge of the spectrum, and suggested an approach towards a proof of (1.6). They also studied the connection between overlaps and mixed matrix moments. Concentration for such moments for more general matrix models was established in [21], together with applications to coupled differential equations with random coefficients. We continue the rigorous analysis of overlaps by deriving the full distribution of the condition numbers for bulk eigenvalues of the complex Ginibre ensembles. We also establish (1.6) and an explicit formula for the correlation between diagonal and off-diagonal overlaps, on any scale including microscopic. These formulas have consequences on the volume of the pseudospectrum and eigenvalues dynamics.
Motivated by our explicit distribution for the overlaps, Fyodorov [24] recently derived the distribution of diagonal overlaps for real eigenvalues of real Ginibre matrices, as well as an alternative proof for the distribution of diagonal overlaps for the complex Ginibre ensemble. Fyodorov’s method relies on the supersymmetry approach in random matrix theory, while our technique is probabilistic, as described below.
1.4 Main results
Equation (1.5) suggests that the overlaps have typical size of order N. For the complex Ginibre ensemble, we confirm that this is indeed the typical behavior, identifying the limiting distribution of \({\mathscr {O}}_{11}\). We recall that a Gamma random variable \(\gamma _{\alpha }\) has density \(\frac{1}{\Gamma (\alpha )}x^{\alpha -1}e^{-x}\) on \({\mathbb {R}}_+\).
Theorem 1.1
(Limiting distribution of diagonal overlaps) Let \(\kappa >0\) be an arbitrarily constant. UniformlyFootnote 2 in \(|z|<1-N^{-\frac{1}{2}+\kappa }\), the following holds. Conditionally on \(\lambda _1=z\), the rescaled diagonal overlap \({\mathscr {O}}_{11}\) converges in distribution to an inverse Gamma random variable with parameter 2 as \(N\rightarrow \infty \), namely
Our proof also gives convergence of the expectation, in the complex Ginibre case, so that it extends (1.5). Equation (1.7) means that for any continuous bounded function f we have
This \(t^{-3}\) asymptotic density was calculated for \(N=2\) in [13, Sect. V.A.2], where this heavy tail was suggested to remain in the large N limit.
Theorem 1.1 requires integrating over all the randomness of the Ginibre ensemble, in this sense this is an annealed result. It derives from a quenched result, when conditioning on all eigenvalues: the overlap \({\mathscr {O}}_{11}\) can then be decomposed as a product of independent random variables, see Theorem 2.2. Very similar results have been recently established for the Quaternionic Ginibre Ensemble [19], as well as for the Spherical and Truncated Unitary Ensembles [20].
We observe that the limiting density in (1.8) vanishes exponentially fast at 0, so that it is extremely unlikely to find any bulk overlap of polynomial order smaller than N: the spectrum is uniformly unstable. This is confirmed by the following bound on the extremes of condition numbers.
Corollary 1.2
(Bounds on the condition numbers) Let \(\kappa ,{\varepsilon }>0\), \(\kappa <\kappa _0\leqslant 1/2\) be fixed and \(\Omega _N\subset \{1-N^{-\frac{1}{2}+\kappa _0}\leqslant |z|<1-N^{-\frac{1}{2}+\kappa }\}\) be deterministic, measurable. Then with probability tending to 1 as \(N\rightarrow \infty \), the following event holds: for any \(\lambda _i\in \Omega _N\),
In particular, all bulk overlaps are in \([N^{1-{\varepsilon }},N^{3/2+{\varepsilon }}]\) with large probability. In terms of polynomial scales in N, the above lower bound is clearly optimal, and we believe the upper bound is also the best possible.
The next result is a rigorous proof of (1.6) in the bulk of the spectrum. It answers Conjecture 4.5 in [54] and gives firm grounds to the heuristic arguments of Chalker and Mehlig [13]. Different heuristics towards Theorem 1.3 for more general ensembles recently appeared in [46], based on diagrammatics. Another recent approach [14] allows to compute the conditional expectation of more general multi-index overlaps when eigenvalues are conditioned to be at macroscopic distance.
Theorem 1.3
(Expectation of off-diagonal overlaps, microscopic and mesoscopic scales) For any \(\kappa \in (0,1/2)\), any \({\varepsilon }>0\) and \(C>0\) the following holds. Uniformly in \(z_1,z_2\) such that \(|z_1|<1-N^{-\frac{1}{2}+\kappa }\), \(\omega =\sqrt{N}|z_1-z_2|\in [N^{-C},N^{\kappa -{\varepsilon }}]\), we have
In particular, under the same hypothesis, in the mesoscopic regime \(|w|\rightarrow \infty \), Eq. (1.9) simplifies to
showing that the expectation of off-diagonal overlaps decreases with the separation of eigenvalues. Our next result, about second moments at any scale, allows to identify the natural size of off-diagonal overlaps, and gives polynomial decay of correlations between condition numbers.
Theorem 1.4
(Correlations of overlaps: microscopic and mesoscopic scales) Let \(\kappa \in (0,1/2)\) and \(\sigma \in (0,\kappa )\). Let \({\varepsilon }>0\). Then uniformly in \(z_1,z_2\) such that \(|z_1|<1-N^{-\frac{1}{2}+\kappa }\), \(\omega =\sqrt{N}|z_1-z_2|\in [N^{-\frac{\kappa }{2}+{\varepsilon }},N^{\sigma }]\), we have
For the mesoscopic scales \(|w|\rightarrow \infty \), the above asymptotics become
Equations (1.10) and (1.13) suggest that for any mesoscopic separation of eigenvalues, \({\mathscr {O}}_{12}\) does not concentrate, because \({\mathbb {E}}(|{\mathscr {O}}_{12}|^2)\) is of larger order than \({\mathbb {E}}({\mathscr {O}}_{12})^2\). Contrary to (1.6), (1.13) therefore identifies the size of off-diagonal overlaps, at mesoscopic scales.
The covariance bounds from Theorem 1.4 yield effective estimates on the volume of the pseudospectrum, defined through \( \sigma _{\varepsilon }(G)=\left\{ z: \Vert z-G\Vert ^{-1}>{\varepsilon }^{-1}\right\} \). We state the result when the pseudospectrum is intersected with a mesoscopic ball, although it clearly holds on any domain within the bulk that is regular enough.
Corollary 1.5
(Volume of the pseudospectrum) Let \(\kappa>a>0\) be any constants and \({\mathscr {B}}_N\subset \{|z|<1-N^{-\frac{1}{2}+\kappa }\}\) be a ball with radius at least \(N^{-\frac{1}{2}+a}\), at most \(N^{-\frac{1}{2}+\kappa -a}\). Then the volume of the pseudospectrum in \({\mathscr {B}}_N\) is deterministic at first order: for any \(c>0\),
Finally, we remark that our results shed some light on natural matrix dynamics on nonnormal matrices, the Dyson-type evolution where all matrix entries follow independent complex Ornstein-Uhlenbeck processes. Under this evolution, the eigenvalues follow the dynamics (see Proposition A.1)
where the martingales \((M_k)_{1\leqslant k\leqslant N}\) have brackets \(\langle M_i,M_j\rangle =0\) and \( \mathrm{d}\langle M_i,\overline{M_j}\rangle _t={\mathscr {O}}_{ij}(t)\frac{\mathrm{d}t}{N}. \) Based on this observation, Theorem 1.1, Theorem 1.3 and some bound from [24], we show that the eigenvalues propagate with diffusive scaling, at equilibrium, but slower when close to the boundary.
Corollary 1.6
(Diffusive exponent for eigenvalues dynamics) Let \(c,a>0\) be arbitrarily. Consider the matrix dynamics (A.1) with initial condition G(0) distributed as (1.1). Let \({\mathscr {B}}\subset \{|z|<1-c\}\) be a ball, and \(t<N^{-c}\). Then as \(N\rightarrow \infty \) we have
Given the time scale in (1.16), we expect that, conditionally on \(\{\lambda _i(0)=z\}\), the process
converges in distribution to a normalized complex Brownian motion as \(N\rightarrow \infty \) (t is any scale \(N^{-1+c}<t<N^{-c}\)), i.e. \( \frac{1}{\sqrt{2}}(B_s^{(i)}+\mathrm {i}\widetilde{B}_s^{(i)})_{0\leqslant s\leqslant 1} \) with \(B^{(i)}\) and \(\widetilde{B}^{(i)}\) independent standard Brownian motions. Moreover, from (1.17), we expect that these limiting processes associated to different eigenvalues are independent.
1.5 About the proofs
Our analysis of the condition numbers starts with the observation of Chalker and Mehlig: the overlaps coincide with those of the Schur form of the original operator, a particularly simple decomposition when the input is a Ginibre matrix. We refer the reader to (2.2) for this key structure at the source of our inductions.
In Sect. 2 our method to prove Theorem 1.1 follows from a simple, remarkable identity in law: the quenched overlap (i.e. conditionally on eigenvalues and integrating only over the complementary randomness of the Ginibre ensemble) is a product of \(N-1\) independent random variables, see Theorem 2.2. In the specific case \(\lambda _1=0\), for the complex Ginibre ensemble, this split of the distribution of \({\mathscr {O}}_{11}\) remains in the annealed setting, as a consequence of a theorem of Kostlan: the radii of eigenvalues of complex Ginibre matrices are independent random variables. For the Ginibre ensemble with conditioned eigenvalue, we give the extension of Kostlan’s theorem in Sect. 5. This concludes a short probabilistic proof of Theorem 1.1 in the case \(\lambda _1=0\).
The extension to general \(\lambda _1\) in the bulk proceeds by decomposition of \({\mathscr {O}}_{11}\) into a long-range factor, which gives the deterministic \(1-|z|^2\) coefficient, and a short-range factor, responsible for the \(\gamma _2^{-1}\) fluctuations. Concentration of the long-range contribution relies on rigidity results for eigenvalues, from [9, 10]. To prove that the short-range contribution is independent of the position of \(\lambda _1\), we need strong form of invariance for our measure, around the conditioned point \(\lambda _1\). While translation invariance for \({\mathbb {P}}\) follows easily from the explicit form of the Ginibre determinantal kernel, the invariance of the conditioned measure requires more involved tools such as the negative association property for determinantal point processes, see Sect. 4. Corollary 1.2 directly follows from our estimates on the speed of convergence to the inverse Gamma distribution, as explained in Sect. 2.
The proof of Theorems 1.3 and 1.4 in Sect. 3 follows the same scheme: first a quenched identity, then an explicit formula obtained for \(z_1=0\) and a localization procedure to isolate the short and long-range contributions. The main difference with the proof of Theorem 1.1 concerns the special case \(z_1=0\), the other step being more robust. Due to conditioning on \(z_2\), rotational invariance of the remaining eigenvalues is broken and there is no analogue of Kostlan’s theorem to obtain the explicit joint distribution of \({\mathscr {O}}_{12}, {\mathscr {O}}_{11}\) and \({\mathscr {O}}_{22}\). Despite this lack of rotational invariance, Chalker and Mehlig had already obtained a closed-form formula for \({\mathbb {E}}({\mathscr {O}}_{12})\) when \(z_1=0\). Remarkably, the second moments are also explicit (although considerably more involved) even for finite N, see Proposition 3.6. Our correlation estimates on overlaps imply Corollary 1.5 by a second moment method. It is plausible yet unclear that Theorem 1.4 admits generalizations to joint moments with an arbitrary number of conditioned eigenvalues. In fact, our second moment calculation involves a P-recursive sequence [see (3.23)], for which explicit solutions are not expected in general. We hope to address the general study of relevant holonomic sequences in future work.
In “Appendix A”, we consider eigenvalues dynamics. After deriving the relevant stochastic differential equations in Proposition A.1, we show that Corollary 1.6 follows from Theorem 1.1 and Theorem 1.3. The diagonal overlaps dictates the eigenvalues quadratic variation, the off-diagonal overlaps their correlations.
Finally, although this article focuses on the eigenvalues condition numbers, the Schur decomposition technique also readily answers the natural question of angles between normalized Ginibre eigenvectors, as explained in “Appendix B”.
1.6 Numerical test for universality of the distribution of overlaps
Universality of eigenvector statistics recently attracted a lot of attention for random Hermitian matrices. For the Gaussian orthogonal and unitary ensembles, the eigenvectors basis is Haar distributed on the corresponding unitary group. As a consequence, projections of eigenvectors on deterministic directions are asymptotically normal, as the projection of the uniform measure on high dimensional spheres (a result due to Lévy and Borel). These eigenvector statistics are now known to be universal [8, 39, 51], holding for generalized Wigner matrices.
The situation is quite different for eigenvectors of dense random non Hermitian matrices: orders of magnitude such as delocalization are known [49] , but universality seems out of reach with current techniques. In Fig. 1, we numerically investigate whether the inverse Gamma distribution from Theorem 1.1 describes the typical behavior of condition numbers. This leads us to conjecture that for any complex random matrix with i.i.d. entries with substantial real and imaginary parts, the normalized condition numbers converge to the inverse Gamma distribution with parameter two.
1.7 Notations and conventions
The partial sums of the exponential are written
and we abbreviate \(e_k=e_k^{(\infty )}\), \(e^{(N)}=e_0^{(N)}\). Throughout the paper, \(0 \leqslant \chi \leqslant 1\) is a smooth cut-off function on \({\mathbb {R}}_+\) such that \(\chi (x)=1\) for \(x<1/2\) and 0 for \(x>1\). We write \(f={{\,\mathrm{O}\,}}(\phi )\) if \(|f|<C|\phi |\) for some \(C>0\) which does not depend on N, and \(f={{\,\mathrm{o}\,}}(\phi )\) if \(|f|/|\phi |\rightarrow 0\) as \(N\rightarrow \infty \). Along this work, the constants C and c are some universal (resp. typically large and small) constants that may vary from line to line.
2 Diagonal overlaps
This section first gives a remarkable identity in law for the diagonal overlap conditioned on the eigenvalues, Theorem 2.2. Eigenvalues are then integrated, first for \(\lambda _1\) at the center of the spectrum thanks to a variant of Kostlan’s theorem, then anywhere in the bulk.
2.1 The quenched diagonal overlap
For all \(i\ne j\) we denote \(\alpha _{ij} = \frac{1}{\lambda _i - \lambda _j},\) and \(\alpha _{ii}=0\). These numbers satisfy
We first recall the analysis of the overlaps given by Chalker and Mehlig, and include the short proof for completeness.
Proposition 2.1
(From [12, 13, 43]) The following joint equality in distribution holds:
where \(b_1= d_2=1\), \(d_1=0\), and the next terms are defined by the recurrence
The \(T_{ij}\) are independent complex Gaussian variables, centered and with variance \(N^{-1}\).
Proof
The overlaps are unchanged by an unitary change of basis, and therefore we can study directly the overlaps of the matrix T. As proved in [44, Appendix 35], this Schur transform T has the eigenvalues of G as diagonal entries, and independently the upper triangle consists in uncorrelated Gaussian random variables, \(T_{ij} {\mathop {=}\limits ^{\mathrm{(d)}}} G_{ij}\):
For T, the right eigenvectors are of type
and the left eigenvectors are denoted
Biorthogonality relations give \(b_1=1\), \(d_1=0\), \(d_2=1\) and \(a=- b_2\). The formulas given for the overlaps follow, and the recurrence formulas proceed from the definition of the eigenvectors, i.e. \(L_j^t T = \lambda _j L_j^t\). \(\square \)
Proposition 2.1 shows that the eigenvectors, and thus the overlaps, are obtained according to a very straightforward random process. Indeed, let us consider the sequences of column vectors:
For any k, \(T_k\) is a k-dimensional centered Gaussian vector with independent coordinates and variance 1 / N. We denote the corresponding \(\sigma \)-algebras
In particular, \(b_2=\frac{T_{12}}{\lambda _1-\lambda _2}\in {\mathcal {F}}_1\). The recurrence formula from Proposition 2.1 becomes
Theorem 2.2
The following equality in law holds conditionally on \(\{\lambda _1,\dots ,\lambda _N\}\):
where the \(X_n\)’s are independent standard complex Gaussian random variables (\(X_n{\mathop {=}\limits ^{\mathrm{(d)}}} \frac{1}{\sqrt{2}}({\mathscr {N}}_1+\mathrm {i}{\mathscr {N}}_2)\) with standard real Gaussians \({\mathscr {N}}_1\), \({\mathscr {N}}_2\)).
Remark 2.3
In particular
where \({\mathbb {E}}_T\) is partial integration in the upper-diagonal variables \((T_{ij})_{j>i}\). We therefore recover the result by Chalker and Mehlig [12, 13, 43].
Proof
For fixed N and \(1\leqslant n\leqslant N\) we shall use the notations
Note that
and, with (2.3),
where \(X_{n+1}:= \sqrt{N} {B_n^{\mathrm{t}}T_n \over \Vert B_n \Vert }\) is a \({\mathcal {F}}_{n}\)-measurable Gaussian with variance 1, independent of \({\mathcal {F}}_{n-1}\). We have therefore proved the expected factorization with independent \(X_n\)’s, by an immediate induction. \(\square \)
2.2 The annealed diagonal overlap at the origin
We recall that a Gamma random variable \(\gamma _{\alpha }\) has density \(\frac{1}{\Gamma (\alpha )}x^{\alpha -1}e^{-x}\) on \({\mathbb {R}}_+\), and a Beta random variable \(\beta _{a,b}\) has density \(\frac{\Gamma (a+b)}{\Gamma (a)\Gamma (b)}x^{a-1}(1-x)^{b-1} \mathbb {1}_{[0,1]}(x)\).
Proposition 2.4
Conditionally on \(\{ \lambda _1 =0 \}\), the following equality in distribution holds:
In particular, \( {\mathbb {E}}\left( {\mathscr {O}}_{11}\mid \lambda _1=0 \right) =N\) and \(N^{-1} {\mathscr {O}}_{11}\) converges weakly to \(\gamma _2^{-1}\).
Proof
With the notations from Theorem 2.2, we have \((|X_2|^2,\dots ,|X_N|^2) \overset{(\mathrm d)}{=} (\gamma _1^{(2)},\dots \gamma _1^{(N)})\), a collection of \(N-1\) independent Gamma random variables with parameter 1. Moreover, still conditionally on \(\lambda _1=0\), from Corollary 5.6 we have \(\{N |\lambda _2|^2,\dots , N |\lambda _N|^2\}\overset{(\mathrm d)}{=}\{\gamma _2,\dots ,\gamma _N\}\), a set of independent Gamma random variables with corresponding parameters. Theorem 2.2 therefore yields, conditionally on \(\lambda _1=0\),
where all random variable are independent. Equation (2.6) then follows immediately from Lemma 2.5 below. This readily implies
The convergence in distribution follows from a simple change of variables: for any bounded test function f,
which clearly converges to the right hand side of (1.8) as \(N\rightarrow \infty \). \(\square \)
Lemma 2.5
The following equalities in distribution hold, where all random variables with different indexes are independent:
Proof
Equation (2.9) is standard, see e.g. [34, Chapter 25]. The equality (2.10) follows by immediate induction from the following property [17]: if \(\beta _{p,q}\) and \(\beta _{p+q,r}\) are independent, their product has distribution \(\beta _{p,q+r}\). \(\square \)
2.3 The annealed diagonal overlap in the bulk
With the following theorem, we first recall how the expectation of diagonal overlaps is accessible, following [12, 13, 43, 54]. We then prove Theorem 1.1.
The following was proved by Chalker and Mehlig for \(z=0\). They gave convincing arguments for any z in the bulk, a result then proved by Walters and Starr. Explicit formulae have also been recently obtained in [1] for the conditional expectation of diagonal and off-diagonal overlaps with respect to any number of eigenvalues. We include the following statement and its short proof for the sake of completeness.
Theorem 2.6
(From [12, 13, 43, 54]) For any \(z \in {\mathbb {D}}\), we have
Proof
From (2.4), we can write
Theorem 5.3 with \(g(\lambda ) =1+\frac{1}{ N|z - \lambda |^2}\) then gives
This is the determinant of a tridiagonal matrix, with entries [we use (5.1)]
Denoting \(x=N |z|^2\) and \(d_k=\det ((M_{ij})_{1\leqslant i,j\leqslant k})\), with the convention \(d_0=1\) we have
so that \(a_k=d_k N^{\frac{k(k+1)}{2}}\) satisfies \(a_0=1\), \(a_1=x+2\),
This gives \(a_k=(k+1)e^{(k+1)}(x)-x e^{(k)}(x)\) by an immediate induction. Thus, we conclude
From the asymptotics
the expected formula follows. \(\square \)
The following proofs make use of special integrability when the conditioned particle is at the center, together with a separation of the short-range and long-range eigenvalues. This separation of scales idea is already present in [13], though not rigorous. To illustrate the main ideas, we first give an alternative proof of Theorem 2.6 (with deteriorated error estimate) which does not rely on explicit formulas, but rather on rigidity and translation invariance. We then prove the main result of this section, Theorem 1.1.
Proof
(Alternative proof of Theorem 2.6) We denote \(\nu _z\) the measure (1.2) conditioned to \(\lambda _1=z\). Note that \(\nu _z\) is a determinantal measure as it has density \(\prod _{2\leqslant i<j\leqslant N}|\lambda _i-\lambda _j|^2e^{-\sum _2^NV(\lambda _i)}\) for some external potential V (which depends on z). With a slight abuse of language, we will abbreviate \({\mathbb {E}}_{\nu _z}(X)={\mathbb {E}}(X\mid \lambda _1=z)\) even for X a function of the overlaps.
The proof consists in three steps: we first show that we can afford a small cutoff of our test function around the singularity, then we decompose our product into smooth long-range and a short-range parts. The long range concentrates, and the short range is invariant by translation.
\({\textit{First step: small cutoff}}\) Let \(g_z(\lambda )=1+ \frac{1}{N|z-\lambda |^2}\). Remember that from (2.4), \( {\mathbb {E}}_{\nu _z}({\mathscr {O}}_{11}) = {\mathbb {E}}_{\nu _z}\left( \prod _{i=2}^{N} g_z(\lambda _i)\right) . \) We denote \(h_z(\lambda )=g_z(\lambda ) \mathbb {1}_{{|z-\lambda |}>N^{-A}}\), with \(A=A(\kappa )\) a large enough chosen constant, and first prove the following elementary equality:
Note that the exponent \(N^{-3}\) here is just chosen for the sake of concreteness. The left hand side coincides with \( {\mathbb {E}}_{N-1}\left( \prod _{i=2}^N(|z{-}\lambda _i|^2{+}N^{-1})\right) \left( {\mathbb {E}}_{N-1}\left( \prod _{i=2}^N(|z{-}\lambda _i|^2)\right) \right) ^{-1}, \) so that (2.12) follows if we can prove
for a constant B sufficiently larger than C. Equation (2.14) follows from Lemma 2.11. For equation (2.13), note that the left hand side has size order
by Cauchy–Schwarz inequality, union bound, and considering that A can be taken as large as needed. This last expectation is bounded by Lemma 2.11, which concludes the proof of (2.12) by choosing A large enough.
Second step: the long-range contribution concentrates We smoothly separate the short-range from a long-range contributions on the right hand side of (2.12). For this, we define:
Note that
To bound the first term on the right hand side, we rely on [10, Lemma 3.2]: for any \(\alpha \) such that \(\alpha \Vert f^{\ell }_z\Vert _\infty <1/3\) (in practice \(\Vert f^{\ell }_z\Vert _\infty <N^{-2\delta }\) so that we will choose \(\alpha =c N^{2\delta }\) for some fixed small c), we have
for some C which does not depend on N. We first bound the above variance. Introduce a partition of type \(1=\chi +\sum _{k\geqslant 1}\xi (2^{-k}x)\) for any \(x>0\), with \(\xi \) smooth, compactly supported. Let \(f^{\ell }_{z,k}(\lambda )=f^{\ell }_{z}(\lambda )\xi (2^{-k}N^ {1/2-\delta }|z-\lambda |)\) and \(K=\min \{k\geqslant 1:2^kN^{-1/2+\delta }>C\}\) where C and therefore K only depend on \(\xi \). Then \(\sum _if^{\ell }_z(\lambda _i)=\sum _{k=1}^{K}\sum _if^{\ell }_{z,k}(\lambda _i)\) with probability \(1-e^{-cN}\) (here we use that there are no eigenvalues \(|\lambda _k|>1+{\varepsilon }\) with probability \(1-e^{-c({\varepsilon })N}\), thanks to the Corollary 5.5). Moreover, from [9, Theorem 1.2], for any \({\varepsilon }>0\) and \(D>0\), there exists \(N_0>0\) such that for any \(N\geqslant N_0\), \(|z|<1\) and \(1\leqslant k\leqslant K\) we have
This implies the same estimate for the conditioned measure by a simple Cauchy–Schwarz inequality:
where we used Lemma 2.11. We conclude that for any \({\varepsilon }>0\) and D we have
so that \({\mathrm{Var}}_{\nu _z}(\sum _{i=2}^N f_z^\ell (\lambda _i))={{\,\mathrm{O}\,}}(N^{-4\delta +{\varepsilon }})\) and \({\mathbb {E}}_{\nu _z}(\sum _{i=2}^N f_z^\ell (\lambda _i))=\bar{f}_z+{{\,\mathrm{O}\,}}(N^{-2\delta +{\varepsilon }})\). As a consequence, (2.20) becomes
The same reasoning yields
The choice \(\alpha =\pm c N^{2\delta }\) in (2.21), (2.22) together with (2.19) implies
This yields, for some \(p=1+c(\kappa )\), \(c(\kappa )>0\) and some \(q,r>1\),
Here we used that the third term has size order \(e^{-c N^{{\varepsilon }}}\), the second one is of order \(e^{q{\bar{f}}_z}={{\,\mathrm{O}\,}}(N^{C})\) from (2.21), (2.22), and so is the first one from Lemma 2.9 (we needed the initial small cutoff changing g into h in order to apply this Lemma). Moreover,
and this last expectation is of order \(e^{-c N^{{\varepsilon }}}\) for the same reason as (2.23). To summarize, with the previous two equations we have proved (up to exponentially small error terms)
Third step: the local part is invariant. With \(p=1\) in (2.33), we have
This yields
From Lemma 2.8, \(e^{{\bar{f}}_z-{\bar{f}}_0}=1-|z|^2\), and from (2.7), (2.12) we have \({\mathbb {E}}_{\nu _0}\left( \prod _{i=2}^Nh_0 (\lambda _i)\right) =N+{{\,\mathrm{O}\,}}(N^{-2})\). This concludes the proof. \(\square \)
Proof of Theorem 1.1
We follow the same method as in the previous proof, except that we won’t need a small a priori cutoff: we are interested in convergence in distribution, not in \({\mathrm{L}}^1\).
First step: the long-range contribution concentrates We smoothly separate the short-range from a long-range contributions in Theorem 2.2:
For the convenience of the reader we recall the notations defined above :
Let \({\mathcal {G}}\) be the distribution of \((X_2,\dots ,X_N)\). For any \({\varepsilon }>0\), by Gaussian tail we have
Moreover,
To bound the first term on the right hand side, we first integrate over the Gaussian variables:
where, for the last inequality, we used (2.21) and (2.22) together with the Cauchy–Schwarz inequality, with \(\alpha =c N^{2\delta }\) for some fixed small enough c being admissible. With (2.25), we obtain
Let \(\xi \in {\mathbb {R}}\) be fixed. From the above bound we have
where we used \(e^{{\bar{f}}_z}/(1-|z|^2)=e^{{\bar{f}}_0}\), from Lemma 2.8.
We now define the function \(a_z\) (omitting the dependence in \(\xi \) in the notation) through
Note that \(a_z\) does not depend on i because the \(X_i\)’s are identically distributed. We want to apply Lemma 4.3. Note that \(a_z\) is supported on \(|z-\lambda |<C N^{-\frac{1}{2}+\delta }\) and \({{\,\mathrm{Re}\,}}(a_z)\leqslant 0\), so that (4.3) and (4.4) are automatically satisfied and \((N\Vert \nu \Vert _1)^r\leqslant C N^{2r\delta }\), hence (4.2) holds for the choice \(r=3\), \(\delta =\kappa /10\). For this choice of \(\delta \), we therefore have
uniformly in \(\xi \). This proves
Together with Proposition 2.4, this concludes the proof of Theorem 1.1. \(\square \)
Proof of Corollary 1.2
We start with the lower bound. From (2.26) we have
We now apply Lemma 2.10 to justify that z can essentially be replaced by \(z=0\) in the above left hand side. From (2.27), the Fourier transforms of \(g_z,g_0\) are exponentially close uniformly in the Fourier parameter \(\xi \). By choosing in Lemma 2.10\(R=T=e^{N^{{\varepsilon }/10}}\) and F smooth bounded equal to 1 on the interval \((-\infty ,(1-{\varepsilon }/2\log N-{\bar{f}}_0)]\), 0 on \([(1-{\varepsilon }/2\log N-\bar{f}_0+1,\infty )\), we have
where this last probability was estimated thanks to Proposition 2.4. For \(|z|<1-N^{-\frac{1}{2}+\kappa }\), this yields
and we conclude by a union bound (an error bound \({{\,\mathrm{o}\,}}(N^{-1})\) above would be enough). For the upper estimate, in the same way as previously, for any \(x>0\) we obtain
For \(x\gg N\), the following is easy to justify:
We obtained
which concludes the proof by a union bound. \(\square \)
Remark 2.7
One may wonder about the true asymptotics of the greatest overlap over the whole spectrum. The above bounds could easily be refined to prove that for any \(C>0\) and \( N\ll x\ll N^C\),
If the overlaps are sufficiently independent (a fact suggested by (1.14)), this hints towards convergence of the maximum to a Fréchet distribution: for any fixed \(y>0\), as \(N\rightarrow \infty \)
Remember that \(0 \leqslant \chi \leqslant 1\) is a smooth cut-off function on \({\mathbb {R}}_+\) such that \(\chi (x)=1\) for \(x<1/2\) and 0 for \(x>1\), and we denote
The following three lemmas were used in the previous proofs.
Lemma 2.8
There exists a constant \(c(\chi )\) such that for any \(|z|<1-N^{-\frac{1}{2}+\delta }\) we have
Proof
For \(z=0\), this is an elementary calculation in polar coordinates, so that we only need to show that for any given \(0<{\varepsilon }<1-|z|\) we have (here \({\mathbb {D}}_a\) is the disk with center a, radius \({\varepsilon }\))
Denote \(x+\mathrm {i}y=r e^{\mathrm {i}\theta }\) and \(a=|z|\). Note that \(\frac{1}{r^2}=\partial _x(x\frac{\log r}{r^2})+\partial _y(y\frac{\log r}{r^2})\), so that by Green’s theorem we have
The second integral clearly does not depend on a. The first integral can be split into
To calculate the first integral above, we expand \(\log |e^{\mathrm {i}\theta }-a|=\mathfrak {R}\sum _{p\geqslant 1}\frac{1}{p}(ae^{-\mathrm {i}\theta })^p\), \(|e^{\mathrm {i}\theta }-a|^{-2}=\sum _{k,\ell \geqslant 0}a^{k+\ell }e^{\mathrm {i}(k-\ell )\theta },\) and obtain
In the same way, we have
To summarize we have proved that
where c does not depend on z, and (2.31) follows. \(\square \)
Lemma 2.9
Let \(h_z^s\) be given by (2.18). For any \(\kappa \in (0,1/2)\), there exists \(c(\kappa ),\, C(\kappa )>0\) such that for any \(|z|<1-N^{-\frac{1}{2}+\kappa }\) and \(p\in [1,1+c(\kappa )]\), we have
Proof
First, the result is true for \(z=0\). Indeed,
where we used Corollary 5.6.
We want to apply Lemma 4.3 to conclude the proof. For this we need to check conditions (4.2), (4.3) and (4.4) for our function \(f=ph^s_z\), and \(\nu =e^f-1\). First note that \(\Vert \nu \Vert _1\leqslant C\int _{N^{-A}<|\lambda |<1}\frac{1}{(N|\lambda ^2|)^p}\leqslant C N^{-p}N^{A(2p-2)}\), so that (4.2) holds by choosing \(p=1+c(\kappa )\) with \(c(\kappa )\) small enough. To prove (4.3) and (4.4), we rely on Lemma 4.2:
where we used (4.2) for the first equation, and the calculation \( \int _{N^{-A}}^{\infty } \frac{1}{x^p}e^{-x}\leqslant N^C. \) We therefore obtained
for any \(1\leqslant p\leqslant 1+c(\kappa )\) with \(c(\kappa )\) small enough. Equations (2.32) and (2.33) conclude the proof. \(\square \)
To quantitatively invert the Fourier transform, we use the following crude bound, see [4, Lemma 2.6] following from [6, Corollary 11.5].
Lemma 2.10
There exists a constant c such that if \(\mu \) and \(\nu \) are probability measures on \({\mathbb {R}}\) with Fourier transforms \({\hat{\mu }}\left( {t}\right) =\int e^{\mathrm {i}{t}{x}}\mu \left( \mathrm{d}x\right) \) and \({\hat{\nu }}\left( {t}\right) =\int e^{\mathrm {i}{t}{x}}\nu \left( { d}x\right) \), then for any \(R,T>0\) and any function \(f:{\mathbb {R}}\rightarrow {\mathbb {R}}\) with Lipschitz constant C,
The following crude a priori estimates are used in this paper. Note that for z strictly in the bulk of the spectrum (\(|z|<1-{\varepsilon }\) for fixed \({\varepsilon }>0\)), the first statement is a simple consequence of the main result in [55].
Lemma 2.11
For any \(p,\kappa >0\), there exists \(C>0\) such that, for large enough N, uniformly in \(|z_1|, |z_2|<1-N^{-\frac{1}{2}+\kappa }\) we have
Proof
We start with the lower bounds, which are elementary: as \({\mathbb {E}}(e^X)>e^{{\mathbb {E}}(X)}\), we have
where for the last inequality we used that the density of states for the Ginibre ensemble is close to the uniform measure on the disk with high accuracy (see e.g. [9, Lemma 4.5]). This proves the lower bounds in (2.35) and the lower bounds for (2.36), (2.37) hold by the same argument.
For the upper bounds, we only need to prove (2.36), as (2.35) will follow by monotonicity, and (2.37) by the Cauchy–Schwarz inequality from (2.35). Remember the notation (2.15) and abbreviate \(\log _N(x)=\log (|x|^2+1/N)\). We can bound
For the first expectation corresponding to the short range, we apply Lemma 4.2, observing that \(N\Vert \nu \Vert _1={{\,\mathrm{O}\,}}(N^{2\delta })\) is negligible for \(\delta \) small enough. We obtain that this first expectation is equivalent to
where the above inequality follows from Corollary 5.5.
The second expectation in (2.38) is the Laplace transform of smooth linear statistics, so that the loop equations techniques apply to prove it is of polynomial order, see [40, Theorem 1.3]. More precisely, [40] applies to the smooth function \((1-\chi _{z_1,\delta })\log \) instead of \((1-\chi _{z_1,\delta })\log _N\), but we can decompose \(\log _N(\lambda )=2\log |\lambda |+\log (1+\frac{1}{N|\lambda ^2|})\). With the Cauchy–Schwarz inequality we separate contribution from these two functions, then the analogue (for the unconditioned measure) of (2.21) shows the Laplace transform of linear statistics of \((1-\chi _{z_1,\delta })\log (1+\frac{1}{N|\lambda ^2|})\) is \({{\,\mathrm{O}\,}}(1)\), and finally [40, Theorem 1.3] bounds the contribution of \((1-\chi _{z_1,\delta })\log |\lambda |\) by \({{\,\mathrm{O}\,}}(N^C)\). \(\square \)
3 Off-diagonal overlaps
In this section we consider the distribution of N Ginibre points conditioned to \(\{ \lambda _1 = z_1, \lambda _2 = z_2\}\). We will successively prove identities for the quenched off-diagonal overlaps, for all \(z_1,z_2\), and then get explicit relations for \(z_1=0\) in the annealed setting. Finally, these new correlation identities are extended to any \(z_1, z_2\) in the bulk of the spectrum by a decomposition of short and long range contributions.
3.1 The quenched off-diagonal overlap
Contrary to the diagonal overlap, the factorization here doesn’t involve independent variables.
Proposition 3.1
The following equality in law holds, conditionally on \(\{ \lambda _1, \dots , \lambda _N\}\):
where, conditionally on \( {\mathcal {F}}_{n-2}\), \(Z_n\) is a product of two (correlated) complex Gaussian random variables, and \({\mathbb {E}}(Z_{n}\mid {\mathcal {F}}_{n-2})=1\).
Proof
As for the diagonal overlap, we simply compute, with the notation (2.5),
and
where
Clearly, conditionally on \({\mathcal {F}}_{n-1}\), \(Z_{n+1}\) is a product of two complex Gaussian random variables, a distribution which depends on \({\mathscr {O}}_{12}^{(n)}\). Moreover, \(B_n,D_n\in {\mathcal {F}}_{n-1}\) and \(T_n\) is independent of \({\mathcal {F}}_{n-1}\), so that \({\mathbb {E}}(Z_{n+1}\mid {\mathcal {F}}_{n-1})=1\). \(\square \)
Remark 3.2
By successive conditional expectations with respect to \({\mathcal {F}}_{n-2},\dots ,{\mathcal {F}}_1\), Proposition 3.1 implies
an important fact already proved in [12, 13, 43].
3.2 The annealed off-diagonal overlap: expectation
Remarkably, the works [12, 13, 43] also explicitly integrated the random variable (3.2) over \(\lambda _3,\dots ,\lambda _N\), in the specific case \(\lambda _1=0\). We state the resulting asymptotics and add the proof from Chalker and Mehlig, for completeness.
Corollary 3.3
(Chalker, Mehlig [12, 13, 43]) For any \({\varepsilon }>0\), there exists \(c>0\) such that uniformly in \(|z|<N^{-{\varepsilon }}\),
Proof
From (3.2), we want to evaluate
From Theorem 5.4 with \(g (\lambda ) =1 - \frac{1}{ N\lambda ( \overline{z - \lambda })}\), we find that
where
This matrix is tridiagonal with entries
Let \(d_k=\det (f_{i,j})_{i,j = 1}^{k}\) and \(x= N|z|^2\). With the convention \(d_0=1\) we have
so that \(a_k=d_k N^{\frac{k(k+3)}{2}}\) satisfies \(a_0=1\), \(a_1=3/2+x/2\),
An immediate induction gives \(a_k=(k+2)x^{-2}e_2^{(k+2)}(x)\). Thus, we conclude
The proof is then concluded by standard asymptotics. \(\square \)
With Corollary 3.3, the expectation of \({\mathscr {O}}_{12}\) is known for \(\lambda _1=0\). To extend the result to anywhere in the bulk of the spectrum, we mimic the alternative proof of Theorem 2.6, from Sect. 2.3.
Proof of Theorem 1.3
We denote \(\nu _{z_1,z_2}\) the measure (1.2) conditioned to \(\lambda _1=z_1,\lambda _2=z_2\). Note that \(\nu _{z_1,z_2}\) is a determinantal measure. With a slight abuse of language, we will abbreviate \({\mathbb {E}}_{\nu _{z_1,z_2}}(X)={\mathbb {E}}(X\mid \lambda _1=z_1,\lambda _2=z_2)\) even for X a function of the overlaps.
We follow the same three steps as in the alternative proof of Theorem 2.6. Strictly speaking, if we were to impose \(|z_1-z_2|>N^{-C}\) for some fixed C, we would not need the first step below, as the singularity 1 / |z| is integrable, contrary to our previous singularity \(1/|z|^2\). However, in Theorem 1.3 we allow \(z_1\) and \(z_2\) to be arbitrarily close, so we first perform an initial small cutoff.
First step: small cutoff Let \(g_{z_1,z_2}(\lambda )= 1+\frac{1}{N(z_1-\lambda )(\overline{z_2-\lambda })}\). Remember that, from (3.2)
We denote \(h_{z_1,z_2}(\lambda )=g_{z_1,z_2}(\lambda ) \mathbb {1}_{\lambda \not \in {\mathscr {B}}}\) where \({\mathscr {B}}=\{|\lambda -z_1|<N^{-A}\}\cup \{|\lambda -z_2|<N^{-A}\}\). For A a large enough constant, the analogue of (2.12) holds:
Indeed, by making explicit the above conditional measures, (3.3) follows from
and
with B much larger than \(C_1\). Lemma 2.11 gives (3.5). The left hand side of (3.4) has size order
by the Cauchy–Schwarz inequality and Lemma 2.11, for some \(C_2\) which does not depend on A. This concludes the proof of (3.3) by choosing A large enough.
Second step: the long-range contribution concentrates We smoothly separate the short-range from a long-range contributions on the right hand side of (3.3):
and we denote \(z=(z_1+z_2)/2\), recall \(|z_1-z_2|<N^{-\frac{1}{2}+\kappa -{\varepsilon }}\), \(\chi _{z,\delta } (\lambda )= \chi \Big (N^{\frac{1}{2}-\delta }|z-\lambda | \Big )\), and choose \(\delta \in (\kappa -{\varepsilon }, \kappa )\). In the definition of \(h_{z_1,z_2}^{s}\), we can choose any branch for the logarithm, this won’t have any impact on the rest of the proof. In the long-range contribution \(h^{\ell }_{z_1,z_2}\), the logarithm is defined by continuity from \(\log (1)=0\). Let \(f^{\ell }_{z_1,z_2}(\lambda )= \frac{1}{N (z_1 - \lambda )\overline{(z_2-\lambda )}} (1 - \chi _{z,\delta } (\lambda ))\) and \(\bar{f}_{z_1,z_2}=\frac{N-2}{N}\frac{1}{ \pi } \int _{{\mathbb {D}}} {1 - \chi _{z,\delta } (\lambda ) \over (z_1-\lambda )\overline{(z_2-\lambda )}} \mathrm {d}m(\lambda )\). Note that
The last two sums are bounded as in (2.22). For the first term on the right hand side, we bound the real and imaginary parts separately: similarly to (2.20), we have
where \(\alpha =c N^{2\delta }\) for some fixed c and C which does not depend on N. We first bound the above variance. Remember we have a partition of type \(1=\chi +\sum _{k\geqslant 1}\xi (2^{-k}x)\) for any \(x>0\), with \(\xi \) smooth, compactly supported. Let \(f^{\ell }_{z_1,z_2,k}(\lambda )=f^{\ell }_{z_1,z_2}(\lambda )\xi (2^{-k}N^{1/2-\delta }|z-\lambda |)\) and \(K=\min \{k\geqslant 1:2^kN^{-1/2+\delta }>10\}\). Then \(\sum _if^{\ell }_z(\lambda _i)=\sum _{k=1}^{K}\sum _if^{\ell }_{z,k}(\lambda _i)\) with probability \(1-e^{-cN}\), and with [9, Theorem 1.2], for any \({\varepsilon }>0\) and \(D>0\), there exists \(N_0>0\) such that for any \(N\geqslant N_0\) and \(1\leqslant k\leqslant K\) we have (we now omit to write the real part, being understanding that f is either \(\mathrm{Re} f\) or \(\mathrm{Im} f\))
The same estimate holds for the conditioned measure by the Cauchy–Schwarz inequality:
for some C which only depends on \(\kappa \), where we used Lemma 2.11. We conclude that for any small \({\varepsilon }>0\) and D we have
so that \({\mathrm{Var}}_{\nu _{z_1,z_2}}(\sum _{i=3}^N f_{z_1,z_2}^\ell (\lambda _i))={{\,\mathrm{O}\,}}(N^{-4\delta +{\varepsilon }})\) and \({\mathbb {E}}_{\nu _{z_1,z_2}}(\sum _{i=3}^N f_{z_1,z_2}^\ell (\lambda _i))=\bar{f}_{z_1,z_2}+{{\,\mathrm{O}\,}}(N^{-2\delta +{\varepsilon }})\). As a consequence, (3.7) becomes
With \(\alpha =c N^{2\delta }\), we obtain
This yields, for some \(p=1+c(\kappa )\), \(c(\kappa )>0\) and some \(q,r>1\),
Here we used that the third term has size order \(e^{-c N^{{\varepsilon }}}\), the second one is of order \(e^{q{\bar{f}}_{z_1,z_2}}={{\,\mathrm{O}\,}}(N^{C})\), and so is the first one from Lemma 3.8 and \(|1+\frac{1}{(z_1-\lambda )\overline{(z_2-\lambda )}}|\leqslant (1+\frac{1}{N|z_1-\lambda |^2})(1+\frac{1}{N|z_2-\lambda |^2})\). Moreover,
and this last expectation is of order \(e^{-c N^{{\varepsilon }}}\) for the same reason as (3.9). To summarize, with the previous two equations we have proved (up to exponentially small additive error terms)
Third step: the local part is invariant. For our test function \(h^s_{z_1,z_2}\), the reader can easily check the conditions of Lemma 4.4, so that
This yields
From Lemma 3.4, \(e^{{\bar{f}}_{z_1,z_2}-\bar{f}_{0,z_2-z_1}}=(1-z_1\overline{z_2})^{\frac{N-2}{N}}\). Together with Corollary 3.3, this concludes the proof. \(\square \)
Lemma 3.4
For any \(\lambda _1, \lambda _2 \in {\mathbb {D}}\),
Proof
We consider the following domains, assuming \(0< |\lambda _1|< |\lambda _2| <1\) and \(\varepsilon >0\) is small enough. The following computation still holds if \(|\lambda _1| = |\lambda _2|\), as long as \(\lambda _1 \ne \lambda _2\). Integrability is clear, as the poles are simple and isolated. Moreover, under these conditions, the integral cancels on the disks \(D_1\) and \(D_2\).
The same type of expansion shows that the integral over \(R_2\) vanishes and the contribution from \(R_4\) is
As the expression is integrable, on the domains \(R_1, R_3\) there is no contribution as \(\varepsilon \rightarrow 0\). Summing (3.11) and (3.12) gives the result. \(\square \)
3.3 The quenched off-diagonal overlap: second moments
The main result of this subsection is the following lemma, which gives the expectation of second moments of overlaps conditionally on the eigenvalues positions. For this, we define
Lemma 3.5
For any \(2\leqslant n\leqslant N-1\) we have
In particular,
Proof
We recall the notation (3.1) and the property \({\mathbb {E}}(Z_{n+1}\mid {\mathcal {F}}_{n-1})=1\). A short calculation also gives \( {\mathbb {E}}(|Z_{n+1}|^2\mid {\mathcal {F}}_{n-1})=1+\frac{\Vert B_n\Vert ^2\Vert D_n\Vert ^2}{|B_n^{\mathrm{t}} \bar{D}_n|^2}\). Abbreviating \(\gamma _k=\gamma _{k,n+1}\), this gives
We now denote \(X=X_{n+1}= \sqrt{N} {B_n^{\mathrm{t}}T_n \over \Vert B_n \Vert }\), \(Y=Y_{n+1}= \sqrt{N}{D_n^{\mathrm{t}}T_n \over \Vert D_n \Vert }\), so that \({\mathbb {E}}(|X_{n+1}|^2 \mid {\mathcal {F}}_{n-1})={\mathbb {E}}(|Y_{n+1}|^2 \mid {\mathcal {F}}_{n-1})=1\) and \( {\mathbb {E}}(|X_{n+1}Y_{n+1}|^2 \mid {\mathcal {F}}_{n-1})=1+\frac{|B_n^{\mathrm{t}} \bar{D}_n|^2}{\Vert B_n\Vert ^2\Vert D_n\Vert ^2} \). This yields
Equations (3.16) and (3.17) together conclude the proof of (3.14). Denoting \(Y_n={\mathbb {E}}_{T}(X_n\mid {\mathcal {F}}_1)\), we obtain
and (3.15) immediately follows. \(\square \)
3.4 The annealed off-diagonal overlap: second moments for \(\lambda _1=0\)
We now want to integrate (3.15) over the eigenvalues \(\lambda _3,\dots ,\lambda _N\), first in the special case \(\lambda _1=0\). This requires some new notations. We abbreviate
and will often use the property
We also define the following rational fractions of x
We can now state the following main proposition on which Theorem 1.4 depends. The reason why such formulas exist relies on two algebraic facts.
- (i)
The \(A_n\)’s from (3.13) are diagonalizable in the same basis, see (3.20). Their commutation was expected, our choice of eigenvalues ordering being arbitrary, while the left hand side of (3.15) is intrinsic.
- (ii)
More surprisingly, the obtained holonomic sequence (3.23) is exactly solvable.
Proposition 3.6
Conditionally on \(\{\lambda _1=0,\lambda _2=z\}\), we have
Proof
Importantly, the matrices \(A_n, 3\leqslant n\leqslant N\), can be diagonalized in the same basis: from (3.13) an elementary calculation based on (2.1) gives
so that its eigenvectors clearly do not depend on n. With these notations, the eigenvalues of \(A_n\) are
and the orthogonal basis
diagonalizes all \(A_n\)’s simultaneously: \(UA_nU^{t}={{\,\mathrm{diag}\,}}(\lambda _+(n),\lambda _-(n))\). Then (3.15), together with the codiagonalization of the \(A_n\)’s, \({\mathbb {E}}|T|^2=1\) and \({\mathbb {E}}(|T|^4)=2\), yields the following simple expression:
Note that we have not yet used \(z_1=0\): the above formula holds for any given \(z_1,z_2\).
The remarkable fact is that in the specific case \(z_1=0\), \({\mathbb {E}}(d_+)\) and \({\mathbb {E}}(d_-)\) can be calculated, as shown below (here \({\mathbb {E}}\) denotes integration over all variables except \(\lambda _1,\lambda _2\)). We start with \({\mathbb {E}}(d_+)\). From Theorem 5.4, the following representation holds:
where \(Z_N^{(0,z)}\) is given by (5.2) and
We expand \(|\lambda -z|^2 = |\lambda |^2 + |z|^2 - {\bar{z}} \lambda - z {\bar{\lambda }}\). The resulting matrix is thus tridiagonal with
With the notation \(d_k=\det (f_{ij})_{1\leqslant i,j\leqslant k}\), the recurrence \(d_{k} = f_{k,k} d_{k-1} - f_{k,k-1} f_{k-1,k} d_{k-2}\) holds, so that defining \(a_k=d_k N^{\frac{k(k+3)}{2}}\) we have
with the convention \(a_0=1\) and \( a_1=2+\frac{\delta }{2}+\frac{1-a^{-1}}{2}. \) Note that for \(z=0\) we have \(\delta =0\) and \(a=1\), hence
As a consequence, \(g_k = {a_k \over a_{k}^{0} }\) satisfies \(g_0=1\), \(g_1=1+\frac{\delta }{4}+\frac{1-a^{-1}}{4}\), and
We cannot see an a priori reason why this equation could be solved, but it can be. We remark that the function \(u_k=u_k(a)\) from (3.19) also satisfies the induction
The reader who would like to check the above equation can substitute \(\delta =a-\frac{1}{a}\) and verify that the Laurent series in a on both sides of (3.24) coincide. This equation implies
Subtracting (3.25) from (3.23) gives
which yields
Together with \({g_k \over u_k} = {g_{k-1} \over u_{k-1}}+ {\delta _k \over u_k}\), from (3.26) we obtain
A calculation gives \(u_0\delta _1=\frac{1}{6}(a-a^{-1})(1+a^{-1})\), so that
Note that
which simplifies (3.27) into
where we used \(\delta =a-a^{-1}\) and (3.19) at several steps. The above formula can be written in terms of (1.18) and further simplified as
By definition of \(g_k\), \(a_k = g_k a_{k}^{0}\), that is
Then, using the normalizing constant (5.2), we obtain
as \(\delta =N|z|^2\). We find
which has been defined as \(d_{N-2}(a, \delta )\). The formula for \({\mathbb {E}}(d_-)\) is obtained in the exact same way, with the only difference that a is replaced by b. Finally, conditionally on \(\{\lambda _1=0,\lambda _2=z\}\), (3.21) gives
We can replace \({\mathbb {E}}(d_+)\) and \({\mathbb {E}}(d_-)\) by their exact expressions to obtain the claimed formula. \(\square \)
Proposition 3.7
Let \(\sigma \in (0,1/2)\). Denoting \(\omega =\sqrt{N}|z|\), uniformly in \(|z|\in [0,N^{-\frac{1}{2}+\sigma }]\) we have
Proof
We consider asymptotics in Proposition 3.6. First, the term \(\frac{\delta ^{N-1}}{(N+1)!}\) is obviously negligible. Second, we always have \(a\geqslant 1\) and \(|b|\geqslant c/\delta \geqslant N^{-2\sigma }\), so that \(u_{N-2}(a)= 1 + {{\,\mathrm{O}\,}}(N^{-1})\), \(u_{N-2}(b) = 1 + {{\,\mathrm{O}\,}}(N^{2\sigma -1})\). Moreover,
From (3.18), we have \( \frac{a^2}{(1+a)^2 (1-a)^2} = \frac{b^2}{(1+b)^2 (1-b)^2} = \frac{1}{\delta ^2},\) so that (3.28) and its analogue for \(d_-\) give
We observe that
and the previous three equations give
The similar computation for \({\mathbb {E}}({\mathscr {O}}_{11} {\mathscr {O}}_{22}\mid \lambda _1=0,\lambda _2=z)\) involves the two terms
Moreover, some algebra gives
Once combined, these four equations yield
which concludes the proof. \(\square \)
3.5 The annealed off-diagonal overlap: second moments in the general case
We can now prove Theorem 1.4. We closely follow the method developed first in our alternative proof of Theorem 2.6, in Sect. 2.3, then in our proof of Theorem 1.3 in Sect. 3.2. In particular, following the proof of Theorem 1.3, we denote \(\nu _{z_1,z_2}\) the measure (1.2) conditioned to \(\lambda _1=z_1,\lambda _2=z_2\).
Proof of Theorem 1.4
Remember the notation (3.22). Assume we can prove (under the hypothesis of Theorem 1.4) that
From (3.31) and (3.32), a calculation gives \({\mathbb {E}}_{\nu _{0,z_2-z_1}}(d_+)>0\), \({\mathbb {E}}_{\nu _{0,z_2-z_1}}(d_-)>0\), so that (3.33), (3.34) together with (3.21) give
Together with Proposition 3.7, this concludes the proof. We therefore only need to show (3.33). The proof for (3.34) is identical up to trivial adjustments.
First step: small cutoff Our test function of interest and its short-range cut version are
where \({\mathscr {B}}=\{|\lambda -z_1|<N^{-A}\}\cup \{|\lambda -z_2|<N^{-A}\}\). We first prove (3.3) for our new definition of \(g_{z_1,z_2},\, h_{z_1,z_2}\). It follows from
and (3.5), for some C which does not depend on A. Equation (3.38) holds as the left hand side is bounded by
where we used the Cauchy–Schwarz inequality and
Indeed, after rescaling and shifting and introducing z so that \(|z|^2=\delta \), the above bound follows from
We used \((1+|\lambda |^2)(1+|z-\lambda |^2)\geqslant \delta \), as proved by a simple optimization. The last expectation in (3.39) has size order at most \(N^{C}\) from Lemma 2.11, which concludes the proof of our initial short-range cutoff by choosing A large enough.
Second step: the long-range contribution concentrates We smoothly separate the short-range from a long-range contributions in (3.36):
and, as earlier in this article, we denote \(z=(z_1+z_2)/2\), recall \(|z_1-z_2|<N^{-\frac{1}{2}+\sigma }\), \(\chi _{z,\delta } (\lambda )= \chi \Big (N^{\frac{1}{2}-\delta }|z-\lambda | \Big )\), and choose \(\delta \in (\sigma , \kappa )\). Note that our notation \(\delta \) in this step of the proof is unrelated to \(\delta = N | \lambda _{1} - \lambda _{2} |^2\) in the previous step. We define \(f^{\ell }_{z_1,z_2}(\lambda )= (\frac{1}{N |z_1 - \lambda |^2}+\frac{1}{N |z_1 - \lambda |^2})(1 - \chi _{z,\delta } (\lambda ))\) and \({\bar{f}}_{z_1,z_2}=(N-2)\frac{1}{ \pi } \int _{{\mathbb {D}}} f^{\ell }_{z_1,z_2}(\lambda ) \mathrm {d}m(\lambda )\). Note that
where we used \(a\leqslant N^{2\sigma }\).
With the exact same reasoning as from (2.20) to (2.21) (with \(\nu _z\) replaced by \(\nu _{z_1,z_2}\)), we obtain that \(\sum _{i=3}^Nf^\ell _{z_1,z_2}(\lambda _i)-\bar{f}_{z_1,z_2}\) is exponentially concentrated on scale \(N^{-2\delta +{\varepsilon }}\). Moreover, similarly, to (2.22), we now have
With \(\alpha =c N^{2\delta }\), we therefore obtain
This yields, for some \(p=1+c(\kappa )\), \(c(\kappa )>0\) and some \(q,r>1\),
Here we used that the third term has size order \(e^{-c N^{{\varepsilon }}}\), the second one is of order \(e^{q{\bar{f}}_{z_1,z_2}}={{\,\mathrm{O}\,}}(N^{C})\), and so is the first one from Lemma 3.8. Moreover,
Following similar arguments as (3.9), (3.10), still relying on Lemma 3.8, we finally obtain
Third step: the local part is invariant. For our test function \(h^s_{z_1,z_2}\), the reader can easily check the conditions of Lemma 4.4: the only new ingredient is
so that in this setting, the existence of \(r>2\), \(d<\kappa \) such that \( (N\Vert \nu \Vert _1)^r\leqslant N^{d} \) means that there exists \({\varepsilon }>0\) such that \( \left( \frac{\log N}{N|z_1-z_2|^2}\right) ^{2+{\varepsilon }}\leqslant N^{2\kappa -{\varepsilon }}, \) i.e. \( |z_1-z_2|\geqslant N^{-\frac{1}{2}-\frac{\kappa }{2}+{\varepsilon }}, \) as we assumed by hypothesis. This gives
This yields
From Lemma 2.8, \(e^{{\bar{f}}_{z_1,z_2}-\bar{f}_{0,z_2-z_1}}=(1-|z_1|^2)(1-|z_2|^2)\). With Proposition 3.7, this concludes the proof. \(\square \)
Lemma 3.8
For any \(0<\sigma<d<\kappa <1/2\), there exists \(p>1\), \(C>0\) such that for any \(|z_1|<1-N^{-1/2+\kappa },|z_1-z_2|<N^{-\frac{1}{2}+\sigma }\), \(z=(z_1+z_2)/2\), we have
Proof
The above left hand side is at most
With Lemma 2.11 and the Cauchy–Schwarz inequality, we therefore just need to prove that
We can apply Lemma 4.2 (for p small enough we have \(N\Vert \nu \Vert _1={{\,\mathrm{O}\,}}(1)\)) to compare it to the case \(z_1=0\), which is easily shown to be \({{\,\mathrm{O}\,}}(N^C)\) by Corollary 5.6. \(\square \)
3.6 Proof of Corollary 1.5
Following a notation from [52], let \(\kappa (\lambda _j)={\mathscr {O}}_{jj}^{1/2}\) be the condition number associated to \(\lambda _j\). As the spectrum of a G is almost surely simple, from [52, Equation (52.11)] we know that
as \(z\rightarrow \lambda _j\). Together with \(\sigma _{\varepsilon }(G)=\left\{ z: \Vert z-G\Vert ^{-1}>{\varepsilon }^{-1}\right\} \), this gives the following almost sure asymptotics:
Denoting \(c_N=N^2\int _{{\mathscr {B}}_N} (1-|z|^2)\frac{\mathrm{d}m(z)}{\pi }\), we therefore have
From Corollary 1.2,
hence we only need to prove
We proceed by bounding the second moment
To bound the first term, note that
Following the same reasoning as (2.29), we have
where we used Proposition 2.4. Denoting a the center of \({\mathscr {B}}_N\), we therefore bounded the right hand side of (3.42) by \(N^{{3+3{\varepsilon }}} (1-|a|^2)^2 m({\mathscr {B}}_N)\ll c_N^2\) because \(m({\mathscr {B}}_N)\geqslant N^{-1+2a}\) and \({\varepsilon }\) is arbitrary.
To bound (3.43), we first consider close eigenvalues and bound \({\mathscr {O}}_{ii}{\mathscr {O}}_{jj}\leqslant \frac{1}{2}({\mathscr {O}}_{ii}^2+{\mathscr {O}}_{jj}^2)\):
In the second estimate, we used the local law for Ginibre matrices: from [9, Theorem 4.1] the above number of close eigenvalues is at most \(C N^{2{\varepsilon }}\) for some large C, with probability at least \(1-N^{-D}\) for arbitrarily large D. The third estimate was obtained in the same way we bounded (3.42).
For eigenvalues at mesoscopic distance in \([N^{-\frac{1}{2}+{\varepsilon }},N^{-\frac{1}{2}+\kappa -a}]\), the contribution of (3.43) is obtained thanks to (1.12):
Finally, the line (3.44) is easily shown to be of order \(-c_N^2(1+N^{-\varepsilon })\) thanks to (1.5) and (2.30). We conclude that the left hand side of (3.42) is at most \(N^{-{\varepsilon }}c_N^2\), which concludes the proof of (3.41) by Markov’s inequality.
4 Translation invariance for conditioned measures
Recall that the Ginibre kernel is
We also denote its bulk limit as \(k_N(z,w)=\frac{N}{\pi }e^{-N(\frac{|z|^2}{2}+\frac{|w|^2}{2}-z\bar{w})}\).
Lemma 4.1
Let \(\kappa >0\). There exists \(c=c(\kappa )>0\) such that for any \(|z|,|w|\leqslant 1-N^{-\frac{1}{2}+\kappa }\), we have
Proof
This is a straightforward adaptation of the proof of [9, Lemma 4.2]. \(\square \)
We denote \({\mathscr {B}}_{a,\delta }\) the ball with center a and radius \(N^{-\frac{1}{2}+\delta }\).
Lemma 4.2
Let \(0<\delta<\kappa <1/2\) be fixed constants. Consider any \({\mathbb {C}}\)-valued measurable function f supported on \({\mathscr {B}}_{0,\delta }\), \(|a|\leqslant 1-N^{-\frac{1}{2}+\kappa }\), and \(\nu (z)=e^{f(z)}-1\). For any \(r>2\) there exists \(c, C>0\) depending only on \(\kappa ,\delta ,r\) such that
Proof
Let \(K_N^{(a)}(z,w)=K_N(z-a,w-a)\). We define \(\Vert K\Vert =\sup _{z,w\in \mathrm{supp}(\nu )}|K(z,w)|\). We successively compare linear statistics for \(K_N^{(a)},k_N^{(a)},k_N \) and \(K_N\). First note that \(k_N\) is the kernel of a translation invariant point process, so that comparison between \(k_N^{(a)}\) and \(k_N\) is trivial. For the other steps, we use [2, Lemma 3.4.5] and obtain
Clearly, \(\Vert K_N\Vert \leqslant \frac{N}{\pi }\) and we bound \(\Vert K_N^{(a)}-K_N\Vert \) with Lemma 4.1. We conclude that for a universal large enough C, and \(1/r+1/s=1\), we have
where in the last inequality we used Hölder’s inequality and \(r>2\). \(\square \)
Lemma 4.3
Remember \(\nu (z)=e^{f(z)}-1\) and \({\mathscr {B}}_{a,\delta }\) is the ball with center a and radius \(N^{-\frac{1}{2}+\delta }\). Let \(0<\delta<\kappa <1/2\) be fixed constants. Consider any \({\mathbb {C}}\)-valued measurable function f supported on \({\mathscr {B}}_{0,\delta }\), and \(|a|\leqslant 1-N^{-\frac{1}{2}+\kappa }\). Assume also that either (i) or (ii) below holds:
- (i)
\(\mathrm{Re}(f)=0\);
- (ii)
there exist \(d>0\), \(p>1\) and \(r>2\) such that \(rd<2\kappa \), \(f=0\) on \(|z|<e^{-N^{d}}\), and (\(f_+=\max (\mathrm{Re} f,0)\))
$$\begin{aligned}&(N\Vert \nu \Vert _1)^r\leqslant N^{d}, \end{aligned}$$(4.2)$$\begin{aligned}&\log \Vert e^{\sum _{i=2}^Nf_+(\lambda _i-a)}\Vert _{{\mathrm{L}}^p} \leqslant N^d, \end{aligned}$$(4.3)$$\begin{aligned}&\log \Vert e^{\sum _{i=2}^Nf_+(\lambda _i)}\Vert _{{\mathrm{L}}^p} \leqslant N^d. \end{aligned}$$(4.4)where the \(L^p\) norm is taken with respect to \({\mathbb {E}}_{N-1}\).
Then for any \(q<2\kappa \), uniformly in f satisfying the above hypothese, we have
Proof
In this proof we first consider the most difficult case (ii), and we will finally mention the simple modifications required for (i). We start with
Fix some constants \(\kappa _1, \kappa _2\) such that \(d<q<\kappa _2<\kappa _1<2\kappa \) and define \(\chi ^a_j(\lambda )=\mathbb {1}_{|\lambda -a|<e^{-N^{\kappa _j}}}\), \(j=1, 2\). We first show we can afford imposing \(\chi ^a_2(\lambda _i)=0\): for some positive q, r such that \(p^{-1}+q^{-1}+r^{-1}=1\), we have
where \({\mathrm{L}}^p={\mathrm{L}}^p({\mathbb {P}}_{N-1})\). By hypotheses (4.3) and (4.4), the first norm is at most \(e^{c N^d}\). The second is at most \(N{\mathbb {P}}_{N-1}(|\lambda _2|\leqslant e^{-N^{\kappa _2}})\leqslant e^{-cN^{\kappa _2}}\). The third norm is at most \(N^{C}\), as a simple consequence of Lemma 2.11. These estimates also hold for \(f=0\), so that we proved
If \(|\lambda _1-a|<e^{-N^{\kappa _1}}\) and \(|\lambda _i-a|>e^{-N^{\kappa _2}}\), \(2\leqslant i\leqslant N\), we have
The expectation in the numerator of (4.7) is therefore (in the first equation below \(\lambda _1\) has distribution U, the uniform measure on the unit disk with center a and radius \(e^{-N^{\kappa _1}}\), with volume \(b_N=\pi (e^{-2N^{\kappa _1}})\)):
We now want to remove the constraint on \((\lambda _i)_{i=2}^N\), i.e. prove
This requires a longer argument. Let \({\mathscr {B}}^a_{i}=\{|z-a|\leqslant e^{-N^{\kappa _i}}\}\), \(i=1\) or 2, \(\xi =\sum _{1}^N\delta _{\lambda _i}\), \(\widetilde{\xi }=\sum _{2}^N\delta _{\lambda _i}\). Then,
To bound the first term, we use the negative association property of determinantal point processes for disjoint sets (see e.g. [42]), using \(f_+\geqslant 0\) and \(f=0\) on \({\mathscr {B}}^a_{2}\):
By (4.3) and (4.4), the first expectation above has size order at most \(e^{cN^{d}}\). The second is of order \(e^{-cN^{\kappa _2}}\) and the third one is bounded by \(Nb_N(1+{{\,\mathrm{o}\,}}(1))\), so that the first term in (4.10) gives an error \({{\,\mathrm{O}\,}}(b_N e^{-c N^{\kappa _2}})\).
For the second term in (4.10), we also use the negative association property and \(f=0\) on \({\mathscr {B}}^a_{2}\):
Together with
we have proved that the second term in (4.10) gives an error \({{\,\mathrm{O}\,}}(b_N e^{-c N^{\kappa _2}})\). This concludes the proof of (4.9), so that the numerator in (4.7) is
where we used \(\frac{e^N Z_N}{Z_{N-1}}\sim c_1 N^{c_2}\) for some \(c_1,c_2\), as obtained from (1.2). In the same way, the denominator in (4.7) is \( \frac{e^N Z_N}{Z_{N-1}} \left( 1+{{\,\mathrm{O}\,}}\left( e^{-cN^{\kappa _2}}\right) \right) \), so that we obtained
where we successively used that fact that f vanishes on \({\mathscr {B}}_1^a\) and symmetrized.
To conclude the proof, we therefore just need
From Lemma 4.2 we know that uniformly on \(|w|<1\) we have
which proves (4.14) by Cauchy’s theorem, and therefore the lemma in case (ii).
Under the assumption (i), up to (4.13) the results hold and the reasoning is simplified as all \(L^p\) norms related to f can be bounded by 1. To justify an analogue of (4.13) and the end of the reasoning, we first replace f by \(\widetilde{f}=f\mathbb {1}_{({\mathscr {B}}_1^a)^c}\) and note that
so that by symmetrizing we now obtain
The rest of the proof is identical to case (i). \(\square \)
We now state and prove an analogue of Lemma 4.3 when conditioning on two points. We will only need case (ii), as we are interested in expectations in Sect. 3, not in convergence in distribution.
Lemma 4.4
Let \(0<\delta<\kappa <1/2\) be fixed constants and \(C>0\) fixed, arbitrarily large. Consider any \({\mathbb {C}}\)-valued measurable function f supported on \({\mathscr {B}}_{0,\delta }\), \(|a|,|b|\leqslant 1-N^{-\frac{1}{2}+\kappa }\), and \(N^{-C}<|b-a|<N^{-1/2+d}\). Assume that there exists \(d<2\kappa \), \(p>1\) and \(r>2\) such that \(f=0\) on \(|z|<e^{-N^{d}}\), on \(|z-(b-a)|<e^{-N^{d}}\), and (4.2), (4.3) and (4.4) hold. Then for any \(q<2\kappa \) we have
Proof
We start similarly to the proof of Lemma 4.3, by writing
Again, we fix some constants \(\kappa _1, \kappa _2\) such that \(d<\kappa _2<\kappa _1<2\kappa \) and define \(\chi ^x_j(\lambda )=\mathbb {1}_{|\lambda -a|<e^{-N^{\kappa _j}}}\), \(j=1, 2\), \(x=a,b\). The strict analogue of (4.6) holds, so that the left hand side of (4.15) can be written
The analogue of (4.8) then holds exactly in the same way: the expectation in the numerator of (4.16) is
Again, we want to remove the constraint on \((\lambda _i)_{i=3}^N\), i.e. prove
With the negative association property, the strict analogues of Eqs. (4.10), (4.11) and (4.12) hold, so that the numerator in (4.16) is
where we used \(\frac{e^{2N} Z_N}{Z_{N-2}}\sim c_1 N^{c_2}\) for some \(c_1,c_2\), as obtained from (1.2). In the same way, the denominator in (4.16) is \( \frac{e^{2N} Z_N}{|a-b|^2Z_{N-2}} \left( 1+{{\,\mathrm{O}\,}}\left( e^{-cN^{\kappa _2}}\right) \right) \), giving
where we successively used that fact that f vanishes on \({\mathscr {B}}_1^a\cup {\mathscr {B}}_1^b\), \({\mathscr {B}}_1^a\cap {\mathscr {B}}_1^b=\varnothing \) (this holds because \(|a-b|>N^{-C}\)) and symmetrized. To conclude the proof, we therefore just need \(\partial _{z_1z_2}f_{a,b}(0,0)=\partial _{z_1z_2} f_{0,b-a}(0,0)+{{\,\mathrm{O}\,}}(e^{-cN^{{q}}})\), where
This follows from Lemma 4.2 and Cauchy’s Theorem, similarly to the end of the proof of Lemma 4.3. \(\square \)
5 Andréief’s identity and Kostlan’s theorem
This section gives applications of Andréief’s identity to the conditioned measures of interest in this work. In particular, it proves some slight extensions of Kostlan’s theorem (Corollary 5.5), following a method from [18]. The common main tool will be the following classical Lemma, by Andréief [3] (see [16] for a short proof). Note that the original proof of Kostlan’s theorem [38] and some of its extensions [32] were based on different arguments.
Lemma 5.1
(Andréief’s identity) On a measured space \((E, {\mathcal {E}}, \mu )\) For any functions \((\phi _i, \psi _i)_{i=1}^N \in L_2(\mu )^{2N}\),
Theorem 5.2
Let \(E={\mathbb {C}}\), \(g \in L_2(\mu )\), and \(\{ \lambda _1, \dots , \lambda _N\}\) eigenvalues from the Ginibre ensemble. Then
Proof
The following is elementary:
The proof then follows from Andréief’s identity. \(\square \)
Theorem 5.3
We have (remember \(\mu =\mu ^{(N)}\))
and
Proof
Using Andréief’s identity with \(\phi _i(\lambda ) = \lambda ^{i-1} g(\lambda ) |z - \lambda |^2 \), \(\psi _j(\lambda )= \lambda ^{j-1}\), we find
where
By expanding \(|z- \lambda |^2 = |z|^2 + |\lambda |^2 - z {\overline{\lambda }} - {\overline{z}} \lambda \), we see that that M is tridiagonal, with entries (remember (5.1))
Denoting \(x=N |z|^2\) and \(d_k=\det ((M_{ij})_{1\leqslant i,j\leqslant k})\), with the convention \(d_0=1\) we have
so that \(a_k=d_k N^{\frac{k(k+1)}{2}}\) satisfies \(a_0=1\), \(a_1=1+x\),
This gives \(a_k=e^{(k)}(x)\) by an immediate induction. \(\square \)
Theorem 5.4
We have
and
Proof
By Andréief’s identity, the result holds with
Expanding \(|z - \lambda |^2 = |z|^2 + |\lambda |^2 - z {\overline{\lambda }} - {\overline{z}} \lambda \), we see that M is tridiagonal with entries
Denoting \(x=N |z|^2\) and \(d_k=\det ((M_{ij})_{1\leqslant i,j\leqslant k})\), with the convention \(d_0=1\) we have
so that \(a_k=d_k N^{\frac{k(k+3)}{2}}\) satisfies \(a_0=1\), \(a_1=1+x/2\),
This gives the expected result by an immediate induction. \(\square \)
Kostlan’s theorem now comes as a corollary, as well as a similar property for the Ginibre ensemble conditioned on \(\lambda _1 =0\).
Corollary 5.5
(Kostlan) The set \(N\{ |\lambda _1|^2, \dots , |\lambda _N|^2 \} \) is distributed as \(\{\gamma _1,\dots ,\gamma _N\}\), a set of (unordered) independent Gamma variables of parameters \(1,2, \dots , N\).
Proof
Let \(g \in {\mathbb {C}}[X]\) and use Theorem 5.2 with the radially symmetric function \(g(|\cdot |^2)\). The relevant matrix is then diagonal, with coefficients
In other words,
Note that these statistics characterize the distribution of a set of unordered points, as such expressions with polynomial g generate all symmetric polynomials, as stated in Lemma 5.7, and the gamma distributions are characterized by their moments. For more details, see [18]. We conclude that \( N\{ |\lambda _1|^2, \dots , |\lambda _N|^2 \} {\mathop {=}\limits ^{d}} \{ \gamma _1, \dots , \gamma _N\}\). \(\square \)
Corollary 5.6
Conditioned on \(\{ \lambda _1 =0\}\), \( \{ N |\lambda _2|^2, \dots , N |\lambda _N|^2 \} \) is distributed as \(\{\gamma _2,\dots \gamma _N\}\), a set of (unordered) independent Gamma variables of parameters \(2,3, \dots , N\).
Proof
Similarly to the proof of Corollary 5.5, we take \(g \in {\mathbb {C}}[X]\) and the radially symmetric function \(g(|\cdot |^2)\). In Theorem 5.3, we have
This together with our expression for \(Z_N^{(z=0)}\) in Theorem 5.3 yields
and we conclude in the same way that \( N\{ |\lambda _2|^2, \dots , |\lambda _N|^2 \} {\mathop {=}\limits ^{d}} \{ \gamma _2, \dots , \gamma _N\} \). \(\square \)
For the proof of the following lemma, we refer to [18]. We define the product symmetric polynomials as the symmetric polynomials given by products of polynomials in one variable:
Lemma 5.7
\(\mathrm {PS}_{{\mathbb {C}}}(N)\) spans the vector space of symmetric polynomials of N variables.
Notes
More precisely, for any smooth, bounded, compactly supported function f and deterministic sequence \((z_N)\) such that \(|z_N|<1-N^{-\frac{1}{2}+\kappa }\) we have \({\mathbb {E}}\left( f({\mathscr {O}}_{11}/(N(1-|z_N|^2)))\mid \lambda _1=z_N\right) \rightarrow {\mathbb {E}}f(\gamma _2^{-1})\).
References
Akemann, G., Tribe, R., Tsareas, A., Zaboronski, O.: On the determinantal structure of conditional overlaps for the complex Ginibre ensemble (2019). arXiv:1903.09016
Anderson, G.W., Guionnet, A., Zeitouni, O.: An Introduction to Random Matrices. Cambridge Studies in Advanced Mathematics, vol. 118. Cambridge University Press, Cambridge (2010)
Andréief, M.C.: Note sur une relation entre les intégrales définies des produits des fonctions. Mémoires de la société des sciences physiques et naturelles de Bordeaux 2, 1–14 (1883)
Arguin, L.-P., Belius, D., Bourgade, P.: Maximum of the characteristic polynomial of random unitary matrices. Commun. Math. Phys. 349, 703–751 (2017)
Belinschi, S., Nowak, M.A., Speicher, R., Tarnowski, W.: Squared eigenvalue condition numbers and eigenvector correlations from the single ring theorem. J. Phys. A 50(10), 105204, 11 (2017)
Bhattacharya, R.N., Ranga Rao, R.: Normal approximation and asymptotic expansions. Wiley Series in Probability and Mathematical Statistics. Wiley, New York (1976)
Bolley, F., Chafaï, D., Fontbona, J.: Dynamics of a planar Coulomb gas. Ann. Appl. Probab. 28(5), 3152–3183 (2018)
Bourgade, P., Yau, H.-T.: The eigenvector moment ow and local quantum unique ergodicity. Commun. Math. Phys. 350(1), 231–278 (2017)
Bourgade, P., Yau, H.-T., Yin, J.: The local circular law II: the edge case. Probab. Theory Rel. Fields 159(3–4), 619–660 (2014)
Breuer, J., Duits, M.: The Nevai condition and a local law of large numbers for orthogonal polynomial ensembles. Adv. Math. 265, 441–484 (2014)
Burda, Z., Grela, J., Nowak, M.A., Tarnowski, W., Warchoł, P.: Dysonian dynamics of the Ginibre ensemble. Phys. Rev. Lett. 113, 104102 (2014)
Chalker, J.T., Mehlig, B.: Eigenvector statistics in non-Hermitian random matrix ensembles. Phys. Rev. Lett. 81(16), 3367–3370 (1998)
Chalker, J.T., Mehlig, B.: Statistical properties of eigenvectors in non-Hermitian Gaussian random matrix ensembles. J. Math. Phys. 41(5), 3233–3256 (2000)
Crawford, N., Rosenthal, R.: Eigenvector correlations in the complex Ginibre ensemble (2018). arXiv: 1805.08993
Davy, M., Genack, A.Z.: Probing nonorthogonality of eigenfunctions and its impact on transport through open systems. Phys. Rev. Res. 1, 033026 (2019)
Deift, P., Gioev, D.: Random matrix theory: invariant ensembles and universality. Courant Lecture Notes in Mathematics, vol. 18, Courant Institute of Mathematical Sciences, New York; American Mathematical Society, Providence, RI (2009)
Diaconis, P., Freedman, D.: A dozen de Finetti-style results in search of a theory. Ann. Inst. H. Poincaré Probab. Statist. 23, 397–423 (1987). (English, with French summary)
Dubach, G.: Powers of Ginibre eigenvalues. Electron. J. Probab. 23, 1–31 (2018)
Dubach, G.: Symmetries of the Quaternionic Ginibre Ensemble. Random Matrices Theory Appl. (to appear, 2019). arXiv:1811.03724
Dubach, G.: On eigenvector statistics in the spherical and truncated unitary ensembles (2019). arXiv:1908.06713
Erdős, L., Krüger, T., Renfrew, D.: Power law decay for systems of randomly coupled differential equations. SIAM J. Math. Anal. 50(3), 3271–3290 (2018)
Edelman, A.: The probability that a random real Gaussian matrix has k real eigenvalues, related distributions, and the circular law. J. Multivar. Anal. 60(2), 203–232 (1997)
Forrester, P.J.: Log-gases and random matrices, London Mathematical Society Monographs Series, vol. 34. Princeton University Press, Princeton (2010)
Fyodorov, Y.V.: On statistics of bi-orthogonal eigenvectors in real and complex ginibre ensembles: combining partial schur decomposition with supersymmetry. Commun. Math. Phys. 363(2), 579–603 (2018)
Fyodorov, Y.V., Mehlig, B.: Statistics of resonances and nonorthogonal eigenfunctions in a model for single-channel chaotic scattering. Phys Rev E 66(4), 045202(R) (2002)
Fyodorov, Y.V., Savin, D.V.: Statistics of resonance width shifts as a signature of eigenfunction nonorthogonality. Phys. Rev. Lett. 108(18), 184101 (2012)
Fyodorov, Y.V., Sommers, H.-J.: Random matrices close to Hermitian or unitary: overview of methods and results. J. Phys. A: Math. Gen. 36(12), 3303–3347 (2003)
Ginibre, J.: Statistical ensembles of complex, quaternion, and real matrices. J. Math. Phys. 6, 440–449 (1965)
Goetschy, A., Skipetrov, S.E.: Non-Hermitian Euclidean random matrix theory. Phys Rev E. 84, 011150 (2011)
Grela, J., Warchoł, P.: Full Dysonian dynamics of the complex Ginibre ensemble. J. Phys. A. Math. Theor. 51(42), 425203 (2018)
Gros, J.B., Kuhl, U., Legrand, O., Mortessagne, F., Richalot, E., Savin, D.V.: Experimental width shift distribution: a test of nonorthogonality for local and global perturbations. Phys. Rev. Lett. 113, 224101 (2014)
Hough, J.B., Krishnapur, M., Peres, Y., Virág, B.: Determinantal processes and independence. Probab. Surv. 3, 206–229 (2006)
Janik, R.A., Noerenberg, W., Nowak, M.A., Papp, G., Zahed, I.: Correlations of eigenvectors for non-Hermitian random-matrix models. Phys. Rev. E 60(3), 2699–2705 (1999)
Johnson, N.L., Kotz, S.: Distributions in statistics. Continuous univariate distributions. 2., Houghton Mifflin Co., Boston, Mass (1970)
Kato, T.: Perturbation Theory for Linear Operators. Springer, Berlin (1980). (corrected second edition)
Keller, J.B.: Multiple eigenvalues. Linear Algebra Appl. 429(8–9), 2209–2220 (2008)
Khoruzhenko, B.A., Sommers, H.J.: Non-Hermitian Ensembles. In: Akemann, G., Baik, J., Francesco, P.D. (eds.) The Oxford Handbook of Random Matrix Theory. Oxford University Press, Oxford (2011)
Kostlan, E.: On the spectra of Gaussian matrices. Linear Algebra Appl. 162(164), 385–388 (1992). Directions in matrix theory (Auburn, AL, 1990)
Knowles, A., Yin, J.: Eigenvector distribution of Wigner matrices. Probab. Theory Rel. Fields 155(3–4), 543–582 (2013)
Lambert, G.: The law of large numbers for the maximum of the characteristic polynomial of the Ginibre ensemble (2019). arXiv:1902.01983
Lehmann, N., Sommers, H.-J.: Eigenvalue statistics of random real matrices. Phys. Rev. Lett. 67(8), 941–944 (1991)
Lyons, R.: Determinantal probability measures. Publ. Math. Inst. Hautes Études Sci. 98, 167–212 (2003)
Mehlig, B., Chalker, J.T.: Eigenvector correlations in non-Hermitian random matrix ensembles. Ann. Phys. 7(5–6), 427–436 (1998)
Mehta, M.L.: Random Matrices, 2nd edn. Academic Press Inc, Boston, MA (1991)
Movassagh, R.: Eigenvalue attraction. J. Stat. Phys. 162(3), 615–643 (2016)
Nowak, M.A., Tarnowski, W.: Probing non-orthogonality of eigenvectors in non-Hermitian matrix models: diagrammatic approach. J. High Energy Phys. 6, 152 (2018)
Overton, M.L., Womersley, R.S.: On minimizing the spectral radius of a nonsymmetric matrix function: optimality conditions and duality theory. SIAM J. Matrix Anal. Appl. 9(4), 473–498 (1988)
Rotter, I.: A non-Hermitian Hamilton operator and the physics of open quantum systems. J. Phys. A: Math. Theor. 42, 153001 (2009)
Rudelson, M., Vershynin, R.: Delocalization of eigenvectors of random matrices with independent entries. Duke Math. J. 164(13), 2507–2538 (2015)
Schomerus, H., Frahm, K.M., Patra, M., Beenakker, C.W.J.: Quantum limit of the laser line width in chaotic cavities and statistics of residues of scattering matrix poles. Physica A 278(3–4), 469–496 (2000)
Tao, T., Vu, V.: Random matrices: universal properties of eigenvectors. Random Matrices Theory Appl. 1(1), 1150001 (2012)
Trefethen, L.N., Embree, M.: Spectra and pseudospectra: the behavior of nonnormai matrices and operators, vol. 1. Princeton University Press, Princeton (2005)
Trefethen, L.N., Trefethen, A.E., Reddy, S.C., Driscoll, T.A.: Hydrodynamic stability without eigenvalues. Science 261, 578,584 (1993)
Walters, M., Starr, S.: A note on mixed matrix moments for the complex Ginibre ensemble. J. Math. Phys. 56(1), 013301, 20 (2015)
Webb, C., Wong, M.D.: On the moments of the characteristic polynomial of a Ginibre random matrix. Proc. Lond. Math. Soc. (3) 118(5), 10171056 (2019)
Acknowledgements
The authors thank the referees for particularly precise and pertinent suggestions which helped improving this article.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This work is supported by the NSF Grant DMS#1513587.
Appendices
Appendix A: Eigenvalues dynamics
This Appendix derives the Dyson-type dynamics for eigenvalues of nonnormal matrices. More precisely, we consider the Ornstein-Uhlenbeck version so that the equilibrium measure is the (real or complex) Ginibre ensemble. These dynamics take a particularly simple form in the case of complex Gaussian addition, where the drift term shows no interaction between eigenvalues: only the correlation of martingale terms is responsible for eigenvalues repulsion.
We also describe natural dynamics with equilibrium measure given by the real Ginibre ensemble. Then, the eigenvalues evolution is more intricate.
It was already noted in [11] that eigenvectors impact the eigenvalues dynamics for nonnormal matrices, and the full dynamics in the complex case have been written down in [30].
1.1 Complex Ginibre dynamics
Let G(0) be a complex matrix of size N, assumed to be diagonalized as \(YGX = \Delta = \mathrm {Diag}(\lambda _1,\dots ,\lambda _N) \), where X, Y are the matrices of the right- and left-eigenvectors of G(0). We also assume that G(0) has simple spectrum, and X, Y invertible. The right eigenvectors \((x_i)\) are the columns of X, and the left-eigenvectors \((y_j)\) are the rows of Y. They are chosen uniquely such that \(XY=I\) and, for any \(1\leqslant k\leqslant N\), \(X_{kk}=1\).
We now consider the complex Dyson-type dynamics: for any \(1\leqslant i,j\leqslant N\),
where the \(B_{ij}\)’s are independent standard complex Brownian motions: \(\sqrt{2}{{\,\mathrm{Re}\,}}(B_{ij})\) and \(\sqrt{2}{{\,\mathrm{Im}\,}}(B_{ij})\) are standard real Brownian motions. One can easily check that G(t) converges to the Ginibre ensemble as \(t\rightarrow \infty \), with normalization (1.1).
In the following, the bracket of two complex martingales M, N is defined by bilinearity: \(\langle M,N\rangle =\langle {{\,\mathrm{Re}\,}}M,{{\,\mathrm{Re}\,}}N\rangle -\langle {{\,\mathrm{Im}\,}}M,{{\,\mathrm{Im}\,}}N\rangle +\mathrm {i}\langle {{\,\mathrm{Re}\,}}M,{{\,\mathrm{Im}\,}}N\rangle +\mathrm {i}\langle {{\,\mathrm{Im}\,}}M,{{\,\mathrm{Re}\,}}N\rangle \).
Proposition A.1
The spectrum \((\lambda _1(t),\dots , \lambda _n(t))\) is a semimartingale satisfying the system of equations
where the martingales \((M_k)_{1\leqslant k\leqslant N}\) have brackets \(\langle M_i,M_j\rangle =0\) and
Remark A.2
As explained below, this equation (in particular the off-diagonal brackets) is coherent with the eigenvalues repulsion observed in (1.2). Contrary to the Hermitian Dyson Brownian motion, all eigenvalues are martingales (up to the Ornstein Uhlenbeck drift term), so that their repulsion is not due to direct mutual interaction, but to correlations between these martingales at the microscopic scale.
For example, assume that G(0) is already at equilibrium. Using physics conventions, for any bulk eigenvalues \(\lambda _1,\lambda _2\) satisfying \(w={{\,\mathrm{O}\,}}(1)\) (remember \(w=\sqrt{N}(\lambda _1-\lambda _2)\)), Proposition A.1 and Theorem 1.4 imply
in the bulk. By considering the real part in this equation and denoting \(\mathrm{d}\lambda _1=\mathrm{d}x_1+\mathrm {i}\mathrm{d}y_1\), \(\mathrm{d}\lambda _2=\mathrm{d}x_2+\mathrm {i}\mathrm{d}y_2\), we have in particular \({\mathbb {E}}(\mathrm{d}x_1\mathrm{d}x_2+\mathrm{d}y_1\mathrm{d}y_2)<0\), and this negative correlation is responsible for repulsion: the eigenvalues tend to move in opposite directions. Moreover, as eigenvalues get closer on the microscopic scale, \(w\rightarrow 0\) and the repulsion gets stronger:
On the other hand, for mesoscopic scale \(N^{-1/2}\ll |\lambda _1-\lambda _2|\), Proposition A.1 and Theorem 1.4 give \( {\mathbb {E}}\left( \mathrm{d}\lambda _1\mathrm{d}\overline{\lambda _2}\right) \sim -\frac{(1-|\lambda _1|^2)}{N^2|\lambda _1-\lambda _2|^4}\mathrm{d}t={{\,\mathrm{o}\,}}(\mathrm{d}t)\), so that increments are uncorrelated for large N.
For a given differential operator \(f\mapsto f'\), we introduce the matrix \(C=X^{-1}X'\). Along the following lemmas, all eigenvalues are assumed to be distinct. In our application, this spectrum simplicity will hold almost surely for any \(t\geqslant 0\) as G(0) has simple spectrum.
Lemma A.3
We have \(X'=XC\) and \(Y'=-CY\).
Proof
The first equality is the definition of C. For the second one, \(XY=I\) gives \(XY'+X'Y=0\), hence \(Y' = - X^{-1} X' Y = -CY.\)\(\square \)
Lemma A.4
The first order perturbation of eigenvalues is given by \(\lambda _k ' = y_k G' x_k\).
Proof
We have \(\Delta ' = (YGX)' = Y'GX + YG'X + YGX' = YG'X + YGXC - C YGX = YG'X + \Delta C - C \Delta = YG'X + [\Delta ,C]\). Therefore \(\lambda _k' = (YG'X)_{kk} + [\Delta ,C]_{kk} = y_k G' x_k\). \(\square \)
Lemma A.5
For any \(i \ne j\), \(C_{ij} = {y_i G' x_j \over \lambda _j - \lambda _i}\).
Proof
For such i, j, \(\Delta '_{ij} = 0\). With the same computation as in the previous lemma, this gives \((YG'X)_{ij} + [\Delta ,C]_{ij}=0\). Thus \((\lambda _i - \lambda _j) C_{ij} = - (YG'X)_{ij} = - y_i G' x_j\), from which the result follows. \(\square \)
Lemma A.6
For any \(1\leqslant k\leqslant N\), \(C_{kk} = - \sum _{l \ne k} X_{kl} {y_l G' x_k \over \lambda _k - \lambda _l}\).
Proof
We use the assumption \( X_{kk}=1\). From this, and the definition of C, we get
As a consequence, \(C_{kk} = - \sum _{l \ne k} X_{kl} C_{lk}\) and we obtain the result thanks to the previous lemma. \(\square \)
From now on the differential operator will be either \(\partial _{{{\,\mathrm{Re}\,}}G_{ab}}\) (\(G' = E_{ab}= \{ \delta _{ia} \delta _{jb} \}_{1\leqslant i,j\leqslant N}\)), or \(\partial _{{{\,\mathrm{Im}\,}}G_{ab}}\), (\(G' = \mathrm {i}E_{ab})\). In both cases, \(G''=0\). We denote \(C^{{{\,\mathrm{Re}\,}}}\) and \(C^{{{\,\mathrm{Im}\,}}}\) accordingly. In particular, for any k and \(i\ne j\) the following holds:
Lemma A.7
We have
Proof
Below is the computation for \( \partial _{{{\,\mathrm{Re}\,}}G_{ab}}X_{ij} \). We use \(X'=XC\) and (A.2):
The case \(\partial _{{{\,\mathrm{Im}\,}}G_{ab}}X_{ij}\) is obtained similarly, as are the formulas for Y. \(\square \)
Lemma A.8
The second order perturbation of eigenvalues is given by
Proof
We compute the perturbation for \( \partial _{{{\,\mathrm{Re}\,}}G_{ab}} \). Differentiating \(\lambda \) a second time gives
Replacing \(X'\) and \(Y'\) with their expressions yields
which concludes the proof, the other cases being similar. \(\square \)
For the proof of Proposition A.1, we need the following elementary lemma.
Lemma A.9
Let \(\tau =\inf \{t\geqslant 0:\exists i\ne j, \lambda _i(t)=\lambda _j(t)\}\). Then \(\tau =\infty \) almost surely.
Proof
The set of matrices G with Jordan form of type
is a submanifold \({\mathcal {M}}_1\) (resp. \({\mathcal {M}}_2\)) of \({\mathbb {C}}^{N^2}\) with complex codimension 1 (resp. 3), see e.g. [36, 47]. Therefore, almost surely, a Brownian motion in \({\mathbb {C}}^{N^2}\) starting from a diagonalizable matrix with simple spectrum will not hit \({\mathcal {M}}_1\) or \({\mathcal {M}}_2\). This concludes the proof. \(\square \)
All derivatives can therefore be calculated, as eigenvalues and eigenvectors are analytic functions of the matrix entries (see [35]).
Proof of Proposition A.1
In our context,the Itô formula will take the following form: for a function f from \({\mathbb {C}}^n\) to \({\mathbb {C}}\) of class \(C^2\), where \(B_t=(B_t^1,\dots ,B_t^n)\) is made of independent standard complex Brownian motions, we have
For any given \(0<{\varepsilon }<\min \{|\lambda _i(0)-\lambda _j(0)|,i\ne j\}\), let
Eigenvalues are smooth functions of the matrix coefficients on the domain \(\cap _{i<j}\{|\lambda _i-\lambda _j|>{\varepsilon }\}\), so that Eq. (A.3) together with Lemmas A.4 and A.8 gives the following equality of stochastic integrals, with substantial cancellations of the drift term:
Taking \({\varepsilon }\rightarrow 0\) in the above equation together with Lemma A.9 yields
The eigenvalues martingales terms are correlated. Their brackets are
This concludes the proof. \(\square \)
1.2 Proof of Corollary 1.6
Let \(A_{t,{\varepsilon }}=\{\sup _{0\leqslant s\leqslant t}|\lambda _1(s)-\lambda _1(0)|<N^{\varepsilon }t^{1/2}\}\). We start by proving that
From Proposition A.1 and Itô’s formula, we have
which is a local martingale. It is an actual martingale because
where in the last equality we used \({\mathbb {E}}({\mathscr {O}}_{11}(s))={{\,\mathrm{O}\,}}(N)\), which follows from (2.11). The estimate (A.7) follows by Doob’s and Markov’s inequalities.
For (1.16), we start with
This implies
Here, we used that \((\mathrm{Re}\int _0^t\overline{e^{\frac{s}{2}}\lambda _1(s)-\lambda _1(0)}e^{\frac{s}{2}}\mathrm{d}M_1(s) )_t\) is an actual martingale, because the expectation of its bracket is
where for the last inequality we used (2.11).
To evaluate the right hand side of (A.11), we would like to change \(\lambda _1(0)\in {\mathscr {B}}\) into \(\lambda _1(s)\in {\mathscr {B}}\). First,
where for the last inequality we used (2.11), again. Moreover, if \(1/p+1/q=1\) with \(p<2\). we have
where we used [24, Theorem 2.3] to obtain that uniformly in the complex plane and in N, \({\mathscr {O}}_{11}/N\) has finite moment of order \(p<2\). Equations (A.11), (A.12) and (A.13) imply
and one concludes the proof of (1.16) with (2.11).
The proof of (1.17) is identical, except that we rely on the off-diagonal bracket \(\mathrm{d}\langle \lambda _1,\bar{\lambda }_2\rangle _s={\mathscr {O}}_{12}(s)\frac{\mathrm{d}s}{N}\), the estimate (1.9), and the elementary inequality
to bound the (first and p-th) moment of \({\mathscr {O}}_{12}\) in the whole complex plane based on those of \({\mathscr {O}}_{11}\), \({\mathscr {O}}_{22}\).
1.3 Real Ginibre dynamics
We now consider G(0) a real matrix of size N, again assumed to be diagonalized as \(YGX = \Delta = \mathrm {Diag}(\lambda _1,\dots ,\lambda _N) \), where X, Y are the matrices of the right- and left-eigenvectors of G(0). We also assume that G(0) has simple spectrum, and X, Y invertible. We keep the same notations for the right eigenvectors \((x_i)\), columns of X, and the left-eigenvectors \((y_j)\), rows of Y. They are again chosen such that \(XY=I\) and, for any \(1\leqslant k\leqslant N\), \(X_{kk}=1\).
In this subsection, the real Dyson-type dynamics are (\(1\leqslant i,j\leqslant N\)),
where the \(B_{ij}\)’s are independent standard Brownian motions. One can easily check that G(t) converges to the real Ginibre ensemble as \(t\rightarrow \infty \).
Note that the real analogue of Lemma A.9 gives weaker repulsion: the set of real matrices with Jordan form of type
is a submanifold \({\mathcal {M}}_1\) of \({\mathbb {R}}^{N^2}\), supported on \(\lambda _{N-1}\in {\mathbb {R}}\), with real codimension 1 (as proved by a straightforward adaptation of [36, Theorem 7]). Denoting \(\tau =\inf \{t\geqslant 0:\exists i\ne j, \lambda _i(t)=\lambda _j(t)\}\), under the dynamics (A.14) for any \(t>0\) we therefore have
so that we can only state the real version of Proposition A.1 up to time \(\tau \). In fact, collisions occur transforming pairs of real eigenvalues into pairs of complex conjugate eigenvalues, a mechanism coherent with the random number of real eigenvalues in the real Ginibre ensemble [22, 41].
The overlaps (1.4) are enough to describe the complex Ginibre dynamics, and so are they for the real Ginibre ensemble, up to the introduction of the following notation: we define \(\bar{i}\in \llbracket 1,N\rrbracket \) through \(\lambda _{\bar{i}}=\overline{\lambda _i}\), i.e. \({\bar{i}}\) is the index of the conjugate eigenvalue to \(\lambda _i\). Note that \({\bar{i}}=i\) if \(\lambda _i\in {\mathbb {R}}\). For real matrices, if \(L_j,R_j\) are eigenvectors associated to \(\lambda _j\), \({\bar{L}}_j,{\bar{R}}_j\) are eigenvectors for \({{\bar{\lambda }}}_j\), so that
Proposition A.10
The spectrum \((\lambda _1(t),\dots , \lambda _n(t))\) evolves according to the following stochastic equations, up to the first collision:
where the martingales \((M_k)_{1\leqslant k\leqslant N}\) have brackets
Note that the real eigenvalues have associated real eigenvectors. For those, \({{\mathscr {O}}}_{k{\bar{l}}}={\mathscr {O}}_{kl}\), and the variation is real: real eigenvalues remain real as long as they do not collide.
Remark A.11
Proposition A.10 is coherent with the attraction between conjugate eigenvalues exhibited in [45]. In fact, if \(\eta ={{\,\mathrm{Im}\,}}(\lambda _k)>0\), the drift interaction term with \(\bar{\lambda }_k\) is \({\mathscr {O}}_{kk}/(\lambda _k-{{\bar{\lambda }}}_k)=-\mathrm {i}{\mathscr {O}}_{kk}/(2\eta )\), so that these eigenvalues attract each other stronger as they approach the real axis.
For the proof, we omit the details and only mention the differences with respect to Proposition A.1. We apply the Itô formula for a \({\mathscr {C}}^2\) function f from \({\mathbb {R}}^n\) to \({\mathbb {C}}\), with argument \(U_t=(U_t^1,\dots ,U_t^n)\) is made of independent Ornstein-Uhlenbeck processes. Together with the perturbation formulas for \(\lambda _k\), Lemmas A.4 and A.8 , we obtain (remember the notation (A.4))
We can take \({\varepsilon }\rightarrow 0\) in the above formulas and the brackets are calculated as follows, concluding the proof:
Appendix B: Normalized eigenvectors
This paper focuses on the condition numbers and off-diagonal overlaps, but the Schur decomposition also easily gives information about other statistics such as the angles between eigenvectors. We include these results for the sake of completeness. We denote the complex angle as
where the phases of \(R_1(1)\) and \(R_2(1)\) can be chosen independent uniform on \([0,2\pi )\). We also define
Proposition B.1
Conditionally on \(\lambda _1=z_1,\lambda _2=z_2\), we have
where \(X \sim {\mathscr {N}}_{{\mathbb {C}}}(0,\frac{1}{2} \mathrm{Id})\).
In particular, for \(\lambda _1, \lambda _2\) at mesoscopic distance, the complex angle converges in distribution to a Dirac mass at 0. Therefore in such a setting eigenvectors strongly tend to be orthogonal: matrices sampled from the Ginibre ensemble are not far from normal, when only considering eigenvectors angles. The limit distribution becomes non trivial in the microscopic scaling \( |\lambda _1 - \lambda _2| \sim N^{-1/2}\), it is the pushforward of a complex Gaussian measure by \(\Phi \).
Proof
From Proposition 2.1 we know that \(R_1^*R_2=R_{T,1}^*R_{T,2}\), \(\Vert R_1\Vert =\Vert R_{T,1}\Vert \) and \(\Vert R_2\Vert =\Vert R_{T,2}\Vert \), where \(R_{T,i}\) (and \(L_{T,i}\)) are the normalized bi-orthogonal bases of right and left eigenvectors for T, defined as (2.2). The first eigenvectors are written \(R_{T,1}=(1,0,\dots ,0)\) and \(R_{T,2}=(a,1,0\dots ,0)\) where \(a=-\bar{b}_2=-\frac{T_{12}}{\lambda _1-\lambda _2}\), with \(T_{12}\) complex Gaussian \({\mathscr {N}}\left( 0,\frac{1}{2N}\mathrm{Id}\right) \), independent of \(\lambda _1\) and \(\lambda _2\). This gives
and concludes the proof. \(\square \)
From Proposition B.1, the distribution of the angle for fixed \(\lambda _1\) and random \(\lambda _2\) can easily be inferred. For example, if \(\lambda _2\) is chosen uniformly among eigenvalues in a macroscopic domain \(\Omega \subset \{|z|<1\}\) with nonempty interior, we obtain the convergence in distribution (\(X_\Omega \) is uniform on \(\Omega \), independent of \({\mathscr {N}}\))
When \(z_1=0\) and \(z_2\) is free, the following gives a more precise distribution, for finite N and in the limit.
Corollary B.2
Conditionally on \(\{ \lambda _1 =0 \}\) we have
where \(U_N\) is an independent random variable uniform on \( \{ 2,\dots ,N\}\), and X has density \( {1 - (1+t) e^{-t} \over t^2} {\mathbf {1}}_{{\mathbb {R}}_+} (t).\)
Proof
From Corollary 5.6, \(N |\lambda _2|^2 \sim \gamma _{U_N}\). Together with Lemma 2.5, this gives
The limiting density then follows from the explicit distribution of \(\beta \) random variables. \(\square \)
Rights and permissions
About this article
Cite this article
Bourgade, P., Dubach, G. The distribution of overlaps between eigenvectors of Ginibre matrices. Probab. Theory Relat. Fields 177, 397–464 (2020). https://doi.org/10.1007/s00440-019-00953-x
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00440-019-00953-x