We prove that the pair correlation function of Sineβ extends holomorphically in β to ℜβ>0 at every fixed nonzero separation. Our proof combines a random walk representation of the stochastic sine phase moments, stretched-exponential Fourier decay obtained by coupling, and weighted energy estimates for complex β.
Introduction
Let Ξβ be the Sineβ point process, normalized to have intensity 1/(2π). Its pair correlation function ρβ is defined by
Ex,y∈Ξβx=y∑F(x,y)=∫R2F(x,y)ρβ(x,y)dxdy
for nonnegative compactly supported Borel functions F, where the sum runs over ordered pairs of distinct points.
Theorem 1 (Analyticity of the pair correlation).For every β>0, the pair correlation function ρβ has a continuous version off the diagonal. For this version and each λ>0, the function
β⟼ρβ(0,λ)
extends holomorphically from (0,∞) to {Reβ>0}. The extension is jointly continuous in (β,λ) on {Reβ>0}×(0,∞).
Figure 1. The holomorphic continuation of 4π2ρβ(0,λ), computed numerically from (5). The black curves on the contour plots are where 4π2∣ρβ(0,λ)∣=1.
Dependence on inverse temperature has long been studied through exact formulas and perturbation theory. For β=1,2,4, the correlation functions of Sineβ have classical explicit formulas: determinantal for β=2 and Pfaffian for β=1,4[10 Chapters 5–7]. Forrester [8, 9] obtained correlation formulas at even integer β, and Ha [12] extended these to rational coupling. Okounkov [19] subsequently obtained
Selberg-type integral formulas for n-point correlations at rational β. Forrester, Jancovici and McAnally (2001) computed first-order perturbations around β=2 and 4, following earlier work on the two-dimensional plasma [14]. More recently, Qu and Valkó (2025) obtained a stochastic representation for all β>0 and proved continuity in β. Theorem 1 strengthens continuity to holomorphic dependence throughout the right half-plane.
Translation and reflection invariance reduce the pair correlation to ρβ(0,λ), λ>0. Our proof uses the stochastic phase representation of Sineβ[16]. We express the pair correlation as the phase density at zero divided by 2π, and use a random walk representation of the phase moments to prove analyticity.
Let Z=B1+iB2, where B1, B2 are independent two-sided standard real Brownian motions. For real β>0 and λ∈R, let αλ,β solve the phase SDE
Equation (1) is the Killip–Stoiciu equation in logarithmic time, as considered in Valkó and Virág [17]. All spectral parameters are driven by the same Z. We often write αλ when β is fixed or clear from context. If U is uniform on [0,2π) and independent of the phase field, then
Ξβ={λ∈R:αλ,β(0)=Umod2π}(2)
has the Sineβ distribution (Killip and Stoiciu [16]). Write T=R/(2πZ) and let pβ,λ denote the density of αλ,β(0)mod2π with respect to dθ on T, for λ>0. For k∈Z, define its Fourier coefficients
qk(β,λ)=Eeikαλ,β(0).
Theorem 2 (Pair correlation and phase density). For real β,λ>0, the pair correlation off the diagonal and the phase density have continuous versions, and
ρβ(0,λ)=2π1pβ,λ(0)=4π21k∈Z∑qk(β,λ),(3)
with absolute convergence.
The existence of correlation functions of all orders for Sineβ, β>0, was proved by Qu and Valkó (2025). More recently, Assiotis and Najnudel [1] represented these functions in terms of the Hua–Pickrell stochastic zeta function introduced by Li and Valkó (2022), and established their continuity and boundedness in the spatial variables. For higher correlations, we obtain the following representation at the level of correlation measures.
Proposition 3 (Phase representation of correlation measures). Let β>0, k≥2, and let I0,…,Ik−1 be pairwise disjoint bounded intervals. Write αλ=αλ,β(0). Then
The proof is in § 5. When k=2, continuity of the phase density turns this limit into (3).
Itô’s formula and the scaling invariance of the phase SDE (1) yield
λ∂λqk=ikλqk+β2k2(qk+1+qk−1−2qk),q0=1,(5)
where all coefficients are evaluated at (β,λ). The derivation is given in § 3. To extend (3) to complex β, we first continue each coefficient and then establish locally uniform
summability. The key tool is a representation by a random walk whose law does not depend on β.
Proposition 4 (Random walk representation). Let (St)t≥0 be the continuous-time nearest-neighbor random walk on Z≥0, absorbed at zero. From k≥1, its jump rate to each of k−1 and k+1 is k2/2. Let Ek denote expectation for the walk started at k. For real β>0, λ>0, and k≥1,
qk(β,λ)=EkeiλAβ,Aβ=4β∫0∞e−βt/4Stdt.(6)
Section 3 proves this formula by applying the countable-state Feynman–Kac formula on [−T,0] and then sending T→∞. The embedded simple random walk is absorbed after finitely many jumps almost surely, so S is nonexplosive and the integral in (6) is almost surely finite. Write τj=inf{t≥0:St=j}. For 1≤k≤n, the finite-space cutoff is
qk,n(β,λ)=Ek[eiλAβ1{τ0<τn+1}].(7)
Gambler’s ruin gives the uniform real-parameter estimate
∣qk(β,λ)−qk,n(β,λ)∣≤n+1k,β>0,λ>0.(8)
The cutoffs are holomorphic on Reβ>0. A weighted energy estimate provides a bound independent of n, allowing Vitali’s theorem to remove the cutoff. The details are in Section 6.
Proposition 5 (Holomorphic Fourier coefficients). For every fixed λ>0 and k∈Z, qk(β,λ) has a unique holomorphic extension to Reβ>0. These extensions satisfy
k=1∑∞k2∣qk(β,λ)∣2≤4(Reβ)2∣β∣2,λ>0,Reβ>0.(9)
Here q0(β,λ)=1 and q−k(β,λ)=qk(β,λ) for $k\geq 1.
To sum these holomorphic extensions in (3), we need locally uniform bounds that are summable in k. For real β>0, a coupling argument for the random walk in Proposition 4 gives stretched-exponential decay (Section 4). Combining this decay with the energy bound (9), we obtain the same type of estimate for complex β with Reβ>0 (Section 7):
Proposition 6 (Stretched-exponential Fourier decay for complex β). For every compact K⊂{Reβ>0} and compact interval L⊂(0,∞), there are constants a,b>0 such that
β∈Kλ∈Lsup∣qk(β,λ)∣≤be−a∣k∣1/3,k∈Z.
Thus the series in (3) converges locally uniformly for Reβ>0 and gives the extension in Theorem 1.
§ 2 records the phase representation and the probabilistic and analytic preliminaries. § 3 and § 4 establish the random walk representation and real Fourier decay. § 5 identifies the pair correlation. § 6 and § 7 complete the analytic continuation argument. Section 8 gives explicit formulas for the Fourier moments and phase density at β=2, recovering the sine-kernel pair correlation. The resulting moment asymptotics show that the decay exponent 1/3 in Proposition 6 cannot be increased beyond 1/2 in general.
Preliminaries
The phase and its boundary angle
Proposition 7 (The stochastic sine phase). For real β>0, the phase SDE (1) has a pathwise unique solution for the prescribed driving Brownian motion. Almost surely λ↦αλ(0) is a real analytic increasing bijection of R, with α0(0)=0 and ∂λαλ(0)>0. Moreover, for λ∈R and t≤0,
αλ(t)=dαλeβt/4(0),Eαλ(t)=λeβt/4.(10)
Let U be uniform on [0,2π) and independent of Z, and define Ξβ by (2). Then Ξβ has the Sineβ distribution. The corresponding normalized spectral weights
wλ=∂λαλ(0)1,λ∈Ξβ,
are mutually independent gamma variables of shape β/2 and mean one, and are independent of Ξβ. These weights are half the right spectral weights in Valkó and Virág (2023, Proposition 3 and Lemma 17). The phase and shooting representations follow from Valkó and Virág (2022, Corollary 51); see also Valkó and Virág (2023, Definition 16 and Lemma 17). Time translation in (1) gives the scaling in (10). The expectation identity follows from Killip and Stoiciu (2009, Proposition 4.5) after the time change s=eβt/4.
The phase representation was introduced by Killip and Stoiciu [16]; its equivalence with the Gaussian bulk limit of Valkó and Virág (2009) was proved by Nakano [18] and Valkó and Virág (2017, Theorem 28).
Lemma 8 (Phase and counting moments). For each 0<L<∞ there is ηL>0 such that
β>0,0≤λ≤L,t≤0supEeηLαλ(t)<∞.(11)
Thus phase moments of every positive order are locally bounded in λ, uniformly in β>0 and t≤0. Counts in bounded intervals have moments of every positive order as well.
Proof. Put Nλ=Eβ((0,λ]). The counting bound in Valkó and Virág (2009, Proposition 9(viii)), applied with total drift λ, gives
P(Nλ≥ak)≤2(2πaλ)k,a,k∈N.
Choose an integer a>L/(2π). This gives a positive exponential moment for Nλ uniformly for λ≤L and β>0. By (2), almost surely,
Nλ=⌊2παλ(0)+2π−U⌋.(12)
Consequently, ∣2πNλ−αλ(0)∣≤2π, so the phase has the same moment property. The scaling identity in (10) gives the bound for t≤0. Stationarity gives the count assertion for any bounded interval.
Probabilistic and analytic tools
Lemma 9 (Feynman–Kac). Let S be a nonexplosive continuous-time Markov chain on a countable state space E, with finite total jump rates and generator G. Fix T>0, and let v:[0,T]×E→C be continuous in time, with Rev(t,k)≤C for some C≥0. Every bounded solution u:[0,T]×E→C of
In particular, there is at most one bounded solution.
Proof. Fix t≤T, and set
Mr=exp(∫0rv(t−s,Ss)ds)u(t−r,Sr),0≤r≤t,
and let σn be the exit times from increasing finite subsets exhausting E. The jump-process product rule shows that Mr∧σn is a martingale, since its drift before σn is proportional to −∂tu+Gu+vu=0. The finite jump rates and boundedness of u justify the generator sums. Moreover,
∣Mr∧σn∣≤eCT∥u∥∞.
Nonexplosion gives σn→∞ almost surely, so bounded convergence yields EkMt=M0, proving (14). See also Fitzsimmons and Pitman (1999) for general Markov-process formulations.
Lemma 10 (Hardy’s inequality in difference form). For z1,…,zn∈C, with z0=zn+1=0,
k=1∑nk2∣zk∣2≤4k=0∑n∣zk+1−zk∣2.(15)
The constant is independent of n.
This is the discrete Hardy inequality in difference form; see Hardy, Littlewood and Pólya (1952, Chapter IX, Theorem 326).
Lemma 11 (Expectations of random analytic functions). Let D⊂C be a domain, and let F(z), z∈D, be a jointly measurable random function whose sample functions are almost surely holomorphic on D. Suppose that, for every compact K⊂D,
z∈KsupE∣F(z)∣<∞.
Then f(z)=E[F(z)] is holomorphic on D.
Proof. Choose a closed disk {∣z−b∣≤r}⊂D. For ∣z−b∣<r, Cauchy’s integral formula and Fubini’s theorem give
The bound on E∣F(w)∣ along the circle justifies Fubini’s theorem. On every smaller disk the denominator stays bounded away from zero, so the last integral can be differentiated with respect to z. Thus f is holomorphic inside the disk. Such disks cover D.
Lemma 12 (Vitali’s theorem). Let fn be holomorphic on a connected open set D⊂C and locally bounded there, uniformly in n. If fn converges on a set with an accumulation point in D, then it converges locally uniformly on D to a holomorphic function.
See Conway (1978, Chapter VII).
Lemma 13 (Two-constants theorem on a slit domain). Let D be a bounded smooth domain and let J be a nondegenerate closed line segment in D such that D∖J is connected. Let ω(z) be the harmonic measure of J in D∖J. If f is holomorphic on a neighborhood of D, and ∣f∣≤b0 on J, ∣f∣≤b1 on ∂D, where b0,b1>0, then
∣f(z)∣≤b0ω(z)b11−ω(z),z∈D∖J.(16)
Moreover ω is positive in D∖J and extends continuously as 1 on J. With this extension, infKω>0 for every compact K⊂D.
This is the two-constants theorem in its harmonic-measure form, applied to log∣f∣; see [21 Sections 4.2–4.3]. The smooth outer boundary and the line-segment slit are regular for the Dirichlet problem, so the stated boundary continuity and positivity of ω apply.
Fourier moments and the random walk
For real β,λ>0 and t≤0, write qk(β,λ,t)=Eeikαλ,β(t). Time translation of the phase SDE (1) gives
qk(β,λ,t)=qk(β,λeβt/4).(17)
Itô’s formula applied to the angular exponential gives
The noise integrand is bounded, so its stochastic integral has mean zero on each finite time interval. The right-hand side is continuous by bounded convergence. By (17), equation (18) gives differentiability in λ>0 and (5). The same scaling identity and the entrance condition imply qk(β,λ)→1 as λ↓0, by bounded convergence.
Proof of Proposition 4. The embedded simple random walk hits zero after finitely many jumps almost surely ([4 Chapter 4]). Thus S is nonexplosive and Aβ is almost surely finite. For T>0, applying Lemma 9 on [−T,0] and using (17) gives
For T≥τ0, the terminal factor is q0=1 and the exponential is eiλAβ. Since τ0<∞ almost surely and the integrand has modulus at most one, bounded convergence as T→∞ gives (6).
The same Feynman–Kac argument gives uniqueness for the moment system among uniformly bounded solutions.
Corollary 14. (Uniqueness of bounded solutions). For every real β>0, the system (5), with q0≡1, has a unique solution (qk)k∈Z of continuously differentiable functions on (0,∞) satisfying
k∈Z,λ>0sup∣qk(λ)∣<∞.
Proof. The phase Fourier coefficients provide such a solution. Let rk=qk−qk be the difference of two solutions, so that r0=0 and ∣rk(λ)∣≤c for some finite c. For fixed λ,T>0, the functions t↦rk(λeβt/4) satisfy (18) on [−T,0] by (5). Lemma 9 therefore gives
Applying the same argument to r−k proves uniqueness for negative indices.
Real Fourier decay from the walk
We use the representation (6) to prove decay of the phase moments on the real parameter intervals needed for analytic continuation.
Proposition 15 (Real Fourier decay). Let J,K⊂(0,∞) be compact sets for λ and β, respectively. There are a,b>0, depending only on J,K, such that
λ∈J,β∈Ksup∣qk(β,λ)∣≤be−a∣k∣1/3,k∈Z.(20)
Consider the walk S from Proposition 4. At height k, the walk fluctuates by order kh over a short time h, producing area fluctuations of order kh3/2. Taking h=k−2/3 makes these fluctuations of order one. This motivates the study of the process on this time scale.
Lemma 16 (Short-time path limit). Put h=k−2/3. If k/2≤j≤2k and j/k→x as k→∞, then under Pj,
(k−2/3(Shu−j):0≤u≤1)⟹(xBu:0≤u≤1),
where B is standard Brownian motion and convergence is in D([0,1]) with its Skorokhod topology.
Proof. Let Y be a simple symmetric continuous-time walk started at zero, with rate 1/2 to each neighbor. Changing its holding times realizes the walk started at j as
St=j+Y∫0tSs2ds
see Ethier and Kurtz [6] for the general random time-change construction. Set
Wk(v)=k−2/3Yk4/3v,ck(u)=∫0u(kShv)2dv.
Donsker’s theorem [4 Theorem 8.1.4], together with the law of large numbers for the Poisson jump count, gives Wk⇒B on every bounded time interval. Let ζk be the exit time of S from [k/4,3k]. Before exit, Ss/k≤3, so the rescaled clock runs at rate at most 9 and
u≤1supk∣Shu∧ζk−j∣≤k−1/3v≤9sup∣Wk(v)∣P0.
Since exit requires displacement at least k/4, Pj(ζk≤h)→0. Thus Shu/k→x uniformly in probability on [0,1], and consequently ck(u)→x2u uniformly in probability. The clock limit is deterministic, so joint convergence and continuity of composition at continuous paths give
k−2/3(Sh−j)=Wk∘ck⇒Bx2=dxB.
Lemma 17 (Area before hitting zero). Let Dn⇒σB in D([0,1]) for some σ>0. Suppose that almost surely Dn cannot change sign without visiting zero. Let ξn be its first zero in [1/2,1], or 1 if there is none. For every nonempty open interval I⊂R,
n→∞liminfP(ξn<1,∫0ξnDn(u)du∈I)>0.
Proof. Choose a and r>0 with [a−r,a+r]⊂I. On [0,1/2] let d be the triangular path from zero to 4a at time 1/4 and back to zero; its integral is a. On [1/2,1] let d run linearly from zero to r at time 3/4 and then to −r at time 1.
On the tube ∥Dn−d∥∞<r/2, the values at 3/4 and 1 have opposite signs, so ξn<1. Since ∣d∣≤r on the second half,
∫0ξnDn(u)du−a≤∥Dn−d∥∞+∫1/21∣d(u)∣du<r.
The Brownian limit has positive probability of lying in this tube: d is polygonal and starts at zero, so this follows from Cameron–Martin and the positive probability of a ball about zero. The map f↦∥f−d∥∞ is continuous at continuous paths in the Skorokhod topology. The continuous mapping theorem and the open-set bound thus give the same positive lower bound for the liminf of the tube probabilities.
If the walk S starts near a large k, then over time h=k−2/3 its area fluctuations are of order one. The next lemma turns these fluctuations into a contraction of the weighted transition operator in the supremum norm, by coupling paths with the same endpoint but different accumulated phases.
Lemma 18 (Uniform short-time contraction). Fix $0<b_0\le b_1<\infty$ and ℓ0≥0. Put h=k−2/3, Ik=[k/2,2k]∩Z, and σk=inf{t≥0:St∈/Ik}. For a real function b on $[0,h]
satisfying b0≤b(t)≤b1 and ∣b(t)−b(r)∣≤ℓ0∣t−r∣, define
There are δ>0 and k0, depending only on b0,b1,ℓ0, such that ∥Tkb∥∞→∞≤1−δ for k≥k0.
Proof. Start two independent walks S1,t,S2,t at j∈Ik. Form coupled walks S1,t,S2,t by letting the second follow the first after their first meeting at or after h/2. On the independent paths put
Dk(u)=b(hu)k−2/3(S1,hu−S2,hu),0≤u≤1,
and let ζk be its first zero in [1/2,1], or 1 if none exists. The coupled phase difference is ∫0ζkDk(u)du. Uniformly over j and b, there is p>0 such that
P(ζk<1,2π/3<∫0ζkDk(u)du<4π/3)≥p
for all sufficiently large k. Indeed, otherwise take a sequence with k→∞ and success probabilities tending to zero. Pass to a subsequence with j/k→x∈[1/2,2] and b(0)→a∈[b0,b1]. Since supu∣b(hu)−b(0)∣≤ℓ0h, Lemma 16 and independence give Dk⇒a2xB. The difference cannot change sign without visiting zero, since the independent walks almost surely never jump simultaneously and b is positive. Lemma 17 contradicts the choice of sequence.
Let G denote the event in the preceding display, so that P(G)≥p, and let σℓ be the exit time of Sℓ from Ik. For ∥f∥∞≤1, extending f by zero outside Ik, set
Vℓ=exp(i∫0hb(t)Sℓ,tdt)f(Sℓ,h)1{σℓ>h},ℓ=1,2.
The marginal laws give EV1=EV2=(Tkbf)(j), and ∣Vℓ∣≤1. On G, the walks agree from time hζk onward. If both survive until h, they have the same endpoint and phase difference
Θ=∫0ζkDk(u)du∈(2π/3,4π/3).
Hence
∣V1+V2∣=∣f(S1,h)∣∣1+eiΘ∣≤2∣cos(Θ/2)∣≤1.
If either walk is killed, at most one contribution is nonzero, so again ∣V1+V2∣≤1. Thus this bound holds throughout G, while ∣V1+V2∣≤2 always. Consequently,
∣(Tkbf)(j)∣≤21E∣V1+V2∣≤21P(G)+P(Gc)≤1−2p.
Taking the supremum over j∈Ik and ∥f∥∞≤1 proves the assertion with δ=p/2.
Lemma 19 (Exponential exit bound).For the walk started at k, every T>0, and 0<θ≤1,
Pk(σk≤T)≤2exp(−2θk+4k2Tθ2).(21)
Proof. Put f(i)=cosh(θ(i−k)). For i∈Ik, the generator satisfies
(Gf)(i)=i2(coshθ−1)f(i)≤4k2θ2f(i),
since coshθ−1≤θ2 for 0≤θ≤1. Dynkin’s formula stopped at σk and Gronwall’s inequality give
Ekf(ST∧σk)≤e4k2θ2T.
On {σk≤T}, the displacement at exit is at least k/2, so f(ST∧σk)≥cosh(θk/2). Therefore
Pk(σk≤T)≤cosh(θk/2)e4k2θ2T≤2exp(−2θk+4k2Tθ2).
Proof of Proposition 15. Since q0=1 and q−k=qk for real β, it suffices to take k≥1. Fix λ∈J and β∈K, and put
h=k−2/3,m=⌊k1/3⌋,T=mh≤k−1/3.
Set b(t)=λ(β/4)e−βt/4 and f(j)=qj(β,λe−βT/4). On [0,1], these b have common positive lower and upper bounds and a common Lipschitz constant, uniformly over λ∈J and β∈K. Also ∥f∥∞≤1. By (19),
qk(β,λ)=Ek[ei∫0Tb(t)Stdtf(ST)].
For r=0,…,m−1, define b[r](t)=b(rh+t) on [0,h]. Each b[r] satisfies the same bounds in Lemma 18. Restricting to paths staying in Ik until T, the Markov property gives
For k≥k0, where k0 depends only on J,K, Lemma 18 therefore yields
∣qk(β,λ)∣≤(1−δ)m+Pk(σk≤T).
Apply Lemma 19 with θ=εk−2/3, where 0<ε<1/8 is fixed. Since T≤k−1/3,
Pk(σk≤T)≤2exp[−(ε/2−4ε2)k1/3].
Also (1−δ)m≤c1e−c2k1/3. Enlarging the prefactor to cover the finitely many remaining k proves (20).
Corollary 20 (Smoothness of the phase density). For β,λ>0, the law of αλ,β(0) mod 2π has a smooth density pβ,λ(θ). Its angular derivatives of every order are jointly continuous in (β,λ,θ) and locally uniformly bounded.
Proof. By Proposition 15, the Fourier series
pβ,λ(θ)=2π1k∈Z∑qk(β,λ)e−ikθ
and all its termwise angular derivatives converge absolutely and locally uniformly. By Fourier inversion [15 Chapter I, §7.1, Corollary, and §7.4], this defines a smooth density of the phase law.
The walk in (6) has a law independent of β and λ. Each path is absorbed after finitely many jumps, so Aβ is almost surely continuous in β>0. Thus, if (βn,λn)→(β,λ) in (0,∞)2, then
eiλnAβn⟶eiλAβalmost surely,eiλnAβn=1.
Bounded convergence in (6) gives joint continuity of each qk in (β,λ). Locally uniform convergence of every termwise angular derivative series gives the claimed joint continuity and local boundedness.
From boundary angles to correlation measures
Proof of Proposition 3. Write αλ=αλ,β(0). For a boundary angle u, put
μu=λ:αλ=umod2π∑∂λαλδλ.
For any continuous compactly supported g on R and bounded measurable h on T, changing variables u=αλmod2π on each inverse branch gives
∫Th(u)∫Rg(λ)μu(dλ)du=∫Rg(λ)h(αλ)dλ,
since du=∂λαλdλ cancels the spectral weight. Thus, as measures on R,
∫Th(u)μu(dλ)du=h(αλ)dλ.
Let U0=U be the uniform angle in (2), and put Uj=U+εVjmod2π, where the Vj are independent uniform variables on [−1,1], independent of U and the phases. For 0<ε<π the angles have joint density
2π(2ε)k−11on{1≤j≤k−1maxdist(uj−u0,2πZ)<ε}
and zero elsewhere. By the substitution above, E∏j=0k−1μUj(Ij) is therefore the integral on the right of (4), including its prefactor, with αλj−αλ0 in place of αλj−λ0.
The two vectors have the same joint law: subtract the equation for λ0 in (1) from the equations for λj. The differences solve the phase SDE (1) at λj−λ0, all driven by the same rotated Brownian motion dZt=e−iαλ0,β(t)dZt.
Almost surely no endpoint of the Ij has angle U, since U is independent and uniform. Only finitely many inverse branches meet each interval near U, and their locations and weights vary continuously. Thus μUj(Ij)→μU(Ij) almost surely. Each Uj is itself uniform and independent of the phases. Given n points in Ij, the rth moment of μU(Ij) is at most nrEwr for r≥1, by convexity, where w is one normalized weight. Proposition 7 and Lemma 8 thus give moments of all orders; Hölder’s inequality bounds the products uniformly in L2. We may therefore pass to expectations. At the common angle U every tuple in the product consists of distinct points because the intervals are disjoint. Given the points, their weights are independent with mean one, so the limiting expectation is E∏j=0k−1Ξβ(Ij).
Proof of Theorem 2. Take k=2 in Proposition 3, with supI0<infI1. Writing λ=λ1−λ0, the quotient in (4) is
2ε1∫−εεpβ,λ(θ)dθ⟶pβ,λ(0)
locally uniformly in λ, by Corollary 20. Thus ρβ(0,λ)=pβ,λ(0)/(2π) gives a continuous version for λ>0. Translation invariance and symmetry give the density off the diagonal, and the Fourier series in Corollary 20 gives (3).
Remark 21. For Nλ=Ξβ((0,λ]), we note in passing that ρβ(0,λ)=21∂λ2Var(Nλ)+1/(4π2), where
Var(Nλ)=61Eν[1−cos(λAβ)]+βπ22E1[Cin(λAβ)].
Here Cin(x)=∫0x(1−cost)dt/t, and ν(k)=6/(π2k2), k≥1, is the initial law of the walk S in the first term. The formula follows by conditioning on the phase in (12), using Fourier expansion, and applying Itô’s formula to αλ2.
Holomorphic Fourier coefficients
We now extend the Fourier moments to complex β with Reβ>0 using the representations (6) and (7). We first construct a holomorphic function for each cutoff, and then remove the cutoff.
Holomorphy before removing the cutoff
Lemma 22 (Holomorphy and ODE for the cutoff moments). For 1≤k≤n, λ>0, and Reβ>0, define qk,n(β,λ) by (7), with Aβ given by the integral in (6). This function is holomorphic in β and continuously differentiable in λ. It satisfies
Proof. Write En={τ0<τn+1}. The walk and this event are independent of β. On a retained path, τ0<∞ and Sv≤n before absorption. The integral
Aβ=4β∫0∞e−βv/4Svdv
is supported on the finite interval [0,τ0] and is entire in β. This follows from Lemma 11 applied to normalized Lebesgue measure on [0,τ0]. On the right half-plane it satisfies the path-independent bound
∣Aβ∣≤4n∣β∣∫0∞e−(Reβ)v/4dv=Reβn∣β∣.(24)
Thus
eiλAβ1En≤exp(Reβnλ∣β∣).
Lemma 11 gives holomorphy of qk,n in β, and dominated convergence gives continuity in λ. The same bound and gambler’s ruin give (23).
For real β>0, let S be the chain on {0,…,n+1} obtained by absorbing S at both endpoints: Sv=Sv∧τ0∧τn+1. Its generator Gn has zero rows at 0 and n+1. For T>0, define the finite-start cutoffs q−T,k(n)(t) on [−T,0] by the finite system
This finite system has a unique solution, which is bounded on [−T,0], with q−T,0(n)≡1 and q−T,n+1(n)≡0; the equations hold at the endpoints as well as in the interior. Lemma 9, with the time origin shifted to −T, gives
Indeed, the integrands agree once an endpoint has been reached, so the difference of their expectations is bounded in modulus by 2Pk(τ0∧τn+1>t+T). Convergence is therefore uniform on compact time intervals. Passing to the integral form of the finite ODE therefore shows that t↦qk,n(β,λeβt/4) satisfies the same time-dependent system. The change of variables s=λeβt/4 gives (22) for real β>0.
To extend the equation, fix 0<a<b and write its integral form:
Both sides are holomorphic in β; the displayed path bound is uniform for s∈[a,b] on compact subsets of the half-plane and justifies the integral by Lemma 11. The identity theorem extends (25) to Reβ>0. Its integrand is continuous in s, so the fundamental theorem of calculus now proves differentiability and (22).
A uniform energy estimate
The pathwise bound used above to prove holomorphy grows exponentially with n. To remove the cutoff, we need a different estimate that is uniform in n; this is the purpose of the following energy argument. A related weighted energy argument is used for real-parameter uniqueness in Qu and Valkó (2025, Proposition 16); here the estimate must be uniform in the cutoff for complex β.
Proof. Write qk=qk,n(β,λ), q=(q1,…,qn), and h(λ)=∥q(λ)∥n2. Multiplying (22) by qk/k2, summing, and taking real parts gives, before summation by parts,
For 0<λ0<λ, Gronwall’s inequality applied in the variable logλ (equivalently, multiplication by the integrating factor λa/2) gives
h(λ)≤(λλ0)a/2h(λ0)+a24b2[1−(λλ0)a/2].
Let λ0↓0. The first term vanishes because h(λ0) is bounded. Since b/a=∣β∣/Reβ, this proves (26).
Removing the cutoff
Proof of Proposition 5. For each fixed k≥1, Lemma 23 gives
∣qk,n(β,λ)∣≤2kReβ∣β∣,n≥k.
Thus the holomorphic cutoff transforms are locally bounded uniformly in n. For real β>0, subtracting (7) from (6) and applying the gambler’s ruin formula to the embedded simple random walk ([4 Chapter 4]) gives
This proves (8) and hence qk,n(β,λ)→qk(β,λ). Vitali’s theorem, Lemma 12, supplies a locally uniform holomorphic limit on the whole half-plane. It is the unique extension of the real Fourier coefficient.
For a fixed integer m, pass to the limit n→∞ in the first m terms of (26), and then let m→∞. We obtain
k=1∑∞k2∣qk(β,λ)∣2≤4(Reβ)2∣β∣2.(29)
For k≥1, define
q−k(β,λ)=qk(β,λ),q0(β,λ)=1.(30)
These are holomorphic and agree with the negative real Fourier moments. Equation (29) proves (9). In particular,
∣qk(β,λ)∣≤2∣k∣Reβ∣β∣,k=0,
uniformly in λ>0.
Analytic transfer and completion of the proof
The complex-parameter energy bound and the real stretched-exponential estimate combine through the two-constants theorem as follows.
Lemma 24 (Transfer of stretched-exponential decay). Let fk, k≥1, be holomorphic on Reβ>0. Suppose that for some a≥0 and every compact K in this half-plane,
β∈Ksup∣fk(β)∣≤cKka.
Suppose also that there is a nondegenerate compact interval J⊂(0,∞) and constants b,c>0 such that
β∈Jsup∣fk(β)∣≤ce−bk1/3.
Then for every compact K⊂{Reβ>0} there are constants bK,cK′>0 such that
β∈Ksup∣fk(β)∣≤cK′e−bKk1/3.(31)
The conclusion is uniform over an additional index whenever the constants in the two hypotheses are uniform over that index.
Proof. Fix K. Choose a bounded smooth convex domain D with K∪J⊂D and D⊂{Reβ>0}. The slit domain D∖J is connected. Let ω be the harmonic measure from Lemma 13, extended as 1 on J, and put ω∗=infKω>0. The bounds on ∂D and J, followed by (16), give
The same bound holds on K∩J directly. Absorbing the polynomial factor gives (31) with bK=bω∗/2. The proof preserves uniformity in an additional index.
Proof of Proposition 6. Fix a compact interval L⊂(0,∞) and apply Lemma 24 with fk(β)=qk(β,λ) and a=1. Proposition 5 supplies the complex bound uniformly in λ>0, and Proposition 15 supplies the real stretched-exponential bound uniformly for λ∈L. This proves the conclusion for positive indices. For negative indices, use (30) and the positive-index estimate on {β:β∈K}. Enlarging the prefactor covers k=0.
Proof of Theorem 1. Fix λ>0. By Proposition 6,
4π21k∈Z∑qk(β,λ)(32)
converges absolutely and locally uniformly on Reβ>0. Every summand is holomorphic by Proposition 5, so the sum is holomorphic. For real β>0, Theorem 2 identifies (32) with ρβ(0,λ). Uniqueness of the holomorphic extension follows from the identity theorem.
To prove joint continuity, let λn→λ>0. By Proposition 6, the holomorphic functions β↦ρβ(0,λn) are locally bounded uniformly in n. For real β>0, they converge to ρβ(0,λ) by Theorem 2. Vitali’s theorem, Lemma 12, therefore gives this convergence locally uniformly on Reβ>0. Together with continuity in β at fixed λ, this proves joint continuity.
Corollary 25 (Holomorphic extension of the phase density). For every fixed λ>0, the Fourier series
pβ,λ(θ)=2π1k∈Z∑qk(β,λ)e−ikθ
extends the real phase density to Reβ>0. It is holomorphic in β and smooth in θ; the series and every angular derivative converge uniformly for β in compact subsets of the half-plane and θ∈T.
Proof. For every j≥0, the bound in Proposition 6 remains summable after multiplication by ∣k∣j. Thus the differentiated series converges locally uniformly, which also proves holomorphy.
Fourier moments and phase density at β=2
We provide an explicit solution to (5) at β=2. The solution shows that the stretched exponential decay exponent 1/3 in Proposition 6 cannot be increased beyond 1/2. It also yields an explicit phase density, with a simple probabilistic representation at integer multiples of 2π.
For a>0, b∈R, and Rez>0, Tricomi’s confluent hypergeometric function is defined by
U(a,b,z)=Γ(a)1∫0∞e−ztta−1(1+t)b−a−1dt.(33)
We use its analytic continuation to C∖(−∞,0], with complex powers on the principal branch, and set U(0,b,z)=1. See DLMF (2026, (13.4.4)).
Proposition 26 (Fourier moments at β=2). For k≥1 and λ>0,
Proof. Let rk(λ) be the right-hand side of (34), with r0=1 and r−k=rk. The recurrence and derivative identities for U in DLMF (13.3.7), (13.3.10), and (13.3.22) give
λrk′=ikλrk+k2(rk+1+rk−1−2rk),k≥1.
The negative-index equations follow by conjugation.
For Rez>0, (33) and a shift in the second integral give
k![U(k,1,z)−ke−zU(k+1,1,z)]=∫0∞e−zthk(t)dt(36)
hk(t)=k[(1+t)ktk−1−1{t>1}tk+1(t−1)k].
For t>1 the ratio of the second term in brackets to the first is (1−t−2)k, so hk≥0. Integrating up to R>1 and substituting s=t−1 in the second term leaves a boundary term bounded by k/R; letting R→∞ gives
∫0∞hk(t)dt=k∫0∞(1+t)k+1tk−1dt=1.
Letting z=ε−iλ with ε↓0 in (36), dominated convergence gives rk(λ)=∫0∞eiλthk(t)dt. Hence ∣rk(λ)∣≤1, and Corollary 14 identifies rk with qk(2,λ).
On the principal branch, the large-parameter expansion of U and the large-argument expansion of the modified Bessel function Kν of the second kind in DLMF (13.8.11) and (10.40.2) give
k!U(k,b,z)∼πez/2z1/4−b/2k1/4+b/2e−2kz,b∈{0,1},
locally uniformly for z=0 with Rez≥0. At z=−iλ the b=1 term dominates unless eiλ=1, when it vanishes. This gives (35). □
The same Fourier coefficients give an explicit phase density. We use the exponential integral
E1(w)=∫1∞te−wtdt,Rew>0,
continued to C∖(−∞,0] on its principal branch; see DLMF (2026, §6.2(i)).
Corollary 27 (Phase density and pair correlation at β=2). For λ>0 and 0<θ<2π, put
w=−1−e−iθiλ=−2λcot2θ−2iλ.(37)
The density of αλ,2(0) mod 2π with respect to dθ is
where X is standard Cauchy and Y is independent exponential with rate πm. To see this, substitute eiλ=1 in (38) and use
ewE1(w)=∫0∞s+we−sds,w∈/(−∞,0],
the analytic continuation of DLMF (6.7.1). Evaluating the real part and setting s=πmt gives (39). Under x=cot(θ/2), the density becomes the convolution of the exponential density with rate πm and the standard Cauchy density, proving (40).
AI use. ChatGPT was used extensively to develop the random walk coupling, the pair correlation identity, and the analytic continuation argument, and to check calculations, locate references, and revise the manuscript. The authors take full responsibility for the mathematics and references.
Acknowledgments. Benedek Valkó was partially supported by the National Science Foundation award DMS-2246435. Bálint Virág was partially supported by an NSERC Discovery Grant.
Yun Li. School of Mathematical Sciences, Nankai University, Tianjin, China. [email protected]
Yahui Qu. Department of Mathematics, University of Wisconsin–Madison, Madison, USA. [email protected]
Benedek Valkó. Department of Mathematics, University of Wisconsin–Madison, Madison, USA. [email protected]
Bálint Virág. Department of Mathematics, University of Toronto, Toronto, Canada. [email protected]
References
[1]T. Assiotis and J. Najnudel (2026). Moments of CβE field partition function, Sineβ correlations and stochastic zeta. arXiv:2602.08739.
[4]R. Durrett (2019). Probability: Theory and Examples, fifth edition. Cambridge University Press.
[5]F. J. Dyson (1962). Statistical theory of the energy levels of complex systems. III. J. Math. Phys. 3, 166–175.DOI
[6]S. N. Ethier and T. G. Kurtz (1986). Markov Processes: Characterization and Convergence. John Wiley & Sons.
[8]P. J. Forrester (1992). Selberg correlation integrals and the 1/r² quantum many-body system. Nuclear Physics B 388(3), 671–699.
[9]P. J. Forrester (1994). Addendum to “Selberg correlation integrals and the 1/r² quantum many-body system”. Nuclear Physics B 416(1), 377–385.DOI
[10]P. J. Forrester (2010). Log-Gases and Random Matrices. London Mathematical Society Monographs 34, Princeton University Press.DOI
[12]Z. N. C. Ha (1995). Fractional statistics in one dimension: view from an exactly solvable model. Nuclear Physics B 435(3), 604–636. doi:10.1016/0550-3213(94)00537-O.DOI
[14]B. Jancovici (1981). Exact results for the two-dimensional one-component plasma. Physical Review Letters 46(6), 386–388. doi:10.1103/PhysRevLett.46.386.DOI
[15]Y. Katznelson (2004). An Introduction to Harmonic Analysis, third edition. Cambridge University Press.
[16]R. Killip and M. Stoiciu (2009). Eigenvalue statistics for CMV matrices: From Poisson to clock via random matrix ensembles. Duke Mathematical Journal 146, 361–399.DOI
[17]Y. Li and B. Valkó (2022). Operator level limit of the circular Jacobi β-ensemble. Random Matrices: Theory and Applications 11(4), 2250043. doi:10.1142/S2010326322500435.
[18]F. Nakano (2014). Level statistics for one-dimensional Schrödinger operators and Gaussian beta ensemble. Journal of Statistical Physics 156, 66–93. doi:10.1007/s10955-014-0987-x.DOI
[19]A. Okounkov (1997). On n-point correlations in the log-gas at rational temperature. arXiv:hep-th/9702001.arxiv.org/abs/hep-th/9702001
[21]T. Ransford (1995). Potential Theory in the Complex Plane. London Mathematical Society Student Texts 28, Cambridge University Press.DOI