Sharp estimates for eigenvalues of localization operators with applications to area laws

Aleksei Kulikov, Martin Dam Larsen · 2026

Read Sharp estimates for eigenvalues of localization operators with applications to area laws by Aleksei Kulikov, Martin Dam Larsen (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-23832.txt.

Introduction

Kulikov and Larsen study eigenvalue counts for localization operators built from Fourier transforms and set projections. They count how many eigenvalues fall in a middle band as one set grows large. For sets made of finitely many boxes, they prove a sharp uniform bound. For general sets, their bound sits one logarithm above the guessed best bound. They apply the results to trace estimates for rough functions. One payoff is an enhanced area law in the box case.

Catalog record

CREATORAleksei Kulikov, Martin Dam Larsen
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE114788
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.23832

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Sharp estimates for eigenvalues of localization operators with applications to area laws Aleksei Kulikov, Martin Dam Larsen Abstract . We study the eigenvalues of the localization operator S A , B = P A ​ ℱ − 1 ​ P B ​ ℱ ​ P A S_{A,B}=P_{A}\mathcal{F}^{-1}P_{B}\mathcal{F}P_{A} , where ℱ \mathcal{F} is the Fourier transform and A = c ​ A 0 , B = B 0 A=cA_{0},B=B_{0} for some fixed sets A 0 , B 0 ⊂ ℝ d A_{0},B_{0}\subset\mathbb{R}^{d} and a large parameter c > 0 c>0 . For the counting function of the eigenvalues | { n : ε < λ n ​ ( A , B ) ≤ 1 − ε } | |\{n:\varepsilon<\lambda_{n}(A,B)\leq 1-\varepsilon\}| we obtain a sharp uniform upper bound if one of the sets is a finite disjoint union of parallelepipeds and a bound which is only a single logarithm off the conjectural optimal bound in the general case. These bounds are applied to the estimation of traces Tr ​ f ​ ( S A , B ) {\rm{Tr}}\,f(S_{A,B}) for functions f f with a very low regularity, in particular establishing an enhanced area law in the former case. 1. Introduction 1.1. Time-frequency localization operator For a measurable set A\subset\mathbb{R}^{d} by P_{A} we denote the projection onto A and by Q_{A} we denote the Fourier projection onto A Q_{A}=\mathcal{F}^{-1}P_{A}\mathcal{F}, where \mathcal{F} is the Fourier transform \mathcal{F}f(\xi)=\hat{f}(\xi)=\int_{\mathbb{R}^{d}}e^{-2\pi ix\cdot\xi}f(x)\,dx, whose inverse is \mathcal{F}^{-1}f(x)=\check{f}(x)=\int_{\mathbb{R}^{d}}e^{2\pi ix\cdot\xi}f(\xi)\,d\xi. For a pair of measurable sets A,B\subset\mathbb{R}^{d} by S_{A,B} we denote the time-frequency localization operator S_{A,B}=P_{A}Q_{B}P_{A}. This is a non-negative definite self-adjoint operator on L^{2}(\mathbb{R}^{d}) with the integral kernel (1.1) S_{A,B}(x,y)=1_{A}(x)\check{1}_{B}(x-y)1_{A}(y),\quad x,y\in\mathbb{R}^{d}. If the measures of A and B are finite, S_{A,B} is a trace class operator with {\rm{Tr}}(S_{A,B})=|A||B| . In particular, it is compact and as such it has a sequence of eigenvalues 1>\|S_{A,B}\|=\lambda_{1}(A,B)\geq\lambda_{2}(A,B)\geq\ldots\geq 0. We note for future reference that S_{A,B} obviously does not change if we modify A and B by sets of measure zero. It can be shown that the eigenvalues remain the same if we swap A and B , that is \lambda_{n}(A,B)=\lambda_{n}(B,A) . The fact that \lambda_{1}(A,B) is strictly less than 1 says that there does not exist a non-zero function such that it and its Fourier transform have supports of finite measure. This is known as the uncertainty principle of Benedicks and Amrein–Berthier [ 2 , 3 ] . In this paper we will be studying the finer properties of the distribution of these eigenvalues. The usual regime that is considered is when the sets A and B are scalings of some fixed sets A_{0},B_{0} . So, we will assume that A=cA_{0} and B=B_{0} for a large parameter c . Note that we do not put a scaling onto B , as it can be transferred to the scaling on A by an affine change of variables without altering the eigenvalues. More general affine changes of variables will play a role in our proofs later. The goal of the present paper is to study the distribution of eigenvalues \lambda_{n}(cA_{0},B_{0}) . We expect that the eigenvalues exhibit a phase transition: the first \approx c^{d}|A_{0}||B_{0}| eigenvalues are very close to 1 , then there are only \asymp c^{d-1}\log c intermediate eigenvalues away from 0 and 1 , and after that the eigenvalues decay to 0 extremely fast. The intermediate part of the spectrum is usually referred to as the plunge region. This picture is particularly clear in dimension 1 in the case when A_{0} and B_{0} are fixed intervals, where the exact limiting behaviour and very precise uniform estimates of the eigenvalues are known [ 18 , 22 , 21 , 6 , 10 ] , see also Section 2 below. In higher dimensions much less is known as the geometry of sets A_{0},B_{0} starts to play a significant role. To describe the results, let us introduce the counting functions for the eigenvalues. (1.2) \displaystyle N_{\varepsilon}(A,B) \displaystyle=|\{n:\lambda_{n}(c)>\varepsilon\}|, \displaystyle\Lambda^{+}_{\varepsilon}(A,B) \displaystyle=\left|\left\{n:1-\varepsilon>\lambda_{n}(c)>\frac{1}{2}\right\}\right|, \displaystyle\Lambda^{-}_{\varepsilon}(A,B) \displaystyle=\left|\left\{n:\frac{1}{2}\geq\lambda_{n}(c)>\varepsilon\right\}\right|, \displaystyle\Lambda_{\varepsilon}(A,B) \displaystyle=\Lambda^{+}_{\varepsilon}(A,B)+\Lambda^{-}_{\varepsilon}(A,B)=|\{n:1-\varepsilon\geq\lambda_{n}(c)>\varepsilon\}|. We chose this arrangement of strict and non-strict inequalities to ensure that for 0<\varepsilon<\frac{1}{2} we have (1.3) N_{1-\varepsilon}(A,B)=N_{1/2}(A,B)-\Lambda_{\varepsilon}^{+}(A,B),\quad N_{\varepsilon}(A,B)=N_{1/2}(A,B)+\Lambda_{\varepsilon}^{-}(A,B). Of course, for all applications whether we have strict or non-strict inequalities is immaterial. Previous results established bounds on N_{\varepsilon}(cA_{0},B_{0}) and \Lambda_{\varepsilon}(cA_{0},B_{0}) under strong enough assumptions on the boundaries \partial A_{0} and \partial B_{0} . In particular, Sobolev [ 26 ] established a two-term asymptotic formula for N_{\varepsilon}(cA_{0},B_{0}) for fixed \varepsilon>0 . There are also several uniform estimates on N_{\varepsilon}(cA_{0},B_{0}) and \Lambda_{\varepsilon}(cA_{0},B_{0}) with \varepsilon\to 0 as c\to\infty [ 16 , 14 , 12 , 23 ] . In this work we establish estimates on \Lambda_{\varepsilon}(cA_{0},B_{0}) which are on the one hand better than all of the previously obtained estimates, in particular being sharp in some cases, and on the other hand requiring much weaker and more natural geometric conditions on the sets A_{0},B_{0} . The form of our results depends on the assumptions we put on the sets A_{0} , B_{0} . If both of them are axis-parallel boxes then we can leverage known estimates in the one-dimensional case to establish precise two-sided estimates on \Lambda^{\pm}_{\varepsilon}(cA_{0},B_{0}) . If one of the sets is a disjoint finite union of parallelepipeds and the other satisfies a very weak boundary regularity condition then we have sharp uniform bounds on \Lambda^{\pm}_{\varepsilon}(cA_{0},B_{0}) from above. This in particular allows us to find two-term asymptotics for {\rm{Tr}}f(S_{A,B}) for very rough functions f . For the general case, the bound we obtain is off the conjectural one by at most a single logarithm. We will now present the cases we consider in the increasing order of generality. 1.2. Case of two boxes In this section we consider the case A=[0,a]^{d} , B=[0,b]^{d} . Since the variables separate, the operator S_{A,B} splits into a tensor power S_{A,B}=\left(S_{[0,a],[0,b]}\right)^{\otimes d} . In particular, the eigenvalues depend only on the product c=ab , and we get that the eigenvalues of S_{A,B} are all the products of d eigenvalues of the operator S_{[0,a],[0,b]} counted with multiplicity. Using this we can generalize all of the estimates on the counting functions from dimension 1 to higher dimensions with a simple combinatorial argument. Theorem 1.1 . Consider A=B=[0,1]^{d} , d\geq 1 , and c\geq 2 . There exists \alpha_{d}\geq 4 such that (1.4) \Lambda_{\varepsilon}^{\pm}(cA,B)\lesssim c^{d-1}\log(\tfrac{1}{\varepsilon})\log\left(\frac{\alpha_{d}\,c}{\log\left(\tfrac{1}{\varepsilon}\right)}\right) uniformly for all \alpha_{d}^{-c}<\varepsilon<1/2 . Moreover, if \varepsilon<c^{-\alpha_{d}} then we also have (1.5) \Lambda_{\varepsilon}^{\pm}(cA,B)\gtrsim c^{d-1}\log(\tfrac{1}{\varepsilon})\log\left(\frac{\alpha_{d}\,c}{\log\left(\tfrac{1}{\varepsilon}\right)}\right) If \varepsilon\leq\alpha_{d}^{-c} then there are no eigenvalues larger than 1-\varepsilon and (1.6) \Lambda_{\varepsilon}^{-}\asymp\left(\frac{\log(\tfrac{1}{\varepsilon})}{\log\left(\tfrac{\log\left(\frac{1}{\varepsilon}\right)}{c}\right)}\right)^{d}. The same tensor product structure can also be employed if A and B are axis-parallel boxes with not necessarily equal side lengths. In this way, by the same combinatorial argument, we get the following slightly more general result. Proposition 1.2 . Let d\geq 1 and I_{1},\dots,I_{d} and J_{1},\dots,J_{d} be fixed closed and finite intervals. Set A=I_{1}\times\dots\times I_{d} and B=J_{1}\times\dots\times J_{d} . Then the conclusion of Theorem 1.1 holds, except \alpha_{d} now also depends on the intervals (I_{k}) , (J_{k}) . 1.3. General case For general sets A,B the operator S_{A,B} no longer splits into a tensor product, and the precision of our results will depend on the geometries of A and B . Specifically, we will either assume that they are disjoint finite unions of parallelepipeds, or that they are bounded and their boundaries have finite upper

Read the complete text