Abstract
Given a probability measure \(\mu \) on the unit circle \({\mathbb {T}}\), consider the reproducing kernel \(k_{\mu ,n}(z_1, z_2)\) in the space of polynomials of degree at most \(n-1\) with the \(L^2(\mu )\)–inner product. Let \(u, v \in {\mathbb {C}}\). It is known that under mild assumptions on \(\mu \) near \(\zeta \in \mathbb {T}\), the ratio \(k_{\mu ,n}(\zeta e^{u/n}, \zeta e^{v/n})/k_{\mu ,n}(\zeta , \zeta )\) converges to a universal limit S(u, v) as \(n \rightarrow \infty \). We give an estimate for the rate of this convergence for measures \(\mu \) with finite logarithmic integral.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
Consider a probability measure \(\mu \) on the unit circle \({\mathbb {T}}= \{z \in {\mathbb {C}}: \; |z| = 1\}\) of the complex plane, \({\mathbb {C}}\). Assume that the support of \(\mu \) is an infinite subset of \({\mathbb {T}}\), so that monomials \(z^k\), \(k \geqslant 0\), are linearly independent in \(L^2(\mu )\). For each integer \(n \geqslant 1\), the set of polynomials of degree at most \(n-1\),
can be viewed as the n-dimensional Hilbert space of analytic functions with respect to \(L^2(\mu )\)-inner product. Denote by \(k_{\mu ,n}(z_1, z_2)\) the reproducing kernel at a point \(z_2 \in {\mathbb {C}}\) in this space, i.e., \(k_{\mu ,n}(\cdot , z_2) \in {\mathcal {P}}_{n}\) and
If \(\mu = m\) is the Lebesgue measure on \({\mathbb {T}}\) normalized by \(m({\mathbb {T}}) = 1\), the reproducing kernel has the following form:
One might check that if \(z_1 = \zeta e^{u/n}\), \(z_2 = \zeta e^{v/n}\) for some \(\zeta \in {\mathbb {T}}\) and \(u, v \in {\mathbb {C}}\) (equivalently, \(z_1\), \(z_2\) are at a distance comparable to 1/n from \(\zeta \)), then
where the remainder is uniform in (u, v) on compact subsets of \({\mathbb {C}}\times {\mathbb {C}}\). Such kind of behaviour of reproducing kernels is universal: under mild assumptions on a measure \(\mu \) near \(\zeta \in {\mathbb {T}}\), we have
uniformly in (u, v) on compact subsets of \({\mathbb {C}}\times {\mathbb {C}}\). Universality of the limiting behaviour of ratios of reproducing kernels attracted major attention in recent years. Several essentially different approaches were developed. Let us mention some of them. First papers dealt with real analytic weights and used the Riemann-Hilbert method, see, e.g., Deift [6] or Kuijlaars and Vanlessen [9]. Lubinsky [10] found a way of reducing a wide class of universality problems to the study of asymptotic behaviour of \(k_{\mu ,n}(z,z)\), \(z \in {\mathbb {T}}\). The latter asymptotic behaviour has been previously identified for general measures of Szegő class by Máté et al. [12]. Global Szegő condition has been weakened to the local one by Findley [8]. Another approach, also pioneered by Lubinsky [11], is based on compactness of normal families of entire functions and properties of \(\frac{\sin x}{x}\) kernel. An overview of this approach and further results can be found in Simon [15] and Totik [16]. Recently, Eichinger et al. [7] found yet another approach to universality based on spectral theory of canonical Hamiltonian systems. While this approach gives extremely general results (even the local Szegő condition can be omitted), it also involves a compactness argument as an essential element of the proof. Most of mentioned papers deal with measures on subsets of the real line due to motivation in the theory of random matrices. However, even in the simplified setting of measures on the unit circle, estimates for the rate of convergence
are missing in the literature. In fact, the rate of convergence in (1.1) is not known even in the case where \(\mu \) is an absolutely continuous measure on \({\mathbb {T}}\) with a smooth non-vanishing weight w. Indeed, compactness arguments, widely used for proving universality, cannot give bounds for the rate of convergence.
As an additional motivation of this work, we mention that Poltoratski [13] recently used universality in the proof of convergence of certain nonlinear Fourier transform (NLFT), and a subsequent development of this area, e.g., bounds for NLFT maximal operators, will require estimates for the convergence of universality limits.
In this paper, we estimate the rate of convergence in (1.1) for probability measures on the unit circle with finite logarithmic integral. For this we use an entropy function of a measure—a powerful instrument that recently found several applications in inverse problems [2, 3], scattering theory [5], and orthogonal polynomials [1, 4]. To be precise, let \(\mu \) be a probability measure on \({\mathbb {T}}\), and let \(\mu = w\,dm + \mu _{{\textbf{s}}}\) be its Radon-Nikodym decomposition into the absolutely continuous and singular parts. The measure \(\mu \) is said to belong to the Szegő class \({\textrm{Sz}}({\mathbb {T}})\) if its logarithmic integral is finite:
Since \(\log x \leqslant x\) for all \(x>0\), the latter condition is equivalent to \(\log w \in L^1(m)\). Let \({\mathbb {D}}= \{z \in {\mathbb {C}}:\; |z| < 1\}\) denote the open unit disk. For a measure \(\mu \in {\textrm{Sz}}({\mathbb {T}})\), its entropy function is given by
The function \({\mathcal {K}}_{\mu }\) is nonnegative in \({\mathbb {D}}\) by Jensen’s inequality. Moreover, for m-almost all \(\zeta \in {\mathbb {T}}\), we have \({\mathcal {K}}_{\mu }(z) \rightarrow 0\) as z non-tangentially approaches \(\zeta \). This follows from well-known properties of the Poisson kernel: we have
for m-almost all \(\zeta \in {\mathbb {T}}\), and
at each Lebesgue point \(\zeta \) of functions \(w, \log w \in L^1(m)\). In case (1.2)–(1.4) are satisfied, we have \({\mathcal {K}}_\mu (r\zeta ) \rightarrow 0\) as \(r \rightarrow 1\). Let \(B(\zeta , r) = \{z \in {\mathbb {C}}:\; |z-\zeta | \leqslant r\}\). The following theorem is the main result of the paper.
Theorem 1.1
Let \(\mu \in {\textrm{Sz}}({\mathbb {T}})\), \(A \geqslant 1\), \(n \geqslant 10A\), \(\zeta \in {\mathbb {T}}\). There exists \(\varepsilon _0 > 0\) depending only on A, such that if \(z_{1,2} \in B(\zeta , A/n)\) and \({\mathcal {K}}_{\mu }(\rho \zeta ) \leqslant \varepsilon _0\) for all \(\rho \in [1-A/n, 1)\), then
where \(\delta = \max \limits _{k=1,2}|z_{k} - \zeta |\), and the constant \(c>0\) is absolute.
Note that \(\rho \in [1- A/n, 1)\) in Theorem 1.1 tends to 1 as \(n \rightarrow \infty \), therefore, the right hand side of (1.5) tends to zero for all \(\zeta \) satisfying (1.2)–(1.4). This gives a nontrivial bound for the rate of convergence in (1.1) for Lebesgue almost all \(\zeta \in {\mathbb {T}}\). If \(\mu \) has some regularity in a neighbourhood of \(\zeta \in {\mathbb {T}}\), its entropy function can be explicitly estimated. For functions \(f, g > 0\), we use notation \(f \lesssim g\) (resp., \(f \gtrsim g\)) if \(f \leqslant cg\) (resp., \(f \leqslant cg\)) for some constant c, and \(f \sim g\) if both relations \(f \lesssim g\) and \(f \gtrsim g\) are satisfied.
Theorem 1.2
Let \(\mu = w\,dm\) be an absolutely continuous probability measure in \({\textrm{Sz}}({\mathbb {T}})\) such that w is positive and continuous in a neighbourhood \(I \subset {\mathbb {T}}\) of \(\zeta \in {\mathbb {T}}\). Assume, moreover, that \(|w(\xi ) - w(\zeta )| \sim |\xi - \zeta |^{s}\) for all \(\xi \in I\) and some \(s> 0\). Then we have
for \(\rho \in (0, 1)\) close enough to 1. The constants involved depend on s, the diameter of I, the value \(w(\zeta )\), and the constants in the relation \(|w(\xi ) - w(\zeta )| \sim |\xi - \zeta |^{s}\).
Let \(\lambda \in {\mathbb {D}}\). For the absolutely continuous probability measure \(\mu = \frac{1 - |\lambda |^2}{|1-\lambda \xi |^2}\,dm\), we have
due to the fact that integration against the Poisson kernel corresponds to harmonic continuation into the unit disk. We see that \({\mathcal {K}}(\mu , (1-1/n)\zeta ) \sim 1/n\) as \(n \rightarrow \infty \). Note that this agrees with bounds in Theorem 1.2 (we have \(s=1\) for this measure). By Theorem 1.1, we then have
for all \(z_1\), \(z_2\) in \(B(\zeta , 1/n)\) and large enough \(n \geqslant 0\), uniformly in \(\zeta \in {\mathbb {T}}\). As we will see in Sect. 4, in fact
This shows that the bound in Theorem 1.1 is not sharp for smooth measures. It seems, however, that this bound cannot be improved in the setting of the whole class \({\textrm{Sz}}({\mathbb {T}})\) of measures with finite logarithmic integral, i.e., there is a measure \(\mu \in {\textrm{Sz}}({\mathbb {T}})\) such that
for all \(n \geqslant 1\). In Sect. 5, we consider the absolutely continuous measures \(w_s\,dm\) such that \(w_s(e^{i\theta }) = c_{s}e^{|\theta |^{s}}\), \(\theta \in [-\pi , \pi ]\), where the constant \(c_s\) is chosen so that \(\int _{{\mathbb {T}}} w_s \,dm = 1\). By Theorem 1.2, we have
We demonstrate numerically that for \(x_n = 1- n^{-1}\), and fixed \(s \in (0, 1/2)\), we have
In other words, for each \(s \in (0,1/2)\), we have
and estimate (1.5) in Theorem 1.1 is sharp on this class of examples. It remains an open problem to give a mathematical proof of this fact.
2 Proof of Theorem 1.1
Let \(\mu \) be a probability measure supported on an infinite subset of \({\mathbb {T}}\), and let \(\{\varphi _n\}_{n \geqslant 0}\) be the family of its orthonormal polynomials obtained by Gram–Schmidt orthogonalization of monomials \(z^n\), \(n \geqslant 0\), in \(L^2(\mu )\). For a polynomial p of degree n, we set \(p^*(z) = z^n\overline{p(1/ {\bar{z}})}\). Note that p is also a polynomial of degree at most n. The polynomials \(\{\varphi _n^*\}_{n \geqslant 0}\) are called reflected orthonormal polynomials. We have the following recurrence relation (see formula (1.5.25), page 58, in [14]):
Here the recurrence coefficients, \(a_n\), \(n\geqslant 0\), belong to \({\mathbb {D}}\). It is also known (see Theorem 2.2.7, p. 124, in [14]) that the reproducing kernel in the n-dimensional space Hilbert space \(({\mathcal {P}}_n, (\cdot , \cdot )_{L^2(\mu )})\) at \(z_2 \in {\mathbb {C}}\) is given by
Note that \(k_{\mu , n}(z_1, z_2)\) is indeed an element of \({\mathcal {P}}_n\), i.e., a polynomial with respect to \(z_1\) of degree at most \(n-1\). It will be convenient to use a different representation of the reproducing kernel. Take \(n \geqslant 1\), \(a \in {\mathbb {D}}\), and define
Lemma 2.1
For all \(z_1, z_2 \in {\mathbb {C}}\), we have
Proof
The proof is a direct computation. At first, let \(z_1 = z_2 = z\). Then the left hand side in (2.3) is equal to
which does not depend on a. Taking a to be the recurrence coefficient \(a_{n-1}\), we see that
This relation holds for all \(z \in {\mathbb {C}}\). Since functions in (2.3) are analytic in \(z_1\) and anti-analytic in \(z_2\), the lemma follows. \(\square \)
The following lemma is Corollary 4 in [4].
Lemma 2.2
For every \(\lambda \in {\mathbb {D}}\) there is \(a \in {\mathbb {D}}\) such that the corresponding polynomial \(\tilde{\varphi }_{n}^*\) in (2.2) defines a probability measure \(\nu _{n,\lambda } = |\tilde{\varphi }_{n}^*|^{-2}\,dm\) on \({\mathbb {T}}\) such that
In the rest of the paper, we use notation \(\tilde{\varphi }_{n}^*\) for the polynomial from Lemma 2.2, where the value of the parameter \(\lambda \in {\mathbb {D}}\) will be clear from the context.
Lemma 2.3
Let \(\lambda \in {\mathbb {D}}\), and let \(\tilde{\varphi }_{n}^*\) be the corresponding polynomial from Lemma 2.2. We have
Proof
By (2.1) and Lemma 2.1, we have
It follows that the function \(|\tilde{\varphi }_{n}^*|^2 - |\tilde{\varphi }_{n}|^2\) is positive in \({\mathbb {D}}\) and is comparable to \(1-|z|\) when z approaches \({\mathbb {T}}\). Therefore, \(\tilde{\varphi }_{n}^*\) has no zeroes in \({\mathbb {D}}\). In fact, it has no zeroes in \(\overline{{\mathbb {D}}} = {\mathbb {D}}\cup {\mathbb {T}}\) (if \(\tilde{\varphi }_{n}^*(z_0) = 0\) at some \(z_0 \in {\mathbb {T}}\), then \(1-|z| \lesssim |\tilde{\varphi }_{n}^*|^2 - |\tilde{\varphi }_{n}|^2 \lesssim |\tilde{\varphi }_{n}^*|^2 \lesssim |z - z_0|^2\) near \(z_0\), leading to a contradiction). It follows that the function \(z \mapsto \frac{\tilde{\varphi }_n^*(\lambda )}{\tilde{\varphi }_n^*(z)}\) is analytic in a neighbourhood of \(\overline{{\mathbb {D}}}\). Then the Poisson formula
for harmonic functions implies
after noting that the function \(u = -2\mathop \textrm{Re}\Bigl (\frac{\tilde{\varphi }_{n}^*(\lambda )}{\tilde{\varphi }_{n}^*}\Bigr ) + 1\) is harmonic in a neighbourhood of \(\overline{{\mathbb {D}}}\), \(u(\lambda ) = -1\). Observe that
because \(|\tilde{\varphi }_{n}^*(\lambda )|^2 = \exp \bigl (\int _{{\mathbb {T}}}\log |\tilde{\varphi }_{n}^*(\xi )|^2 \frac{1-|\lambda |^2}{|1 - \bar{\xi }\lambda |^2}\,dm\bigr )\) (we use again formula (2.8), this time – for the harmonic function \(u = \log |\tilde{\varphi }_{n}^*|^2\)). The lemma now follows from (2.9) and (2.10). \(\square \)
Given two points \(\xi _{\pm } \in {\mathbb {T}}\), \(|\xi _+ - \xi _-| < 2\), and a number \(r<1\), we denote by \(\Gamma (\xi _{\pm }, r)\) the path in \({\mathbb {D}}\) formed by the union of two line segments \(\{\rho \xi _{\pm }, \; \rho \in [r, 1)\}\) and the smaller arc of the circle \(|z| = r\) with endpoints \(r\xi _-\), \(r\xi _{+}\). We also let \(z^* = 1/{\bar{z}}\) for \(z \in {\mathbb {C}}\setminus \{0\}\), and
The union \(\Gamma (\xi _{\pm }, r) \cup \{\xi _\pm \} \cup \Gamma ^*(\xi _{\pm }, r)\) is then the boundary of a domain to be denoted by \(\Omega (\xi _{\pm }, r)\). See Fig. 1.
Lemma 2.4
Suppose that \(h \in L^1(m)\), \(\eta >0\), and \(\lambda \in {\mathbb {D}}{\setminus } B(0, 3/4)\) are such that
Then there are \(\xi _{\pm } \in {\mathbb {T}}\) such that for every \(f \in H^1\) satisfying \(|f| \leqslant h\) on \({\mathbb {T}}\) we have \(|f(z)| \lesssim \eta \) on \(\Gamma (\xi _{\pm }, r)\), where r: \(1-r = 2(1-|\lambda |)\). Moreover, \(\xi _{\pm }\) are such that \(\zeta = \lambda /|\lambda |\) belongs to the arc of \({\mathbb {T}}\) with endpoints \(\xi _{\pm }\), and \(1-r \leqslant |\xi _{\pm }-\zeta | \leqslant 2(1-r)\).
Proof
Let \(\lambda \in {\mathbb {D}}{\setminus } B(0, 3/4)\), \(r=2|\lambda |-1\), \(\zeta = \lambda /|\lambda |\). Consider the arc of the unit circle \(G = {\mathbb {T}}\cap B(\zeta ,2(1-r))\), and define \(m_{G} = \frac{\chi _{G}}{m(G)}dm\). We have
The set G is the metric space with respect to the usual distance in \({\mathbb {C}}\). The measure \(m_G\) has doubling property on this space:
for every \(\xi \in G\), \(\rho > 0\). It follows that the weak norm of Hardy-Littlewood maximal operator on \(L^1(G, m_G)\) is bounded by a constant that does not depend on G. In other words, for every \(t>0\), \(g \in L^1(G, m_G)\) we have
where the constant involved does not depend on G, t, g. Taking \(g = h\), \(t = \varepsilon ^{-1}\eta \) for some \(\varepsilon > 0\), we obtain
It follows that
Indeed, this follows from the fact that for each \(\xi \in {\mathbb {T}}\), \(\rho \geqslant 0\), the Poisson kernel \(u \mapsto \frac{1-\rho ^2}{|1 - \rho {\bar{u}} \xi |^2}\) can be uniformly approximated on \({\mathbb {T}}\) by positive convex combinations of functions of the form \(\frac{\chi _{B(\xi , \delta )}}{m(B(\xi , \delta ))}\), \(\delta > 0\). Therefore, if \(\xi \in G\) is such that for some \(\rho \in [0, 1)\) we have
then
for some \(\delta > 0\), and so
proving (2.11). Let us now take \(\varepsilon \in (0, 1)\) so small that the left hand side of (2.11) does not exceed 1/10. Then there are \(\xi _{\pm } \in G\) such that \(\zeta \) belongs to the arc of \({\mathbb {T}}\) with endpoints \(\xi _{\pm }\), we have \(|\xi _{\pm } - \zeta | \geqslant 1-r\), and, moreover,
For \(u \in {\mathbb {T}}\setminus G\), we have
It follows that
for every \(\rho \in [r, 1)\) with absolute constants. We also note that
therefore,
for \(\xi \in G\). Collecting (2.12) and (2.13), we see that \(|f(z)|\!\lesssim \!\eta \) for \(z \in \Gamma (\xi _{\pm },\! r)\).\(\square \)
Define \( \tilde{b}_n = \frac{\tilde{\varphi }_n}{\tilde{\varphi }_n^*}. \) Formula (2.7) shows that \(|\tilde{b}_n|\leqslant 1\) on \({\mathbb {D}}\). In fact, \(\tilde{b}_n\) is a Blaschke product of order n.
Lemma 2.5
Let \(\lambda \in {\mathbb {D}}\), denote by \(\eta = e^{{\mathcal {K}}_{\nu _{n,\lambda }}(\lambda )} - 1\) the number in the right hand side of (2.6). Set \(\alpha = \overline{\tilde{\varphi }_n^*(\lambda )}/\tilde{\varphi }_n^*(\lambda )\). We have
Proof
Consider the sets
On \(E_l\), the difference \(|\tilde{b}_n(\xi ) - \alpha \xi ^n| \leqslant 2\) could be large, but the measure of this set is small. Let us use Chebyshev’s inequality and Lemma 2.3 to estimate the corresponding integral:
On \(E_s\), the difference \(|\tilde{b}_n(\xi ) - \alpha \xi ^n|\) is small. Indeed, for \(\xi \in E_s\), we write
For \(z \in {\mathbb {C}}\) such that \(|1-z| < 1/2\), we have \(|z^{-1}| \leqslant 2\), so \(|\tilde{\varphi }_n^*(\xi )/\tilde{\varphi }_n^*(\lambda )| \leqslant 2\) on \(E_s\). Then
Using \(ab -1 = a(b-1) + a-1\) for \(a = \overline{\varphi _n^*(\xi )/\varphi _n^*(\lambda )}\), \(b = \varphi _n^*(\lambda )/\varphi _n^*(\xi )\), we see that
on \(E_s\). The claim now follow from Lemma 2.3. \(\square \)
Lemma 2.6
Let \(\lambda \in {\mathbb {D}}\), \(A \geqslant 1\), \(\eta = e^{{\mathcal {K}}_{\nu _{n,\lambda }}(\lambda )} - 1\). Define r so that \(1-r = 2(1-|\lambda |)\), and assume that \(1/2 \leqslant 1- A/n \leqslant r < 1\). There exists a number \(\eta _0 \in (0,1)\) depending only on A, such that if \(\eta \leqslant \eta _0\), then there are \(\xi _{\pm } \in {\mathbb {T}}\) such that
for all \(z \in \Gamma (\xi _{\pm }, r) \cup \Gamma ^*(\xi _{\pm }, r)\). Moreover, \(\xi _{\pm }\) are such that \(\zeta = \lambda /|\lambda |\) belongs to the arc of \({\mathbb {T}}\) with endpoints \(\xi _{\pm }\), and \(1-r \leqslant |\xi _{\pm }-\zeta | \leqslant 2(1-r)\).
Proof
Consider the function \(h = |\tilde{b}_n - \alpha \xi ^n|^2 + \left| \frac{\tilde{\varphi }_n^*(\lambda )}{\tilde{\varphi }_n^*} -1 \right| ^2\) on \({\mathbb {T}}\). By Lemma 2.5 and Lemma 2.3, we have
Then, by Lemma 2.4, applied to the functions \((\tilde{b}_n(z) - \alpha z^n)^2\), \((\tfrac{\tilde{\varphi }_n^*(\lambda )}{\tilde{\varphi }_n^*(z)} -1)^2\), there exists a contour \(\Gamma (\xi _{\pm }, r)\) such that
for all \(z \in \Gamma (\xi _{\pm }, r)\). Moreover, \(\xi _{\pm }\) are such that \(\zeta \) belongs to the arc of \({\mathbb {T}}\) with endpoints \(\xi _{\pm }\), and \(1-r \leqslant |\xi _{\pm }-\zeta | \leqslant 2(1-r)\). Choosing \(\eta _0 \in (0,1)\) sufficiently small, one can guarantee that the left hand sides in these inequalities are smaller than \(e^{-4A}/4\) for all \(z \in \Gamma (\xi _{\pm }, r)\). In particular, for all \(z \in \Gamma (\xi _{\pm }, r)\) we have
where we have used the elementary inequality \(\log (1-x) \geqslant -2x\), \(x \in [0, 1/2]\). Then the identity
for the inner functions \(\theta = \tilde{b}_n\), \(\theta = \alpha z^n\), gives (2.14) on \(\Gamma ^*(\xi _{\pm }, r)\):
To estimate \(\bigl |\frac{\tilde{\varphi }_n^*(\lambda )}{\tilde{\varphi }_n^*} - 1\bigr |\) on \(\Gamma ^*(\xi _{\pm }, r)\), we use relation
Then, formula \(\overline{ab} - 1 = \overline{a(b-1) + a-1}\) for \(a = \frac{\tilde{\varphi }_n^*(\lambda )}{\tilde{\varphi }_n^*(z)}\), \(b=\frac{\alpha z^n}{\tilde{b}_n(z)}\) implies (2.15) on \(\Gamma ^*(\xi _{\pm }, r)\). \(\square \)
Lemma 2.7
Let \(\lambda , A, n, r, \zeta , \eta \) be as in Lemma 2.6 and let \(z_{1,2} \in B(\zeta , 1-|\lambda |)\). Then
Proof
By Lemma 2.6, relations (2.14), (2.15) hold on some contour \(\Gamma (\xi _{\pm }, r) \cup \Gamma ^*(\xi _{\pm }, r)\), where \(\xi _{\pm }\) are such that \(\zeta = \lambda /|\lambda |\) belongs to the arc of \({\mathbb {T}}\) with endpoints \(\xi _{\pm }\), and \(1-r \leqslant |\xi _{\pm }-\zeta | \leqslant 2(1-r)\). By the maximum modulus principle, (2.14) holds in the bounded domain \(\Omega (\xi _{\pm }, r)\) such that \(\partial \Omega (\xi _{\pm }, r) = \Gamma (\xi _{\pm }, r) \cup \Gamma ^*(\xi _{\pm }, r) \cup \{\xi _{\pm }\}\). In particular, (2.14) holds for all \(z \in B(\zeta , 1-r)\). Now pick two points \(z_1, z\) such that
We have
Note that \(|1-{\bar{z}}_1 z| \gtrsim 1-|\lambda | \geqslant \frac{1-r}{2} \gtrsim n^{-1},\) because \(A \geqslant 1\). Recall that the maximum principle and (2.14) imply
Next, we have
Since \(\eta \leqslant \eta _0 \leqslant 1\), we can conclude that
The function \(z \mapsto \frac{\overline{\tilde{b}_n(z_1)}\tilde{b}_n - \overline{z_1}^n z^n}{1-{\bar{z}}_1 z}\) is analytic, hence the same estimate holds for all \(z \in B(\zeta , 1-r)\) by the maximum modulus principle. In particular, it holds for all points \(z_{1}\), \(z_2\) in \(B(\zeta , 1-|\lambda |)\). The lemma follows. \(\square \)
Proof of Theorem 1.1
Let \(\mu \in {\textrm{Sz}}({\mathbb {T}})\), \(\zeta \in {\mathbb {T}}\), \(A \geqslant 1\). Let us choose \({\tilde{N}}_{0}\) such that \(A/{\tilde{N}}_{0} < 1/4\) and
where \(\eta _0 \in (0, 1)\) is the number in Lemma 2.6. For \(n \geqslant {\tilde{N}}_{0}\), consider \(z_{1,2} \in B(\zeta , A/n)\) and let \(\lambda \in {\mathbb {D}}\) be defined by
Define \(\eta = e^{{\mathcal {K}}_{\nu _{n,\lambda }}(\lambda )} - 1\) and observe that \(\eta < \eta _0\) by (2.16) and Lemma 2.2. We are in assumptions of Lemma 2.6. Estimate (2.15) and the maximum modulus principle give
for all \(z \in B(\zeta , 1-|\lambda |)\). By Lemma 2.7, for all \(z_{1}, z_2 \in B(\zeta , 1-|\lambda |)\) we have
with \(|R_2(z_1, z_2)| \lesssim e^{4A}\sqrt{\eta }\), so
Taking \(z_1 = z_2 = \zeta \), we get
It remains to write
Note that
for an absolute constant c. We see that one can take \(\varepsilon _0\in (0, \eta _0)\) so that \(ce^{4A}\sqrt{\varepsilon _0} \leqslant 1/10\), then the required estimate will hold. \(\square \)
3 Proof of Theorem 1.2
Proof of Theorem 1.2
Let I be an arc of \({\mathbb {T}}\) with center \(\zeta \) such that \(|w(\xi ) - w(\zeta )| \sim |\xi - \zeta |^{s}\) for all \(\xi \in I\) and some \(s > 0\). By the assumption of the theorem, we have \(w(\zeta ) > 0\). Replacing I by a smaller interval, we can assume that
for all \(\xi \in I\) and some \(\delta > 0\). Denote \(u = \log w - \log w(\zeta )\). We have
We also have
By construction, we have
Let us set \(\varepsilon = 1-\rho \), assume that \(\varepsilon < m(I)/2\), and estimate
and
We have
as \(\varepsilon \rightarrow 0\). The constants in these relations depend on s. We see that
as \(\rho \rightarrow 1\). On the other hand, for \(\varepsilon = 1-\rho \) we have
Let \(c_1\), \(c_2\) be such that \(c_1|\xi - \zeta |^{s} \leqslant u(\xi ) \leqslant c_2|\xi - \zeta |^{s}\) for all \(\xi \in I\). Choose \(\delta \in (0, 1)\) so small that
From (3.3) we see that there are \(\xi _1 \in B(\zeta , \delta \varepsilon )\), \(\xi _2 \in B(\zeta , \varepsilon ) {\setminus } B(\zeta , (1-\delta )\varepsilon )\) such that
Then
and, simultaneously,
We see that \({\mathcal {K}}_{\mu }(\rho \zeta ) \gtrsim (1-\rho )^{2s}\) for all \(\rho \) close enough to 1. A more accurate estimate is needed for \(s \geqslant 1/2\). Let \(\delta > 0\) be such that
Consider \(j \in {\mathbb {Z}}\) satisfying \((1+\delta )^{j+3} < |I|/2\), and let \(\xi _1\), \(\xi _{2}\) be such that
We claim that either \(|u(\xi _2) - c| \gtrsim |\xi _2 - \zeta |^s\) or \(|u(\xi _1) - c| \gtrsim |\xi _1 - \zeta |^s\) for all such \(\xi _1\), \(\xi _2\), j. Indeed, if \(|u(\xi _2) - c| \leqslant c_1|\xi _2 - \zeta |^s/2\), then
proving the claim. We see that for each \(j \in {\mathbb {Z}}\) such that \((1+\delta )^{j+3} < |I|/2\), we have \(|u(\xi ) - c| \gtrsim |\xi - \zeta |^{s}\) on a half of the set \({\mathbb {T}}\cap B(\zeta , (1+\delta )^{j+3}) {\setminus } B(\zeta , (1+\delta )^{j})\). It follows that
and estimate (3.2) shows that
\(\square \)
4 Example: The Poisson Kernel
In the following example, the asymptotic behaviour of ratios of reproducing kernels could be explicitly analysed.
Example
Let \(\lambda \in {\mathbb {D}}\setminus \{0\}\). Consider the probability measure \(\mu = \frac{1 - |\lambda |^2}{|1 - \lambda \xi |^2}\,dm\) on the unit circle \({\mathbb {T}}\). For \(\zeta \in {\mathbb {T}}\), \(u, v \in {\mathbb {R}}\), \(z_1 = \zeta e^{u/n}\), \(z_2 = \zeta e^{v/n}\), we have
where \(\sup \nolimits _{|u|,\,|v| \leqslant 1}|\delta _n(u, v)|\) is comparable to 1/n.
Proof
We have (see Sect. 1.6 in [14])
It follows that
Then
and we see that
is comparable to
if \(\lambda \ne 0\), \(|\zeta -z| \sim 1-|z| \sim 1/n\), and n is large. It follows that
Now take \(z_1 = \zeta e^{u/n}\), \(z_2 = \zeta e^{v/n}\), with \(|u|,\,|v| \leqslant 1\). We have
where
In particular, we have
Here
and
Note that \(\theta _\lambda (z_1, z_2) =1\) in the case \(z_1 {\bar{z}}_2 = 1\). Considering \(z_1\), \(z_2\) such that \(|1 - z_1 {\bar{z}}_2| \sim 1/n\) and using maximum modulus principle for analytic functions, we see that
for all \(z_1\), \(z_2\) such that \(|z_{1,2} -\zeta | \lesssim 1/n\). It follows that \(\sup \nolimits _{|u|,\,|v| \leqslant 1}|\delta _n(u, v)| \lesssim 1/n\).
5 Sharpness of Theorem 1.1
Sharpness of Theorem 1.1 is an open problem. As we have seen in the previous section, one can have
if the measure \(\mu \) is very regular. Our aim in this section is to present some numerical results for the measures of the form \(w_s\,dm\), where \(w_s(e^{i\theta }) = c_{s}e^{|\theta |^{s}}\), \(\theta \in [-\pi , \pi ]\), and the constant \(c_s\) is chosen so that \(\int _{{\mathbb {T}}} w_s \,dm = 1\). We will see that
in the case \(s=0.1\), \(s = 0.2\), and \(s = 0.4\) (similar results can be obtained for other values of \(s \in (0, 1/2)\)). In particular, the upper bound in Theorem 1.1 in these cases coincides with the lower bound up to a multiplicative factor comparable to 1. We use the following MATLAB code to produce our examples (note that the weight \(w_s\) and the orthogonal polynomials in the script are not normalized, because the normalization plays no role when we consider ratios of reproducing kernels).
\(\square \)
This script produces the array D consisting of differences
for the value \(s = 0.4\) of Hölder continuity index. One can plot the values of this array (see below) and observe that it behaves like \(C_{\alpha }n^{-\alpha }\) for certain values \(\alpha >0\), \(C_{\alpha }>0\). To find these values, we observe that
which gives a reasonable recipe to compute the array alphaCand of candidates for the power index \(\alpha \):
When printed, the last 7 elements of this array (corresponding to values \(n=7880,7900\), ..., 8000) look as follows:
which is pretty close to \(s = 0.4\). So it seems plausible that
for all n large enough. We can even suggest a candidate for the constant \(C_{0.4}\) from our numerical data. For this we use approximation
The plots of the resulting functions
\(\square \) are given on Fig. 2. We also consider the cases \(s = 0.1\) and \(s = 0.2\).
The graphs demonstrate the fact that \(f_1 \geqslant f_2\) for \(s \in \{0.1, 0.2, 0.4\}\) and large integers n. Similar numerical results can be obtained for other values \(s \in (0,1/2)\).
Data Availability Statement
Not applicable.
References
Bessonov, R.: Entropy function and orthogonal polynomials. J. Approx. Theory 272, 105650 (2021)
Bessonov, R., Denisov, S.: A spectral Szegő theorem on the real line. Adv. Math. 359, 106851 (2020). https://doi.org/10.1016/j.aim.2019.106851
Bessonov, R., Denisov, S.: De Branges canonical systems with finite logarithmic integral. Anal. PDE 14(5), 1509–1556 (2021)
Bessonov, R., Denisov, S.: Zero sets, entropy, and pointwise asymptotics of orthogonal polynomials. J. Funct. Anal. 280(12), 109002 (2021)
Bessonov, R., Denisov, S.: Szegő condition, scattering, and vibration of Krein strings. Invent. Math. 234(1), 291–373 (2023)
Deift, P.A.: Orthogonal polynomials and random matrices: a Riemann-Hilbert approach. In: Volume 3 of Courant Lecture Notes in Mathematics. New York University, Courant Institute of Mathematical Sciences, New York; American Mathematical Society, Providence (1999)
Eichinger, B., Lukic, M., Simanek, B.: An approach to universality using Weyl \(m\)-functions. preprint arXiv:2108.01629
Findley, E.: Universality for locally Szegő measures. J. Approx. Theory 155(2), 136–154 (2008)
Kuijlaars, A.B.J., Vanlessen, M.: Universality for eigenvalue correlations from the modified Jacobi unitary ensemble. Int. Math. Res. Not. 30, 575–1600 (2002)
Lubinsky, D.: A new approach to universality limits involving orthogonal polynomials. Ann. Math. (2) 170(2), 915–939 (2009)
Lubinsky, D.S.: Universality limits in the bulk for arbitrary measures on compact sets. J. Anal. Math. 106, 373–394 (2008)
Máté, A., Nevai, P., Totik, V.: Szegő’s extremum problem on the unit circle. Ann. Math. (2) 134(2), 433–453 (1991)
Poltoratski, A.: Pointwise convergence of the non-linear Fourier transform. Preprint arXiv:2103.13349 (2021)
Simon, B.: Orthogonal Polynomials on the Unit Circle. Part 1: Classical Theory. Colloquium Publications. American Mathematical Society (2004)
Simon, B.: The Christoffel–Darboux kernel. In: Perspectives in Partial Differential Equations, Harmonic Analysis and Applications, Volume 79 of Proc. Sympos. Pure Math., pp. 295–335. American Mathematical Society, Providence (2008)
Totik, V.: Universality under Szegő’s condition. Can. Math. Bull. 59(1), 211–224 (2016)
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
The author declares no conflict of interest.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
The work is supported by grant RScF 19-71-30002 of the Russian Science Foundation. The author is a Young Russian Mathematics award winner and would like to thank its sponsors and jury.
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Bessonov, R. On Rate of Convergence for Universality Limits. Integr. Equ. Oper. Theory 96, 6 (2024). https://doi.org/10.1007/s00020-024-02757-8
Received:
Revised:
Accepted:
Published:
DOI: https://doi.org/10.1007/s00020-024-02757-8