1. Introduction
This paper is concerned with the asymptotic expansion for the partition function and the multilinear statistics of $\beta $ matrix models. These laws represent a generalisation of the joint distribution of the N eigenvalues of the Gaussian Unitary Ensemble [Reference MehtaMeh04]. The convergence of the empirical measure of the eigenvalues is well known (see, for example, [Reference Boutet de Monvel, Pastur and ShcherbinadMPS95]), and we are interested in the allorder finite size corrections to the moments of this empirical measure. Much attention has been paid to this problem in the regime when the eigenvalues condense on a single segment, usually referred to as a onecut regime. In this case, a central limit theorem for linear statistics was proved by Johansson [Reference JohanssonJoh98], while a full $\frac {1}{N}$ expansion was derived first for $\beta = 2$ [Reference Albeverio, Pastur and ShcherbinaAPS01, Reference Ercolani and McLaughlinEM03, Reference Bleher and ItsBI05] and then for any $\beta> 0$ in [Reference Borot and GuionnetBG11]. However, the multicut regime was, until recently, poorly understood at the rigorous level, except for $\beta = 2$ , which is related to integrable systems and can be treated with the powerful asymptotic analysis techniques for Riemann–Hilbert problems; see, for example, [Reference Deift, Kriecherbauer, McLaughlin, Venakides and ZhouDKM+99b]. Nevertheless, a heuristic derivation of the asymptotic expansion for the multicut regime has been proposed to leading order by Bonnet, David and Eynard [Reference Bonnet, David and EynardBDE00] and extended to all orders in [Reference EynardEyn09], in terms of Theta functions and their derivatives. It features oscillatory behaviour, whose origin lies in the tunneling of eigenvalues between the different connected components of the support. This heuristic, originally written for $\beta = 2$ , can be trivially extended to $\beta> 0$ ; see, for example, [Reference BorotBor11].
More recently, M. Shcherbina has established this asymptotic expansion up to terms of order $1$ [Reference ShcherbinaShc11, Reference ShcherbinaShc12]. This allows us to observe, for instance, that linear statistics do not always satisfy a central limit theorem (this fact was already noticed for $\beta = 2$ in [Reference PasturPas06]). In this work, we go beyond the $O(1)$ and put the heuristics of [Reference EynardEyn09] to all orders on a firm mathematical ground. Our strategy is to first study the asymptotics in the model with fixed filling fractions and then reconstruct the asymptotics in the original model via a finitedimensional analysis. As a consequence, we obtain a replacement for the central limit theorem for linear statistics and for filling fractions. Besides, we treat uniformly soft and hard edges, while [Reference ShcherbinaShc12] assumed soft edges.
For $\beta = 2$ , we can establish the full asymptotic expansion outside of the bulk for the orthogonal polynomials with realanalytic potentials and the allorder asymptotic expansion of certain solutions of the Toda lattice in the continuum limit. The same method allows us to rigorously establish the asymptotics of skeworthogonal polynomials ( $\beta = 1$ and $4$ ) away from the bulk, derived heuristically in [Reference EynardEyn01]. To our knowledge, the Riemann–Hilbert analysis of skeworthogonal polynomials is possible in principle but is cumbersome and has not been done before, so our method provides the first proof of these asymptotics. After this work was released, this method was extended to treat more general Coulomblike interactions in [Reference Borot, Guionnet and KozlowskiBGK15]. We also note that a proof of the asymptotics up to $o(1)$ with $\beta = 2$ was obtained by the Riemann–Hilbert approach in the twocuts situation in [Reference Claeys, Grava and McLaughlinCGMcL15] and in the kcut situation with $k \geq 2$ in [Reference Charlier, Fahs, Webb and WongCFWW].
Since the first release of this work, several authors have considered asymptotic questions in the multicut regime of $\beta $ ensembles. A recent approach to central limit theorems inspired by Stein’s method was proposed in [Reference Lambert, Ledoux and WebbLLW19], but it is restricted to the onecut regime. The transport method introduced in [Reference Bekerman, Guionnet and FigalliBGF15] allowed the rigidity of eigenvalues [Reference LiLi16] and universality [Reference BekermanB18] in the multicut regime to be established. In [Reference Bekerman, Leblé and SerfatyBLS18], the validity of central limit theorems for linear fluctuations has also been extended to include test functions with weaker regularity assumptions and to critical cases (and then test functions in the range of the socalled ‘master operator’). Beyond being a source of inspiration for these works, and the first rigorous article where Dyson–Schwinger equations were used to derive central limit theorem in the multicut regime, the present article contains results that still did not appear anywhere else, such as the asymptotics of to (skew) orthogonal polynomials and integrable systems (see Section 2), a discussion about the relation with Chekhov–Eynard–Orantin topological recursion (see Section 1.5), and the detailed use of precise estimates of beta ensembles with fixed filling fractions to estimate the free energy in multicut models and the reconstruction of the Theta function (see Section 8). Besides, Shcherbina derives in [Reference ShcherbinaShc12] via operator methods and for soft edges an expression of the order N in the free energy in terms of the entropy of the equilibrium measure and a universal constant. Our work proves a similar formula both with soft and hard edges and with a different method based on complex analysis.
Our results on the asymptotics of the partition function have been used (e.g., to study the asymptotics of the determinant of Töplitz matrices in [Reference MarchalMar20, Reference MarchalMar21]). The ideas that we introduce to handle the multicut regime are extended in a work in progress [Reference Borot, Guionnet and GorinBGG] to study the fluctuations of discrete $\beta $ ensembles appearing in random tiling models in nonsimply connected domains (with holes and/or frozen regions).
For Coulomb gases in dimension $d> 1$ , carrying out the asymptotic analysis when the support of the equilibrium measure has several connected components remains, in general, an open problem. Some specific $d = 2$ , $\beta = 2$ situations have been treated in [Reference Ameur, Charlier and CronwallACC, Reference Ameur, Charlier, Cronvall and LenellsACCL] relying on the determinantal structure of these models. In general, probabilistic methods in the spirit of this article that do not rely on integrability, and therefore could address arbitrary $\beta> 0$ (where integrability is absent), are still insufficiently developed.
1.1. Definitions
1.1.1. Model and empirical measure
We consider the probability measure $\mu _{N,\beta }^{V;\mathsf {B}}$ on $\mathsf {B}^N$ given by
$\mathsf {B}$ is a finite disjoint union of closed intervals of $\mathbb {R}$ possibly with infinite endpoints, $\beta $ is a positive number and $Z_{N,\beta }^{V;\mathsf {B}}$ is the partition function so that (1.1) has total mass $1$ . This model is usually called the $\beta $ ensemble [Reference MehtaMeh04, Reference Dumitriu and EdelmanDE02, Reference ForresterFor10]. We introduce the unnormalised empirical measure $M_N$ of the eigenvalues
and we consider several types of statistics for $M_N$ . We sometimes denote $\mathbb {L} = \mathrm {diag}(\lambda _1,\ldots ,\lambda _N)$ .
1.1.2. Correlators
We introduce the Stieltjes transform of the nth order moments of the empirical measure, called disconnected correlators:
They are holomorphic functions of $x_i \in \mathbb {C}\setminus \mathsf {B}$ . It is more convenient to consider the correlators to study large N asymptotics:
By construction, the coefficients of their expansions as a Laurent series in the variables $x_i $ (sufficiently large) give the nth order cumulants of $M_N$ . If I is a set, we introduce the notation $x_I = (x_i)_{i \in I}$ for a set of variables indexed by I; their order will not matter as we insert them only in symmetric functions of their variables (like $W_n$ , $\widetilde {W}_n$ , etc.). The two types of correlators are related by
where $\dot {\cup }$ stands for the disjoint union. If $\varphi _n$ is an analytic (symmetric) function in n variables in a neighbourhood of $\mathsf {B}^n$ , then the nlinear statistics can be deduced as contour integrals of the disconnected correlators:
We remark that the knowledge of the correlators for an analytic family of potentials $(V_{t})_{t}$ determines the partition function up to an integration constant since
where $W_1^{t}$ is the first correlator in the model with potential $V_t$ , and the notation $\oint _{\mathsf {B}} \mathrm {d} \xi \cdots $ means integration along a contour in $\mathbb {C} \setminus \mathsf {B}$ surrounding $\mathsf {B}$ with positive orientation. If the integrand has poles in $\mathbb {C} \setminus \mathsf {B}$ (e.g., it depends on extra variables $x_i \in \mathbb {C} \setminus \mathsf {B}$ that are not integrated upon and has poles at $\xi = x_i$ ), the contour should be chosen (unless stated otherwise) so that the poles remain outside. The notation should not be confused with $\int _{\mathsf {B}} \mathrm {d} \xi \cdots $ , which is the Lebesgue integral on $\mathsf {B} \subseteq \mathbb {R}$ .
1.1.3. Kernels
Let $\mathbf {c}$ be a ntuple of nonzero complex numbers. We introduce the npoint kernels:
When $c_j$ are integers, the kernels are holomorphic functions of $x_j \in \mathbb {C}\setminus \mathsf {B}$ . When $c_j$ are not integers, the kernels are multivalued holomorphic functions of $x_j$ in $\mathbb {C}\setminus \mathsf {B}$ , with monodromies around the connected components of $\mathsf {B}$ and around $\infty $ . The righthand side of (1.4), where we used $\ln $ , has the same multivalued nature. Alternatively, both sides of (1.4) can be defined as singlevalued functions of $x_1,\ldots ,x_n$ by choosing a determination of the logarithm in a domain $\mathsf {D}$ of the form $\mathbb {C} \setminus \ell $ , where $\ell $ is a smooth path in $\mathbb {C}$ from $0$ to $\infty $ , and using $z^{c} = e^{c \ln z}$ for the lefthand side.
In particular, for $\beta = 2$ , $\mathsf {K}_{1,(1)}(x)$ is the monic Nth orthogonal polynomial associated to the weight $\mathbf {1}_{\mathsf {B}}(x)\,e^{N\,V(x)}\mathrm {d} x$ on the real line, and $\mathsf {K}_{2,(1,1)}(x,y)$ is the Nth Christoffel–Darboux kernel associated to those orthogonal polynomials; see Section 2.
1.2. Equilibrium measure and multicut regime
By standard results of potential theory and large deviations – see [Reference JohanssonJoh98, Reference Ben Arous and GuionnetBAG97] or the textbooks [Reference DeiftDei99, Theorem 6] or [Reference Anderson, Guionnet and ZeitouniAGZ10, Theorem 2.6.1 and Corollary 2.6.3] (note there that $\mathsf {B}=\mathbb R$ , but the generalisation to integration over general sets $\mathsf {B}$ is straightforward) – we have the following:
Theorem 1.1. Assume that $V\,:\, \mathsf {B} \rightarrow \mathbb {R}$ is a continuous function, and if V depends on N, assume also that $V $ converges towards $V^{\{0\}}$ when N goes to infinity in the space of continuous functions over $\mathsf {B}$ for the sup norm. Moreover, for $\tau \in \{\pm 1\}$ with $\tau \infty \in \mathsf {B}$ , assume that
We consider the normalised empirical measure $L_N=N^{1}\,M_N$ in the space $\mathcal {P}(\mathsf {B})$ of probability measures on $\mathsf {B}$ equipped with its weak topology. Then, the law of $L_N$ under $\mu _{N,\beta }^{V;\mathsf {B}}$ satisfies a large deviation principle with scale $N^2$ and good rate function J given by
As a consequence, $L_N$ converges almost surely and in expectation to the unique probability measure $\mu _\mathrm{{eq}}^{V}$ on $\mathsf {B}$ which minimises E. $\mu _\mathrm{{eq}}^{V}$ has compact support, denoted $\mathsf {S}$ . It is characterised by the existence of a constant $C^V$ such that
with equality realised $\mu _\mathrm{{eq}}^V$ almost surely.
The goal of this article is to establish an allorder expansion of the partition function, the correlators and the kernels in all such situations.
1.3. Assumptions
We will refer throughout the text to the following set of assumptions. An integer number $g\ge 0$ is fixed.
Hypothesis 1.1.

○ (Regularity) $V\,:\,\mathsf {B} \rightarrow \mathbb {R}$ is continuous, and if V depends on N, it has a limit $V^{\{0\}}$ in the space of continuous functions on $\mathsf {B}$ for the sup norm.

○ (Confinement) For $\tau \in \{\pm 1\}$ so that $\tau \infty \in \mathsf {B}$ , $\liminf _{x \rightarrow \tau \infty } \frac {V(x)}{2\ln x}> 1$ . If V depends on N, we require its limit $V^{\{0\}}$ to satisfy this condition.

○ ( $(g + 1)$ cut regime) The support of $\mu _\mathrm{{eq}}^{V}$ is of the form $\mathsf {S} = \bigcup _{h = 0}^{g} \mathsf {S}_h$ , where $\mathsf {S}_h = [\alpha _{h}^{},\alpha _{h}^{+}]$ are pairwise disjoint and $\alpha _{h}^{} < \alpha _{h}^+$ for any .

○ (Control of large deviations) The effective potential $U^{V;\mathsf {B}}_{\mathrm{eq}}(x) = V(x) 2\int _{\mathsf {B}} \ln x\xi \mathrm {d}\mu _\mathrm{{eq}}^{V}(\xi )$ for $x \in \mathsf {B}$ achieves its minimum value for $x \in \mathsf {S}$ only.

○ (Offcriticality) $\mu _\mathrm{{eq}}^{V}$ has a density of the form
(1.7) $$ \begin{align} \frac{\mathrm{d}\mu_{\mathrm{eq}}^{V}}{\mathrm{d} x} = \frac{S(x)}{\pi}\,\prod_{h = 0}^{g} (\alpha_h^{+}  x)^{\rho_h^{+}/2}(x  \alpha_h^{})^{\rho_h^{}/2}, \end{align} $$where $\rho _{h}^{\bullet }$ is $+1$ (resp. $1$ ) if the corresponding edge is soft (resp. hard), and $S(x)> 0$ for $x \in \mathsf {S}$ . Hard edges must be boundary points of $\mathsf {B}$ .
Note that if $V^{\{0\}}$ is realanalytic in a neighbourhood of $\mathsf {B}$ , the $(g + 1)$ cut regime hypothesis is always satisfied (the support consists of a finite disjoint union of segments) and S is analytic in a neighbourhood of $\mathsf {S}$ . We will hereafter say that V is regular and confining in $\mathsf {B}$ if it satisfies the two first assumptions above. We will also require a stronger regularity for the potential.
Hypothesis 1.2.

○ (Analyticity) V extends to a holomorphic function in some open neighbourhood $\mathsf {U}$ of $\mathsf {S}$ .

○ ( $\frac {1}{N}$ expansion of the potential) There exists a sequence $(V^{\{k\}})_{k \geq 0}$ of holomorphic functions in $\mathsf {U}$ and constants $(v^{\{k\}})_{k \geq 1}$ such that, for any $K \geq 0$ ,
(1.8) $$ \begin{align} \sup_{\xi \in \mathsf{U}} \BigV(\xi)  \sum_{k = 0}^{K} N^{k}\,V^{\{k\}}(\xi)\Big \leq v^{\{K + 1\}}\,N^{(K + 1)}. \end{align} $$
In Section 6, we shall weaken Hypothesis 1.2 by allowing complex perturbations of order $\frac {1}{N}$ and harmonic functions instead of analytic functions.
Hypothesis 1.3. $V\,:\,\mathsf {B} \rightarrow \mathbb {C}$ can be decomposed as $V = \mathcal {V}_1 + \overline {\mathcal {V}_2}$ where:

○ For $j = 1,2$ , $\mathcal {V}_j$ extends to a holomorphic function in some neighbourhood $\mathsf {U}$ of $\mathsf {B}$ . There exists a sequence of holomorphic functions $(\mathcal {V}_{j}^{\{k\}})_{k \geq 0}$ and constants $(v_{j}^{\{k\}})_{k \geq 1}$ so that, for any $K \geq 0$ ,
$$ \begin{align*}\sup_{\xi \in \mathsf{U}} \Big\mathcal{V}_j(\xi)  \sum_{k = 0}^{K} N^{k}\,\mathcal{V}_j^{\{k\}}(\xi)\Big \leq v_j^{\{K + 1\}}\,N^{(K + 1)}. \end{align*} $$ 
○ $V^{\{0\}} = \mathcal {V}_1^{\{0\}} + \overline {\mathcal {V}_2^{\{0\}}}$ is realvalued on $\mathsf {B}$ .
The topology for which we study the large N expansion of correlators is described in § 5 and amounts to controlling the (moments of order p) $\times C^p$ uniformly in p for a constant $C> 0$ . We now describe our strategy and announce our results.
1.4. Main result with fixed filling fractions: partition function and correlators
Before coming to the multicut regime, we analyse a different model where the number of $\lambda $ s in a small enlargement of $\mathsf {S}_h$ is fixed. Let $\mathsf {A} = \bigcup _{h = 0}^{g} \mathsf {A}_h$ , where $\mathsf {A}_h = [a_{h}^{},a_h^{+}]$ are pairwise disjoint segments such that $a_{h}^{} \leq \alpha _{h}^{} < \alpha _{h}^+ \leq a_{h}^+$ , where the inequalities are equalities if the corresponding edge is hard and are strict if the corresponding edge is soft. We introduce the set
If $\boldsymbol {N}=(N_1,\ldots ,N_g)$ is an integer vector such that $\boldsymbol {\epsilon }=\frac {\boldsymbol {N}}{N} \in \mathcal {E}$ , we denote $N_0 = N  \sum _{h = 1}^{g} N_h$ and consider the probability measure on $\prod _{h = 0}^{g} \mathsf {A}_h^{N_h}$ :
The empirical measure $M_{N}$ and the correlators $W_{n;\boldsymbol {N}/N}(x_1,\ldots ,x_n)$ for this model are defined as in § 1.1 with $\mu _{N,\beta }^{V;\mathsf {A}}$ replaced by $\mu _{N,\beta ;\boldsymbol {N}/N}^{V;\mathsf {A}}$ . We call $\epsilon _h=\frac {N_h}{N}$ the filling fraction of $\mathsf {A}_h$ . It follows from the definitions that
for $x_2,\ldots ,x_n \in \mathbb {C} \setminus \mathsf {A}$ . Indeed, from the definition of the correlators (1.2), $W_{n;\boldsymbol {N}/N}(x_{1},x_2,\ldots ,x_n) $ for $n \geq 2$ can be expressed as a sum of products of moments of products of the ntuple of random variables $\big (\sum _{i = 1}^{N} \frac {1}{x_j  \lambda _{i}}\mu _{N,\beta ;\boldsymbol {\epsilon }}^{V;\mathsf {A}}[\sum _{i = 1}^{N} \frac {1}{x_j  \lambda _{i}}] \big )_{j = 1}^{n}$ which are linear in each of these variables. Therefore, we can integrate over the variable $x_{1}$ in each of these terms by Fubini’s theorem. The key observation is that $\oint _{\mathsf {A}_h} \sum _{i = 1}^{N} \frac {\mathrm{d} \xi }{2\mathrm{i}\pi }\,\frac {1}{\xi  \lambda _i}$ is the number $N_{h}$ of $\lambda _i$ s belonging to $\mathsf {A}_h$ . Since $N_h$ is deterministic in the fixed filling fraction model, it is equal to its expectation, and therefore, each of these terms vanish which implies (1.11) for $n\ge 2$ . When $n=1$ , the cumulant is simply equal to the expectation of $ \sum _{i = 1}^{N} \frac {1}{\xi  \lambda _i}$ , and the previous remark proves (1.11).
We will refer to (1.1) as the initial model and to (1.10) as the model with fixed filling fractions. Standard results from potential theory or a straightforward generalisation of [Reference Anderson, Guionnet and ZeitouniAGZ10, Theorem 2.6.1 and Corollary 2.6.3] imply the following:
Theorem 1.2. Assume V regular and confining on $\mathsf {A}$ . We consider the normalised empirical measures $L_{N,h}=\frac {1}{N_h}\sum _{i=1}^{N_h} \delta _{\lambda _{h,i}}\in \mathcal {P}(\mathsf {A}_h)$ for . Take a sequence $\boldsymbol {N} = (N_1,\ldots ,N_g)$ of gtuple of integers, indexed by N, such that $\sum _{h = 1}^g N_h \leq N$ , and such that $\boldsymbol {N}/N$ converges to a given $\boldsymbol {\epsilon }\in \mathcal {E}$ when $N \rightarrow \infty $ . Then, the law of $(L_{N,h})_{0\le h\le g}$ under $\mu _{N,\beta ;\boldsymbol {N}/N}^{V;\mathsf {A}}$ satisfies a large deviation principle with scale $N^2$ and good rate function
where $\epsilon _0=1\sum _{h=1}^g \epsilon _h$ , $N_0=N\sum _{h=1}^g N_h$ and E is defined in Equation (1.5). As a consequence, the empirical measure $L_{N;\boldsymbol {\epsilon }}=\sum _{h=0}^g\frac {N_h}{N} L_{N,h}$ converges almost surely and in expectation towards the unique probability measure $\mu _\mathrm{{eq};\boldsymbol {\epsilon }}^{V;\mathsf {A}}$ on $\mathsf {A}$ which minimises E among probability measures with fixed mass $\epsilon _h$ on $\mathsf {A}_h$ for any . It is characterised by the existence of constants $C_{\boldsymbol {\epsilon },h}^{V,\mathsf {A}}$ such that
with equality realised $\mu _\mathrm{{eq};\boldsymbol {\epsilon }}^{V;\mathsf {A}}$ almost surely. $\mu _\mathrm{{eq};\boldsymbol {\epsilon }}^{V;\mathsf {A}}$ can be decomposed as a sum of positive measures $\mu _\mathrm{{eq};\boldsymbol {\epsilon },h}^{V}$ having compact support in $\mathsf {A}_h$ , denoted $\mathsf {S}_{\boldsymbol {\epsilon },h}$ . Moreover, if $V^{\{0\}}$ is realanalytic in a neighbourhood of $\mathsf {A}$ , the support $\mathsf {S}_{\boldsymbol {\epsilon },h}$ consists of a finite union of segments.
Later in the text, we shall consider $\mu _\mathrm{{eq};\boldsymbol {N}/N}^{V;\mathsf {A}}$ with $\boldsymbol {N}=(N_1,\ldots ,N_g)$ a vector of positive integers so that $\sum _{h = 1}^{g} N_h < N$ : this will denote the unique solution of (1.12) with $\boldsymbol {\epsilon }=\boldsymbol {N}/N$ . $\mu _\mathrm{{eq}}^{V;\mathsf {A}}$ appearing in Theorem 1.1 coincides with $\mu _\mathrm{{eq};\boldsymbol {\epsilon }_{\star }}^{V}$ for the optimal value $\boldsymbol {\epsilon }_{\star } = (\mu _\mathrm{{eq}}^{V;\mathsf {A}}(\mathsf {A}_{h}))_{1 \leq h \leq g}$ , and in this case, $\mathsf {S}_{\boldsymbol {\epsilon }_{\star },h}$ is actually the segment $[\alpha ^{}_h,\alpha _{h}^+]$ . The key point – justified in Appendix 1 – is that, for $\boldsymbol {\epsilon }$ close enough to $\boldsymbol {\epsilon }_{\star }$ , the support $\mathsf {S}_{\boldsymbol {\epsilon },h}$ remains connected, and the model with fixed filling fractions enjoys a $\frac {1}{N}$ expansion.
Theorem 1.3. If V satisfies Hypotheses 1.1 and 1.3 on $\mathsf {A}$ , there exists $t> 0$ such that, uniformly for integers $\boldsymbol {N}=(N_1,\ldots ,N_g)$ such that $\boldsymbol {N} /N\in \mathcal {E}$ and $\boldsymbol {N}/N \boldsymbol {\epsilon }_{\star }_1 < t$ , we have an expansion for the correlators, for any $K \geq 0$ ,
Up to a fixed $O(N^{(K + 1)})$ and for a fixed n, Equation (1.13) holds uniformly for $x_1,\ldots ,x_n$ in compact regions of $\mathbb {C}\setminus \mathsf {A}$ . The $W_{n;\boldsymbol {\epsilon }}^{\{k\}}$ can be extended into smooth functions of $\boldsymbol {\epsilon }\in \mathcal {E}$ close enough to $\boldsymbol {\epsilon }_{\star }$ .
We prove this theorem, independently of the nature soft/hard of the edges, in Section 5 for realanalytic potential (i.e., Hypothesis 1.2 instead of 1.3). For $\beta = 2$ and potential V independent of N, the coefficients of expansion $W_{n;\boldsymbol {N}/N}^{\{k\}} = 0$ are zero for $k = (n + 1) \,\,\mathrm{mod} 2$ , as is well known for hermitian random matrix models (see (1.16) and the remarks on $\beta $ dependence in Section 1.5). The result is extended to harmonic potentials (i.e., Hypothesis 1.3) in Section 6.1. In Proposition 5.6, we provide an explicit control of the errors in terms of the distance of $x_1,\ldots ,x_k$ to $\mathsf {A}$ , and its proof makes clear that the expansion of the correlators is not expected to be uniform for $x_1,\ldots ,x_n$ chosen in a compact of $\mathbb {C}\setminus \mathsf {A}$ independently of n and K (namely, it is uniform only for K fixed). Note that we will sometimes omit to specify the dependence in $\mathsf {A},V$ , etc. in the notations (e.g., for the equilibrium measure, for the correlators and their coefficient of expansions), but we will at least include it when this dependence is of particular importance.
We then compute in Section 7 the expansion of the partition function, thanks to the expansion of $W_{1;\boldsymbol {N}/N}$ and $W_{2;\boldsymbol {N}/N}$ , by an interpolation that reduces the strength of pairwise interactions between eigenvalues in different segments while preserving the equilibrium measure. At the end of the interpolation, we are left with a product of $(g + 1)$ partition functions in a onecut regime, for which the asymptotic expansion was established in [Reference Borot and GuionnetBG11].
Theorem 1.4. If V satisfies Hypotheses 1.1 and 1.3 on $\mathsf {A}$ , there exists $t> 0$ such that, uniformly for gdimensional vectors of positive integers $\boldsymbol {N}$ such that $\boldsymbol {N} /N\in \mathcal {E}$ and $\boldsymbol {N}/N \boldsymbol {\epsilon }_{\star }_1 < t$ , we have for any $K \geq 0$ ,
with
Besides, $F^{\{k\};V}_{\beta ;\boldsymbol {\epsilon }}$ extends to a smooth function of $\boldsymbol {\epsilon }$ close enough to $\boldsymbol {\epsilon }_{\star }$ , and at the value $\boldsymbol {\epsilon } = \boldsymbol {\epsilon }_{\star }$ , the first derivatives of $F^{\{2\};V}_{\beta ;\boldsymbol {\epsilon }}$ vanish and its Hessian is negative definite.
We can identify explicitly the following:
where
is the entropy. The formula for $F^{\{2\};V}_{\beta ;\boldsymbol {\epsilon }}$ is obvious from potential theory, while the formula for $F^{\{1\};V}_{\beta ;\boldsymbol {\epsilon }}$ is established in Proposition 7.1 (the first term comes from the fact that we let the potential depend on N). The appearance of the entropy in the term of order N in the free energy is well known in the onecut case, and here we prove that it appears in the same way for the multicut case with fixed filling fractions, and we determine the additional constant. The term $\frac {\beta }{2} N \ln N$ is universal, while the term $\varkappa \ln N$ only depends only on the nature of the endpoints of the support. These logarithmic corrections can already be observed in the asymptotic expansion of Selberg integrals for large N computing the partition function of the classical Jacobi, Laguerre or Gaussian $\beta $ ensembles, corresponding to a onecut regime [Reference Borot and GuionnetBG11]. The fact that the coefficient of $\ln N$ shadows in some way the geometry of the support was observed in other contexts (see, for example, [Reference Cardy and PeschelCP88]) and is not specific to twodimensional Coulomb gases living on a line. Their identification in the multicut regime and fixed filling fractions results from an interpolation with a product of one such model for each cut, which changes only the coefficients of powers of N. Up to a given $O(N^{K})$ , all expansions are uniform with respect to the parameters of the potential and of $\boldsymbol {\epsilon }$ chosen in a compact set so that the assumptions hold. Theorems 1.3–1.4 are the generalisations to the fixed filling fractions model of our earlier results about existence of the $\frac {1}{N}$ expansion in the onecut regime [Reference Borot and GuionnetBG11] (see also [Reference JohanssonJoh98, Reference Albeverio, Pastur and ShcherbinaAPS01, Reference Ercolani and McLaughlinEM03, Reference Bleher and ItsBI05, Reference Guionnet and MaurelSegalaGMS07, Reference Kriecherbauer and ShcherbinaKS10] for earlier results concerning the onecut regime in $\beta = 2$ or general $\beta $ ensembles).
1.5. Relation with Chekhov–Eynard–Orantin topological recursion
Once these asymptotic expansions are shown to exist, by consistency, their coefficients $W_{n;\boldsymbol {\epsilon }}^{\{k\}}$ are computed by the $\beta $ topological recursion of Chekhov and Eynard [Reference Chekhov and EynardCE06]. As a matter of fact, the asymptotic expansion
has a finer structure so that for $n \geq 1$ and $k \geq 1$ , we can write
where $\mathcal {W}_{n;\boldsymbol {\epsilon }}^{[G,l]}$ are the quantities computed by the topological recursion of [Reference Chekhov and EynardCE06]. The initial data consists of the nondecaying terms in the correlators – namely,
All these quantities have an analytic continuation in the variables $x_i$ on the same Riemann surface $\mathcal {C}_{\boldsymbol {\epsilon }}$ called spectral curve. The curve $\mathcal {C}_{\boldsymbol {\epsilon }}$ can, in fact, be defined as the maximal Riemann surface on which $W_{1;\boldsymbol {\epsilon }}^{\{1\}}(x)$ , initially defined for $x \in \mathbb {C} \setminus \mathsf {A}$ , admits an analytic continuation (cf. Section 1.7 for a continued discussion on geometry of spectral curves). The information carried by the decomposition (1.16) is that, if V is chosen independent of $\beta $ and N, all the $\mathcal {W}_{n;\boldsymbol {\epsilon }}^{[G,K]}$ are also independent of $\beta $ and N (except perhaps through the implicit dependence in N of $\boldsymbol {\epsilon }$ ), and thus, the coefficients of the expansions of the correlators display a remarkable structure of Laurent polynomial in $\frac {\beta }{2}$ . This property comes from the structure of the Dyson–Schwinger equations.
From the same initial data, Chekhov and Eynard also define numbers $W_{0;\boldsymbol {\epsilon }}^{[G,K]} = \mathcal {F}_{\boldsymbol {\epsilon }}^{[G,K]}$ , which give the coefficients of the asymptotic expansion of the free energy $\ln Z_{N,\beta ;\boldsymbol {N}/N}^{V;\mathsf {A}}$ up to an integration constant independent of the potential, and which are independent of $\beta $ provided V is chosen independent of $\beta $ . More precisely, we mean that for any two potentials V and $\tilde {V}$ satisfying the assumptions of Theorem 1.4 and leading to a $(g + 1)$ cut regime, we must have for $k \geq 2$ , by consistency with [Reference Chekhov and EynardCE06],
In particular, the topological recursion defines $\mathcal {F}^{[0,0];V}_{\boldsymbol {\epsilon }} = E[\mu _{\mathrm{eq};\boldsymbol {\epsilon }}^V]$ and $\mathcal {F}^{[0,1];V}_{\boldsymbol {\epsilon }} = \mathrm{Ent}[\mu _{\mathrm{eq};\boldsymbol {\epsilon }}^{V}]$ . By comparison with (1.15), we arrive to an absolute comparison (here, assume the potential to be independent of N – i.e., $V = V^{\{0\}}$ ):
The constant in the second line was not computed in [Reference Chekhov and EynardCE06]. To our knowledge, the absolute – including a $\beta $ dependent, possibly gdependent but otherwise Vindependent constant – comparison between the coefficients $F_{\beta ;\boldsymbol {\epsilon }}^{\{k\};V}$ of the asymptotic expansion of the $\beta $ ensembles and the invariants $\mathcal {F}^{[G,m]}$ for $(G,m) \neq (0,0),(0,1)$ produced by the topological recursion has not been performed in full generality. It is only known for $\beta = 2$ for all G in the onecut regime; see [Reference MarchalMar17, Proposition 2.5].
When $\beta = 2$ , only $\mathcal {W}_{n;\boldsymbol {\epsilon }}^{[G]} = \mathcal {W}_{n;\boldsymbol {\epsilon }}^{[G,0]}$ and $\mathcal {F}_{\boldsymbol {\epsilon }}^{[G]} = \mathcal {F}^{[G,0]}_{\boldsymbol {\epsilon }}$ appear. These are the quantities defined by the Chekhov–Eynard–Orantin topological recursion [Reference Eynard and OrantinEO07], and we retrieve the usual asymptotic expansions
involving only powers of $\frac {1}{N}$ with parity $(1)^n$ in the npoint correlators and powers of $\frac {1}{N^2}$ in the free energy.
1.6. Main results in the multicut regime: partition function
Let us come back to the initial model (1.1). We can always take $\mathsf {A} = \bigcup _{h = 0}^{g} \mathsf {A}_h \subseteq \mathsf {B}$ to be a small enlargement of the support $\mathsf {S}$ respecting the setup of § 1.4. It is indeed well known that the partition function $Z_{N,\beta }^{V;\mathsf {B}}$ can be replaced by $Z_{N,\beta }^{V;\mathsf {A}}$ up to exponentially small corrections when N is large (see [Reference Pastur and ShcherbinaPS11, Reference Borot and GuionnetBG11] for results in this direction, and we give a proof for completeness in § 3.1 below). The latter can be decomposed as a sum over all possible ways of distributing the $\lambda $ s between the segments $\mathsf {A}_h$ – namely,
where we have denoted $N_0 = N  \sum _{h = 1}^{g} N_h$ the number of $\lambda $ s put in the segment $\mathsf {A}_0$ . So we can use our results for the model with fixed filling fractions to analyse the asymptotic behaviour of each term in the sum and then find the asymptotic expansion of the sum taking into account the interference of all contributions. This is carried out in Section 8.1.
Before stating the results, we need two ingredients. First, we let $\mathfrak {Z}_{N,\beta ;\boldsymbol {\epsilon }}^{V;\mathsf {A}}$ be the (truncated at an arbitrary order K) asymptotic series depending on a gdimensional vector with positive entries, at least when its coefficients are defined:
If we substitute $\boldsymbol {\epsilon } = \boldsymbol {N}/N$ as in Theorem 1.4, it gives the asymptotic expansion of the partition function of the fixed filling fractions model with unordered eigenvalues, and we recall that $F^{\{k\};V}_{\beta ;\boldsymbol {\epsilon }}$ exists as a smooth function of $\boldsymbol {\epsilon }$ in some nonempty open set. We shall denote $(F^{\{k\};V}_{\beta ;\boldsymbol {\epsilon }})^{(j)}$ the tensor of jth derivatives with respect to $\boldsymbol {\epsilon }$ .
Second, we introduce the Siegel Theta function with characteristics $\boldsymbol {\mu },\boldsymbol {\nu } \in \mathbb {C}^{g}$ . If $\boldsymbol {\tau }$ is a symmetric $g \times g$ matrix of complex numbers such that $\mathrm {Im}\,\boldsymbol {\tau }> 0$ , the Siegel Theta function is the entire function of $\boldsymbol {v} \in \mathbb {C}^{g}$ defined by the exponentially fast converging series
Among its essential properties, we mention the following:

○ for any characteristics $\boldsymbol {\mu },\boldsymbol {\nu }$ , it satisfies the diffusionlike equation $4\mathrm{i}\pi \partial _{\tau _{h,h'}}\vartheta = \partial _{v_h}\partial _{v_{h'}}\vartheta $ .

○ it is a quasiperiodic function with lattice $\mathbb {Z}^{g} \oplus \boldsymbol {\tau }(\mathbb {Z}^{g})$ : for any $\boldsymbol {m}_0,\boldsymbol {n}_0 \in \mathbb {Z}^{g}$ ,
$$ \begin{align*}\vartheta\!\left[\begin{array}{@{\hspace{0.02cm}}l@{\hspace{0.02cm}}} \boldsymbol{\mu} \\ \boldsymbol{\nu} \end{array}\right]\!\!(\boldsymbol{v} + \boldsymbol{m}_0 + \boldsymbol{\tau}\cdot\boldsymbol{n}_0\boldsymbol{\tau}) = \exp\big(2\mathrm{i}\pi\boldsymbol{m}_0\cdot\boldsymbol{\mu}  2\mathrm{i}\pi \boldsymbol{n_0}\cdot(\boldsymbol{v} + \boldsymbol{\nu})  \mathrm{i}\pi\boldsymbol{n}_0\cdot\boldsymbol{\tau}\cdot\boldsymbol{n}_0\big)\,\vartheta\!\left[\begin{array}{@{\hspace{0.02cm}}l@{\hspace{0.02cm}}} \boldsymbol{\mu} \\ \boldsymbol{\nu} \end{array}\right]\!\!(\boldsymbol{v}\boldsymbol{\tau}). \end{align*} $$ 
○ it has a nice transformation law under $\boldsymbol {\tau } \rightarrow (\boldsymbol {A\tau } + \boldsymbol {B})(\boldsymbol {C\tau } + \boldsymbol {D})^{1}$ , where $\boldsymbol {A},\boldsymbol {B},\boldsymbol {C},\boldsymbol {D}$ are the $g \times g$ blocks of a $2g \times 2g$ symplectic matrix [Reference MumfordMum84].

○ when $\boldsymbol {\tau }$ is the matrix of periods of a genus g Riemann surface, it satisfies the Fay identity [Reference FayFay70].
We define the gradient operator $\nabla _{\boldsymbol {v}}$ acting on the variable $\boldsymbol {v}$ of this function. For instance, the diffusion equation takes the form $4\mathrm{i}\pi \partial _{\boldsymbol {\tau }}\vartheta = \nabla ^{\otimes 2}_{\boldsymbol {v}}\vartheta $ .
Theorem 1.5. Assume Hypotheses 1.1 and 1.3. Let $\boldsymbol {\epsilon }_{\star } = (\mu _\mathrm{{eq}}^V[\mathsf {S}_h])_{1 \leq h \leq g}$ – we shall replace all indices $\boldsymbol {\epsilon }$ by $\star $ in our notations to indicate a specialisation at $\boldsymbol {\epsilon } = \boldsymbol {\epsilon }_\star $ . Then, the partition function has an asymptotic expansion of the form, with $\mathsf {C}=\mathsf {B}$ or $\mathsf {A}$ , for any $K \geq 2$ ,
In this expression, $\mathfrak {Z}_{N,\beta ;\star }^{V;\mathsf {A}}$ is the asymptotic series defined in Equation (1.19) and evaluated at $\boldsymbol {\epsilon } = \boldsymbol {\epsilon }_{\star }$ . If $\boldsymbol {X}$ is a vector with g components, we set $T^{\{0\}}_{\beta ;\boldsymbol {\epsilon }}[\boldsymbol {X}] = 1$ , and for $k \geq 1$ ,
where $\cdot $ denotes the standard scalar product on the tensor space. We have also introduced
Being more explicit but less compact, we may rewrite
For $\beta = 2$ , this result has been derived heuristically to leading order in [Reference Bonnet, David and EynardBDE00] and to all orders in [Reference EynardEyn09]. These heuristic arguments can be extended straightforwardly to all values of $\beta $ ; see, for example, [Reference BorotBor11]. Our work justifies their heuristic argument. To prove this result, we exploit the Dyson–Schwinger equations for the $\beta $ ensemble with fixed filling fractions taking advantage of a rough control on the large N behaviour of the correlators. The result of Theorem 1.5 has been derived up to $o(1)$ by Shcherbina [Reference ShcherbinaShc12] for realanalytic potentials, with different techniques, based on the representation of $\prod _{h < h'} \prod _{i,j} \lambda _{h,i}  \lambda _{h',j}^{\beta }$ , which is the exponential of a quadratic statistic, as expectation value of a linear statistics coupled to a Brownian motion. The rough a priori controls on the correlators do not allow at present the description of the $o(1)$ by such methods. The results in [Reference ShcherbinaShc12] were also written in a different form: $F^{\{0\};V}_{\beta ;\boldsymbol {\epsilon }}$ appearing in $\mathfrak {Z}$ was identified with a combination of Fredholm determinants (see also the physics paper [Reference Wiegmann and ZabrodinWZ06]), while this representation does not come naturally in our approach. Also, the steps undertaken in Section 8 where we replace the sum over nonnegative integers such that $N_0 + \cdots + N_g = N$ in Equation (1.18), by a sum over $\boldsymbol {N} \in \mathbb {Z}^{g}$ , thus reconstructing the Siegel Theta function, was not performed in [Reference ShcherbinaShc12].
The $2\mathrm{i}\pi $ appears because we used the standard definition of the Siegel Theta function,and should not hide the fact that all terms in Equation (1.23) are realvalued. Here, the matrix
involved in the Theta function has purely imaginary entries, and $\mathrm {Im}\,\boldsymbol {\tau }_{\beta ;\star }$ is definite positive according to Theorem 1.4; hence, the Theta function in the righthand side makes sense. Notice also that for it is $\mathbb {Z}^g$ periodic with respect to $\boldsymbol {\mu }$ ; hence, we can replace $N\boldsymbol {\epsilon }_{\star }$ by $N\boldsymbol {\epsilon }_{\star } + \lfloor N\boldsymbol {\epsilon }_{\star } \rfloor $ , and this is responsible for modulations in the asymptotic expansion, and thus breakdown of the $\frac {1}{N}$ expansion. Still, the model has ‘subsequential’ asymptotic expansions in $\frac {1}{N}$ . For instance, for an even potential with two cuts ( $g = 1$ ) model, we have $\epsilon _{\star } = \frac {1}{2}$ , so $(N\epsilon _{\star }\,\,\mathrm{mod}\,\,\mathbb {Z})$ appearing as characteristic in the Theta function only depends on the parity of N, and for each fixed parity, we get an asymptotic expansion in $\frac {1}{N}$ . In fact, having an even potential implies that the fixedfilling fraction model is invariant under $\epsilon \rightarrow 1  \epsilon $ , so only the terms with even numbers $j_i$ of derivatives with respect to filling fractions contribute in $T^{\{k\}}_{\beta ;\boldsymbol {\epsilon }}$ . If, furthermore, $\beta = 2$ , only the $(F^{\{k\};V}_{\beta = 2;\star })^{(j)}$ with k even survive, and we deduce that the same is true for $T^{\{k\}}_{\beta =2;\star }$ , so that the logarithm of the partition function has an asymptotic expansion in $\frac {1}{N^2}$ for N odd and different asymptotic expansion in $\frac {1}{N^2}$ for N even (of course, up to the universal logarithmic corrections $\frac {\beta }{2}N\ln N + \varkappa \ln N$ ).
Let us give the two first orders of Equation (1.23):
and:
For $\beta = 2$ , unlike the onecut regime where the asymptotic expansion was in $\frac {1}{N^2}$ up to constants independent of the potential, the multicut regime features an asymptotic expansion with nontrivial terms in powers of $\frac {1}{N}$ . For instance, we have a contribution at order $\frac {1}{N}$ of
In a twocuts regime ( $g = 1$ ), a sufficient condition for all terms of order $N^{(2k + 1)}$ to vanish (again, up to integration constants already present in $\mathfrak {Z}$ ) is that $\epsilon _{\star } = \frac {1}{2}$ and $Z_{N,\beta = 2;\epsilon }^{V;\mathsf {A}} = Z_{N,\beta = 2;1  \epsilon }^{V;\mathsf {A}}$ , for the same reasons that we mentioned for the case of an even potential with two cuts. In such a case, we have an expansion in powers of $\frac {1}{N^2}$ for the partition function, whose coefficients depend on the parity of N. In general, we also observe that $\boldsymbol {v}_{\beta = 2;\star } = \boldsymbol {0}$ (i.e., Thetanullwerten appear in the expansion).
Using the fact that the nth correlator is the nderivative of the free energy of the partition function for a perturbed potential or order $1/N$ , and our asymptotic results are uniform for small perturbations of this kind, it is pure algebra to derive from (1.5) an asymptotic expansion for the correlators $W_n$ for the initial model in the multicut regime. For $\beta = 2$ , the resulting expression can be found, for instance, in [Reference Borot and EynardBE11, Section 6.2] up to $O(\frac {1}{N})$ and a systematic diagrammatic for all orders is given in [Reference Borot and EynardBE12, Appendix A]. This can be straightforwardly extended to the $\beta \neq 2$ case simply by including halfinteger genera g (in our conventions, k not having fixed parity).
1.7. Comments relative to the geometry of the spectral curve
We now stress facts from the theory of the topological recursion [Reference Chekhov and EynardCE06, Reference Eynard and OrantinEO07] which are relevant in the present case – for further details on the geometry compact Riemann surfaces, see, for instance, [Reference EynardEyn18]. When V is a polynomial and $\boldsymbol {\epsilon }$ is close enough to $\boldsymbol {\epsilon }_{\star }$ , the density of the equilibrium measure can be analytically continued to a hyperelliptic curve of genus g, denoted $\mathcal {C}_{\boldsymbol {\epsilon }}$ (the spectral curve). Its equation is
and $\mathcal {C}_{\boldsymbol {\epsilon }}$ is the compactification of the locus of such $(x,y)$ obtained by adding the two points at $\infty $ , where $y \sim x^{g +1}$ (first sheet) and $y \sim x^{g+1}$ (second sheet). Let $\mathcal {A}_h$ be the cycle in $\mathcal {C}_{\boldsymbol {\epsilon }}$ surrounding $\mathsf {A}_{\boldsymbol {\epsilon },h} = [\alpha ^{}_{\boldsymbol {\epsilon },h},\alpha ^{+}_{\boldsymbol {\epsilon },h}]$ . The family $\boldsymbol {\mathcal {A}} = (\mathcal {A}_h)_{1 \leq h \leq g}$ can be completed by a family of cycles $\boldsymbol {\mathcal {B}}$ so that $(\boldsymbol {\mathcal {A}},\boldsymbol {\mathcal {B}})$ is a symplectic basis of homology of $\mathcal {C}_{\boldsymbol {\epsilon }}$ . More precisely, the cycle $\mathcal {B}_h$ travels from $\alpha _{\boldsymbol {\epsilon },h}^{}$ to $\alpha _{\boldsymbol {\epsilon },h  1}^+$ in the second sheet and $\alpha _{\boldsymbol {\epsilon },h  1}^+$ to $\alpha _{\boldsymbol {\epsilon },h}^$ in the first sheet. The correlators $W_{n;\boldsymbol {\epsilon }}^{[G,K]}$ are meromorphic functions on $\mathcal {C}^n_{\boldsymbol {\epsilon }}$ , computed recursively by a residue formula on $\mathcal {C}_{\boldsymbol {\epsilon }}$ .
In particular, the analytic continuation of
is the unique meromorphic bidifferential, denoted $\Omega $ , on $\mathcal {C}_{\boldsymbol {\epsilon }}$ , which has vanishing $\boldsymbol {\mathcal {A}}$ periods and has for only singularity a double pole at coinciding point with leading coefficient $1$ and without residue. This $\Omega $ plays an important role for the geometry of the spectral curve and is called fundamental bidifferential of the second kind. It sometimes appears under the name of ‘Bergman kernel’, although it does not coincide with (but it is related to) the kernel introduced by Bergman in [Reference Bergman and SchifferBS53]. It can be explicitly computed by the formula
where

○ $\theta = \vartheta \big [\begin {smallmatrix} \mathbf {0} \\ \mathbf {0} \end {smallmatrix}\big ]$ is the Riemann Theta function.

○ $\boldsymbol {\varpi }(z)\mathrm {d} x(z)$ is the basis of holomorphic oneforms dual to the $\boldsymbol {\mathcal {A}}$ cycles – that is, characterised by
(1.28) 
○ $\boldsymbol {\tau }^{\mathcal {C}_{\boldsymbol {\epsilon }}}$ is the Riemann matrix of periods of the spectral curve $\mathcal {C}_{\boldsymbol {\epsilon }}$ :

○ $\mathbf {c} = \frac {1}{2}(\mathbf {r} + \boldsymbol {\tau}^{\mathcal{C}_{\boldsymbol{\epsilon}}}(\mathbf {s}))$ with $\mathbf {r},\mathbf {s} \in \mathbb {Z}^{g}$ such that $\mathbf {r} \cdot \mathbf {s}$ is odd, is a nonsingular characteristic for the Theta function (i.e., such that $\theta \big (\int _{z_1}^{z_2} \boldsymbol {\varpi }\mathrm {d} x + \mathbf {c}\,\big \,\boldsymbol {\tau}^{\mathcal{C}_{\boldsymbol{\epsilon}}}\big )$ is not identically $0$ when $z_1,z_2 \in \mathcal {C}_{\boldsymbol {\epsilon }}$ ). Such a $\mathbf {c}$ exists and the result then does not depend on which such $\mathbf {c}$ is chosen.
It is a property of the topological recursion that the derivatives of $F_{\beta ;\boldsymbol {\epsilon }}^{\{k\};V}$ can be computed as $\boldsymbol {\mathcal {B}}$ cycle integrals of the correlators:
This relation extends as well to derivatives of correlators:
where it is understood that we differentiate keeping x fixed. In particular,
Besides, the matrix to use in the Theta function appearing in Theorem 1.5 is
This simple dependence in $\beta $ of $W_{2;\boldsymbol {\epsilon }}^{\{0\}}$ can be traced back to the fact that, as a consequence of the Dyson–Schwinger equations, we have
and this equation (together with the properties of the analytic continuation of $W_{2;\boldsymbol {\epsilon }}^{\{0\}}$ on $\mathcal {C}_{\boldsymbol {\epsilon }}$ and the constraint of vanishing $\boldsymbol {\mathcal {A}}$ periods) fully characterises $W_{2;\boldsymbol {\epsilon }}^{\{0\}}$ .
This relation has a long history and follows from the identification of $F^{\{2\};V}_{\beta ;\boldsymbol {\epsilon }} = \frac {\beta }{2} \mathcal {F}^{[0,0];V}_{\boldsymbol {\epsilon }}$ (cf. Equation (1.17)) with the prepotential of the Hurwitz space associated to the family of curves (1.25) – considered as a Frobenius manifold – computed by Dubrovin [Reference DubrovinDub91], as well as with the tau function of the Whitham hierarchy as shown by Krichever [Reference KricheverKri92]. A derivation in the context of matrix model is, for instance, given in [Reference Chekhov and MironovCM02]. Although a priori differentiability of $F^{\{2\};V}_{\beta ;\boldsymbol {\epsilon }}$ is not justified in [Reference Chekhov and MironovCM02], it is guaranteed by our results of Section A.2.
Equation (1.29) at $\boldsymbol {\epsilon } = \boldsymbol {\epsilon }_{\star }$ can be used to compute $T_{\beta ;\star }^{\{k\}}[\boldsymbol {X}]$ appearing in Equation (1.22). The derivation with respect to $\boldsymbol {\epsilon }$ is not a natural operation in the initial model when N is finite since $N\epsilon _h$ are forced to be integers in Equation (1.10). Yet we show that the coefficients of expansion themselves are smooth functions of $\boldsymbol {\epsilon }$ , and thus, $\partial _{\boldsymbol {\epsilon }}$ makes sense.
1.8. Central limit theorems for fluctuations and their breakdown
In Section 8.2, we describe the fluctuation of the number of particles $N_h$ in each segment $\mathsf {A}_h$ : when $N \rightarrow \infty $ , its law is approximated by the law of a Gaussian conditioned to live in a shifted integer lattice. The shift of the lattice oscillates with N by an amount $\lfloor N\epsilon _{\star ,h} \rfloor $ . Note that since $N \epsilon _{\star ,h}$ is for general N not an integer, strictly speaking, one cannot say that it converges in law to a discrete Gaussian random variable. This is, however, true along subsequences of N in case $\epsilon _{\star ,h} = \mu _{\mathrm{eq}}^{V}(\mathsf {A}_h)$ is a rational number.
Theorem 1.6. Assume Hypotheses 1.1 and 1.3, and let $\boldsymbol {N} = (N_1,\ldots ,N_g)$ be the vector of filling fractions as above. If $\boldsymbol {P}$ is a gtuple of integers depending on N and such that $\boldsymbol {P}  N\boldsymbol {\epsilon }_{\star } = o(N^{\frac {1}{3}})$ when $N \rightarrow \infty $ , we have
In Section 8.3, we describe the fluctuations of linear statistics in the multicut regime.
Theorem 1.7. Assume Hypotheses 1.1 and 1.3. Let $\varphi $ be an analytic test function in a neighbourhood of $\mathsf {A}$ , and $s \in \mathbb {R}$ . We have when $N \rightarrow \infty $ ,
where
We recall that the $\varpi _h(x)\mathrm {d} x$ are the holomorphic oneforms from Equations (1.28)–(1.30), while $W_{1;\boldsymbol {\epsilon }}^{\{0\}}$ and $W_{2;\boldsymbol {\epsilon }}^{\{0\}}$ appear in the asymptotic expansion of the correlators in the model with fixed filling fractions (Theorem 1.3), and here they must be specialised at $\boldsymbol {\epsilon } = \boldsymbol {\epsilon }_{\star }$ .
Remark 1.4. In particular, $\boldsymbol {u}_{\beta ;\star }$ is a linear map associating to a test function $\varphi $ a gdimensional vector. When $\varphi $ is such that $\boldsymbol {u}_{\beta ;\star }[\varphi ] = 0$ , the Theta functions cancel out, and we deduce that the random variable
converges in law to a Gaussian random variable with mean $M_{\beta ;\star }[\varphi ]$ and covariance $Q_{\beta ;\star }[\varphi ,\varphi ]$ . We remark that we have the alternative formula from (8.10):
showing that $\boldsymbol {u}_{\beta ;\star }[\varphi ] $ vanishes when $ \boldsymbol {\epsilon }_{\star }$ is a critical point of $\int _{\mathsf {S}} \varphi (\xi )\,\mathrm {d}\mu _{\mathrm{eq};\boldsymbol {\epsilon }}^V(\xi )$ . Even though our results are obtained for analytic potentials and test functions, this condition clearly makes sense with less regularity. In fact, it is possible to generalise our results and techniques to consider sufficiently smooth potential and test functions instead of analytic ones. We refer the interested reader to [Reference GuionnetG19, Sections 4 and 6] to such a generalisation in the onecut case.
When $\boldsymbol {u}_{\beta ;\star }[\varphi ] \neq 0$ , the central limit theorem does not hold anymore. Instead, from the shape of the righthand side, $\Phi _N[\varphi ]$ is approximated when $N \rightarrow \infty $ by the sum of two independent random variables: the first one is a Gaussian random variable with mean $M_{\beta ;\star }[\varphi ]$ and covariance $Q_{\beta ;\star }[\varphi ,\varphi ]$ , and the second one is the scalar product with $2\mathrm{i}\pi \boldsymbol {u}_{\beta ;\star }[\varphi ]$ (which is a vector in $\mathbb {R}^g$ when $\varphi $ is realvalued) of a random Gaussian vector conditioned to live on the lattice $\lfloor N\boldsymbol {\epsilon }_{\star } \rfloor + \mathbb {Z}^{g}$ . This also displays Ndependent oscillations. These oscillations can be interpreted in physical terms from tunnelling of particles between different segments. One sees, indeed, than moving a single $\lambda _i$ from $\mathsf {A}_h$ to $\mathsf {A}_{h'}$ changes $\Phi _N[\varphi ]$ by a quantity of order $1$ , which is already the typical order of fluctuation of linear statistics when filling fractions are fixed.
The next term in the asymptotic expansion of the lefthand side of (1.32) is of relative order $O(\frac {1}{N})$ , which therefore gives the speed of convergence of the associated linear statistics of the empirical measure.
1.9. Asymptotic expansion of kernels and correlators
Once the result on large N expansion of the partition function is obtained, we can easily infer the asymptotic expansion of the correlators and the kernels by perturbing the potential by terms of order $\frac {1}{N}$ , maybe complexvalued, as allowed by Hypothesis 1.3.
1.9.1. Leading behaviour of the correlators
Although we could write down the expansion for the correlators as a corollary of Theorem 1.5, we bound ourselves to point out their leading behaviour. Whereas $W_n$ behaves as $O(N^{2  n})$ in the onecut regime or in the model with fixed filling fractions, $W_n$ for $n \geq 3$ does not decay when N is large in a $(g + 1)$ cut regime with $g \geq 1$ . More precisely, we have the following.
Theorem 1.8. Assume Hypothesis 1.1 and 1.3 and that the number of cuts $(g + 1)$ is greater or equal to $2$ . When $N \rightarrow \infty $ , we have, uniformly when $x_1,\ldots ,x_n$ belongs to any compact of $(\mathbb {C}\setminus \mathsf {A})^n$ ,
and for any $n \geq 3$ ,
Integrating this result over $\boldsymbol {\mathcal {A}}$ cycles provides the leading order behaviour of nth order moments of the filling fractions $\boldsymbol {N}$ , and the result agrees with Theorem 1.6.
1.9.2. Kernels
We explain in § 6.3 that the following result concerning the kernel – defined in Equation (1.4) – is a consequence of Theorem 1.3:
Corollary 1.9. Assume Hypothesis 1.1 and 1.3. There exists $t> 0$ such that, for any sequence of $\boldsymbol {N} = (N_1,\ldots ,N_{g})$ such that $\boldsymbol {N}/N  \boldsymbol {\epsilon }_{\star }_1 < t$ , the npoint kernels in the model with fixed filling fractions have an asymptotic expansion when $N \rightarrow \infty $ of the form, for any $K \geq 0$ ,
where $\mathcal {L}_{\boldsymbol {x},\boldsymbol {c}}$ is the linear form
The error terms in this expansion are uniform for $x_1,\ldots ,x_n$ in any compact of $\mathbb {C}\setminus \mathsf {A}$ .
The $(r,k) = 1$ term in (1.33) depends on choices for the path of integration from $\infty $ to $x_j$ (the other terms do not and are also unaffected by the difference between f and $\check {f}$ in (1.34)), and $\chi _j \in \mathbb {Z}$ . These two features are a manifestation of the fact that the definition of the kernel depends on a choice of determination for the complex logarithm; resolving them by the choice of suitable determinations and domain of definition leads to specific integer values for $\chi _j$ . These subtleties are explained in details in § 6.3 and can be ignored if all $c_j \in \mathbb {Z}$ (in that case, the definition of the kernel does not depend on choices).
Hereafter, if $\gamma $ is a smooth path in $\mathbb {C}\setminus \mathsf {S}_{\boldsymbol {\epsilon }}$ , we set $\mathcal {L}_{\gamma } = \int _{\gamma }$ , and $\mathcal {L}_{\gamma }^{\otimes r}$ is given by
A priori, the integrals in the righthand side of Equation (1.33) depend on the relative homology class in $\mathbb {C}\setminus \mathsf {A}$ of paths between $\infty $ to $x_i$ . A basis of homology cycles in $\mathbb {C}\setminus \mathsf {A}$ is given by $\overline {\boldsymbol {\mathcal {A}}} = (\mathcal {A}_h)_{0 \leq h \leq g}$ , and we deduce from Equation (1.11) that
Therefore, the only multivaluedness of the righthand side comes from the first term $N \mathcal {L}_{\boldsymbol {x},\mathbf {c}}[W_{1;\boldsymbol {\epsilon }}^{\{1\}}]$ , and given Equation (1.35) and observing that $N_h = N\epsilon _h$ are integers, we see that it exactly reproduces the monodromies of the kernels depending on $c_j$ .
We now come to the multicut regime of the initial model. If $\boldsymbol {X}$ is a vector with g components, and $\mathcal {L}$ is a linear form on the space of holomorphic functions on $\mathbb {C}\setminus \mathsf {S}_{\boldsymbol {\epsilon }}$ , let us define
where we took as convention $W_{n = 0;\boldsymbol {\epsilon }}^{\{k\}} = F_{\beta ;\boldsymbol {\epsilon }}^{\{k\}}$ and the derivatives are computed for fixed xs. Then, as a consequence of Theorem 1.5, we have the following.
Corollary 1.10. Assume Hypothesis 1.1 and 1.3. With the notations of Corollary 1.9, the npoint kernels have an asymptotic expansion, for any $K \geq 0$ ,
The first factor comes from evaluation of the righthand side of Equation (1.33) at $\boldsymbol {\epsilon } = \boldsymbol {\epsilon }_{\star }$ , ${\mathcal {L}}_{\boldsymbol {x},\boldsymbol {c}} = \sum _{j = 1}^{n} c_j\int _{\infty }^{x_j}$ and $\boldsymbol {\varpi }\mathrm{{d}} x$ is the basis of holomorphic oneforms.
A diagrammatic representation for the terms of such expansion was proposed in [Reference Borot and EynardBE12, Appendix A].
1.10. Strategy of the proof
The key idea of this article is to establish an asymptotic expansion for the partition functions of our models for fixed filling fractions:
for any $K \geq 0$ . Indeed, such an expansion allows to estimate the free energy of the original model $\ln Z_{N,\beta }^{V;\mathsf {A}}$ up to errors of order $O(N^{K1 + \delta })$ ; see (1.18) and Theorem 1.5. It also allows to analyse the asymptotic distribution of the filling fractions $\boldsymbol {\epsilon } = \boldsymbol {N}/N$ (see Theorem 1.6) since this distribution is given as the following ratio of partition functions:
In particular, if (1.36) is known up to $o(1)$ , the leading behaviour of (1.37) when $N \rightarrow \infty $ can be computed. This analysis is detailed in Section 8.2.
To handle fluctuations of linear statistics, we use the wellknown approach of considering the free energy for perturbations of order $\frac {1}{N}$ of the potential. In fact, if we denote by $\Phi _N[\varphi ] := \sum _{i = 1}^N \varphi (\lambda _i)  N\int _{\mathsf {S}} \varphi (\xi )\mathrm {d}\mu _{\mathrm{eq}}^{V}(\xi ) $ , as in Remark 1.4, we see that for any real number s,
Again, the expansion of the free energies up to $o(1)$ allows to derive the asymptotics of the Laplace transform of $\Phi _{N}[\varphi ]$ and hence the central limit theorem; see Section 8.3. Note in passing that another way to study these fluctuations is to first condition the law $\mu _{N,\beta }^{V;\mathsf {A}}$ by fixing its filling fractions to be equal to some $\boldsymbol {N}$ . Indeed, we can also recover the fluctuations of the linear statistics from those under the conditioned law (that can be deduced from the ratio of the partition functions of Theorem 1.4 and lead to classical central limit theorems with Gaussian limits), together with the fluctuations of the filling fractions. Then, one easily sees that the term $\boldsymbol {u}_{\beta ;\star }$ comes from the fluctuations of the filling fractions and more precisely from the difference of centerings $N(\int _{\mathsf {S}} \varphi (\xi )\mathrm {d}\mu _{\mathrm{eq}}^{V}(\xi )\int _{\mathsf {S}} \varphi (\xi )\mathrm {d}\mu _{\mathrm{eq};\boldsymbol {N}/N}^{V}(\xi ))$ for varying $\boldsymbol {N}/N$ .
Therefore, the central result of this article is Theorem 1.5. To prove this theorem, we shall as in [Reference Borot and GuionnetBG11] interpolate between the partition functions we are interested in and explicitly computable reference partition functions. For the latter, we take a product of partition functions of onecut models with Gaussian, Laguerre or Jacobi weight (depending on the nature of edges, soft or hard, of the equilibrium measure one wishes to match) that are evaluated as Selberg integrals. Such reference partition functions were already used in [Reference Borot and GuionnetBG11]. One important new element of the present analysis is the interpolation from a model with several cuts to independent onecut models. This is realised by considering the sdependent model
for $s \in [0,1]$ . We choose to take the sdependent potential $V_h(x) = T^s_h(x)$ on the hsegment
where V is the potential of the original model. This choice is such that the equilibrium measure associated with the model $Z_{N,\beta ;\boldsymbol {\epsilon }}^{T_{s};\mathsf {A}}(s) $ is the equilibrium measure of the original model; see Section 7.4. Moreover, $Z_{N,\beta ;\boldsymbol {\epsilon }}^{V;\mathsf {A}}=Z_{N,\beta ;\boldsymbol {\epsilon }}^{T_{1};\mathsf {A}}(1) $ , whereas $Z_{N,\beta ;\boldsymbol {\epsilon }}^{T_{0};\mathsf {A}}(0)$ is a product of models whose equilibrium measure has only onecut (they are the restriction of the equilibrium measure of the original model to each of the connected pieces of its support), which we can compute by [Reference Borot and GuionnetBG11] (see Section 7.1). Interpolating along this family yields
It is important to note that in the first equality, the singularity of the logarithm is away from the range of integration as it involves variables in distinct segments, so we could express (1.38) in terms of analytic linear and quadratic statistics, which, in turn, can be expressed in terms of the correlators $W_{n;\boldsymbol {\epsilon }}^{s}$ of the model associated with $Z_{N,\beta ;\boldsymbol {\epsilon }}^{T_{s};\mathsf {A}}(s)$ . Lemma 7.5 gives the large N expansion of these correlators.
These expansions are based on the socalled Dyson–Schwinger equations (4.1); see also (7.36) for the correlators of the interpolating models. These equations are exact equations satisfied by the correlators for any fixed N and obtained simply by integration by parts. They are a priori not closed, but the idea is to show that they are asymptotically closed so that if we can show that the correlators have a large N expansion of topological type, their coefficients will satisfy a closed system of equations. The latter is based on the fact that coefficients beyond the leading order satisfy an inhomogeneous linear equation, with inhomogeneous term involving coefficients of lower order only. Hence, solving the linear equation allows to define uniquely and recursively all the coefficients in the expansion of the correlators. The linear equation is described by a linear operator, called the masteroperator, that we denote $\mathcal K$ (see (5.6)) and which is the same for all orders. An inversion of this operator (continuously on some function space) precisely allows to solve the linear equation.
The central point of our approach is therefore to invert the operator $\mathcal K$ . In fact, the operator is not invertible but rather has a kernel of dimension at least g, where $(g + 1)$ is the number of cuts (i.e., connected components of the support of the equilibrium measure). However, its extension $\hat {\mathcal K}$ , where we also record the periods around the cuts, is invertible in an offcritical situation; see Section 5.2.3. Fixing the filling fractions exactly amounts to use the extended operator $\widehat {\mathcal {K}}$ instead of $\mathcal {K}$ , and this is why we first consider the model with fixed filling fractions. The invertibility of the extended operator indeed allows us not only to formally solve the Dyson–Schwinger equations but also to show the existence of this asymptotic expansion to all orders in $\frac {1}{N}$ . To this end, it is necessary to use a priori rough estimates on the correlators, which we obtain by classical methods of concentration of measure and large deviations; see Section 3. These estimates can be improved iteratively with the Dyson–Schwinger equations (see, for example, Section 5.3) to obtain optimal estimates and eventually reach the allorder asymptotic expansion. This bootstrap strategy was first introduced in [Reference Borot and GuionnetBG11] for the onecut model. We detail these computations in the case where $s=1$ in Section 5. We also need to carry this out for the interpolating sdependent model in order to have asymptotic expansions to insert in (1.38). In that case, the extended operator does not have an explicit inverse, but we can nevertheless show by Fredholm arguments that it is invertible. Then we indicate in Section 7 the modifications to take into account the previous bootstrap argument for $s\in [0,1]$ .
We stress again that we cannot use the inversion and bootstrap strategy in the Dyson–Schwinger equations for the correlators of the original model in the multicut regime because the relevant master operator is not invertible. This is the reason why we need the detour through the partition function with fixed filling fractions (via (1.36)), from which any desired expansion of the correlators of the original model can be obtained by looking at $\frac {1}{N}$ perturbations of the potential.
2. Application to (skew) orthogonal polynomials and integrable systems
The onehermitian matrix model (i.e., $\beta = 2$ ) is related to the Toda chain and orthogonal polynomials (see, for example, [Reference DeiftDei99]). Similarly, the onesymmetric (resp. quaternionic selfdual) matrix model corresponds to $\beta = 1$ (resp. $\beta = 4$ ) and is related to the Pfaff lattice and skeworthogonal polynomials [Reference EynardEyn01, Reference Adler and van MoerbekeAvM02, Reference Adler, Horozov and van MoerbekeAHvM02]. Therefore, our results establish the allorder asymptotics of certain solutions (those related to matrix integrals) of the Toda chain and the Pfaff lattice in the continuum limit, and the allorder asymptotics of (skew) orthogonal polynomials away from the bulk. We illustrate it for orthogonal polynomials with respect to an analytic weight defined on the whole real line. It could be applied equally well to orthogonal polynomials with respect to an analytic weight on a finite union of segments of the real axis. We review with fewer details in § 2.4 the definition of skeworthogonal polynomials and the way to obtain them from Corollary 1.10.
The leading order asymptotic of orthogonal polynomials is well known since the work of Deift et al. [Reference Deift, Kriecherbauer, McLaughlin, Venakides and ZhouDKM+97, Reference Deift, Kriecherbauer, McLaughlin, Venakides and ZhouDKM+99b, Reference Deift, Kriecherbauer, McLaughlin, Venakides and ZhouDKM+99a], using the asymptotic analysis of Riemann–Hilbert problems which was pioneered in [Reference Deift and ZhouDZ95]. In principle, it is possible to push the Riemann–Hilbert analysis beyond leading order, but because this approach is very cumbersome, it has not been performed yet to our knowledge. Notwithstanding, the allorder expansion has a nice structure and was heuristically derived by Eynard [Reference Eynard, Brézin, Kazakov, Serban, Wiegmann and ZabrodinEyn06] based on the general works [Reference Bonnet, David and EynardBDE00, Reference EynardEyn09]. In this article, we provide a proof of those heuristics.
Unlike the Riemann–Hilbert technique, which becomes cumbersome to study the asymptotics of skeworthogonal polynomials (i.e., $\beta = 1$ and $4$ ) and thus has not been performed up to now, our method could be applied without difficulty to those values of $\beta $ and would allow to justify the heuristics of Eynard [Reference EynardEyn01] formulated for the leading order and describe all subleading orders. In other words, it provides a purely probabilistic approach to address asymptotic problems in integrable systems. It also suggests that the appearance of Theta functions is not intrinsically related to integrability. In particular, we see in Theorem 2.2 that for $\beta = 2$ , the Theta function appearing in the leading order is associated to the matrix of periods of the hyperelliptic curve $\mathcal {C}_{\boldsymbol {\epsilon }_{\star }}$ defined by the equilibrium measure. Actually, the Theta function is just the basic block to construct analytic functions on this curve, and this is the reason why it pops up in the Riemann–Hilbert analysis. However, for $\beta \neq 2$ , the Theta function is associated to $\frac {\beta }{2}$ times the matrix of periods of $\mathcal {C}_{\boldsymbol {\epsilon }_{\star }}$ , which might or might not be the matrix of period of a curve, and anyway is not that of $\mathcal {C}_{\boldsymbol {\epsilon }_{\star }}$ . So the monodromy problem solved by this Theta function is not directly related to the equilibrium measure, which makes, for instance, for $\beta = 1$ or $4$ , its construction via Riemann–Hilbert techniques a priori more involved.
Contrary to Riemann–Hilbert techniques, however, we are not yet in position within our method to consider the asymptotic in the bulk or at the edges, or the doublescaling limit for varying weights close to a critical point, or the case of complexvalues weights which has been studied in [Reference Bertola and MoBM09]. It would be very interesting to find a way out of these technical restrictions within our method.
2.1. Setting
We first review the standard relations between orthogonal polynomials on the real line, random matrices and integrable systems see, for example, [Reference Claeys and GravaCG12, Section 5]. In this section, $\beta = 2$ , and we omit to precise it in the notations. Let $V_{\mathbf {t}}(\lambda ) = V(\lambda ) + \sum _{k = 1}^{d} t_k\lambda ^{k}$ . Let $(P_{n,N}(x))_{n \geq 0}$ be the monic orthogonal polynomials associated to the weight $\mathrm {d} w(x) = \mathrm {d} x\,e^{NV_{\mathbf {t}}(x)}$ on $\mathsf {B} = \mathbb {R}$ . We choose V and restrict in consequence $t_k$ so that the weight increases quickly at $\pm \infty $ . If we denote $h_{n,N}$ the $L^2(\mathrm {d} w)$ norm of $P_{n,N}$ , the polynomials $\hat {P}_{n,N} = P_{n,N}/\sqrt {h_{n,N}}$ are orthonormal. They satisfy a threeterm recurrence relation:
The recurrence coefficients are solutions of a Toda chain: if we set
we have
and the coefficients $t_{k}$ generate higher Toda flows. The recurrence coefficients also satisfy the string equations
where $\mathbf {Q}_N$ is the semiinfinite matrix:
The equations 2.2 determine in terms of V the initial condition for the system (2.1). The partition function $\mathcal {T}(\mathbf {t}) = Z_{N}^{V_{\mathbf {t}};\mathbb {R}}$ is the Tau function associated to the solution $(u_{n,N}(\mathbf {t}),v_{n,N}(\mathbf {t}))_{n \geq 1}$ of Equation (2.1). The partition function itself can be computed as [Reference MehtaMeh04, Reference Pastur and ShcherbinaPS11]:
We insist on the dependence on N and V by writing $h_{j,N} = h_{j}(NV)$ . Therefore, the norms can be retrieved as
The regime where $n,N \rightarrow \infty $ but $s = \frac {n}{N}$ remains fixed and positive corresponds to the small dispersion regime in the Toda chain, where $\frac {1}{n}$ plays the role of the dispersion parameter.
2.2. Small dispersion asymptotics of $h_{n,N}$
When $V_{\mathbf {t}_0}/s_0$ satisfies Hypotheses 1.1 and 1.2 for a given set of times $(s_0,\mathbf {t}_0)$ , $V_{\mathbf {t}}/s$ satisfies the same assumptions at least for $(s,\mathbf {t})$ in some neighbourhood $\mathcal {U}$ of $(s_0,\mathbf {t}_0)$ , and Theorem 1.5 determines the asymptotic expansion of $\mathcal {T}_{N}(\mathbf {t})=Z_N^{V_{\mathbf {t}};\mathbb R}$ up to $O(N^{\infty })$ . Besides, we can apply Theorem 1.5 to study the ratio in the righthand side of Equation (2.3) when $n \rightarrow \infty $ up to $o(n^{\infty })$ . For instance, we record below the expansion up to order $O(n^{2})$ .
Theorem 2.1. In the regime $n,N \rightarrow \infty $ , $s = \frac {n}{N}> 0$ fixed, and Hypotheses 1.1 and 1.2 are satisfied with soft edges, we have the following asymptotic expansion:
We used the shortcut notations
and in Equation (2.4), it is understood that the argument $\boldsymbol {v}$ is specialised to $\boldsymbol {0}$ after application of the $\nabla = \nabla _{\boldsymbol {v}}$ . Besides,
When $V_{\mathbf {t}}/s$ leads to a multicut regime, this asymptotic expansion features oscillations. Numerical evidence for such oscillations first appeared in [Reference JurkiewiczJur91], where plots of $h_{n  1,N}/h_{n,N}$ displaying the phase transitions from a onecut to a multicut regime can be found for a sextic potential.
We recall that all the quantities $\mathcal {W}_{m;\star }^{[G]}$ can be computed from the equilibrium measure associated to the potential $V_{\mathbf {t}}$ , so making those asymptotic explicit just requires to solve the scalar Riemann–Hilbert problem for $\mu _\mathrm{{eq}}^{sV_{\mathbf {t}}}$ . Notice that the number $(g + 1)$ of cuts a priori depends on $(s_0,\mathbf {t}_0)$ , and we do not address the issue of transitions between regimes with different number of cuts (because we cannot relax at present our offcriticality assumption), which are expected to be universal [Reference DubrovinDub08].
2.3. Asymptotic expansion of orthogonal polynomials away from the bulk
The orthogonal polynomials can be computed thanks to Heine formula [Reference SzegöSze39]:
Hence, as a consequence of Corollary 1.10, we obtain their asymptotic expansion away from the bulk. We first collect some notations that appeared throughout the introduction, specialised to the case $\beta = 2$ relevant here:
and
where
Theorem 2.2. In the regime $n,N \rightarrow \infty $ , $s = \frac {n}{N}> 0$ fixed, and Hypotheses 1.1 and 1.2 are satisfied, for $x \in \mathbb {C}\setminus \mathsf {S}$ , we have the asymptotic expansion, for any $K \geq 0$ ,
where $\mathcal {L}_{x} = \int _{\infty }^{x}$ . For a given K, this expansion is uniform for x in any compact of $\mathbb {C}\setminus \mathsf {S}$ .
We remark that $\mathcal {L}_{x}[\boldsymbol {\varpi }] = \int _{\infty }^{x} \boldsymbol {\varpi }$ is the Abel map evaluated between the points x and $\infty $ . The variable $s = \frac {n}{N}$ rescales the potential, and therefore, the equilibrium measure and all the coefficient of expansions depend on s.
As such, the results presented in this article do not allow the study of the asymptotic expansion of orthogonal polynomials in the bulk (i.e., for $x \in \mathsf {S}$ ). Indeed, this requires perturbing the potential $V(\lambda )$ by a term $\frac {1}{n}\,\ln (\lambda  x)$ having a singularity at $x \in \mathsf {S}$ , a case going beyond our Hypothesis 1.3. Similarly, we cannot address at present the regime of transitions between a gcut regime and a $g'$ cut regime with $g \neq g'$ because offcriticality was a key assumption in our derivation. Although it is the most interesting in regard of universality, the question of deriving uniform asymptotics, even at the leading order, valid for the crossover around a critical point is still open from the point of view of our methods.
2.4. Asymptotic expansion of skeworthogonal polynomials
The expectation values of $\prod _{i = 1}^N (x  \lambda _i)$ in the $\beta $ ensembles for $\beta = 1$ and $4$ are skeworthogonal polynomials. Let us review this point and just mention that the application of Corollary 1.10 implies allorder asymptotic for skeworthogonal polynomials away from the bulk. Here, the relevant skewsymmetric bilinear products are
A family of polynomials $(P_{N}(x))_{N \geq 0}$ is skeworthogonal if
For a given skewsymmetric product, the family of skeworthogonal polynomials is not unique since one can add to $P_{2N + 1}$ any multiple of $P_{2N}$ , and this does not change the skewnorms $h_N$ . If we add the requirement that the degree $2N$ term in $P_{2N + 1}$ vanishes, the skeworthogonal polynomials are then unique. The generalisation of Heine formula was proved in [Reference EynardEyn01]:
Theorem 2.3. Let $P_{N;n,\beta }$ be a set of monic skeworthogonal polynomials associated to (2.5). We can take
Corollary 1.10 then determines the asymptotics of the righthand side. The partition function itself can be deduced from the skewnorms [Reference MehtaMeh04]
and conversely,
It has been shown that this partition function for $\beta = 1$ is a taufunction of the Pfaff lattice [Reference Adler, Horozov and van MoerbekeAHvM02, Reference Adler and van MoerbekeAvM02]. Here, we obtain its asymptotic expansion from Theorem 1.5.
3. Large deviations and concentration of measure
3.1. Restriction to a vicinity of the support
Our first step is to show that the interval of integration in Equation (1.1) can be restricted to a vicinity of the support of the equilibrium measure up to exponentially small corrections when N is large. The proofs are very similar to the onecut case [Reference Borot and GuionnetBG11], and we recall briefly their idea in § 3.2. Let V be a regular and confining potential, and $\mu _\mathrm{{eq}}^{V;\mathsf {B}}$ the equilibrium measure determined by Theorem 1.1. We denote by $\mathsf {S}$ its (compact) support. We define the effective potential by
when $x \in \mathsf {B}$ , and $+\infty $ otherwise.
Lemma 3.1. If V is regular, is confining, and converges uniformly to $V^{\{0\}}$ on $\mathsf {B}$ , then we have large deviation estimates: for any $\mathsf {F} \subseteq \overline {\mathsf {B}\backslash \mathsf {S}}$ closed in $\mathsf {B}$ and $\mathsf {O} \subseteq \mathsf {B}\backslash \mathsf {S}$ open in $\mathsf {B}$ ,
Definition 3.1. We say that V satisfies a control of large deviations on $\mathsf {B}$ if $\tilde {U}_{\mathrm{eq}}^{V;\mathsf {B}}$ is positive on $\mathsf {B}\setminus \mathsf {S}$ .
Note that $\tilde {U}_{\mathrm{eq}}^{V;\mathsf {B}}$ vanishes at the boundary of $ \mathsf {S}$ . According to Lemma 3.1, such a property implies that large deviations outside $\mathsf {S}$ are exponentially small when N is large.
Corollary 3.2. Let V be regular, confining and satisfying a control of large deviations on $\mathsf {B}$ . Let $ \mathsf {A} \subseteq \mathsf {B}$ be a finite union of segments which contains $\{x\in \mathsf {B}: d(x,\mathsf {S}) \le \delta \}$ for some positive $\delta $ . There exists $\eta (\mathsf {A})> 0$ so that
and for any $n \geq 1$ , there exists a universal constant $\gamma _n> 0$ so that, for any $x_1,\ldots ,x_n \in (\mathbb {C}\setminus \mathsf {B})^{n}$ ,
Note that if all edges are hard, we have $\mathsf {B} = \mathsf {S}$ , and Lemma 3.1 and Corollary 3.2 are useless.
It is useful to have a local version of this result, saying that we can vary endpoints of the segments which are not hard edges for the equilibrium measure, up to exponentially small corrections.
Corollary 3.3. Let V be regular, confining and satisfying a control of large deviations on $\mathsf {B}$ . Let $\mathsf {A} \subseteq \mathsf {B}$ be a finite union of segments which contains $\{x\in \mathsf {B}: d(x,\mathsf {S}) \le \delta \}$ for some positive $\delta $ . If $a_0$ is the left edge of a connected component of $\mathsf {A}$ and $a < a_0$ and is not in $\mathsf {S}$ , let us define $\mathsf {A}_{a} = \mathsf {A} \cup [a,a_0]$ . For any $\varepsilon> 0$ small enough, there exists $\eta _{\varepsilon }> 0$ so that, for N large enough and any $a \in (a_0  \varepsilon ,a_0) \subseteq \mathsf {B}$ , we have
and for N large enough and any $n \geq 1$ and $x_1,\ldots ,x_n \in (\mathbb {C}\setminus \mathsf {A}_{a})$ ,
A similar result holds at the right endpoint of a connected component of $\mathsf {A}$ .
From now on, even though we initially want to study the model on $\mathsf {B}^N$ , we are first going to study the model on $\mathsf {A}^N\kern1.3pt$ , where $\mathsf {A}$ is a small (but fixed) enlargement of $\mathsf {S}$ within $\mathsf {B}$ , as allowed above. In particular, when $\mathsf {S}$ is a disjoint union of finite segments $(\mathsf {S}_{h})_{h = 0}^g$ , we can take $\mathsf {A}$ to be a disjoint union of finite segments $(\mathsf {A}_{h})_{h = 0}^g$ such that $\mathsf {A}_h$ is a neighbourhood of $\mathsf {S}_h$ in $\mathsf {B}$ . More precisely, we can take as endpoints of $\mathsf {A}$ points close enough to the soft edges of the equilibrium measure but outside of its support, while the hard edges must remain endpoints common to $\mathsf {S},\mathsf {A}$ and $\mathsf {B}$ . We next state similar results for the fixed filling fractions model of Section 1.4. Recall that part of the data defining this model is a sequence (indexed by N) of guple of positive integers $\boldsymbol {N} = (N_1,\ldots ,N_g)$ such that $N_0 = N  \sum _{h = 1}^{g} N_h \geq 0$ and such that $\boldsymbol {\epsilon } = \boldsymbol {N}/N$ converges to a point in
In this context, the effective potential is defined for $x \in \mathsf {A}_{h}$ by the formula
and for $x \notin \mathsf {A}$ , we declare $U^{V;\mathsf {A}}_{\mathrm{eq};\boldsymbol {\epsilon }} = \tilde {U}^{V;\mathsf {A}}_{\mathrm{eq};\boldsymbol {\epsilon }} = +\infty $ .
Proposition 3.4. If V is regular, confining and uniformly to $V^{\{0\}}$ on $\mathsf {A}$ , then for any closed set $\mathsf {F}$ and open set $\mathsf {O}$ of $\mathbb R$ ,
Moreover, Corollaries 3.2 and 3.3 also extend to this setting.
We may omit the superscript $\mathsf {A}$ in the equilibrium measure, the effective potential, etc. when it is clear that we work with the compact set $\mathsf {A}$ .
3.2. Sketch of the proof of Lemma 3.1
We only sketch the proof since it is similar to [Reference Borot and GuionnetBG11] as well as [Reference Anderson, Guionnet and ZeitouniAGZ10, section 2.6.2]. The only technical difference is that the lower bound is achieved here by introducing the functions $H_{x,\varepsilon }$ and $\phi _{x,K}$ below rather than localising $L_{N1}$ to probability measures on some smaller sets than $\mathsf {B}$ in [Reference Borot and GuionnetBG11]. We first give the proof for the initial model and at the end of the proof precise the necessary changes to deal with the model with fixed filling fractions.
Recall that $L_N = N^{1}\sum _{i = 1}^N\delta _{ \lambda _i}$ denotes the normalised empirical measure. We observe that
where, for any measurable set $\mathsf {X}$ ,
We shall hereafter estimate $\frac {1}{N}\ln \Upsilon _{N,\beta }^{V;\mathsf {B}}(\mathsf {X})$ .
We first prove a lower bound for $\Upsilon _{N,\beta }^{V;\mathsf {B}}(\mathsf {X})$ with $\mathsf {X}$ open in $\mathsf {B}$ . For any $x\in \mathsf {X}$ , we can find $\varepsilon>0$ such that $(x\varepsilon ,x+\varepsilon ) \cap \mathsf {B} \subset \mathsf {X}$ . Let
Using twice Jensen inequality and the convention $V(\xi ) = +\infty $ for $\xi \notin \mathsf {B}$ , we get
where we have set
and $\phi _{x,K}$ is a continuous function vanishing outside of a large compact K that includes the support of $\mu _\mathrm{{eq}}^{V}$ , is equal to $1$ on a ball around x with radius $1+\varepsilon $ and on the support of $\mu _\mathrm{{eq}}^{V}$ , and takes values in $[0,1]$ . For any fixed $\varepsilon> 0$ , $\phi _{x,K} \cdot H_{x,\varepsilon }$ is bounded continuous, so we have by Theorem 1.1
with $\lim _{N \rightarrow \infty } R(\varepsilon ,N) = 0$ for all $\varepsilon>0$ . Letting $N \rightarrow \infty $ , we deduce since
and since V converges uniformly towards $V^{\{0\}}$ , that
Exchanging the integration over $\xi $ and $\lambda $ , observing that $\xi \rightarrow \int _{\mathsf {B}} \mathrm {d}\mu _\mathrm{{eq}}^{V}(\lambda )\,\ln \xi \lambda $ is continuous and then letting $\varepsilon \rightarrow 0$ , we conclude that for all $x\in \mathsf {X}$ ,
where we have recognised the effective potential of Equation (3.1). We finally optimise over $x\in \mathsf {X}$ to get the desired lower bound. To prove the upper bound, we note that for any $M>0$ ,
Observe that there exists $C_0$ and $c>0$ and d finite such that for $\xi  \geq C_0$ and all probability measures $\mu $ on $\mathsf {B}$ ,
by the confinement Hypothesis 1.1. As a consequence, if $\mathsf {X}\subset \mathsf {B} \setminus [C,C]$ for some C large enough, we deduce that
where the last bound holds for N large enough. Combining Equations (3.7), (3.8) and (3.6) shows that
Hence, we may restrict ourselves to $\mathsf {X}$ bounded. Moreover, the same bound extends to $ \mu _{N  1,\beta }^{\frac {NV}{N  1};\mathsf {B}}$ so that we can restrict the expectation over $L_{N1}$ to probability measures supported on $[C,C]$ up to an arbitrary small error $e^{N e(C)}$ , provided C is large enough and where $\lim _{C \rightarrow +\infty } e(C) = +\infty $ . The confinement hypothesis also guarantees that $V(\xi )2\int _{\mathsf {B}} \mathrm {d} L_{N  1}(\lambda )\ln \mathrm {max}\big (\xi  \lambda ,M^{1}\big )$ is uniformly bounded from below by a constant D. As $\lambda \mapsto \ln \mathrm {max}\big (\xi  \lambda ,M^{1}\big )$ is bounded continuous on compacts and MLipschitz on $\mathbb {R}$ , we can then use the large deviation principles of Theorem 1.1 to deduce that for any $\varepsilon>0$ , any $C\ge C_0$ ,
with
In terms of the Vaserstein distance between two probability measures,
Moreover, $\xi \mapsto V(\xi )  2\int _{\mathsf {B}} \mathrm {d}\mu _\mathrm{{eq}}^{V}(\lambda ) \ln \mathrm {max}\big (\xi  \lambda ,M^{1}\big )$ is bounded continuous so that a standard Laplace method yields, as V goes to $V^{\{0\}}$ ,
We finally choose C large enough so that the first term is larger than the second. Then, by the monotone convergence theorem, we deduce that $\int _{\mathsf {B}} \mathrm {d}\mu _\mathrm{{eq}}^{V}(\lambda ) \ln \mathrm {max}\big (\xi  \lambda ,M^{1}\big )$ increases as M goes to infinity towards $\int _{\mathsf {B}} \mathrm {d}\mu _\mathrm{{eq}}^{V}(\lambda ) \ln \xi  \lambda $ . This completes the proof of the large deviation in the initial model.
For the fixed filling fractions model, we make the decomposition
with
and
where $\boldsymbol {\epsilon }1_h/N$ corresponds to the filling fraction where one eigenvalue has been suppressed from $\mathsf {A}_h$ . The estimates for $\Upsilon _{N,\beta ,h}^{V;\mathsf {B}}(\mathsf {X}\cap \mathsf {A}_h)$ are done exactly as above and the result follows since the logarithm of a finite sum of exponentially small terms is asymptotically equivalent to the logarithm of the maximal term.
3.3. Concentration of measure and consequences
We will need rough a priori bounds on the correlators, which can be derived by purely probabilistic methods. This type of result first appeared in the work of [Reference Boutet de Monvel, Pastur and ShcherbinadMPS95, Reference JohanssonJoh98] and more recently [Reference Kriecherbauer and ShcherbinaKS10, Reference Maïda and MaurelSegalaMMS12]. Given their importance, we find useful to prove independently the bound we need by elementary means.
Hereafter, we will say that a function $f\,:\,\mathbb R\rightarrow \mathbb {C}$ is bHölder if
Our final goal is to control $\int _{\mathsf {A}} \varphi (x)\mathrm {d}[L_N  \mu _\mathrm{{eq}}^{V}](x)$ for a class of functions $\varphi $ which is large enough and, in particular, contains analytic functions on a neighbourhood of the interval of integration $\mathsf {A}$ . This problem can be settled by controlling the ‘distance’ between $L_N$ and $\mu _\mathrm{{eq}}^{V}$ for an appropriate notion of ‘distance’. We introduce the pseudodistance $\mathfrak {D}$ between probability measures $\mu ,\nu $ given by
It can be represented in terms of Fourier transform of the measures:
Since $L_N$ has atoms, its pseudodistance to another measure is, in general, infinite. There are several methods to circumvent this issue, and one of them, that we borrow from [Reference Maïda and MaurelSegalaMMS12], is to define a regularised measure $\widetilde {L}_N^\mathrm{{u}}$ (see the beginning of § 3.4.1 below) from $L_N$ . Then, the result of concentration takes the following form:
Lemma 3.5. Let V be regular, $\mathcal {C}^3$ , confining, satisfying a control of large deviations on $\mathsf {A}$ and satisfying (1.8) for $K=0$ (namely, $N(VV^{\{0\}})$ is uniformly bounded by a constant $v^{\{1\}}$ on $\mathsf {A}$ ). There exists $C> 0$ so that, for t small enough and N large enough,
Moreover, for any $\boldsymbol {N}=(N_1,\ldots ,N_g)$ so that $\boldsymbol {\epsilon }=\boldsymbol {N}/N\in \mathcal {E}$ ,
We prove it in § 3.4.1 below. The assumption V of class $\mathcal {C}^3$ ensures that the effective potential (3.1) defined from the equilibrium measure is a $\frac {1}{2}$ Hölder function (and even Lipschitz if all edges are soft) on the compact set $\mathsf {A}$ , as one can observe on Equation (A.9) given in Appendix A. This lemma allows an a priori control of expectation values of test functions.
Corollary 3.6. Let V be regular, $\mathcal {C}^3$ , confining, satisfying a control of large deviations on $\mathsf {A}$ and satisfying (1.8) for $K=0$ (namely, $N(VV^{\{0\}})$ is uniformly bounded by a constant $v^{\{1\}}$ on $\mathsf {A}$ ). Let $b> 0$ and assume $\varphi \,:\,\mathbb {R} \rightarrow \mathbb {C}$ is a bHölder function with constant $\kappa _{b}[\varphi ]$ such that
Then, there exists $C_3> 0$ such that, for t small enough and N large enough,
and for any $\boldsymbol {N}=(N_1,\ldots ,N_g)$ so that $\boldsymbol {\epsilon }=\boldsymbol {N}/N\in \mathcal {E}$ ,
As a special case, we can obtain a rough a priori control on the correlators. Recall the notation, for $\boldsymbol {\epsilon }\in \mathcal {E}$ ,
Corollary 3.7. Let V be regular, $\mathcal {C}^3$ , confining and satisfying a control of large deviations on $\mathsf {A}$ . Let $D'> 0$ and
There exists a constant $\gamma _1(\mathsf {A},D')> 0$ so that, for N large enough, for any $\boldsymbol {N}=(N_1,\ldots ,N_g)$ so that $\boldsymbol {\epsilon }=\boldsymbol {N}/N\in \mathcal {E}$
Similarly, for any $n \geq 2$ , there exist constants $\gamma _n(\mathsf {A},D')> 0$ so that, for N large enough,
In the $(g + 1)$ cut regime with $g \geq 1$ , we denote $(\mathsf {S}_h)_{0 \leq h \leq g}$ the connected components of the support of $\mu _\mathrm{{eq}}^{V}$ , and we take $\mathsf {A} = \bigcup _{h = 0}^{g} \mathsf {A}_h$ , where $\mathsf {A}_h = [a_h^,a_h^+] \subseteq \mathsf {B}$ are pairwise disjoint bounded segments such that $\mathsf {S}_h \subset \mathring {\mathsf {A}}_h$ . For any configuration $\lambda \in \mathsf {A}^N$ , we denote $N_h$ the number of $\lambda _i$ s in $\mathsf {A}_h$ , and $\boldsymbol {N} = (N_h)_{1 \leq h \leq g}$ . The following result gives an estimate for large deviations of $\boldsymbol {N}$ away from $N\boldsymbol {\epsilon }_{\star }$ in the large N limit.
Corollary 3.8. Let $\mathsf {A}$ be as above, and V be $\mathcal {C}^3$ , confining, satisfying a control of large deviations on $\mathsf {A}$ and leading to a $(g + 1)$ cut regime. There exists positive constants $C,C'$ such that, for N large enough and uniformly in t,
As an outcome of this article, we will obtain in Section 8.2 a stronger large deviation statement for filling fractions when the potential satisfies the stronger Hypotheses 1.1–1.3.
3.4. Concentration of $L_N$ : proof of Lemma 3.5
Throughout this section, proofs will be given for the initial model. They are exactly the same for the fixed filling fractions model.
3.4.1. Regularisation of $L_N$
We start by following an idea introduced by Maïda and MaurelSegala [Reference Maïda and MaurelSegalaMMS12, Proposition 3.2]. Let $\sigma _N,\eta _N \rightarrow 0$ be two sequences of positive numbers. To any configuration of points $\lambda _1 \leq \ldots \leq \lambda _N$ in $\mathsf {A}$ , we associate another configuration $\widetilde {\lambda }_1,\ldots ,\widetilde {\lambda }_N$ by the formula
It has the properties
Let us denote by $\widetilde {L}_N = \frac{1}{N} \sum _{i = 1}^N \delta _{\widetilde {\lambda }_i}$ the new counting measure. Then, we define $\widetilde {L}_N^\mathrm{{u}}$ be the convolution of $\widetilde {L}_N$ with the uniform measure on $[0,\eta _N\sigma _N]$ .
We are going to compare the (opposite of the) logarithmic energy of $L_N$ to that of $\widetilde {L}_N^\mathrm{{u}}$ , which has the advantage of having no atom. We first have
because the logarithm is increasing and the spacings of $\tilde {\lambda }$ s are larger than the spacings of $\lambda $ s. Let
denote the (opposite of the) logarithmic energy of a probability measure $\mu $ . Then,
Thanks to the minimal distance $\sigma _N$ enforced between the $\widetilde {\lambda }_i$ s in Equation (3.16), $\sigma _N\,\big (u  v)/(\tilde \lambda _i\tilde \lambda _j)\big $ is bounded by $1$ , so that for $\eta _N\le \frac {1}{2}$ (thus for N large enough),
Since