Asymptotics of Symmetric Polynomials: A Dynamical Point of View Authors: Alice Guionnet, Jiaoyang Huang Published: 2026 ABSTRACT AbstractIn this paper we study the asymptotic behavior of the (skew) Macdonald and Jack symmetric polynomials as the number of variables grows to infinity. We characterize their limits in terms of certain variational problems. As an intermediate step, we establish a large deviation principle for the θ analogue of non-intersecting Bernoulli random walks. When θ=1, these walks are equivalent to random Lozenges tilings of strip domains, where the variational principle (with general domains and boundary conditions) has been proven in the seminal work by Cohn et al. (J Am Math Soc 14(2):297–346, 2001). Our result gives a new argument of this variational principle, and also extends it to non-intersecting θ-Bernoulli random walks for any θ∈(0,∞). Remarkably, the rate functions remain identical, differing only by a factor of 1/θ. Introduction Macdonald symmetric functions, initially discovered by Macdonald in the late 1980s [46], are central to a number of key developments in mathematics and mathematical physics. They are indexed by Young diagrams and implicitly depend on two parameters q,t∈(0,1). Their degeneracy, by taking t=qθ and sending q to approach one, gives the Jack polynomials. This family of symmetric functions depends on the positive parameter θ>0, and was first introduced in [39, 40]. Notably, when θ=1, the Jack polynomials coincide with the well-known Schur polynomials, up to some multiplicative constants. These symmetric polynomials appear in algebraic combinatorics as generating functions, and provide a systematic way to enumerate combinatorial objects. In representation theory, these symmetric polynomials offer a crucial bridge between algebraic structures, combinatorial objects and group representations. Recently, symmetric polynomials have become central tools in the development of integrable stochastic models. This theory has found numerous applications in areas such as random partitions, random matrix theory, and directed polymers. In this article, we study the asymptotic behaviors of symmetric polynomials, namely (skew) Macdonald and Jack polynomials, as the number of parameters goes to infinity. In this introduction, we discuss the results for skew Jack polynomials. The results on skew Macdonald polynomials are more involved and are presented in Sect. 1.4. Recall that skew Jack polynomials Jλ\μ(x;θ) are symmetric functions in infinitely many variables x=(xi)1⩽i⩽∞ and parametrized by two Young diagrams μ⊂λ, where μ=(μ1⩾μ2⩾⋯⩾μℓ(μ)),λ=(λ1⩾λ2⩾⋯⩾λℓ(λ)) are arrays of integer numbers with length ℓ(μ),ℓ(λ) respectively. We refer to Sects. 6.1 and 6.2 for more detailed discussion on (skew) Jack polynomials. We consider the limit of skew Jack polynomials along a sequence of Young diagrams λ(N),μ(N) and parameters b(N), where the column and row sizes of Young diagrams and the number of parameters grow linearly in N. We show the limit N-2lnJλ(N)\μ(N)(b(N);θ) exists and give an explicit characterization of the limit in terms of a variational problem. Nonintersecting θ-Bernoulli Walk Ensembles Understanding the asymptotic behavior of symmetric polynomials is fundamental to modern combinatorics, statistical mechanics and representation theory. However, there are few results in this direction, as symmetric polynomials tend to have complicated structures and thus are not easily approachable by algebraic combinatorics techniques. In this article, we will study the asymptotic behaviors of symmetric polynomials via a dynamical approach. Many symmetric polynomials arise naturally as the partition functions of interacting particle systems. Some notable examples are The partition function of nonintersecting Brownian bridges is given by the Harish–Chandra–Itzykson–Zuber integral formula [29]. It is closely related with Schur polynomials. The partition function of nonintersecting Brownian bridges is given by the Harish–Chandra–Itzykson–Zuber integral formula [29]. It is closely related with Schur polynomials. Macdonald processes, introduced by Borodin and Corwin [8], are probability measures on sequences of partitions whose weights are expressed in terms of Macdonald symmetric functions (typically products of Macdonald polynomials and/or skew Macdonald polynomials evaluated at specializations). Their normalization constants are computed using the Macdonald Cauchy identity. Degenerations of Macdonald processes include Schur processes [52, 54], Jack processes [28, 37], Hall–Littlewood processes [7], and Whittaker processes [8]. Macdonald processes, introduced by Borodin and Corwin [8], are probability measures on sequences of partitions whose weights are expressed in terms of Macdonald symmetric functions (typically products of Macdonald polynomials and/or skew Macdonald polynomials evaluated at specializations). Their normalization constants are computed using the Macdonald Cauchy identity. Degenerations of Macdonald processes include Schur processes [52, 54], Jack processes [28, 37], Hall–Littlewood processes [7], and Whittaker processes [8]. Partition functions of vertex models give families of symmetric rational functions, which generalize Schur symmetric polynomials, as well as some of their variations [1, 2, 6, 11]. Partition functions of vertex models give families of symmetric rational functions, which generalize Schur symmetric polynomials, as well as some of their variations [1, 2, 6, 11]. On one hand, the asymptotic behavior of symmetric polynomials provides a valuable tool for characterizing the dynamics of interacting particle systems, where these polynomials serve as partition functions. On the other hand, this also offers an alternative approach to investigating the asymptotic behaviors of symmetric polynomials by exploring the asymptotic behaviors of the associated interacting particle systems using dynamical approaches. Fix large N, and take Young diagrams λ=(λ1⩾λ2⩾⋯⩾λN),μ=(μ1⩾μ2⩾⋯⩾μN), and parameters b=(b0,b1,b2,⋯,bT-1). The skew Jack symmetric polynomials Jλ′\μ′(b;θ-1) (where λ′,μ′ are the transposes of λ,μ), can be interpreted as the partition functions of N-particle nonintersecting θ-Bernoulli walks. For any Young diagrams ν=(ν1⩾ν2⩾⋯⩾νN) with at most N rows, we encode it by a particle configuration x=x(ν)=(x1,x2,⋯,xN), where From the construction, this particle configuration lives on the following θ-dependent lattice Definition 1.1 (Non-intersecting θ-Bernoulli walk ensembles) An n-particle non-intersecting θ-Bernoulli walk from time 0 to time S is a sequence of particle configurations p=(x(0),x(1),…,x(S))∈(Wθn)S such that x(t)=(x1(t),x2(t),⋯,xn(t)), and xi(t+1)-xi(t)∈{0,1} for each t∈[0,S] and i∈{1,…,n}; viewing {xi(t)}0⩽t⩽S as the space-time trajectory for the i-th particle, which may either jump to the right or not move at each step. From the construction of the lattice (1.2), the paths are non-intersecting, since we have xi(t)>xj(t) whenever 1⩽iy2>⋯>yN),z=(z1>z2>⋯>zN)∈WθN We denote the set of non-intersecting θ-Bernoulli walks from y=(y1>y2>⋯>yN)∈WθN to z=(z1>z2>⋯>zN)∈WθN as The set P(y,z;T) is nonempty if there exists a non-intersecting Bernoulli walk {x(t)}0⩽t⩽T from x(0)=y to x(T)=z, which is equivalent to Indeed, if (1.5) holds we can take Then we can rewrite the skew Jack polynomial Jλ′/μ′(·;θ-1) evaluated at (b0,b1,⋯,bT-1), in terms of the weights (1.7) as follows (we remark that this is not the standard definition of skew Jack polynomials in [46]; rather, it is obtained by taking the limit q→1 in the corresponding formula for skew Macdonald polynomials (see Claim C.4).) where λ′,μ′ are the transposes of λ,μ, and for non-intersecting θ-Bernoulli walk p={x(t)}0⩽t⩽T and b=(b0,b1,⋯,bT-1), we define their weights as where e(t)=(e1(t),e2(t),⋯,eN(t)) is the increment at time t, The formula (1.6) is derived from the construction of the Jack process, as explained in more detail in Sect. 6. On the right-hand side of (1.6), there are explicit formulas for the Jack symmetric polynomials evaluated at 1N, i.e. Jμ(1N;θ),Jλ(1N;θ), making them amenable for asymptotic analysis. The primary challenge lies in analyzing the sum of the weights W(p;b). In the special case where bi=1, this sum of weights can be interpreted as the partition function of non-intersecting θ-Bernoulli random walks from y to z, with the transition probability given by for 0⩽t⩽T-1, where V is the Vandermonde determinant, and e=(e1,e2,⋯,eN)∈{0,1}N, see Fig. 1. We remark that the above transition probability is given by pairwise interactions, and the interactions between adjacent particles are singular to prevent colliding. More precisely, from the above transition probability, if x∈WθN and xi-xi+1⩾θ+1, then (xi+ei)-(xi+1+ei+1)∈θ+Z⩾0. If xi-xi+1=θ, then (1.8) is nonzero only if (ei,ei+1)∈{(0,0),(1,0),(1,1)}. In these cases, (xi+ei)-(xi+1+ei+1)∈θ+Z⩾0. Therefore the Markov process (1.8) stays in the lattice WθN. Shown above is the transition from the Young diagram μ=(5,3,2,1,1,1) to the Young diagram λ=(6,4,2,2,1,1), which are encoded by the particle systems through the relation (1.1). Our first main result establishes a large deviation principle for the non-intersecting θ-Bernoulli random walks (1.8). Notably, when θ=1, these walks reduce to non-intersecting Bernoulli random walks, which is equivalent to random Lozenge tiling of strip domains. Cohn, Kenyon, and Propp [20] previously proved a variational principle for random tiling models (domino tiling and lozenge tiling) with general boundary conditions. Our result extends this variational principle to non-intersecting θ-Bernoulli random walks for any θ∈(0,∞). Remarkably, the rate functions remain identical, differing only by a factor of 1/θ. To state the large deviation principle, we need to introduce the notations of height function and surface tension. Height function and surface tension Given any N-particle nonintersecting θ-Bernoulli walk ensemble p=(x(0),x(1),…,x(T))∈(WθN)T from time 0 to time T, we denote the increment at time t as e(t)=(e1(t),e2(t),⋯,eN(t)), We denote the rescaled particle configuration (by rescaling space and time by 1/N), In the rest of the paper, we will use the scaling above (we will denote x(t) for particle configurations before scaling, and x(t) for after scaling). For t/N⩽t⩽(t+1)/N, we linearly interpolate the particle configurations We encode the particle densities of x(t) as and define the associated height function We notice that the total mass of ρ in (1.10) is θ for any 0⩽t⩽T/N, and H(x, t) is a non-decreasing function of x from 0 to θ. The walk ensemble p=(x(0),x(1),…,x(T)), the rescaled particle configuration x(t) can be recovered easily by either the empirical particle density ϱ(x;x(t)), or the height function H(x, t). In the rest of the paper, we will not distinguish them. Given an N-particle nonintersecting θ-Bernoulli walk ensemble p={x(t)}0⩽t⩽T, we recall its weight W(p;b) from (1.7). For b=1T, we simply denote the weight as W(p), Under the transition probability (1.8), the probability of the walk p is given by 2-NTW(p). We also define the weight of the height function H associated with p={x(t)}0⩽t⩽T as W(H)=W(p). From the construction of the height function (1.11), H(x, t) is 2-Lipschitz, and almost surely, To analyze the limits of height functions of these interacting particle systems, it will be useful to introduce continuum analogs of the height functions considered in (1.11). So, set and its closure T¯={(s,t)∈[0,1]×R⩽0:s+t⩾0}. We interpret T¯ as the set of possible gradients, also called slopes, for a continuum height function. The height functions H(x, t) associated with particle configurations (from (1.11)) is 2-Lipschitz and satisfies ∇H(x,t)∈T¯ for (x,t)∈R×[0,T/N]. Definition 1.2 Fix time T>0, and denote R=R×[0,T]. We say that a function H:R→R is admissible if H is 2-Lipschitz and ∇H(u)∈T¯ for almost all u∈R. We further say that the boundary height function h=(h(x,0),h(x,T)):∂R→R admits an admissible extension to R if Adm(R;h), the set of admissible functions H:R→R with H|R×{0,T}=h, is not empty. For any x∈R⩾0 and (s,t)∈T¯ we denote the Lobachevsky function L:R⩾0→R and the surface tension σ:T¯→R2 by For any admissible height function H on R, we further denote the entropy functional The entropy functional (1.15) also governs the logarithm of the number of lozenge tilings for a given domain and boundary height function [20]. The surface tension σ is strictly concave within the interior of T and linear along the boundary of T¯. As a consequence, the maximizer of the energy functional (1.15) possesses what are known as “liquid regions", where the solution is real analytic, and “frozen regions", where the solution is piecewise linear. Variational problems associated with (1.15), and in more general setting, have been explored in previous studies [4, 42, 60]. Large Deviation Principle for Non-Intersecting θ-Bernoulli Random Walks We prove a large deviation for non-intersecting θ-Bernoulli random walks (1.8) with given boundary condition at time 0 and time T. We make the following assumption on the height profile of the limiting boundary condition. Definition 1.3 Fix any θ>0. We denote by Admθ∂(R) the set of boundary height functions h=(h(x,0),h(x,T)):∂R↦R such that ∂xh(x,0),∂xh(x,T) are two compactly supported positive measures with density bounded by 1 and total mass θ. ∂xh(x,0),∂xh(x,T) are two compactly supported positive measures with density bounded by 1 and total mass θ. Moreover, h=(h(x,0),h(x,T)) admits an admissible extension to R. Moreover, h=(h(x,0),h(x,T)) admits an admissible extension to R. We equip Adm(R,h) with the uniform topology. First, we remark that since we assumed H is Lipschitz, the uniform topology in the spatial variable is equivalent to the weak topology on its derivative, as can be easily checked by integration by parts. Moreover, it is evident that Adm(R,h) is a compact space. Moreover, if H∈Adm(R,h), then for every t∈[0,T], ∂xH(.,t) is compactly supported, bounded by one and , because ∇H∈T¯, ∂xH(.,t) is a non-negative measure for all time t. Remark 1.4 We remark that Adm(R,h) depends on θ, since h∈Admθ∂(R). We emphasize this dependence because it has an important consequence: for each t∈[0,T], the function x↦H(x,t) is non-decreasing and takes values in [0,θ]. Indeed, ∂tH(x,t)≤0 for every x, hence Moreover, H(x,T)=h(x,T) and H(x,0)=h(x,0), and since h∈Admθ∂(R) we have h(x,0)=h(x,T)=0 for x sufficiently small and h(x,0)=h(x,T)=θ for x sufficiently large. Therefore, for every t∈[0,T], H(x,t)=0 for x sufficiently small and H(x,t)=θ for x sufficiently large. To state our main result, let us precisely state our running assumption concerning the particle configurations. Definition 1.5 Let θ,T>0 and h∈Admθ∂(R). Given two sequences of particle configurations (y(N),z(N)), we say that (y(N),z(N),T) are (h,θ,T)-admissible iff There exists a constant C>0, |yi(N)|,|zi(N)|⩽CN for 1⩽i⩽N. There exists a constant C>0, |yi(N)|,|zi(N)|⩽CN for 1⩽i⩽N. The empirical density (recall from (1.10)) of y(N),z(N) converges, namely when N→∞1.16ϱ(x;y(N)/N)→∂xh(x,0),ϱ(x;z(N)/N)→∂xh(x,T), in distribution. The empirical density (recall from (1.10)) of y(N),z(N) converges, namely when N→∞1.16ϱ(x;y(N)/N)→∂xh(x,0),ϱ(x;z(N)/N)→∂xh(x,T), in distribution. As N goes to infinity, T/N goes to T. As N goes to infinity, T/N goes to T. We further assume that for N large enough the set of non-intersecting θ-Bernoulli walks of length T from y(N) to z(N) is nonempty, namely P(y(N),z(N);T)≠∅ (see (1.4) ). We remark that a height function G∈P(y(N),z(N);T) is defined on R×[0,T/N]. If T/N0 and h∈Admθ∂(R). Consider two sequences of particle configurations and time (y(N),z(N),T) which are (h,θ,T)-admissible, and non-intersecting θ-Bernoulli random walks (1.8) starting from y(N) at time 0. Then, for any H∈Adm(R,h), the following holds 1.17limε→0lim supN→∞1N2lnP({G∈P(y(N),z(N);T):‖G-H‖∞⩽ε})=-1θJh(H)-Tln2, where Jh is the functional defined for H∈Adm(R,h) by 1.18Jh(H)=-∬Rσ(∇H(x,t))dxdt-12∬Rln|x-y|dh(x,t)dh(x,t)0T. Moreover, (1.17) holds if we replace the lim sup by a lim inf. Then, for any H∈Adm(R,h), the following holds 1.17limε→0lim supN→∞1N2lnP({G∈P(y(N),z(N);T):‖G-H‖∞⩽ε})=-1θJh(H)-Tln2, where Jh is the functional defined for H∈Adm(R,h) by 1.18Jh(H)=-∬Rσ(∇H(x,t))dxdt-12∬Rln|x-y|dh(x,t)dh(x,t)0T. Moreover, (1.17) holds if we replace the lim sup by a lim inf. Jh has compact level sets in Adm(R,h). Jh has compact level sets in Adm(R,h). The law of the height functions H(N) conditioned to remain in P(y(N),z(N);T) satisfies a large deviation principle with speed N2 and good rate function Ih, in the following sense. For any Borel set B of height functions (with respect to the sup-norm topology on R), -infH∈B∘Ih(H)⩽lim infN→∞1N2logPH(N)∈B⩽lim supN→∞1N2logPH(N)∈B⩽-infH∈B¯Ih(H). Moreover, Ih is a good rate function, i.e. it is lower semicontinuous and its sublevel sets {H:Ih(H)⩽M} are compact. The rate function Ih is infinite outside the admissible class Adm(R,h), and for H∈Adm(R,h) it is given by 1.19Ih(H)=1θ(Jh(H)-infG∈Adm(R,h)Jh(G)). Moreover, Ih is minimized at a unique minimizer. The law of the height functions H(N) conditioned to remain in P(y(N),z(N);T) satisfies a large deviation principle with speed N2 and good rate function Ih, in the following sense. For any Borel set B of height functions (with respect to the sup-norm topology on R), -infH∈B∘Ih(H)⩽lim infN→∞1N2logPH(N)∈B⩽lim supN→∞1N2logPH(N)∈B⩽-infH∈B¯Ih(H). Moreover, Ih is a good rate function, i.e. it is lower semicontinuous and its sublevel sets {H:Ih(H)⩽M} are compact. The rate function Ih is infinite outside the admissible class Adm(R,h), and for H∈Adm(R,h) it is given by 1.19Ih(H)=1θ(Jh(H)-infG∈Adm(R,h)Jh(G)). Moreover, Ih is minimized at a unique minimizer. The surface tension σ is the same as that of lozenge tiling. It has been proven in [23, Theorem 7.5], that Jh is lower semicontinuous over Adm(R,h) with the uniform topology. Since Adm(R,h) is compact, it follows that Jh has compact level sets in Adm(R,h). The uniqueness of the minimizer for the variational principle (1.19) has been proven in [60, Proposition 4.5]. We remark that, once the boundary data h∈Admθ∂(R) is fixed, the functional Jh(H) and the admissible class Adm(R,h) over which it is minimized do not depend on θ. Our second main result derives the large deviation asymptotics of skew Jack symmetric polynomials as the number of variables goes to infinity. As discussed after (1.6), the main challenge lies in analyzing the sum of the weights W(p;b) from (1.7). It can be interpreted as the partition function of non-intersecting θ-Bernoulli random walks from y to z, with time dependent drift. Explicitly, the transition probability is given by for 0⩽t⩽T-1 and e=(e1,e2,⋯,eN)∈{0,1}N. The large deviation of the aforementioned non-intersecting θ-Bernoulli random walks with time-dependent drift can be deduced from Theorem 1.6 through the application of Varadhan’s lemma. An immediate consequence of this is the following result concerning the asymptotics of skew Jack polynomials following from the formula (1.6). The proof for this result is provided in Sect. 6.4. We next state our large deviation results under the law Pb and deduce the asymptotics of skew Jack polynomials in the related scaling. Theorem 1.7 Let θ,T>0 and h∈Admθ∂(R). We consider a sequence of Young diagrams such that (y(N),z(N)) are given by (1.3). Assume (y(N),z(N),T) are (h,θ,T)-admissible and for 0⩽t⩽T-1 take bt=ef(t/N) for a real continuously differentiable function f. The asymptotics of the skew Jack polynomials are given by limN→∞1N2lnJ(λ(N))′\(μ(N))′(b0,b1,⋯,bT-1;θ-1)=1θJhf, where 1.21Jhf:=supH∈Adm(R,h)∬Rσ(∇H(x,t))dxdt+Ff(H), and the linear functional Ff(H) is given by 1.22Ff(H):=-∬Rf(s)∂sH(y,s)dyds. The asymptotics of the skew Jack polynomials are given by limN→∞1N2lnJ(λ(N))′\(μ(N))′(b0,b1,⋯,bT-1;θ-1)=1θJhf, where 1.21Jhf:=supH∈Adm(R,h)∬Rσ(∇H(x,t))dxdt+Ff(H), and the linear functional Ff(H) is given by 1.22Ff(H):=-∬Rf(s)∂sH(y,s)dyds. The law of the height functions H(N) under Pb defined in (1.20), conditioned to remain in P(y(N),z(N);T), satisfies a large deviation principle with speed N2 and good rate function Ihf/θ, in the following sense. For any Borel set B of height functions (with respect to the sup-norm topology on R), -infH∈B∘Ihf(H)θ⩽lim infN→∞1N2logPbH(N)∈B⩽lim supN→∞1N2logPbH(N)∈B⩽-infH∈B¯Ihf(H)θ. Moreover, Ihf/θ is a good rate function, i.e. it is lower semicontinuous and its sublevel sets {H:Ihf(H)/θ≤M} are compact. The rate function is infinite outside the admissible class Adm(R,h), and for H∈Adm(R,h) it is given by 1.23Ihf(H)=-∬Rσ(∇H(x,t))dxdt+Ff(H)+Jhf. Moreover, Ihf achieves its minimal value at a unique height function. The law of the height functions H(N) under Pb defined in (1.20), conditioned to remain in P(y(N),z(N);T), satisfies a large deviation principle with speed N2 and good rate function Ihf/θ, in the following sense. For any Borel set B of height functions (with respect to the sup-norm topology on R), -infH∈B∘Ihf(H)θ⩽lim infN→∞1N2logPbH(N)∈B⩽lim supN→∞1N2logPbH(N)∈B⩽-infH∈B¯Ihf(H)θ. Moreover, Ihf/θ is a good rate function, i.e. it is lower semicontinuous and its sublevel sets {H:Ihf(H)/θ≤M} are compact. The rate function is infinite outside the admissible class Adm(R,h), and for H∈Adm(R,h) it is given by 1.23Ihf(H)=-∬Rσ(∇H(x,t))dxdt+Ff(H)+Jhf. Moreover, Ihf achieves its minimal value at a unique height function. Asymptotics for Skew Macdonald Polynomials The goal of this section is to study the asymptotics of skew Macdonald polynomials Pλ/μ(x;q,t). Formally, Pλ/μ is a symmetric function in the algebra of symmetric functions (equivalently, in infinitely many variables x=(xi)i≥1), but throughout this paper we only consider its specialization to finitely many variables, e.g. x=(x1,⋯,xT). We refer to “Appendix C” for a more detailed discussion of (skew) Macdonald polynomials. Fix θ>0, and take t=qθ. Assume λ,μ are Young diagrams with at most N rows, and take any b=(b0,b1,⋯,bT-1). Then the same as in (1.3), we identify μ,λ as particle configurations y,z. For non-intersecting Bernoulli walks p={x(t)}0⩽t⩽T from x(0)=y to x(T)=z, we define their weights as Then we can rewrite the skew Macdonald polynomial evaluated at (b0,b1,⋯,bT-1) as follows (we remark that this is not a standard definition of skew Macdonald polynomial as in [46], we give a proof in Claim C.4) where λ′,μ′ are the transposes of λ,μ. Similarly to the skew Jack polynomials, there are explicit formulas for the Macdonald symmetric polynomials evaluated at the principal specialization, i.e. Pμ(1,t,t2,⋯,tN-1;q,t),Pλ(1,t,t2,⋯,tN-1;q,t), making them amenable for asymptotic analysis. The sum of the weights W~(p;b) can be interpreted as the partition function of non-intersecting θ-Bernoulli random walks from y to z, with the transition probability given by for 0⩽t⩽T-1, and e=(e1,e2,⋯,eN)∈{0,1}N. The above Markov process (1.26) is a special case of ascending Macdonald process as constructed in [8], we refer to “Appendix C.2” for more details. The Macdonald processes (1.26) are different from non-intersecting θ-Bernoulli random walks (1.8), and also have singular interactions among adjacent particles. However, by taking their ratios, we observe the cancellation of singularities. This observation allows us to establish the large deviation principle for the Macdonald process through the application of Varadhan’s lemma. The following result on the asymptotics of skew Macdonald polynomials is a consequence of the large deviation principle of the Macdonald process (1.26). The proof is given in Sect. 6.4. Theorem 1.8 Let θ,T>0 and h∈Admθ∂(R). We consider a sequence of Young diagrams (λ(N),μ(N)) such that (y(N),z(N)) are given by (1.3). Assume (y(N),z(N),T) are (h,θ,T)-admissible. For 0⩽t⩽T-1, we take bt=ef(t/N) for a real continuously differentiable function f. Moreover, we take t=qθ,q=eκ/N for some κ<0. The asymptotics of the skew Macdonald polynomials are given by =limN→∞1N2lnP(λ(N))′\(μ(N))′(b0,b1,⋯,bT-1;q,t)=1θJhf, where Jhf is defined in (1.21). The asymptotics of the skew Macdonald polynomials are given by =limN→∞1N2lnP(λ(N))′\(μ(N))′(b0,b1,⋯,bT-1;q,t)=1θJhf, where Jhf is defined in (1.21). The law of height functions under Pb,q defined in (1.26) conditioned to remain in P(y(N),z(N);T) satisfies a large deviation principle with speed N2 and good rate function Ihf(H)/θ which is infinite outside Adm(R,h) and otherwise given by (1.23). Moreover, Ihf achieves its minimal value at a unique height function. The law of height functions under Pb,q defined in (1.26) conditioned to remain in P(y(N),z(N);T) satisfies a large deviation principle with speed N2 and good rate function Ihf(H)/θ which is infinite outside Adm(R,h) and otherwise given by (1.23). Moreover, Ihf achieves its minimal value at a unique height function. Remark 1.9 (Independence of κ at speed N2) In Theorem 1.8 we take q=eκ/N, so q→1 as N→∞ at rate 1/N. Although the underlying measure Pb,q depends on κ for each finite N, the large deviation rate function at speed N2 does not: both the normalizing constant Jhf and the functional Ihf are independent of κ. Heuristically, changing q from 1 to eκ/N perturbs the logarithm of the Radon–Nikodym derivative by at most order N per mesoscopic time step, hence by order N (or, at most, NlogN) after summing over all time steps and particles. After dividing by the LDP speed N2, this contribution vanishes, so the leading-order variational problem is unchanged. In other words, at the N2 scale the tilt produced by q=eκ/N is too weak to modify the macroscopic limit shape and the corresponding rate function. This does not rule out κ-dependence at finer scales (e.g. order N corrections to the free energy), nor does it address other regimes of q. For instance, if 1-q decays more slowly, one expects the q-dependence to contribute already at the N2 level and hence change the rate function. Establishing such regimes would require large deviation estimates that control the q-dependent weights more sharply than what is available via a direct application of Varadhan’s lemma in the present setting. Related Works on the Asymptotics of Symmetric Polynomials The asymptotic behavior of symmetric polynomials can be studied under other scaling regimes. One intensively studied regime is to take limit along a sequence of Vershik–Kerov partitions, where the size and rows of the partitions all grow linearly in N. In this way, the normalized Schur functions approximate a character of the infinite unitary group. The asymptotics of normalized Schur functions in this scaling regime were originally derived by Vershik and Kerov in [63]. Generalizations of these results to the asymptotics of Jack and Macdonald polynomials under the same scaling can be found in [21, 22, 53]. While the Vershik–Kerov partitions involve thin partitions, this paper explores a different scaling regime. Specifically, we consider the limit along a sequence of partitions where the number of rows and columns grows linearly in N, but the size grows quadratically in N2. Gorin and Panova’s work in [26] presents an example within this regime. They obtained the asymptotics of Schur polynomials with all but finitely many parameters set to one, see also [32, 33]. These asymptotic results played an important role in the study of extreme characters of the infinite-dimensional unitary group [16, 26], as well as applications in random lozenge tilings [17, 18] and the six-vertex model [25, 26]. The structure constants of symmetric polynomials, such as Kostka number, Kronecker coefficients and Littlewood–Richardson coefficients are fundamental quantities in algebraic combinatorics. Despite the absence of explicit formulas, the asymptotics of specific extreme Kronecker and Littlewood–Richardson coefficients have been derived in [55–57, 62]. Using the asymptotics of Schur symmetric polynomials, the large deviation asymptotics of Kostka numbers and the large deviation upper bound for the Littlewood–Richardson coefficients have been obtained in [5]. These results were generalized in [38] to the asymptotic behavior of weight multiplicities of irreducible representations of compact or complex simple Lie algebras in the limit of large rank. The asymptotics of Jack polynomials and Macdonald polynomials should allow to address similar questions for the large deviations of their structure constants. Ideas of the Proofs In the continuous setting, the asymptotics of the Harish–Chandra–Itzykson–Zuber integral has been discovered through studying the large deviations of Dyson’s Brownian motion by the first named author and Zeitouni [30, 34, 35]. Subsequently, analogous findings were extended to the rectangular spherical integral and Generalized Bessel Functions in [31, 38]. Large deviations are established thanks to the standard tilting argument by martingales. However, these martingales are constructed thanks to stochastic calculus and the specific structures of Dyson’s Brownian motion. In this paper, we focus on the discrete setting, where tools from stochastic calculus are not available. We summarize the main ideas of the proof of our main Theorems, Theorem 1.6. Theorems 1.7 and 1.8, can then be deduced by Varadhan’s lemma. In the special case when θ=1, the non-intersecting θ-Bernoulli random walks (1.8) are equivalent to random Lozenge tiling of strip domains, where the variational principle of the height function has been proven in [20]. In fact their results apply to random tilings of more general domains. The proof of the variational principle for tilings in [20] relies on the exact computation of the partition function for tilings on a large torus, utilizing Kasteleyn’s matrix. Unfortunately, this method is not applicable in our context when θ≠1. Instead, our proof relies on the recently introduced dynamical loop equations by the second author and Gorin [24], based on Nekrasov’s equations [10, 50, 51]. These equations have proven effective for analyzing the fluctuations of large families of two-dimensional interacting particle systems in both discrete and continuous settings. Examples include nonintersecting Bernoulli/Poisson random walks [27, 36, 45], β-corner processes [9, 28], measures on Gelfand-Tsetlin patterns [19, 58, 59], and Macdonald processes [8]. In particular, a version of dynamical loop equations has been derived for the non-intersecting θ-Bernoulli random walks. We establish the large deviation principle using Cramér’s method, by tilting the measure of nonintersecting θ-Bernoulli walk using exponential martingales. A key point is that these martingales can be identified as smooth functions of the height functions thanks to dynamical loop equations. Our proof is divided into an upper bound and a lower bound. The normalization constant for these exponential martingales can be determined using dynamical loop equations. Subsequently, the large deviation upper bound can be obtained straightforwardly by applying Markov’s inequality. The measure tilted by the exponential martingale coincides with the drifted non-intersecting θ-Bernoulli random walk. Utilizing the dynamical loop equation once again, we establish a limit shape theorem for the drifted non-intersecting θ-Bernoulli random walk. As a consequence, for any targeting measure process, we can construct an exponential martingale such that the tilted measure concentrates around it. This provides the large deviation lower bound. In the above outlined proof, we utilize dynamical loop equations to investigate nonintersecting θ-Bernoulli walks with drift. However, two challenges hinder their direct application. Firstly, dynamical loop equations require that the drift is analytic. To overcome this, we convolve the targeting measure process with a small Cauchy distribution (which is reminiscent of the strategy followed in [34]). This convolution enables the analytical extension of the density to a strip region around the real axis, allowing for a similar extension of the drift terms. As a trade-off, we are required to analyze measure-valued processes supported on the entire real axis. This challenge is addressed through precise truncations and the utilization of the explicit form of the rate function. Secondly and more critically, dynamical loop equations require certain non-criticality conditions. Specifically, the particle system cannot be too dense, meaning the density cannot be too close to one. To tackle this challenge, rather than studying the nonintersecting θ-Bernoulli walks as a whole, we partition space-time into small regions of parallelogram shape. Within each region, we ensure that both the density and velocity remain nearly constant. In each region, the system resembles nonintersecting θ-Bernoulli walks but with a reduced number of particles and possibly left and right boundaries. If the density and velocity are non-extremal, we can establish a large deviation principle using the dynamical loop equation as outlined above. However, if the density or velocity are extremal, we prove a large deviation upper bound by directly analyzing the walk. Importantly, it is observed that the contribution from these regions is negligible. The original nonintersecting θ-Bernoulli walks have long-range pairwise interactions. After partitioning, particles from different regions also interact with each other. Fortunately, these interactions occur between particles which are far from each other and are not singular, which can be approximated as smooth weights. In this manner, we can integrate all regions together and derive the variational principle for the original nonintersecting θ-Bernoulli walks. Loop equations and Nekrasov’s equations have been key to establish central limit theorems [12–15, 41]. To our knowledge, it is the first time they have been used to derive large deviations. Organization of the Paper In Sect. 2, we collect various facts about height functions, Vandermonde determinants and the free entropy. They will be used repeatedly in the rest of this paper. Section 3 provides an overview of the proof for Theorem 1.6, which relies on both a large deviation upper bound (Proposition 3.2) and a corresponding lower bound (Proposition 3.5) for non-intersecting θ-Bernoulli random walks with a constant slope. The proof for Proposition 3.2 is detailed in Sect. 4, while the proof for Proposition 3.5 is presented in Sect. 5. Moving forward, Sect. 6 contains the proof for Theorem 1.7, on the asymptotics of skew Jack polynomials. Similarly, the proof for Theorem 1.8, which explores the asymptotics of skew Macdonald polynomials, is outlined in “Appendix C”. Finally in “Appendix A”, we collect the results on dynamical loop equations from [24]. Notations We use λ,μ to represent Young diagrams. For macro particle locations we use mathsf letters: xi(t),x(t),y,x, where time t ranges from 0 to T. For micro particle locations we use standard letters xi(t),x(t),y,z, where the time t ranges from 0 to T. We use H and H∗ for arbitrary limiting Height functions, that is an element of Adm(R,h) that arises as the limit of height functions. H~ stands for the smoothed version of H. And we use H for the associated height functions of non-intersecting random walks. For two quantities X and Y depending on N, we write that X=O(Y) or X≲Y if there exists some universal constant C>0 such that |X|⩽CY . We write X=o(Y), or X≪Y if the ratio |X|/Y→∞ as N goes to infinity. We write X≍Y if there exists a universal constant C>0 such that Y/C⩽|X|⩽CY. We denote [[a,b]]=[a,b]∩Z and [[n]]=[[1,n]]. Setup and Preliminary Results In this section, we collect various facts about height functions, the Vandermonde determinant, and free entropy. These facts will be used repeatedly throughout the rest of this paper. In the rest of this paper, for simplicity of notations, we will assume that T=T/N. The more general case that T/N converges to T as in Definition 1.5 can be proven in the same way, by noticing that the rate functions in Theorems 1.6, 1.7, 1.8 are all continuous in T. Approximating Height Function For any height function H∗∈Adm(R;h), the following lemma states that if P(y(N),z(N);T)≠∅, then there exists a non-interesecting θ-Bernoulli walk from y(N) to z(N), and its height function is close to H∗. The proof is a consequence of a careful discretization of H∗, and we postpone it to “Appendix B”. Lemma 2.1 Let θ,T>0 and h∈Admθ∂(R). Consider two sequences of particle configurations and time T: (y(N),z(N),T) which are (h,θ,T)-admissible as defined in Definition 1.5. For any ε>0, and any H∗∈Adm(R;h), there exists a non-intersecting Bernoulli paths {x(t)}0⩽t⩽T from x(0)=y(N)/N to x(T)=z(N)/N with height function H, such that on R provided N is large enough. We introduce the following ℓ-mesh, and Lemma 2.3 states that on most of the ℓ-mesh, H∗ has an approximate linear approximation. Definition 2.2 . Let h be a boundary height function as in Definition 1.3, there exists a large constant A>0, such that for any H∈Adm(R;h), Take a small ℓ>0 such that A/ℓ,T/ℓ∈Z. We cover the region [-A,A]×[0,T] by rectangles of size 6ℓ×ℓ (see Fig. 2): where α,β are integer numbers so that -A/ℓ⩽α⩽A/ℓ,0⩽β⩽T/ℓ. Shown above is the ℓ-mesh. Lemma 2.3 Let H be an asymptotic height function, namely a given element of Adm(R,h), and let ε0>0. If ℓ is sufficiently small then on at least a 1-ε0 fraction of the rectangles R(α,β) from the ℓ-mesh (in Definition 2.2) we have the following two properties On R(α,β), H has a linear approximation within an error bounded by ℓε0, namely there exists (ϱ,-ϱv)∈T¯ such that for (x,t)∈R(α,β)|H(x,t)-(H(αℓ,βℓ)+ϱ(x-αℓ)-ϱv(t-βℓ))|⩽ε0ℓ. On R(α,β), H has a linear approximation within an error bounded by ℓε0, namely there exists (ϱ,-ϱv)∈T¯ such that for (x,t)∈R(α,β)|H(x,t)-(H(αℓ,βℓ)+ϱ(x-αℓ)-ϱv(t-βℓ))|⩽ε0ℓ. For at least a 1-ε0 fraction (in measure) of the points x of R(α,β), the gradient ∇H(x) exists and is within ε0 of (ϱ,-ϱv). For at least a 1-ε0 fraction (in measure) of the points x of R(α,β), the gradient ∇H(x) exists and is within ε0 of (ϱ,-ϱv). Proof The proof is the same as in [20, Lemma 2.2], so we omit. □ Lemma 2.3 states that any height function H∈Adm(R,h) locally can be approximated by height functions with constant slope. We introduce the following height functions with constant slope. Definition 2.4 Given a slope (ϱ,-ϱv)∈T¯, we construct the constant slope height function A on R×[0,ℓ]: where P is the parallelogram region: We say A is the height function with slope given by (ϱ,-ϱv). Vandermonde Determinant and Free Entropy Lemmas 2.5 and 2.6 concern certain useful identity and estimates of Vandermonde determinants. Lemma 2.5 For any particle configuration x∈Wθn, and 0⩽k⩽n, the following holds Proof Since Vandermonde determinant is anti-symmetric, it follows that as a function of x1,x2,⋯,xn, is anti-symmetric. Therefore, it has the Vandermonde determinant V(x) as a factor. We notice that for every θ, the coefficient of the monomial x1n-1x2n-2⋯xn-1 of degree n(n-1)/2 in V(x+θe) is 1. The claim (2.4) follows from that the coefficient of x1n-1x2n-2⋯xn-1 in (2.5) is nk, and in V(x) is 1. □ Lemma 2.6 There exists a constant C>0 such that for any particle configuration x∈Wθn, and any y=x+e∈Wθn (this is equivalent to that V(x+θe)≠0), the following hold Proof The upper bound in (2.6) follows from Lemma 2.5 which implies and the fact that all terms in the left hand side are non-negative. For the lower bound, we first show that there exists a constant C=C(θ) such that for every configuration x, If ei-ej=0, there is nothing to prove, the lefthand side is 0. If xi-xj=(j-i)θ, i.e. the particle configuration on {xi,xi+1,⋯,xj} is tightly packed. The quantity V(x+θe) is nonzero only if ei-ej∈{0,1}. Thus, Otherwise, xi-xj⩾1+θ, and we have The statement (2.8) follows from the inequality below which follows from the fact that for x∈[-(1+θ)-1,θ-1] Summing over i, j in (2.8), we deduce By the same argument as for (2.9) (replacing e by -e), we also have that x=y-e satisfies where the upper bound by nln2 comes from (2.6). Combining (2.9) and (2.10), we obtain the lower bound in (2.6) □ In the following lemma we collect certain estimates on the limit of the Vandermonde determinants, and on the free entropy. Lemma 2.7 Fix κ<0 and A>0. There exists a constant C=C(κ,A)>0, such that the following holds. For any particle configuration x∈WθN, let x∈x/N. Assume x is supported on [-A,A], and we denote its empirical density as ϱ(x;x), then For any two height functions h1,h2 such that ‖h1-h2‖∞⩽ε with 0⩽∂xh1,∂xh2⩽1 and their supports supp(∂xh1),supp(∂xh2)∈[-A,A], we have Proof We will only prove the first statements in (2.11) and (2.12), and the second statements follow from the same arguments. For any i0 such that Lemma 2.3 holds and L=ℓN∈N. We introduce auxiliary parameters 0<ε,δ,ξ≪ℓ and 0<ζ≪1, and assume they satisfy These constraints are compatible. For instance, one may take ζ=ζ(N)→0 and choose in which case (3.1) holds as ζ→0. We denote and Bε(H∗) the set of non-intersecting θ-Bernoulli walks with height function close to H∗ Upper Bound To prove the large deviation upper bound for P(Bε(H∗)), we partition Bε(H∗) according to the possible particle configurations at times NL: where y(0)=y(N) and y(T)=z(N). By our assumption that |yi(N)|,|zi(N)|⩽CN, then there exists A>0 (say A=C+T) such that the particle configuration y(t) is supported on [-AN,AN] for all 0⩽t⩽T. The number of total choices of {y(t)}t∈{0,L,2L,⋯,T} is given by which is negligible. In what follows we fix one such sequence {y(t)}t∈{0,L,2L,⋯,T}. Then any path has the same fixed particle configurations at these mesh times, i.e. t∈{0,L,2L,⋯,T}. We will group paths according to their behavior on each ℓ-mesh as introduced in Definition 2.2, and show that on each such patch the height function is close to a linear profile with an appropriate slope. We fix integer numbers α,β, and recall the ℓ-mesh and rectangle R(α,β)=[(α-3)ℓ,(α+3)ℓ]×[βℓ,(β+1)ℓ] from Definition 2.2. We consider the restriction of the particle configuration y(βℓ) on the interval [αℓ,(α+1)ℓ) We denote the index set of the above restricted particle configuration as I(α,β)={i,i+1,i+2,⋯,j} and denote n:=|I(α,β)|=j-i+1. For any p∈Bε(H∗:{y(t)}t∈{0,L,2L,⋯,T}), its restriction on the index set I(α,β) and time interval [βℓ,(β+1)ℓ], {xk(t)}k∈I(α,β),t∈[βℓ,(β+1)ℓ] forms an n-particle nonintersecting Bernoulli random walk inside the rectangle R(α,β). We remark that during the interval [βℓ,(β+1)ℓ], it is in principle possible that additional paths with labels j) enter the spatial window [(α-3)ℓ,(α+3)ℓ). In the lemma below we do not attempt to track all particles that may visit this window. Instead, we focus on the sub-ensemble of paths with labels in I(α,β)={i,⋯,j}. The associated height function Hs records the contribution of this sub-ensemble. The main conclusion of Lemma 3.1 is that Hs is uniformly close on R(α,β) to a constant-slope height function A. Lemma 3.1 There exists a large finite constant C>1 so that the following holds. Assume that on R(α,β), the height function H∗ has a linear approximation with slope (ϱ,-ϱv)∈T and error ε0=ε/ℓ in the sense of Lemma 2.3. For any {x(t)}0⩽t⩽T∈Bε(H∗:{y(t)}t∈{0,L,2L,⋯,T}), we denote the shifted and rescaled particle configuration {zk(t)}1⩽k⩽n,0⩽t⩽ℓ Then {zk(t)}1⩽k⩽n,0⩽t⩽ℓ form an n-particle nonintersecting Bernoulli random walk, where and its height function Hs (defined as in (1.11)) satisfies where A is the height function with slope given by (ϱ,-ϱv) from Definition 2.4. Proof We denote the height function of the rescaled particle configuration {x(t)}0⩽t⩽T as H, then from the construction (3.5), Hs is a shifted version of H. More precisely, for any 0⩽t⩽ℓ, And for x⩽zn(t) we have Hs(x,t)=0; for x⩾z1(t)+θ/N we have Hs(x,t)=nθ/N. Next, we show (3.6). Since H∈Bε(H∗), we have The claim (3.6) follows by rearranging. We will prove (3.7), by dividing it into different regions. For x⩽min{zn(t),tv}, Hs(x,t)=A(x,t)=0. For x⩽min{zn(t),tv}, Hs(x,t)=A(x,t)=0. For x⩾min{z1(t)+θ/N,tv+ℓ}, Hs(x,t)=θn/N and A(x,t)=ℓϱ, and the claim (3.7) follows from (3.6) . For x⩾min{z1(t)+θ/N,tv+ℓ}, Hs(x,t)=θn/N and A(x,t)=ℓϱ, and the claim (3.7) follows from (3.6) . For zn(t)⩽x⩽z1(t)+θ/N, we have 3.10=Hs(x,t)=H(x+αℓ,t+βℓ)-(n-j)θN=H(x+αℓ,t+βℓ)-H(αℓ,βℓ)=H∗(x+αℓ,t+βℓ)-H∗(αℓ,βℓ)+O(ε)=xϱ-tϱv+O(ε). It follows that for max{zn(t),tv}⩽x⩽min{z1(t)+θ/N,tv+ℓ}, we have |Hs(x,t)-A(x,t)|=O(ε). For zn(t)⩽x⩽z1(t)+θ/N, we have 3.10=Hs(x,t)=H(x+αℓ,t+βℓ)-(n-j)θN=H(x+αℓ,t+βℓ)-H(αℓ,βℓ)=H∗(x+αℓ,t+βℓ)-H∗(αℓ,βℓ)+O(ε)=xϱ-tϱv+O(ε). It follows that for max{zn(t),tv}⩽x⩽min{z1(t)+θ/N,tv+ℓ}, we have |Hs(x,t)-A(x,t)|=O(ε). If tv⩽zn(t), then for tv⩽x⩽zn(t) we have 0=Hs(x,t)⩾H(x+αℓ,t+βℓ)-H(αℓ,βℓ)=xϱ-tϱv+O(ε)=A(x,t)+O(ε), where in the second statement we used the construction of (3.10); in the third statement we used that H∗ and H are close, and H∗ has a linear approximation with slope (ϱ,-ϱv). The claim (3.7) follows. If tv⩽zn(t), then for tv⩽x⩽zn(t) we have 0=Hs(x,t)⩾H(x+αℓ,t+βℓ)-H(αℓ,βℓ)=xϱ-tϱv+O(ε)=A(x,t)+O(ε), where in the second statement we used the construction of (3.10); in the third statement we used that H∗ and H are close, and H∗ has a linear approximation with slope (ϱ,-ϱv). The claim (3.7) follows. If zn(t)⩽tv, then for zn(t)⩽x⩽tv we have A(x,t)=0, and 0⩽Hs(x,t)=xϱ-tϱv+O(ε)⩽O(ε), where in the second statement we used (3.10); in the last statement we used x⩽tv. If zn(t)⩽tv, then for zn(t)⩽x⩽tv we have A(x,t)=0, and 0⩽Hs(x,t)=xϱ-tϱv+O(ε)⩽O(ε), where in the second statement we used (3.10); in the last statement we used x⩽tv. If z1(t)+θ/N⩽tv+ℓ, then for z1(t)+θ/N⩽x⩽tv+ℓ we have θnN=Hs(x,t)⩽H(x+αℓ,t+βℓ)-H(αℓ,βℓ)=xϱ-tϱv+O(ε)=A(x,t)+O(ε), where in the second statement we used the construction of (3.10); in the third statement we used that H∗ and H are close, and H∗ has a linear approximation with slope (ϱ,-ϱv). The claim (3.7) follows from the above estimate together with (3.9). If z1(t)+θ/N⩽tv+ℓ, then for z1(t)+θ/N⩽x⩽tv+ℓ we have θnN=Hs(x,t)⩽H(x+αℓ,t+βℓ)-H(αℓ,βℓ)=xϱ-tϱv+O(ε)=A(x,t)+O(ε), where in the second statement we used the construction of (3.10); in the third statement we used that H∗ and H are close, and H∗ has a linear approximation with slope (ϱ,-ϱv). The claim (3.7) follows from the above estimate together with (3.9). If tv+ℓ⩽z1(t)+θ/N, then for tv+ℓ⩽x⩽z1(t)+θ/N we have A(x,t)=ϱℓ, and 3.11θnN⩾Hs(x,t)=xϱ-tϱv+O(ε)⩾ℓϱ+O(ε). where in the second statement we used (3.10); in the last statement we used x⩾ℓ+tv. The claim (3.7) follows from combining (3.11) with (3.9). □ If tv+ℓ⩽z1(t)+θ/N, then for tv+ℓ⩽x⩽z1(t)+θ/N we have A(x,t)=ϱℓ, and 3.11θnN⩾Hs(x,t)=xϱ-tϱv+O(ε)⩾ℓϱ+O(ε). where in the second statement we used (3.10); in the last statement we used x⩾ℓ+tv. The claim (3.7) follows from combining (3.11) with (3.9). □ With the above decomposition, we can rewrite the probability P(Bε(H∗:{y(t)}t∈{0,L,2L,⋯,T})) as where the summation is over paths {x(t)}0⩽t⩽T∈Bε(H∗:{y(t)}t∈{0,L,2L,⋯,T}). The large deviation upper bound in Theorem 1.6 is a consequence of the following two statements. We postpone their proofs in Sect. 4. Proposition 3.2 Adopt the notations in Theorem 1.6. Fix any H∗∈Adm(R,h), and a particle configuration {y(t)}0⩽t⩽T∈Bε(H∗), with y(0)=y(N) and y(T)=z(N). There exists a constant C>0, where the summation is over {xk(t)}k∈I(α,β),t∈[βL,(β+1)L] with xk(βL)=yk(βL). If we assume that on R(α,β), the height function H∗ has a linear approximation with slope (ϱ,-ϱv)∈T and error ε0=ε/ℓ in the sense of Lemma 2.3, then where the summation is over {xk(t)}k∈I(α,β),t∈[βL,(β+1)L] with xk(βL)=yk(βL) and (3.7) holds. Lemma 3.3 Adopt notations in Theorem 1.6. Fix any H∗∈Adm(R,h), and a particle configuration {y(t)}0⩽t⩽T∈Bε(H∗), with y(0)=y(N) and y(T)=z(N). For any {x(t)}0⩽t⩽T∈Bε(H∗:{y(t)}t∈{0,L,2L,⋯,T}), the following holds Proof of Large Deviation Upper bound in Theorem 1.6 We recall from (3.4) that the number of total choices of {y(t)}t∈{0,L,2L,⋯,T} is negligible. In the rest of the proof we fix any choice of {y(t)}t∈{0,L,2L,⋯,T}. Thanks to Lemma 3.3, we can bound (3.12) as where in the second inequality we upper bound the global sum by collecting, for each patch (α,β), the contribution of the paths restricted to R(α,β) and then applying Proposition 3.2 to the resulting patchwise partition function. Moreover, among all patches (α,β), the proportion for which the linear approximation of H∗ fails is at most ε0; on these exceptional patches we use the crude bound (3.13), so their total contribution to the normalized logarithm is O(ε0). In the last line we used Lemma 2.7. The large deviation upper bound in Theorem 1.6 follows by using the choice of parameters from (3.1). □ Lower Bound To prove the large deviation lower bound, we recall the height function H as constructed in (2.1) and the associated path configuration {y(t)}0⩽t⩽T. We recall the ℓ-mesh from Definition 2.2. Take any R(α,β), the same as in the upper bound, we consider the restriction of the particle configuration y(βℓ) on the interval [αℓ,(α+1)ℓ) Denote the index set I(α,β)={i,i+1,i+2,⋯,j}. The restriction of {y(t)}0⩽t⩽T on the index set I(α,β):=n and time interval [βℓ,(β+1)ℓ], {yk(t)}k∈I(α,β),t∈[βℓ,(β+1)ℓ] form an n-particle nonintersecting Bernoulli random walk inside R(α,β). The shifted particle configuration form an n-particle nonintersecting Bernoulli random walk on the parallelogram shaped region Fix a small parameter 0<ζ≪1 satisfying (3.1), we define the sub-region for slopes Tζ To obtain a large deviation lower bound, we will restrict to the following set of N-particle nonintersecting Bernoulli random walks {x(t)}0⩽t⩽T: for any α,β, Suppose that on R(α,β), the height function H∗ has a linear approximation with slope (ϱ,-ϱv)∈Tζ (recall from (3.19)) and error ε0=ε/ℓ in the sense of Lemma 2.3. We denote the shifted particle configuration {zk(t)}1⩽k⩽n,0⩽t⩽ℓ3.20zk(t)=yk+i-1(t+βℓ)-αℓ,0⩽t⩽ℓ,1⩽k⩽n. Then zk(t) form an n-particle nonintersecting Bernoulli walk inside P~ from (3.18). Recall that A is the height function with slope given by (ϱ,-ϱv) from Definition 2.4. Then A is a linear approximation of the height function of {zk(t)}1⩽k⩽n,0⩽t⩽ℓ, with L∞ distance bounded by Cε. We restrict to x(t) such that the height function Hs of {xk+i-1(t+βℓ)-αℓ}1⩽k⩽n,0⩽t⩽ℓ satisfies 3.21(xk+i-1(t+βℓ)-αℓ,t)∈P~,Hs(x,t)-A(x,t)⩽Cε,0⩽t⩽ℓ. Suppose that on R(α,β), the height function H∗ has a linear approximation with slope (ϱ,-ϱv)∈Tζ (recall from (3.19)) and error ε0=ε/ℓ in the sense of Lemma 2.3. We denote the shifted particle configuration {zk(t)}1⩽k⩽n,0⩽t⩽ℓ3.20zk(t)=yk+i-1(t+βℓ)-αℓ,0⩽t⩽ℓ,1⩽k⩽n. Then zk(t) form an n-particle nonintersecting Bernoulli walk inside P~ from (3.18). Recall that A is the height function with slope given by (ϱ,-ϱv) from Definition 2.4. Then A is a linear approximation of the height function of {zk(t)}1⩽k⩽n,0⩽t⩽ℓ, with L∞ distance bounded by Cε. We restrict to x(t) such that the height function Hs of {xk+i-1(t+βℓ)-αℓ}1⩽k⩽n,0⩽t⩽ℓ satisfies 3.21(xk+i-1(t+βℓ)-αℓ,t)∈P~,Hs(x,t)-A(x,t)⩽Cε,0⩽t⩽ℓ. Otherwise if the height function H∗ does not admit a linear approximation on R(α,β), then we require that the path configuration x(t) to satisfy 3.22{xk(t)}k∈I(α,β),t∈[βℓ,(β+1)ℓ]={yk(t)}k∈I(α,β),t∈[βℓ,(β+1)ℓ]. Otherwise if the height function H∗ does not admit a linear approximation on R(α,β), then we require that the path configuration x(t) to satisfy 3.22{xk(t)}k∈I(α,β),t∈[βℓ,(β+1)ℓ]={yk(t)}k∈I(α,β),t∈[βℓ,(β+1)ℓ]. Lemma 3.4 Fix a path configuration {y(t)}0⩽t⩽T with height function H satisfying (3.15). For any N-particle nonintersecting Bernoulli random walks {x(t)}0⩽t⩽T satisfying (3.21) and (3.22), its height function H^ satisfies Proof For any α,β, we recall the index set I(α,β)={i,i+1,i+2,⋯,j} from (3.16). There are two cases. If on R(α,β), H∗ has a linear approximation with slope (ϱ,-ϱv)∈Tζ, then for βℓ⩽t⩽(β+1)ℓ and yj(t)⩽x⩽yi(t), we have where we used (3.15). And using (3.21), the same argument as in (3.10) The claim (3.23) follows from combining (3.24) and (3.25). In the second case that on R(α,β), either H∗ has a linear approximation with slope (ϱ,-ϱv)∉Tζ, or H∗ does not have a linear approximation. Then our construction gives and (3.23) holds trivially. □ The large deviation lower bound in Theorem 1.6 is a consequence of the following statement. We postpone its proof in Sect. 5. Proposition 3.5 Adopt notations from Theorem 1.6. Fix any H∗∈Adm(R,h), and denote {y(t)}0⩽t⩽T the particle configuration constructed in Lemma 2.1. There exists a constant C>0, where the summation is over {xk(t)=xk(Nt)/N}k∈I(α,β),t∈[βℓ,(β+1)ℓ] such that (3.21) holds. If we assume that on R(α,β), the height function H∗ has a linear approximation with slope (ϱ,-ϱv)∈Tζ (recall from (3.19)) and error ε0=ε/ℓ in the sense of Lemma 2.3. where the summation is over {xk(t)=xk(Nt)/N}k∈I(α,β),t∈[βℓ,(β+1)ℓ] such that (3.21) holds. Proof of Large Deviation Lower bound in Theorem 1.6 The same as in (3.12), the large deviation lower bound in Theorem 1.6 follows from a lower bound of the following quantity where the constant C in BCε(H∗) is from (3.23). Fix a path configuration {y(t)}0⩽t⩽T with height function H satisfying (3.15). We recall from Lemma 3.4 that the height function H^ of any paths configurations {x(t)}0⩽t⩽T satisfying (3.21) and (3.22) is inside the ball BCε(H∗). We have where in the first inequality, we used Lemma 3.3; in the second equality, we lower bound the sum over BCε(H∗) by path configurations {x(t)}0⩽t⩽T satisfying (3.21) and (3.22); we also used that the fraction of (α,β) such that on R(α,β), H∗ does not have a linear approximation is bounded by ε0; in the last line we used Lemma 2.7. □ Large Deviation Upper Bound: Constant Slope Case In this section we study the following n-particle non-intersecting θ-Bernoulli walk from time 0 to L=ℓN (x(0),x(1),…,x(L))∈(Wθn)L (see Definition 1.1) Then we will prove Proposition 3.2, the large deviation upper bound for the non-intersecting θ-Bernoulli walks with height function approximately linear. We recall the height function A(x,t) with constant slope ∇A(x,t)=(ϱ,-ϱv)∈T¯ from Definition 2.4, where P is the parallelogram region We assume that the density satisfies (3.6) Fix a small parameter 0<ζ≪1 satisfying (3.1), we recall Tζ from (3.19) There are two cases, Extreme slope case where (ϱ,-ϱv)∉Tζ. Extreme slope case where (ϱ,-ϱv)∉Tζ. Interior slope case where (ϱ,-ϱv)∈Tζ. Interior slope case where (ϱ,-ϱv)∈Tζ. The following proposition gives the large deviation upper bound for n-particle non-intersecting θ-Bernoulli walks with height function having approximately constant slope. Its proofs are given in Sects. 4.2 and 4.4. Proposition 4.1 We recall the rate function σ from (1.14), and parameters ζ,δ,ε,ℓ from (3.1). Fix (ϱ,-ϱv)∈T¯ such that (4.4) holds. Take the height function A(x,t) with constant slope ∇A=(ϱ,-ϱv) on the parallelogram region P, as in (2.2). The height function H of the n-particle non-intersecting θ-Bernoulli walk (4.1) satisfies If (ϱ,-ϱv)∉Tζ, then we have 4.51(ℓN)2lnP(‖H-A‖∞⩽ε)=-nℓNln(2)+O(ζ1/2ln(1/ζ)). If (ϱ,-ϱv)∉Tζ, then we have 4.51(ℓN)2lnP(‖H-A‖∞⩽ε)=-nℓNln(2)+O(ζ1/2ln(1/ζ)). If (ϱ,-ϱv)∈Tζ, then we have 1(ℓN)2lnP(‖H-A‖∞⩽ε)⩽σ(ϱ,-ϱv)θ-nℓNln(2)+Oεδζ2+(δ/ℓ)log2(ℓ/δ). If (ϱ,-ϱv)∈Tζ, then we have 1(ℓN)2lnP(‖H-A‖∞⩽ε)⩽σ(ϱ,-ϱv)θ-nℓNln(2)+Oεδζ2+(δ/ℓ)log2(ℓ/δ). Proof of Proposition 3.2 The first statement (3.13) in Proposition 3.2 follows from Lemma 2.6. For the second statement (3.14), we recall from Lemma 2.3 that at least 1-ε0 fraction of the points x of R(α,β), the gradient ∇H∗ exists and is within ε0 of (ϱ,-ϱv). As a consequence, If (ϱ,-ϱv)∈Tζ, then the claim (3.14) follows from (4.6) and the second statement of Proposition 4.1. If (ϱ,-ϱv)∉Tζ, without loss of generality we assume that ϱ⩽ζ. The Lobachevsky function (recall from (1.14)) is π-periodic, namely L(x+π)=L(x) and -L(π-x)=L(x). Moreover, because of the logarithmic singularity, the Lobachevsky function satisfies and L(ζ)=O(ζlog(1/ζ)) for ζ→0+. Therefore The claim (3.14) follows from (4.5) and (4.8). □ Outline of the Proof of Proposition 4.1 For the extreme slope case, to estimate P(‖H-A‖∞⩽ε), we will directly upper bound the number of height profiles satisfying ‖H-A‖∞⩽ε by a combinatorics argument, and the probability P(H) for each of these height profile by a delicate analysis of Vandermonde determinants. It turns out both parts are negligible. The proof is given in Sect. 4.2. For the interior slope case, we recall the constant slope height function A from (4.2). We convolve it with a Cauchy distribution. Take small δ>0, denote Since for (x,t)∈P, ∇A(x,t)=(ϱ,-ϱv)∈Tζ, namely, ζ⩽ϱ,ϱv,ϱ(1-v)⩽1-ζ, it follows that ∇H~(x,t)∈T for (x,t)∈R. Properties of the smoothed height function H~(x,t) are collected in Sect. 4.3. We define the associated complex slope f~:R→H- by setting, for any (x,t)∈R, f~t(x)=f~(x,t)∈H- to be the unique complex number satisfying where for any z∈H- we have set arg∗z=θ∈(-π,0) to be the unique number in (-π,0) satisfying e-iθz∈R; see Fig. 3 for a depiction. By the law of sines (4.10) implies that From (4.11), we can solve the complex slope f~t(x), Shown above is the complex slope f~=f~(x,t). We take a drift function gt(x) given by As we will show in Sect. 4.3, the complex slope f~t(x), and the drift function gt(x) can both be extended analytically to a strip neighborhood of the real axis, and on this strip region they are related by We notice by Fubini, For Imz>0, the inner integral equals 1z+iδ-y (a standard Cauchy/Poisson identity), and therefore To prove the large deviation upper bound for the interior case in Proposition 3.2, we tilt the law of the Markov process (4.1) by an exponential martingale. For any 0⩽t⩽L-1, let and the following is a Martingale In particular it gives the following identity The large deviation upper bound follows from restricting the above identity to the event {‖H-A‖∞⩽ε)}. In Sect. 4.4, we will show that on the event, {‖H(x,t)-A(x,t)‖∞⩽ε)}, up to a small error, the numerator in the exponential martingale (4.15) can be written in terms of A(x,t). For the denominator, we will use the dynamical loop equation Theorem A.5 to compute its expectation. They together lead to the following large deviation upper bound. Proposition 4.2 Adopt notations and assumptions from Proposition 4.1. Take the height function A(x,t) with constant slope ∇A=(ϱ,-ϱv) on the parallelogram region P, as in (2.2). If (ϱ,-ϱv)∈Tζ, then the height function H of the n-particle non-intersecting θ-Bernoulli walk (4.1) satisfies where gt(x) is the drift function from (4.13). There exists a large constant K>0, gt(x) can be extended analytically over the following annulus region and S(A;g) in (4.18) is given by where the contour ω⊂Λ encloses [-3ℓ,3ℓ] and mt∗(z) is the Stieltjes transform of ∂xA(x,t), We will then analyze the rate function S(A;g) in (4.20) in Sect. 4.5. Miraculously, after simplifying the rate function, we recover a combination of Lobachevsky functions (see (1.14)), leading to the surface tension. Proposition 4.3 Adopt the notations and assumptions in Proposition 4.2. The rate function satisfies Proof of Proposition 4.1: Extreme Slope Case There are three cases for the extreme slope (ϱ,-ϱv)∉Tζ (recall from (3.19)): i) ϱv⩽ζ ii) ϱ(1-v)⩽ζ iii) ϱ⩾1-ζ. We recall from (3.1) that ε≪ζℓ. We start with the first two cases. Their proofs are the same, so we will only give the proof of the first case. Proof of the First and Second Cases In the first case -ζ⩽∂tA⩽0, and most of ei(t) are zero. In fact, we have by integration by parts where we used that ε≪ζℓ from (3.1). We denote the time set Then (4.23) gives that |I∁|=O(ζ1/2ℓN). There are ℓNO(ζ1/2ℓN)=eO(-ζ1/2ln(ζ)(ℓN)) ways to choose the set I∁. Recall the definition (1.12) of the weight W. Once we fix the sets I,I∁, the total weights of non-intersecting Bernoulli walks satisfying (4.24) is where we used that from Lemma 2.5. It follows from (4.25) and the fact that there at most eO(-ζ1/2ln(ζ)(ℓN)) ways to choose the set I∁, we conclude (4.5) for -ϱv⩽ζ. In the second case that 1-ζ⩽∂xA+∂tA=ϱ(1-v)⩽1, we have that most of ei(t) are one. By the same argument as the first case, we will have that (4.5) holds. □ For the last case, we have 1-ζ⩽∂xA=ϱ⩽1, and (1-ζ)ℓN⩽n⩽ℓN, we need to bound the total weight of the paths p corresponding to H, We remark that the normalization constant 2-nL gives the term -(n/ℓN)ln2 in (4.5). We denote the index sets For ‖H-A‖∞⩽ε⩽ζℓ, we have that for any index i∈I1, xi(t)∈[tv,ℓ+tv]. Fix any time 0⩽t⩽T. For i∈I0∪I2, there are 2|I0|+|I2|=22ζℓN possible choices for ei(t). For the particle configuration {xi(t)+ei(t)}i∈I1, using again that ‖H-A‖∞⩽ε⩽ζℓ, the number of empty sites is bounded by Therefore, the number of configurations p satisfying ‖H-A‖∞⩽ε is at most The following lemma bounds each summand in (4.26) Lemma 4.4 For non-intersecting Bernoulli walks p={xi(t)}1⩽i⩽n,0⩽t⩽L satisfying that ‖H-A‖∞⩽ε, the following holds Proof To simplify the notation, we will omit the time t dependence. We recall from (2.8) Next, we show that for ‖H-A‖∞⩽ε, Let I={i∈[[1,n]]:ei=1} and J={j∈[[1,n]]:ej=0}. The particle configuration {xi}1⩽i⩽n consists of at most O(ζℓN) tightly packed pieces. Namely, there exists 1=i0x2>⋯>x|I1|, we add an empty site between xk and xk+1, and get the particle configuration x1+1>x2+1>⋯>xk+1>xk+1>⋯x|I1|, then where we used that the multiset {|i-j|:i⩽k,j⩾k+1} contains at most m copies of m for any 1⩽m⩽ℓN. □ Complex Slope Estimates In this section, we prove that the complex slope f~t(x) and the drift function gt(x) (as constructed in (4.10) and (4.13)) can be extended analytically to a strip neighborhood of the real axis in Lemma 4.7. This step is crucial to use the dynamical loop equations, since for instance Theorem A.3 requires the functions ϕ± to be holomorphic. In Claim 4.5 and Lemma 4.6, we collect some estimates of the smoothed height function H~(x,t) (as constructed in (4.9)) . Their proofs follow from explicit computations, so we postpone them to Sect. B. Lemma 4.5 Fix any 0⩽t⩽ℓ, δ≪ℓ, and denote then The function κt(x)∈[0,1] and for x∈[tv,ℓ+tv], it holds for x∉[tv,ℓ+tv], it holds We recall that the Hilbert transform Hib(u) of a function u:R↦R is given by Hib(u)(t)=PV∫Ru(x)/(t-x)dx where PV denotes the principal value. Lemma 4.6 There exists a large constant C>0, for any small 0<δ≪ℓ, recall κt(x) from (4.37), the following holds κt(x) can be extended analytically to any z=x+iη with Im[η]⩽δ/2, and 4.40|κt(z)-κt(x)|⩽Cηδmin{κt(x),1-κt(x)}. κt(x) can be extended analytically to any z=x+iη with Im[η]⩽δ/2, and 4.40|κt(z)-κt(x)|⩽Cηδmin{κt(x),1-κt(x)}. The Hilbert transform Hib(κt) extends analytically to any z=x+iη with Im[η]⩽δ/2, 4.41Hib(κt)(z)⩽Cδdist(z,{tv,ℓ+tv})ifdist(z,{tv,ℓ+tv})⩾ℓln(ℓ/δ)+Cfor allz,ImHib(κt)(z)⩽Cηδ. The Hilbert transform Hib(κt) extends analytically to any z=x+iη with Im[η]⩽δ/2, 4.41Hib(κt)(z)⩽Cδdist(z,{tv,ℓ+tv})ifdist(z,{tv,ℓ+tv})⩾ℓln(ℓ/δ)+Cfor allz,ImHib(κt)(z)⩽Cηδ. If dist(z,{tv,ℓ+tv})≳ℓ, we have We recall the complex slope f~t(x) and drift function gt(x) from (4.12) and (4.13) The following lemma states that the complex slope f~t(x) and the drift function gt(x) can be extended analytically to a strip neighborhood of the real axis. Its proof follows from explicit computations, so we postpone it to “Appendix B”. Lemma 4.7 There exists a large constant C>0, for any 0<δ≪ℓ, ζ≪1, extreme slope (ϱ,-ϱv)∉Tζ (recall from (3.19)), f~t(x) and gt(x) as in (4.43) can be extended analytically to the strip region D=D(C)={x+iη:|η|⩽C-1δ}, and for any z∈D the following holds The norm of f~t(z) satisfies: |f~t(z)|⩽C/ζ. The norm of f~t(z) satisfies: |f~t(z)|⩽C/ζ. The norm of gt(z) satisfies: |gt(z)|⩽ln(1/ζ)+ln(ℓ/δ)+C. If dist(z,{tv,ℓ+tv})⩾ℓ, we have 4.44lnsin(πϱvκt(z))sin(πϱ(1-v)κt(z))=lnsin(v)sin(1-v)+Oδdist(z,{tv,ℓ+tv}). The imaginary part of gt(z) satisfies: |Im[gt(z)]|⩽CIm[z]/δ. The norm of gt(z) satisfies: |gt(z)|⩽ln(1/ζ)+ln(ℓ/δ)+C. If dist(z,{tv,ℓ+tv})⩾ℓ, we have 4.44lnsin(πϱvκt(z))sin(πϱ(1-v)κt(z))=lnsin(v)sin(1-v)+Oδdist(z,{tv,ℓ+tv}). The imaginary part of gt(z) satisfies: |Im[gt(z)]|⩽CIm[z]/δ. The derivative of gt(x) satisfies: for any x∈R, |∂xgt(x)|,|∂tgt(x)|⩽C/δ. The derivative of gt(x) satisfies: for any x∈R, |∂xgt(x)|,|∂tgt(x)|⩽C/δ. Proof of Proposition 4.2: Interior Slope Case In this section we prove Proposition 4.2, which implies Proposition 3.2 for the interior slope case. It is a consequence of the following two lemmas. Lemma 4.8 Adopt the notations and assumptions in Proposition 4.2. Let H be a height function with ‖H-A‖∞⩽ε, and denote {x(t)}0⩽t⩽ℓ the particle configuration associated with H(x,t), then Lemma 4.9 Adopt the notations and assumptions in Proposition 4.2. Let H(x,t) be a height function with ‖H-A‖∞⩽ε, and denote {x(t)}0⩽t⩽ℓ the particle configuration associated with H(x,t), then for any 0⩽t=tN⩽L-1, where ω is a contour enclosing the interval [-3ℓ,3ℓ], mt∗(z) is defined in (4.21) and mt(z) is the Stieltjes transform of the empirical measure of x(t), Proof of Proposition 4.2 Recall from (4.17) that Proposition 4.2 follows from plugging Lemma 4.8 and Lemma 4.9 into (4.47), the error term simplifies using the relation (3.1). □ Proof of Lemma 4.8 Let Gt(z)=∫0zgt(x)dx. We first notice that, for t=t/N, where in the first equality follows from ρ(x;x(t))=∑i=1n1(x∈[xi(t),xi(t)+θ/N]); in the second equality we used n⩾N. Thus for the lefthand side of (4.45), we have This gives the first statement in (4.45). Next we replace H by A, we can do an integration by part in t and use ‖H-A‖∞⩽ε to deduce that Thanks to Lemma 4.7, we have ‖g‖∞⩽ln(1/ζ)+ln(ℓ/δ)+C and |∂tgt(x)|⩽C/δ, and the second statement of (4.45) follows from (4.49). □ In the following, we prove Lemma 4.9. For simplicity of notations, we will omit the dependence on time t. We can rewrite the lefthand side of (4.46) as which is the expectation of ∑i=1neig(xi/N) under the following deformed measure where Zτ is the normalization factor. We denote Eτ[·] the expectation with respect to the measure (4.51). We will use the dynamical loop equation Theorem A.3 and Theorem A.5 to compute the expectation Take a large constant 8C/(δζ)⩽K⩽ζ/(2ε), where the constant C>0 is from Lemma 4.7. This is possible thanks to our choice of parameters from (3.1). We denote the annulus region as in Proposition 4.2 We need to verify the conditon that on Λ, the following are well defined, and lower and upper bounded Lemma 4.10 Take 8C/(δζ)⩽K⩽ζ/(2ε), then on Λ from (4.19), we have and ln(1+f(τ)(z)) is well defined. Proof We first show that for any z∈Λ Since Im[m(z)]=-Im[m(z¯)], we will only prove (4.54) for z∈Λ∩H¯. By the assumption of our initial data (4.1), and the construction of A in (4.2), for |x|⩾3ℓ, ∂xH(x,t)=∂xA(x,t)=0. Therefore, for any z∈Λ, we find where in the third statement we did an integration by part in x; in the last inequality we used that K⩽ζ/(2ε). We can also compute m∗(z) and Im[m∗(z)] explicitly and The claim (4.54) follows from combining (4.55) and (4.57). From Lemma 4.7, |Img(z)|⩽CIm[z]/δ⩽2C/(δK). Using (4.54), for any z∈Λ, we have |Img(z)|⩽2C/(δK)⩽ζ/4, it follows that It follows that argem(z)+τg(z)∈(-(1-ζ/10)π,(1-ζ/10)π). Thus 1+em(z)+τg(z) takes value in the sector region given by {1+reiu:u∈(-(1-ζ/10)π,(1-ζ/10)π)}, which does not include the negative axis. Therefore, we can take ln(·) the branch defined on C\R⩽0 and ln(1+em(z)+τg(z)) is well defined. Moreover we conclude that for z∈Λ, it holds |1+em(z)+τg(z)|⩾sin(ζπ/10)≳ζ. The last statement of (4.53) also follows. On Λ, since dist(z,[-3ℓ,3ℓ])⩾1/K, we have where we used (4.56) to bound |Rem∗(z)| and (4.55) to bound |m(z)-m∗(z)|. From the second statement of Lemma 4.7, for any z∈Λ, we have |g(z)|⩽ln(1/ζ)+ln(ℓ/δ)+C, thus |1+em(z)+τg(z)|≲ℓ2K/(ζδ). This finishes the proof of (4.53). □ Proof of Lemma 4.9 Lemma 4.10 with K=8C/δζ verifies that the measure (4.51) satisfies the assumption in Theorem A.5 (by taking ε,n there to be 1/N, n here). Let y=x+e/N. We conclude that where the contour ω-⊂Λ encloses [-3ℓ,3ℓ], but not z, the implicit constant in the error term depends on K,δ,ℓ. Moreover, ΔM(z) is a random variable with mean zero under Eτ. By a contour integral with respect to G(z):=∫0zg(w)dw on both sides and taking expectation we get where ω is a contour inside Λ. For the lefthand side of (4.59), we have It follows that This gives the first statement in (4.46) by integrating from τ=0 to τ=1 and using (4.50). Next we replace m(z) in the integral (4.60) by m∗(z). Thanks to (4.55), on the contour we have |m(z)-m∗(z)|⩽εKπ⩽1/2, thus where we used Lemma 4.10 and take K=8C/δζ in the last inequality. The second statement in (4.46) follows by plugging (4.61) into (4.60), and noticing that the contour integral over ω gives a factor O(ℓ). □ Rate Function and Proof of Proposition 4.3 We recall the rate function S(A;g) from (4.20) where the contour ω⊂Λ (from (4.52)) encloses [-3ℓ,3ℓ]. In this section, we simplify S(A;g) and prove Proposition 4.3. Lemma 4.11 We recall from (4.14) and (4.56) If dist(z,{tv,ℓ+tv})≳δ, then Lemma 4.12 We introduce a new complex slope as ft∗(z)=emt∗(z)+gt(z)∈H-. Then for x∈[tv,tv+ℓ], Proof of Lemma 4.11 The expressions of mt∗(z) and m~t(z) are explicit as in (4.62), from which the result follows from |ln(1+x)|≲|x|. □ Proof of Lemma 4.12 The first claim follows from (4.62) For the second statement, we recall κt(x) from (4.37), gt(x) from (4.13), and f~t(x)=em~t(x)+gt(x) from (4.12). The three vertices {0,-1,f~t} form a triangle, with three angles give by ϱκt(x)π,ϱvκt(x)π,ϱ(1-v)κt(x)π. From Claim 4.5, for x∈[tv,ℓ+tv], It follows that |1+f~t(x)|≳ϱ. We can rewrite ft∗(x) in terms of f~t(x) mt∗(x) and m~t(x) are explicit, satisfy (4.64) and Lemma 4.11. For dist(x,{tv,ℓ+tv})≳δ we have where in the last equality we used (4.66) and Lemma 4.11. □ Proof of Proposition 4.3 We examine the second term on the righthand side of (4.20). We recall the dilogarithm function (see [64]), which can be analytically extended to the cut plane C\[1,∞]. By a change of variable, which is analytic on the strip {w∈C:-πx^ξN+1(t) and x^n-ξN(t)>x^n-ξN+1(t). x^ξN(t)>x^ξN+1(t) and x^n-ξN(t)>x^n-ξN+1(t). From Assumption 5.1, Adm(y(0),y(ℓ),a,b)≠∅, so yi(0)⩾Na(0)+(n-i)θ, yn-i+1(L)⩽Nb(L)-(n-i+1)θ and yi(0)⩾yi(L)-L. Thus at time t=0, x^i(0)=yi(0); at time t=L, x^i(L)=yi(L). In particular x^i(t) is a Bernoulli path from yi(0) to yi(L). Also from the construction (5.12), we have x^i(L)>x^j(L) for 1⩽i1 To lower bound (5.20), by the same argument as in proof of the large deviation upper bound, we tilt the Markov chain by the exponential Martingale (4.15). The following lemma collects some estimates for the numerator and denominator of the exponential Martingale on the event ‖H(x,t)-H~(x,t)‖∞⩽ε, Lemma 5.5 Adopt Assumption 5.1. Let H(x,t) be the height function associated with the particle configuration {x(t)}0⩽t⩽ℓ, with supp(x(t))∈[-3ℓ,3ℓ] and ‖H(x,t)-H~(x,t)‖∞⩽ε then Dynamical loop equation gives where the contour ω⊂Λ (from (4.52)) encloses [-3ℓ,3ℓ], and mt∗(z) is the Stieltjes transform of the empirical measure of ∂xA(x,t) With Lemma 5.5, we can lower bound the righthand side of (5.20) as where S(A;g) is from (4.20) and Proposition 4.3 gives To lower bound the second term on the righthand side of (5.23), we introduce the tilded measure Pg(·) as defined below, Under Pg(·), the Markov process (4.1) becomes The large deviation lower bound follows from the following limit shape result Proposition 5.6 Adopt Assumption 5.1. We recall the parameter ε from (3.1). For any ε>0, there exists a small c(ε)>0, such that if the height function H(x,0) of the initial data x(0) satisfies ‖H(x,0)-H~(x,0)‖∞⩽c(ε), then provided N is large enough. Proof of Proposition 5.3 Take x with height function H(x,0) as in Proposition 5.6. Proof of Proposition 5.3 follows from plugging (5.24) and Proposition 5.6 into (5.23). □ Proof of Lemma 5.5 Proof From the construction of H~ in (4.9), and thanks to (4.39) from Claim 4.5 it follows that Using the above estimate we can rewrite the lefthand side of (5.21) as By the same argument as in (4.48), (5.29) leads to Next we replace H(x,t) by H~(x,t). Using the second and third statement of Lemma 4.7, and integrating by part Finally, we replace H~(x,t) by A(x,t) where we used Claim 4.5 for the second inequality. The claim (5.21) follows from combining (5.29), (5.30), (5.31) and (5.32). The same as in Lemma 5.5, by using the dynamical loop equation, we have where mt(z) is the Stieltjes transform of the empirical measure of x(t), and the implicit constant in the error term O(1) depends on δ,ζ. We can deform the contour ω such that it consists of two parallel lines {z:Im[z]=±1/K}, with K≍1/δζ. Recall from Lemma 4.7 and use (4.42), if dist(z,{tv,ℓ+tv})⩾ℓ, we have We can replace gt(z) in (5.33) by gt(z)-ln(sin(v)/sin(1-v)) and contour integral does not change. To replace mt(z) in (5.33) by mt∗(z), the error is given by For the difference |ln(1+emt∗(z)+τgt(z))-ln(1+emt(z)+τgt(z))|, we need to upper bound |mt∗(z)-mt(z)|. Recall from Lemma 4.11, for dist(z,{tv,ℓ+tv})⩾δ Next we bound |mt(z)-m~t(z)|. Since Im[z]⩾1/K, the same as in (4.55), we have |mt(z)-m~t(z)|⩽επK. They together give a simple bound |mt(z)-mt∗(z)|⩽δ/dist(z,{tv,ℓ+tv})+εKπ≪1 for dist(z,{tv,ℓ+tv})≫δ. It follows that where in the second to last inequality we used (4.53). Next we will prove an improved estimate of the difference |mt(z)-m~t(z)| when z is far away from 0. For dist(z,{tv,ℓ+tv})⩾3ℓ, By symmetry, we will only prove (5.38) for Re[z]⩽-ℓ, then dist(z,{tv,ℓ+tv})≍Re[z]⩾ℓ. where the last term is bounded by an integration by part and |z-x|≳|z|. For the first term on the righthand side of (5.39) and we have where we used Claim 4.5 to bound κt in the last term. The claim (5.38) follows from combining (5.39) and (4.38). We decompose the contour integral ω={z:Im[z]=±1/K} on the righthand side of (5.35) into the following three parts For dist(z,{tv,ℓ+tv})⩽δ, using Lemma 4.10 and (3.1) that |ln(1+emt∗(z)+τg(z))-ln(1+emt(z)+τg(z))|≲ln(ℓ2K/δζ)≲ln(ℓ/ζδ), we can bound the integral as 5.41∮dist(z,{tv,ℓ+tv})⩽δln(ℓ/ζδ)(ln(1/ζ)+ln(ℓ/δ)+C)|dz|≲δln2(ℓ/δ). For dist(z,{tv,ℓ+tv})⩽δ, using Lemma 4.10 and (3.1) that |ln(1+emt∗(z)+τg(z))-ln(1+emt(z)+τg(z))|≲ln(ℓ2K/δζ)≲ln(ℓ/ζδ), we can bound the integral as 5.41∮dist(z,{tv,ℓ+tv})⩽δln(ℓ/ζδ)(ln(1/ζ)+ln(ℓ/δ)+C)|dz|≲δln2(ℓ/δ). For δ⩽dist(z,{tv,ℓ+tv})⩽10ℓ, using (5.37) and (5.36), we can bound the integral as 5.42∮δ⩽dist(z,{tv,ℓ+tv})⩽10ℓεK+δ/dist(z,{tv,ℓ+tv})ζ(ln(1/ζ)+ln(ℓ/δ)+C)|dz|≲εℓln(ℓ/δ)/(δζ2)+δln2(ℓ/δ)/ζ. For δ⩽dist(z,{tv,ℓ+tv})⩽10ℓ, using (5.37) and (5.36), we can bound the integral as 5.42∮δ⩽dist(z,{tv,ℓ+tv})⩽10ℓεK+δ/dist(z,{tv,ℓ+tv})ζ(ln(1/ζ)+ln(ℓ/δ)+C)|dz|≲εℓln(ℓ/δ)/(δζ2)+δln2(ℓ/δ)/ζ. For dist(z,{tv,ℓ+tv})⩾10ℓ, using (5.37), (5.36), (5.38) and (5.34), we can bound the integral as 5.43∮dist(z,{tv,ℓ+tv})⩾10ℓ1ζδln(|z|K)|z|+∫|x|⩾|z|/2∂xH(x,t)dx|z-x|δ|z||dz|≲δ2ln(ℓ/ζδ)ζℓ+∫|x|⩾3ℓ∂xH(x,t)∫6ℓ⩽|z|⩽2|x|δζ|z-x||z||dz|dx, where we used that K≍1/ζδ, the inner integral is integrable 5.44∫6ℓ⩽|z|⩽2|x|δζ|z-x||z||dz|≲δln(ℓK)ℓζ, and it follows from combining (5.43) and (5.44), and using (5.28) 5.45∮dist(z,{tv,ℓ+tv})⩾3ℓ1ζδln(|z|K)|z|+∫|x|⩾Re[z]/2∂xH(x,t)dx|z-x|δ|z||dz|≲δ2ln(ℓ/ζδ)ℓζ. For dist(z,{tv,ℓ+tv})⩾10ℓ, using (5.37), (5.36), (5.38) and (5.34), we can bound the integral as 5.43∮dist(z,{tv,ℓ+tv})⩾10ℓ1ζδln(|z|K)|z|+∫|x|⩾|z|/2∂xH(x,t)dx|z-x|δ|z||dz|≲δ2ln(ℓ/ζδ)ζℓ+∫|x|⩾3ℓ∂xH(x,t)∫6ℓ⩽|z|⩽2|x|δζ|z-x||z||dz|dx, where we used that K≍1/ζδ, the inner integral is integrable 5.44∫6ℓ⩽|z|⩽2|x|δζ|z-x||z||dz|≲δln(ℓK)ℓζ, and it follows from combining (5.43) and (5.44), and using (5.28) 5.45∮dist(z,{tv,ℓ+tv})⩾3ℓ1ζδln(|z|K)|z|+∫|x|⩾Re[z]/2∂xH(x,t)dx|z-x|δ|z||dz|≲δ2ln(ℓ/ζδ)ℓζ. By plugging (5.41), (5.42) and (5.45) into (5.35), we get The claim (5.22) follows from combining (5.46) and (5.33), and δ≪ζℓ from (3.1). □ Proof of Proposition 5.6 In this section we prove Proposition 5.6. We recall the smoothed height function H~ from (4.9), the associated complex slope f~t(z) from (4.10), and the drift gt(z) from (4.13). The following lemma collects some properties of f~t(z) and gt(z) from Sect. 4.3 Lemma 5.7 The Markov process (5.26) satisfies There exists a constant A=3ℓ>0, such that supp(x(t))⊂[-A,A]. There exists a constant A=3ℓ>0, such that supp(x(t))⊂[-A,A]. The height function H~(x,t)∈Adm(R,h) such that for x∈R, and a complex slope ft(x)∈H- which satisfies -argf~t(x)=π∂xH~(x,t),arg(1+f~t(x))=π∂tH~(x,t). The height function H~(x,t)∈Adm(R,h) such that for x∈R, and a complex slope ft(x)∈H- which satisfies -argf~t(x)=π∂xH~(x,t),arg(1+f~t(x))=π∂tH~(x,t). There exists a neighborhood Λ of [-A,A], such that f~t(x) extends continuously to Λ∩H¯ as f~t(z)=em~t(z)+gt(z), where m~t(z) is the Stieltjes transform of ϱ~t(x)=∂xH~(x,t), 5.47m~t(z)=∫Rϱ~t(x)dxx-z, and gt(z) is analytic on Λ, and gt(z¯)=gt(z)¯. There exists a neighborhood Λ of [-A,A], such that f~t(x) extends continuously to Λ∩H¯ as f~t(z)=em~t(z)+gt(z), where m~t(z) is the Stieltjes transform of ϱ~t(x)=∂xH~(x,t), 5.47m~t(z)=∫Rϱ~t(x)dxx-z, and gt(z) is analytic on Λ, and gt(z¯)=gt(z)¯. As a function of z, ln(1+f~t(z)) is well defined on Λ\[-A,A] and is uniformly bounded on any compact subset of Λ\[-A,A]. As a function of z, ln(1+f~t(z)) is well defined on Λ\[-A,A] and is uniformly bounded on any compact subset of Λ\[-A,A]. Proof The second statement is from the definition (4.10); the third and fourth statement follows from Lemma 4.7. □ From the second the third statement in Lemma 5.7, we have where the contour ω-∈Λ encloses [-A,A], but not w. We can deform the contour ω- to enclose w, and where the contour ω⊂Λ encloses [-A,A] and w. In the rest of this section we prove the uniqueness of the solutions of the equation (5.48). Assume that there is another solution f^t(z)=em^t(z)+gt(z) satisfying (5.48) where m^t(z) is the Stieltjes transform of ϱ^t(s) supported inside [-A,A]. Proposition 5.8 If m^0(z)=m~0(z) for any z∈C\[-A,A], then m^t(z)=m~t(z) for any t⩾0 and z∈C\[-A,A]. Proof of Proposition 5.8 Local uniqueness implies the global uniqueness. We only need to show that the statement of Proposition 5.8 holds for t∈[0,δ] for some small δ>0. For the equation (5.48), we define the characteristic flow By plugging the characteristic flow (5.51) into (5.49), we have We take a contour C0∈Λ surrounding the interval [-A,A], such that the characteristic flow (5.51) maps z0∈C0 to zt∈Ct for t⩾0. For t⩽δ, by the continuity of the characteristic flow, we have that Ct is inside an annulars region S, see Fig. 5 Shown above is the contour Ct inside the annulars region S. For δ>0 sufficiently small, there exists some constant 00 sufficiently small, we can assume Γt0 be a parameter (indeterminante), and let Q(θ) denote the field of all rational functions of θ with rational coefficients. Define a scalar product ⟨·,·⟩ on the vector space Sym⊗Q(θ) of all symmetric functions over the field Q(θ) by the condition We recall that a partial order on Young diagrams is obtained by declaring λ⪯μ if |λ|=|μ| and λ1+λ2+⋯+λi⩽μ1+μ2+⋯+μi for all i. The following fundamental result is due to Macdonald, which characterizes Jack symmetric functions. Let θ>0 be a parameter (indeterminante). Jack symmetric functions Jλ(x;θ) are elements of the algebra Sym of the symmetric functions in infinitely many variables (xi)i=1∞ uniquely determined by the following two properties: Jλ, |λ|=m, can be expressed in terms of the monomial symmetric functions via a strictly upper unitriangular transition matrix: Jλ=mλ+∑μ<λ∈YmRλμmμ, where Rλμ are functions of θ and μ<λ is comparison in the dominance order on the set Ym of all partitions of m (equivalently, Young diagrams with m boxes). Jλ, |λ|=m, can be expressed in terms of the monomial symmetric functions via a strictly upper unitriangular transition matrix: Jλ=mλ+∑μ<λ∈YmRλμmμ, where Rλμ are functions of θ and μ<λ is comparison in the dominance order on the set Ym of all partitions of m (equivalently, Young diagrams with m boxes). They are pairwise orthogonal with respect to the scalar product defined on the power sums via 6.2⟨pλ,pμ⟩θ=δλμzλθ-ℓ(λ),zλ=∏1⩽i⩽λ1imi(λ)∏1⩽i⩽λ1mi(λ)!, where λ=1m12m2⋯, i.e. mi is the multiplicity of i in λ, ℓ(λ) is the number of rows, and pk=(x1)k+(x2)k+⋯, k⩾1. They are pairwise orthogonal with respect to the scalar product defined on the power sums via 6.2⟨pλ,pμ⟩θ=δλμzλθ-ℓ(λ),zλ=∏1⩽i⩽λ1imi(λ)∏1⩽i⩽λ1mi(λ)!, where λ=1m12m2⋯, i.e. mi is the multiplicity of i in λ, ℓ(λ) is the number of rows, and pk=(x1)k+(x2)k+⋯, k⩾1. We recall the scalar product on Sym as defined in (6.1), and introduce the following notation, The number jλ(θ) is explicitly given by See [61, Theorem 5.8] for a proof. We denote the dual polynomial J~λ(x;θ)=Jλ(x;θ)/jλ(θ). Finally, the skew Jack polynomials Jλ/μ,J~λ/μ are defined through the expansions: Theorem 6.1 We have Proof See [61, Proposition 2.3]. □ The following lemma gives the asymptotics of the Jack symmetric polynomials at the principal specialization Lemma 6.2 Given a Young diagram λ∈YN, we identify it as a particle configuration then Given a sequence of Young diagrams such that There exists a constant C>0, λ1(N)⩽CN There exists a constant C>0, λ1(N)⩽CN There exists a 1-Lipschitz nondecreasing function h:R↦[0,θ], and when N→∞θN∑i=1Nδλi(N)-(i-1)θN→∂xh(x), in distribution. Then There exists a 1-Lipschitz nondecreasing function h:R↦[0,θ], and when N→∞θN∑i=1Nδλi(N)-(i-1)θN→∂xh(x), in distribution. Then Proof We can reorganize the numerator and denominator in (6.5), and rewrite it in terms of the particle configuration x from (6.6). and The claim (6.7) follows from combining the above two expressions. For the asymptotics (6.8), simply write λi(N)-(i-1)θ=xi. We notice that there exists C=C(θ)>0, Stirling’s formula implies that the log Gamma function satisfies (z-1/2)lnz-z-C⩽lnΓ(z)⩽(z-1/2)lnz-z+C for z⩾θ, and |∂zlnΓ(z)-lnz|⩽C/z for z⩾θ. So we have and where we used that xi-xj⩾θ(i-j) for the second inequality and in the last equality we used Lemma 2.7. The claim (6.8) follows from plugging (6.9) and (6.10) into (6.7). □ Jack Process Computations in the algebra of symmetric functions Sym can be converted into numeric identities by means of specializations, which are algebra homomorphism from Sym to the set of complex numbers. A specialization ρ is uniquely determined by its values on any set of algebraic generators of Sym and we use (pk)k=1∞ as such generators. The value of ρ on a symmetric function f is denoted f(ρ). Given two specializations ρ,ρ′, we define their union (ρ,ρ′) through the formula: A specialization ρ is an algebra homomorphism from Sym to the set of complex numbers. A specialization ρ is called Jack-positive if its values on all (skew) Jack symmetric functions with a fixed θ are real and non-negative, i.e., for all λ∈Y. The set of Jack positive specializations are classified by Kerov, Okounkov and Olshanski in [43, Theorem A] (We slightly modify the statement (our βi is θβi in [43, Theorem A]) such that it matches with Theorem C.1). Theorem 6.3 For any fixed θ>0, Jack positive specializations can be parameterized by triplets (α,β,γ), where α,β are sequences of real numbers with and γ is a non-negative real number. The specialization corresponding to a triplet (α,β,γ) is given by its values on Newton power sums pk, Following [28, Section 2.3], we create Markov chains out of the Jack-positive specializations: Definition 6.4 Given two specializations ρ and ρ′ we define the ascending transition through where λ and μ are partitions with μ⊂λ and the normalization constant Hθ(ρ;ρ′) is given by We further use the wβ automorphism of the algebra of symmetric functions Sym defined on the power sums by: As shown in [46, Chapter VI, Section 10, (10.6)], where λ′ and μ′ are transposed Young diagrams λ and μ, respectively. Hence, with the specialization β=(b0,b1,⋯,bT-1), as in (6.11), we have To obtain the asymptotics of skew Jack polynomials, we study the following special Jack ascending process. We fix N,T⩾1, and a sequence of positive numbers b0,b1,b2,⋯,bT-1. The transition probability at time 0⩽t⩽T-1 is given in the notations of Theorem 6.3 as with all other parameters set to 0. The transition probability at time 0⩽t⩽T-1 is given by Then the transition probability gives the skew Jack polynomials The following claim states that we can encode the Jack ascending process 6.13 of Young diagrams as an N-particle non-intersecting θ-Bernoulli random walk. Claim 6.5 The transition probability of the Jack process 6.13 is non-degenerate only for partitions λ,μ with at most N parts, i.e. λ=(λ1⩾λ2⩾⋯⩾λN) and μ=(μ1⩾μ2⩾⋯⩾μN). Further, if we identify then ei∈{0,1} and the transition probability is Proof of Theorem 1.7 Thanks to the relation (6.14), the large deviation asymptotics of the skew Jack polynomials can be obtained from the large deviation principle of the dynamics (6.16), where we used Lemma 6.6. The dynamics (6.16) and the non-intersecting θ-Bernoulli random walks (1.8) differ by the following drift Theorem 1.7 follows from plugging the following lemma to (6.17). Lemma 6.6 Adopt the notations and assumptions of Theorem 1.7, and denote the height function of {x(t)}0⩽t⩽T as H(x, t), then the following holds with Moreover, Ff is continuous on the space of Lipschitz functions with uniform norm. Proof We notice that since ei(t)=xi(t+1)-xi(t) and bt=ef(t/N), we have On the other hand, for t∈[t/N,(t+1)/N], we have set x(t)=x(t)/N+(Nt-t)e(t)/N, so that for all integer number t, Therefore, we deduce that where the O(N) error is given as Nθ-1∫0T|∂sf(s)|ds. Now it is easy to check that By integration by parts we also have since H(x;x(t))=∫-∞xρ(y;x(t))dy, where we notice that we can restrict the integral on the left to a compact set where the particles lives, say [-K,K] that can be chosen independent of the time (provided the particles at time 0 are bounded and T is also bounded). Then [yH(y;x(t))]-KK=θK is independent of t. As a conclusion, we get: Noticing that H(y, s) is Lipschitz, thus differentiable. The claim (6.18) follows from an integration by part. Finally, it is clear that if ∂sf is bounded, Ff is continuous for the uniform norm. □ Acknowledgements The research of J.H. is supported by NSF grant DMS-2331096 and DMS-2337795, and the Sloan research award. A.G. has received funding from the European Research Council (ERC) under the European Union Horizon 2020 research and innovation program (grant agreement No. 884584). The authors thank Ofer Zeitouni for helpful discussions at an early stage of this project, and two anonymous reviewers for valuable comments. Dynamical Loop Equation We define the particle configuration: Given a particle configuration x=(x1,x2,⋯,xn)∈Wn, an interaction function b(z), and weight functions ϕ+(z),ϕ-(z). Our central object is the following transition probability where e=(e1,e2,⋯,en)∈{0,1}n and Z(x) is a normalization constant. Assumption A.1 Fix constants N>0 and l,r. We introduce a small parameter ε=N/n≪1, and assume the particle configuration x=(x1,x2,⋯,xn)∈W~θn satisfies l⩽εxn⩽εx1⩽r-ε. We further assume There exists an open complex neighborhood Λ of [l,r], such that for z=εξ, A.3b(ξ)=b(z)andϕ±(ξ)=φ±(z),z∈Λ. where b(z) and φ±(z) are holomorphic functions on Λ. In addition, we require b(z) to be conformal (injective and biholomorphic) and b(z¯)=b(z)¯. There exists an open complex neighborhood Λ of [l,r], such that for z=εξ, A.3b(ξ)=b(z)andϕ±(ξ)=φ±(z),z∈Λ. where b(z) and φ±(z) are holomorphic functions on Λ. In addition, we require b(z) to be conformal (injective and biholomorphic) and b(z¯)=b(z)¯. The functions b(z), [∂zb(z)]-1, and φ±(z) are uniformly bounded, namely there exists a universal1 constant C>0 such that A.4|b(z)|⩽C,|∂zb(z)|⩾1/C,|φ±(z)|⩽C,z∈Λ. The functions b(z), [∂zb(z)]-1, and φ±(z) are uniformly bounded, namely there exists a universal1 constant C>0 such that A.4|b(z)|⩽C,|∂zb(z)|⩾1/C,|φ±(z)|⩽C,z∈Λ. Remark A.2 In [24], the particle configurations are assumed to be inside the lattice: Here we relaxed the particle configuration to (A.1). The dynamical loop equation states that the following observable is a holomorphic function. Theorem A.3 (Dynamical loop equation). Choose an open set U⊂C, a particle configuration x=(x1>x2>⋯>xn)∈Rn such that the interval [xn,x1]⊂U, a parameter θ>0, two holomorphic functions ϕ+(z), ϕ-(z) on U and a conformal (i.e., holomorphic and injective) function b(z) on U. Assume that the random n–tuple e=(e1,e2,⋯,en)∈{0,1}n is distributed according to the transition probability (A.2). Then the following observable is a holomorphic function of z∈U: Proof See [24, Proof of Theorem 1.1] □ Our last assumption for the asymptotic analysis involves the following functions defined in terms of the empirical density (1.10) and functions ϕ±(z) of Assumption A.1: Note that B(z) is a holomorphic function for z∈Λ\[l,r]. Assumption A.4 There exists annular set S⊂(Λ\[l,r]) (containing a contour surrounding [l,r]), a small universal constant c>0 such that for z∈S we have c<|B(z)|θ. Formally, {x(t)}0⩽t⩽T is given as for any 1⩽i⩽N and t∈N-1[[0,T]] We denote the height function of {x(t)}0⩽t⩽T as H. Next we check that {x(t)}0⩽t⩽T forms a non-intersecting θ-Bernoulli walk from y(N)/N to z(N)/N. Under assumption (1.16), for N large enough It follows that for t∈{0,T}, And thus γ(θ(i+m)/N,t)⩽yi(t)=xi(t)⩽γ(θ(i-m)/N,t) for t∈{0,T}. We need to show xi(t+1/N)-xi(t)∈{0,1/N} and xi-1(t)-xi(t)⩾θ/N for any t=t/N∈N-1[[0,T]]. By symmetry, we assume yi(t)⩾γi(t), so it is far from γi+m(t). In fact we have Notice that under (B.2) we only need to deal with the possible truncation caused by γi-m(t). There are still several cases If γ(θ(i+m)/N,t)⩽yi(t)⩽γ(θ(i-m)/N,t), then xi(t)=yi(t). If also xi(t+1/N)=yi(t+1/N), then xi(t+1/N)-xi(t)=yi(t+1/N)-yi(t)∈{0,1/N}; otherwise yi(t+1/N)>γ(θ(i-m)/N,t+1/N). Recall that ∂sγ(y,s)∈[0,1], we must have yi(t+1/N)>γ(θ(i-m)/N,t+1/N)⩾γ(θ(i-m)/N,t)⩾yi(t), and it follows that yi(t+1/N)=yi(t)+1/N. In this case xi(t+1/N)=yi(t)+1/N, and xi(t+1/N)-xi(t)=1/N. If xi-1(t)=yi-1(t), then xi-1(t)-xi(t)=yi-1(t)-yi(t)⩾θ/N; otherwise yi-1(t)>γ(θ(i-1-m)/N,t). Recall that γ(y-θ/N,s)-γ(y,s)⩾θ/N, we must have yi-1(t)>γ(θ(i-1-m)/N,t)⩾γ(θ(i-m)/N,t)+θ/N⩾yi(t)+θ/N. In this case xi-1(t)⩾yi(t)+θ/N=xi(t)+θ/N. If γ(θ(i+m)/N,t)⩽yi(t)⩽γ(θ(i-m)/N,t), then xi(t)=yi(t). If also xi(t+1/N)=yi(t+1/N), then xi(t+1/N)-xi(t)=yi(t+1/N)-yi(t)∈{0,1/N}; otherwise yi(t+1/N)>γ(θ(i-m)/N,t+1/N). Recall that ∂sγ(y,s)∈[0,1], we must have yi(t+1/N)>γ(θ(i-m)/N,t+1/N)⩾γ(θ(i-m)/N,t)⩾yi(t), and it follows that yi(t+1/N)=yi(t)+1/N. In this case xi(t+1/N)=yi(t)+1/N, and xi(t+1/N)-xi(t)=1/N. If xi-1(t)=yi-1(t), then xi-1(t)-xi(t)=yi-1(t)-yi(t)⩾θ/N; otherwise yi-1(t)>γ(θ(i-1-m)/N,t). Recall that γ(y-θ/N,s)-γ(y,s)⩾θ/N, we must have yi-1(t)>γ(θ(i-1-m)/N,t)⩾γ(θ(i-m)/N,t)+θ/N⩾yi(t)+θ/N. In this case xi-1(t)⩾yi(t)+θ/N=xi(t)+θ/N. If yi(t)>γ(θ(i-m)/N,t), then xi(t)=yi(t)-{1/N,2/N,⋯}. For yi(t+1/N) there are two cases: If yi(t+1/N)⩽γ(θ(i-m)/N,t+1/N), recalling that ∂sγ(y,s)∈[0,1], then we must have yi(t+1/N)⩽γ(θ(i-m)/N,t+1/N)⩽γ(θ(i-m)/N,t)+1/Nγ(θ(i-1+m)/N,t), recall that γ(y-θ/N,s)-γ(y,s)⩾θ/N, so xi(t)+θ/N⩽γ(θ(i-m)/N,t)+θ/N⩽γ(θ(i-1-m)/N,t), and xi-1(t)⩾xi(t)+θ/N. If yi(t)>γ(θ(i-m)/N,t), then xi(t)=yi(t)-{1/N,2/N,⋯}. For yi(t+1/N) there are two cases: If yi(t+1/N)⩽γ(θ(i-m)/N,t+1/N), recalling that ∂sγ(y,s)∈[0,1], then we must have yi(t+1/N)⩽γ(θ(i-m)/N,t+1/N)⩽γ(θ(i-m)/N,t)+1/Nγ(θ(i-1+m)/N,t), recall that γ(y-θ/N,s)-γ(y,s)⩾θ/N, so xi(t)+θ/N⩽γ(θ(i-m)/N,t)+θ/N⩽γ(θ(i-1-m)/N,t), and xi-1(t)⩾xi(t)+θ/N. Finally we check that the height function H of {x(t)}0⩽t⩽T as constructed in (B.1) satisfies (2.1). For any xi(t)⩽x1 large enough. Similarly, for the second term on the left of (B.13), we have The two estimates (B.14) and (B.15) together gives that If κt(x)⩽1/2, then If κt(x)⩾1/2, then using ζ⩽ϱv⩽1-ζ, We conclude from plugging (B.16), (B.17) and (B.18) into (B.13) The above estimate (B.19) and the first statement of Lemma 4.6 together imply that we can extend ft(z) to the strip region, and The estimate (B.19) and the second statement of Lemma 4.6 together imply that we can extend gt(z) to the strip region, and For dist(z,{tv,ℓ+tv})⩾ℓ, it follows from (4.40) and (4.39) in Claim 4.5 Then by Taylor expansion This gives the claim (4.44). Next, we estimate Im[gt(z)] for z=x+iη. By taking imaginary part of (B.13), and using (B.14) and (B.15) The above estimate together with the second statement (4.41) of Lemma 4.6 imply that Hib(κt)(z) can be computed explicitely, see (B.12), and its derivative at z=x can be bounded as Since gt(z) is analytic, we have |∂xgt(x)|=|∂ηgt(x+iη)|η=0|. As a consequence of (B.14) and (B.15), we have The claim |∂xgt(x)|≲1/δ follows from (B.21) and (B.20). Recall our construction of gt from (4.43), since H~(x-tv,t) is constant for 0⩽t⩽ℓ, so is gt(x-tv). By taking derivative with respect to t we get Thus |∂tgt|⩽v|∂xgt|⩽C/δ. □ Asymptotics for Macdonald Polynomials In this section, we explain the correspondence between non-intersecting θ-Bernoulli walk ensembles and certain Macdonald ascending process (C.4). This correspondence will be used to derive large deviation asymptotics for (skew) Macdonald polynomials. We collect some basic properties of Macdonald symmetric functions in Sect. C. We recall the Macdonald ascending process in Sect. C.2. Our main references are [47, 61]. In Sect. C.3, we prove Theorem 1.8. Macdonald Polynomials and Specializations We use the following notations: The next class of Markov chains is built out of the specializations of Macdonald polynomials. We refer to [46, Section VI] and [8] for definitions and properties of Macdonald symmetric functions and use them as a black box in this section. Macdonald symmetric functions P and Q are indexed by partitions and implicitly depend on two parameters q,t∈(0,1). The coefficients for the symmetric functions are in Q[q,t]. Macdonald symmetric functions Pλ(x;q,t) and Qλ(x;q,t) are elements of the algebra Λ of the symmetric functions in infinitely many variables (xi)i=1∞ uniquely determined by the following two properties: Pλ, |λ|=m, can be expressed in terms of the monomial symmetric functions via a strictly upper unitriangular transition matrix: Pλ=mλ+∑μ<λ∈YmRλμPμ, where Rλμ are functions of q, t and μ<λ is comparison in the dominance order on the set Ym of all partitions of m (equivalently, Young diagrams with m boxes). Pλ, |λ|=m, can be expressed in terms of the monomial symmetric functions via a strictly upper unitriangular transition matrix: Pλ=mλ+∑μ<λ∈YmRλμPμ, where Rλμ are functions of q, t and μ<λ is comparison in the dominance order on the set Ym of all partitions of m (equivalently, Young diagrams with m boxes). They are pairwise orthogonal with respect to the scalar product defined on the power sums via ⟨pλ,pμ⟩q,t=δλμzλ(q,t),pλ=∏i=1∞pλi,zλ(q,t)=∏i⩾1imi(mi)!∏i=1ℓ(λ)1-qλi1-tλi, where λ=1m12m2⋯, i.e. mi is the multiplicity of i in λ, ℓ(λ) is the number of rows, and pk=(x1)k+(x2)k+⋯, k⩾1. They are pairwise orthogonal with respect to the scalar product defined on the power sums via ⟨pλ,pμ⟩q,t=δλμzλ(q,t),pλ=∏i=1∞pλi,zλ(q,t)=∏i⩾1imi(mi)!∏i=1ℓ(λ)1-qλi1-tλi, where λ=1m12m2⋯, i.e. mi is the multiplicity of i in λ, ℓ(λ) is the number of rows, and pk=(x1)k+(x2)k+⋯, k⩾1. We further define Qλ=Pλ⟨Pλ,Pλ⟩q,t. Finally, the skew Macdonald polynomials Pλ/μ and Qλ/μ are defined through the expansions: Macdonald Ascending Processes Computations in the algebra of symmetric functions Λ can be converted into numeric identities by means of specializations, which are algebra homomorphism from Λ to the set of complex numbers. Specialization ρ is uniquely determined by its values on any set of algebraic generators of Λ and we use (pk)k=1∞ as such generators. The value of ρ on a symmetric function f is denoted f(ρ). Given two specializations ρ,ρ′, we define their union (ρ,ρ′) through the formula: A specialization ρ is called Macdonald nonnegative if its values on all (skew) Macdonald symmetric functions are non-negative, i.e., if for all partitions λ and μ, The description of Macdonald nonnegative specializations was conjectured in [44] and proven in [48]: Theorem C.1 [48]. For any fixed q,t∈(0,1), Macdonald nonnegative specializations can be parameterized by triplets (α={αi}i⩾1,β={βi}i⩾1,γ) of nonnegative numbers satisfying ∑i=1∞(αi+βi)<∞. The specialization ρ corresponding to a triplet (α,β,γ) is defined by for k⩾2 Following [8, Section 2.3], we create Markov chains out of the Macdonald-positive specializations: Definition C.2 Given two specializations ρ and ρ′ we define the ascending transition through where λ and μ are partitions with μ⊂λ and Π(ρ;ρ′) is the result of applying ρ to the xi variables and ρ′ to the yj variables in the infinite product To obtain the asymptotics of Skew Macdonald polynomials, we study the following special Macdonald ascending process. We fix N,T⩾1, and a sequence of positive numbers b0,b1,b2,⋯,bT-1. The transition probability at time 0⩽t⩽T-1 is given in the notations of Theorem C.1 as with all other parameters set to 0. This gives a Markov process on Young diagrams λ(0),λ(1),⋯,λ(T) Then the transition probability gives the skew Macdonald polynomials The evaluation of the Macdonald polynomial under the principle specialization ρ=(1,t,⋯,tN-1) is explicit, see [46, Chapter VI, (6.11)]: for λ=λ1⩾λ2⋯⩾λN⩾0; Pλ(ρ;q,t)=0 if λN+1>0. We further use the wu,v automorphism of the algebra of symmetric functions Λ defined on the power sums by: As shown in [46, Chapter VI, Section 7], where λ′ and μ′ are transposed Young diagrams λ and μ, respectively. Hence, with the specialization β=(b0,b1,⋯,bT-1), as in (6.12), we have The following claim states that we can encode the Macdonald ascending process C.4 of Young diagrams as an N-particle non-intersecting θ-Bernoulli random walk. Claim C.3 [24, Corollary 3.11]. The transition probability of Definition C.2 under the specializations (C.3) is non-degenerate only for partitions λ,μ with at most n parts, i.e. λ=(λ1⩾λ2⩾⋯⩾λn) and μ=(μ1⩾μ2⩾⋯⩾μn). Further, if we set t=qθ and identify then ei∈{0,1} and the transition probability is given by The following statement establishes (1.25). Claim C.4 For non-intersecting Bernoulli walks p={x(t)}0⩽t⩽T from x(0)=y to x(T)=z, we define their weights as Then we can rewrite the skew Macdonald polynomial evaluated at (b0,b1,⋯,bT-1) as follows where λ′,μ′ are the transposes of λ,μ. Proof Thanks to (C.7) and (C.5), We set t=qθ and use the specialization (6.12C.3). The second factor on the righthand side of (C.11) can be expressed as a weighted sum, exactly given by (C.10). □ Lemma C.5 Given a Young diagram λ∈YN, we identify it as an particle configuration then Given sequences a sequence of Young diagrams such that There exists a constant C>0, λ1(N)⩽CN There exists a constant C>0, λ1(N)⩽CN There exists a 1-Lipschitz nondecreasing function h:R↦[0,θ], and when N→∞C.14θN∑i=1Nδλi(N)-(i-1)θN→∂xh(x), in distribution. Then There exists a 1-Lipschitz nondecreasing function h:R↦[0,θ], and when N→∞C.14θN∑i=1Nδλi(N)-(i-1)θN→∂xh(x), in distribution. Then Proof We can reorganize (C.6), and rewrite it in terms of the q-Gamma functions The claim (C.13) follows. For the asymptotics (6.8), simply write λi(N)-(i-1)θ=xi. By our assumption (C.14), the exponent in the first term on the righthand side of (C.13) is We recall the following estimates for the q-Gamma function for q⩽1, the first statement follows from [49, Corollary], and the second follows from [49, Theorem 1]. So we have and where we used that xi-xj⩾θ(i-j) for the second inequality; in the last equality we used Lemma 2.7. The claim (C.15) follows from plugging (C.16), (C.17) and (C.18) into (C.13). □ Proof of Theorem 1.8 Thanks to the relation (C.5), the large deviation asymptotics of the skew Macdonald polynomials can be obtained from the large deviation principle of the dynamics (C.9), where we used Lemma C.5. The dynamics (C.9) and the non-intersecting θ-Bernoulli random walks (1.8) differ by the drift E({x(t)}0⩽t⩽T)D({x(t)}0⩽t⩽T) with Theorem 1.8 follows from plugging Lemma 6.6 and the following lemma to (C.19). We remark that all the terms involving κ cancels out perfectly. Lemma C.6 Adopt the notations and assumptions of Theorem 1.8, the following holds Proof We observe that so that where for a real number β, Eβ denotes the expectation over the variable α∈[0,1] given, for any bounded measurable function f on [0, 1], by We recall that q=eκ/N, and denote t=t/N, so that where O(1/N2) is uniform over the xi(t),xi(t) and over θ in a bounded set. Recall that ei(t)=xi(t+1)-xi(t). Hence, we deduce that if f(y):=Eey[α], For the first term on the righthand side of (C.20), we have For the second term on the righthand side of (C.20), let so that F′(x)=κf(κx), and where the error is again uniform since f is uniformly Lipschitz as |∂yf(y)|≤2. Therefore, for the second term on the righthand side of (C.20), we have where we finally used that F is uniformly Lipschitz on compacts. Finally, we can see that the above right hand side is a continuous function of H(y;x(T)) and H(y;x(0)) equipped with the uniform topology, provided its derivative is a probability measure. The proof follows the ideas from Lemma 2.7, so we omit the details. It follows from plugging (C.21) and (C.22) into (C.20), □ Data Availibility No datasets were generated or analyzed during the current study. Declarations Competing interests The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Footnotes By “universal” we mean not depending on n, ε, or (x1,⋯,xn). Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement No datasets were generated or analyzed during the current study.