The main result of this paper is an upper bound on the correlations of multiplicative functionals of determinantal point processes. In particular, for the sine process, the correlations of a multiplicative functional associated with a rapidly decaying function decay quadratically. It follows that the spectral density of such functionals belongs to the Sobolev space H1.
Introduction and statement of the main result
Multiplicative functionals of determinantal point processes
Let (E,d) be a complete separable metric space, let μ be a σ-finite measure on E, and let Π:E×E→C be a Borel function of two variables that is bounded on compact sets, and such that the operator Πf(x)=∫Π(x,y)f(y)dμ(y), initially defined on bounded Borel functions with compact support, extends to an orthogonal projection acting on L2(E,μ) (all Hilbert spaces considered in this paper are complex). By the Macchi–Soshnikov–Shirai–Takahashi theorem [10, 15, 14], the projection Π determines a determinantal measure on the space Conf(E) of locally finite point configurations on E.
If f is a bounded Borel function with compact support on E, and Ψ1+f is the multiplicative functional associated with f on the space of configurations Conf(E) defined by the formula
Ψ1+f(X)=x∈X∏(1+f(x)),
then (cf. [5, 4]) the determinantal measure PΠ is uniquely determined by the condition
EPΠΨ1+f=det(1+ΠfΠ)=det(1+fΠf).
Examples of measures that would subsequently be called determinantal already appear in Andréiéf’s work [1].
The notation fΠf is well-defined for both real- and complex-valued functions f. Either square root of f may be chosen; the only requirement is that the same square root be used on the left and on the right hand side. Now set
Ψ1+f=Ψ1+f⋅exp(−tr(ΠfΠ))=Ψ1+fexp(−tr(fΠf)).
We next introduce the normalized multiplicative functionals. Let HS to be the space of Hilbert–Schmidt operators acting on L2(E,μ). Consider the space HS(Π) of Borel functions f for which the operator fΠ is Hilbert–Schmidt, and equip this space with the Euclidean structure defined by ∥f∥HS(Π)=∥fΠ∥HS; by definition, we have
⟨f,h⟩HS(Π)=tr(fΠh)=⟨fΠ,hΠ⟩HS.
For the Hilbert space obtained by completing the space of bounded Borel functions with compact support with respect to the Euclidean structure <,>HS(Π), we retain, with a slight abuse of notation, the symbol HS(Π).
Recall that the regularized Hilbert–Carleman determinant [8] is defined for a finite-rank bounded operator A by the formula
2det(1+A)=det(1+A)exp(−trA)
and is then extended by continuity to the entire space of Hilbert–Schmidt operators.
Proposition 1.The map f→Ψ1+f extends to a continuous map HS(Π)→L1(Conf(E),PΠ), and for every f∈HS(Π), we have:
EPΠΨ1+f=2det(1+ΠfΠ)=2det(1+fΠ).
A general theorem on decay of correlation
The main result of this paper is an upper bound on the covariance of two normalized multiplicative functionals.
We use L2(Π) to denote the space of square-integrable functions with respect to the measure Π(x,x)dμ(x).
We will also need another embedding of the space of bounded Borel functions with compact support into the space of Hilbert–Schmidt operators, namely f→ΠfΠ. Define the Hilbert space H(2)(Π) as the completion of the space of bounded Borel functions with compact support in the norm
Remark 1. The same determinantal measure can be represented by different kernels, but the space H(2)(Π) is defined in terms of the determinantal measure itself, since the measure ∣Π(x,y)∣2dμ(x)dμ(y) is uniquely determined by the measure PΠ.
For r1>0,r2>0,r3>0, let EΠ(r1,r2,r3) be the family of functions f∈L∞(E,μ) defined by the formula
Theorem 1. For every r1>0,r2>0,r3>0, there exist δ>0 and A>0 such that the following holds. For any locally trace-class projection Π and the corresponding determinantal point process PΠ, any f,h∈EΠ(r1,r2,r3) and (f,h)∈NΠ(δ), we have
1.3. Correlations of multiplicative functionals of stationary determinantal point processes. Theorem 1 yields estimates on the rate of mixing for stationary determinantal point processes. Our convention for the Fourier transform is as follows:
f(λ)=∫Re−iλtf(t)dt;
f(t)=2π1∫Rf(λ)eiλtdλ.
As usual, we denote the inverse Fourier transform of f by fˇ.
We now consider stationary determinantal point processes on R corresponding to Fourier projections onto bounded Borel sets. Let B⊂R be a bounded Borel set of positive measure, and let the projection ΠB be defined by the formula
ΠBf=(1Bf)∨.
By definition, the kernel of the operator ΠB is given by the formula
ΠB(x,y)=1ˇB(x−y).
In particular, for B=[−π,π], we obtain the sine process, a determinantal point process with kernel
S(x,y)=π(x−y)sinπ(x−y).
the kernel of the projection onto the Paley–Wiener space. Let gs denote the action of the translation group on the real line, gs(t)=t+s; by a slight abuse of notation, we use the same symbol gs for the induced action of the translation group on Conf(R). Recall that the tail σ-algebra of a determinantal process associated with an orthogonal projection is trivial (see [11] in the discrete case and [7] and [12] in the general case). In particular, the determinantal process PΠB also has a trivial tail σ-algebra. Consequently, the translation group gs induces a Kolmogorov flow with respect to the determinantal measure PΠB and, consequently, has Lebesgue spectrum of countable multiplicity. Our aim is to give explicit regularity estimates for the spectral measures of multiplicative functionals with respect to the determinantal measure PΠB.
To a bounded Borel set B assign a positive increasing function ΦB by setting
ΦB(t)=(r:∣r∣≥∣t∣sup1ˇB(r)2)−1.
Note that positivity of ΦB follows from the Heisenberg uncertainty principle (the Fourier transform of the indicator function 1B(⋅) is not compactly supported, and hence ΦB(t) never takes the value +∞). Our key assumption is
∫0+∞ΦB(t)dt<+∞.(1)
Introduce the class KB of bounded Borel functions f on R such that
t∈RsupΦB(t)∣f(t)∣<+∞.
and set ∥f∥KB=supt∈RΦB(t)∣f(t)∣.
Corollary 1. Let B⊂R be a bounded Borel set satisfying condition (1). For every C>0 and c>0, there exists A>0 such that for any Borel functions fi∈KB, i=1,2, such that
Thus, the multiplicative functional corresponding to a function f satisfying conditions (3) has, under the sine process, continuous spectral density with square-integrable derivative; in other words, the spectral density belongs to the Sobolev space H1.
The proofs of Corollaries 1 and 2 are given in § 4.
Outline of the proof of Theorem 1
It will be more convenient for us to work with operator norms rather than determinants and to estimate the multiplicative rather than the additive correlation. We therefore reformulate Theorem 1 as follows. For r>0, let CΠ(r) be the family of functions f∈L∞(E,μ), defined by
CΠ(r)={f∈L∞(E,μ):∥f∥L∞+∥fΠ∥HS+(1+ΠfΠ)−1<r}.
Proposition 2. For every r>0, there exist δ>0 and A>0 such that the following holds. For any locally trace-class projection Π and
the corresponding determinantal point process PΠ, any f,h∈GΠ(r) and (f,h)∈NΠ(δ), we have
Note that the upper bound requires only that f,h belong to CΠ(r), whereas the lower bound relies on the condition (f,h)∈NΠ(δ0). The proof of Lemma 2 is given in Section 2.
After estimating the regularized Hilbert–Carleman determinant
2det(1+KΠ(f,h)),
in order to prove Proposition 2 and the main Theorem 1, it remains to estimate the trace tr(KΠ(f,h)).
Lemma 3.For anyr>0, there existsC4(r)>0, depending only onr, such that for allf,h∈CΠ(r), we have
Remark 2. Recall here that multiplying a determinantal measure by a regularized multiplicative functional again yields a determinantal point process; see [3].
To obtain the upper bound, it suffices to bound the Hilbert–Schmidt norm of the operator (ΠfhΠ−ΠfΠhΠ) from above and apply the classical Carleman inequality. Indeed, the upper bound follows directly from Carleman’s inequality (cf. (17)), the identity
∥ΠfhΠ∥HS2=⟨∣f∣2,∣h∣2⟩L2(Π)
and the inequality
∥ΠfΠhΠ∥HS2≤⟨∣f∣2,∣h∣2⟩H(2)(Π).
To obtain the lower bound, apply Lemma 5 to A=KΠ(f,h). In view of condition (13), for every δ1>0 there exists δ0>0 such that (f,h)∈NΠ(δ0) implies ∥KΠ(f,h)∥HS<δ1 and, a fortiori, ∥KΠ(f,h)∥HS<δ1. The second part of Lemma 2 now follows directly from Lemma 5.
Estimating the trace in terms of Hilbert–Schmidt norms and proof of Lemma 3
Proposition 4. For every r>0, there exists C4(r)>0, depending only on r, such that for all f,h∈CΠ(r), we have
Corollary 3 follows directly from Proposition 4 and it in turn implies Lemma 3. We proceed to the proof of Proposition 4. The proof of Proposition 4 is based on the following simple observation: if Q:L2(E,μ)→L2(E,μ) is a bounded operator, then we have
The absolute value of the first integral is at most A∥h∥L1/Φ(R/2), while the absolute value of the second is at most A∥f∥L1/Φ(R/2). The proof of Lemma 4 is now complete. ■
Next, for α>0, let KB,α be the set of Borel functions f on R such that supt∈RΦB(αt)∣f(t)∣<+∞, and define
∥f∥KB,α=t∈RsupΦB(αt)∣f(t)∣.
Recall that we make the key condition (1), as above. For any f,h∈KB, we now have f∗h∈KB,1/2, and
∥f∗h∥KB,1/2≤(∥f∥KB+∥h∥KB)⋅(∫RΦB(t)dt).
The following simple proposition, which estimates the inner product ⟨f,h(⋅+R)⟩H(2)(ΠB) in terms of ΦB, plays a key rôle in deriving Corollaries 1 and 2 from Theorem 1.
Proposition 5. If a bounded Borel set B satisfies assumption (1), then for all f,h∈KB we have
Define f by the formula f(k)=f(−k). Set g=f∗h. By Lemma 4, g∈KB,1/2 and we have
∥g∥KB,1/2≤(∥f∥KB+∥h∥KB)⋅∥ΦB−1∥L1.
Write
⟨f,h(⋅+R)⟩=∫Rg(R+k)∣ΠB(k)∣2dk.
Applying Lemma 4 once again proves the proposition.
Finally, note that if f,h∈KB, then for every δ>0, the condition (f,h(⋅+R))∈NΠB(δ) automatically holds for all sufficiently large R.
For the correlation coefficient of multiplicative functionals from Theorem 1, we conclude that for any bounded Borel set B⊂R satisfying condition (1), and for any r>0 and δ>0, there exists A>0 such that, if
We compare our estimates for the decay of correlation of multiplicative functionals with the formulas for the spectral densities of additive functionals obtained by A. B. Soshnikov (cf. (3.11) in [15]). To a bounded Borel function f with compact support on E assign the additive functional Sf on Conf(E) by the formula
Define the Hilbert space H˙1/2(Π) as the completion of the family of all bounded Borel functions with compact support on E with respect to the Euclidean norm ∥f∥H˙1/2(Π), defined by
In the last step, we used the fact that the function ∣1B∣2 is even. Set
QB=∣1B∣2.
Let ρB,f stand for the spectral density of the regularized additive functional Sf with respect to the determinantal measure PΠB. Following A. B. Soshnikov, observe that equation (16) implies the identity
ρB,f(λ)=∣f(λ)∣2(QB(0)−QB(λ)).
Using Lemma 4, one can directly verify that if f∈KB, then we have
R∈RsupΦB(R/4)EPΠBSf⋅Sf(⋅+R)<+∞.
Therefore, Theorem 1 gives the same rate of correlation decay for multiplicative functionals as Soshnikov’s spectral formula does for additive functionals.
Question. What lower bounds can be obtained for the correlation ratio of multiplicative functionals of determinantal point processes?
Bounds for Fredholm and Hilbert–Carleman determinants
For completeness, this section provides proofs of the elementary estimates for Fredholm and Hilbert–Carleman determinants used in the paper. We first recall Carleman’s inequality for the regularized Hilbert–Carleman determinant; see, for example, Gohberg, Goldberg, and Krupnik [9], Theorem 7.4, formula (7.10), p. 69: for every Hilbert–Schmidt operator A,
2det(1+A)≤exp(21∥A∥HS2).(17)
We now obtain a simple bound for the Hilbert–Schmidt norm of the inverse operator in terms of the Hilbert–Schmidt norm of the original operator and its Hilbert–Carleman determinant.
Proposition 6. Let A be a Hilbert–Schmidt operator such that 1+A is invertible. Then
1+AAHS≤∣det2(1+A)∣exp(∥A∥HS2+1).
The proof is a straightforward adaptation of the proof of Corollary 5.2 in Simon [13].
Set
B=1+AAHS(1+AA)∗.
Clearly, trBA(1+A)−1=1+AAHS.
Consider the function
F(λ)=2det(1+A+λB)=
2det(1+A)⋅2det(1+λB(1+A)−1)⋅exp(−λ1+AAHS).
We have
dλdF(λ)λ=0=−2det(1+A)1+AAHS.
By Cauchy’s estimate for the derivative of an entire function,
dλdF(λ)λ=0≤exp(∥A∥HS2+1).
Proposition 6 is proved. ■
Next, let J1 be the ideal of trace-class operators acting on the Hilbert space L2(E,μ):
J1={A:trAA∗<+∞}.
As usual, write ∥A∥J1=trAA∗.
Recall that if A∈J1, and the operator 1+A is invertible, then we have
Corollary 4. For every R>0, there exist positive constants c13 and C13 depending only on R such that, for every Hilbert–Schmidt operator A whose Hilbert–Schmidt norm does not exceed R and such that the operator 1+A is invertible, the following inequalities hold:
If we additionally assume that our Hilbert–Schmidt operator is a contraction, then the regularized Hilbert–Carleman determinant admits a lower bound in terms of the Hilbert–Schmidt norm of the operator:
Lemma 5. Let A be a Hilbert–Schmidt operator such that ∥A∥<1. Then
Note added in proof. After this paper was completed, I became aware of the beautiful paper [6] in which speed of mixing is investigated for the sine-beta process.
Acknowledgements
I am deeply grateful to A.S. Arakcheev, M.A. Borovikov and S.M. Gorbunov for fruitful discussions. I am deeply grateful to the anonymous referees whose detailed comments greatly helped me to improve the presentation.
The author is a winner of the “Leader” competition of the BA-SIS Foundation for the Advancement of Mathematics and Theoretical Physics, and would like to thank the Jury and sponsors.
Part of this paper was written while I was a guest at the Sirius Mathematics Centre, and I am deeply grateful to the Centre for its warm hospitality.
Steklov Mathematical Institute of Russian Academy of Sciences, Moscow
[4]Alexander I. Bufetov. “Quasi-symmetries of determinantal point processes”. The Annals of Probability, 46:2 (2018), 956–1003.arxiv.org/abs/1409.2068
[5]A. I. Bufetov. “Conditional measures of determinantal point processes”. Functional Analysis and Its Applications, 54:1 (2020), 11–28.
[6]L. Dumaz and M. Malvy “Long-Range Correlation of the Sineβ point Process”. arXiv, 2026, arXiv:2603.15289.arxiv.org/abs/2603.15289
[7]Alexander Bufetov, Y. Qui and Alexander Shamov “Kernels of conditional determinantal measures and the Lyons-Peres completeness conjecture”. Jour. Eur. Maths Soc (JEMS), 23:5 (2021), 1477–1519.
[8]Carleman, Torsten. Zur Theorie der linearen Integralgleichungen. Mathematische Zeitschrift. 9 (3–4): 196–217.
[9]Israel Gohberg, Seymour Goldberg, Nahum Krupnik, Traces and Determinants of Linear Operators, Birkhäuser Verlag 2000.
[10]O. Macchi. “The coincidence approach to stochastic point processes”. Adv. Appl. Prob., 7 (1975), 83–122.DOI
[12]Osada, H., Osada, S. Discrete Approximations of Determinantal Point Processes on Continuous Spaces: Tree Representations and Tail Triviality. J Stat Phys 170, 421–435 (2018). https://doi.org/10.1007/s10955-017-1892-2
[13]Simon B. Trace Ideals and Their Applications. — 2nd ed. — Providence, RI : American Mathematical Society, 2005. — viii, 150 p. — (Mathematical Surveys and Monographs ; vol. 120).
[14]T. Shirai and Y. Takahashi. “Fermion process and Fredholm determinant”. Proceedings of the Second ISAAC Congress, vol. 1, 15–23, Kluwer 2000.DOI
[15]A. B. Soshnikov. “Determinantal random point fields”. Uspekhi Mat. Nauk, 55:5(335) (2000), 107–160.