37.08.04 · probability / 08-random-matrices

Determinantal Point Processes and Sine-Kernel Bulk Universality

shipped3 tiersLean: none

Anchor (Master): Anderson-Guionnet-Zeitouni, An Introduction to Random Matrices (Cambridge, 2010) Ch. 3-4; Mehta, Random Matrices 3e (Elsevier, 2004) Ch. 5-7; Deift, Orthogonal Polynomials and Random Matrices: A Riemann-Hilbert Approach (AMS, 1999); Tracy-Widom, Commun. Math. Phys. 159 (1994); Soshnikov, Russian Math. Surveys 55 (2000) (determinantal random point fields)

Intuition Beginner

The eigenvalues of a random matrix are not scattered like raindrops. Raindrops land independently: one drop tells you nothing about where the next will fall, and two can land arbitrarily close together. Eigenvalues behave oppositely. They push apart. Find one eigenvalue and you have learned something — the spots right next to it are nearly empty, because a second eigenvalue strongly avoids sitting on top of the first. This mutual avoidance is called level repulsion, and it is the single most important fact about how eigenvalues are spaced.

There is a clean bookkeeping device that captures all of this avoidance at once. To each pair of locations we attach one number, and the chance of finding eigenvalues at any list of spots is computed as a single determinant of the little table of those paired numbers. Determinants are built to vanish when two rows agree, so the moment two requested spots coincide the answer drops to zero — the repulsion is wired directly into the arithmetic. A cloud of points whose spacing statistics all come from one such table is called a determinantal process.

The surprise is that when you zoom far into the dense middle of the spectrum, rescaling so neighbouring eigenvalues sit about one unit apart, every reasonable random matrix produces the same paired-number table in the limit. That universal limiting table is the sine pattern, and it governs the fine-grained spacing of eigenvalues no matter what random rule built the matrix.

Visual Beginner

Picture two rows of dots on a line. The top row is a batch of independent raindrop points: they clump, leave big gaps, and sometimes two nearly touch. The bottom row is a batch of eigenvalues: they are eerily evenly spread, each keeping a polite distance from its neighbours, with no near-collisions and no large empty stretches. The eye can tell them apart instantly — the eigenvalue row looks "combed."

The little curve on the right is the punchline. It records the chance of finding a second eigenvalue a distance away from a chosen one. At zero distance the chance is zero — that is the repulsion. As the distance grows the chance climbs, gently overshoots, and levels off at the average rate. For the raindrops this curve would be flat at one everywhere, because independent points never notice each other.

Worked example Beginner

We compute the repulsion curve from the universal sine table in the simplest way, to see the zero-at-coincidence built in.

Step 1. The universal table assigns to a pair of spots separated by a distance $r$ the number $sin (π r) \div (π r)$ . When the two spots are the same, $r = 0$ , this expression is one (the value of $sin (π r) \div (π r)$ as $r$ shrinks to zero). When the spots differ, it is the familiar wavy function that crosses zero at every whole-number distance.

Step 2. The chance of finding eigenvalues at two spots, divided by the chance for two independent points, is one minus the square of that paired number. Call the paired number $s = sin (π r) \div (π r)$ ; the two-point answer is $1 - s^{2}$ .

Step 3. Put the two spots on top of each other, $r = 0$ . Then $s = 1$ , and the answer is $1 - 1^{2} = 1 - 1 = 0$ . The chance of two eigenvalues at the very same place is zero. That is level repulsion, stated as a single subtraction.

Step 4. Now separate the spots by one unit, $r = 1$ . Then $sin (π \times 1) = sin (π) = 0$ , so $s = 0$ and the answer is $1 - 0 = 1$ . At distance one — the average spacing — the points behave as if independent again.

Step 5. What this tells us: the one paired number $s = sin (π r) \div (π r)$ controls the whole spacing story. It equals one at zero distance, forcing the two-point chance down to zero (points cannot collide); and it dies away at larger distances, letting far-apart eigenvalues ignore each other. The single formula $1 - s^{2}$ is the repulsion curve drawn in the visual.

Check your understanding Beginner

Formal definition Intermediate+

Let $(E, E, λ)$ be a locally compact Polish space with a reference (Lebesgue-type) measure $λ$ . A simple point process on $E$ is a random locally finite configuration of points; its $k$ -point correlation function $ρ_{k} : E^{k} \to [0, \infty)$ is defined by requiring, for disjoint Borel sets $A_{1}, \dots, A_{k}$ , $$ \mathbb{E}\Big[\prod_{i=1}^k N(A_i)\Big] = \int_{A_1\times\cdots\times A_k} \rho_k(x_1,\dots,x_k), d\lambda(x_1)\cdots d\lambda(x_k), $$ where $N (A)$ counts points in $A$ . Informally $ρ_{k} (x_{1}, \dots, x_{k}) d λ (x_{1}) \dots d λ (x_{k})$ is the probability of finding a point in each infinitesimal neighbourhood of $x_{1}, \dots, x_{k}$ .

A simple point process is determinantal with correlation kernel $K : E \times E \to C$ if for every $k$ and all $x_{1}, \dots, x_{k}$ , $$ \rho_k(x_1,\dots,x_k) = \det\big[K(x_i, x_j)\big]_{i,j=1}^{k}. $$ The kernel $K$ is assumed to define a locally trace-class, locally self-adjoint integral operator (so the determinants are nonnegative and the $ρ_{k}$ are genuine correlation functions). The defining identity makes repulsion automatic: $\rho_2(x,x) = \det\begin{psmallmatrix} K(x,x) & K(x,x)\\ K(x,x) & K(x,x)\end{psmallmatrix} = 0$ , so no two points coincide.

The GUE eigenvalue process is the central example. The Gaussian Unitary Ensemble of $n \times n$ Hermitian matrices has joint eigenvalue density $$ p_n(\lambda_1,\dots,\lambda_n) = \frac{1}{Z_n}\prod_{i<j}(\lambda_i - \lambda_j)^2 \exp\Big(-\tfrac12\sum_{i=1}^n \lambda_i^2\Big), $$ the $β = 2$ log-gas developed in 37.08.03. Let ${p_{k}}_{k \geq 0}$ be the Hermite polynomials orthonormal for the weight $e^{- x^{2} /2}$ , and set the Hermite functions $ϕ_{k} (x) = p_{k} (x) e^{- x^{2} /4}$ , so $\int ϕ_{k} ϕ_{ℓ} d x = δ_{k ℓ}$ . The Hermite kernel is $$ K_n(x,y) = \sum_{k=0}^{n-1} \phi_k(x),\phi_k(y), $$ the kernel of the orthogonal projection onto $span {ϕ_{0}, \dots, ϕ_{n - 1}}$ . The eigenvalue process of the GUE is determinantal on $(R, d x)$ with kernel $K_{n}$ .

The Christoffel-Darboux formula collapses the sum into two terms: $$ K_n(x,y) = \sum_{k=0}^{n-1}\phi_k(x)\phi_k(y) = \sqrt{n},\frac{\phi_n(x)\phi_{n-1}(y) - \phi_{n-1}(x)\phi_n(y)}{x - y}, $$ with the diagonal value $K_{n} (x, x) = n (ϕ_{n}^{'} (x) ϕ_{n - 1} (x) - ϕ_{n - 1}^{'} (x) ϕ_{n} (x))$ by l'Hôpital. The constant $n$ is the ratio of leading coefficients $κ_{n - 1} / κ_{n}$ of the orthonormal Hermite polynomials in this normalisation.

The sine kernel is the bulk scaling limit. Fix a point $E$ in the open interval $(- 2 n, 2 n)$ where the semicircle density 37.08.01 is positive, let $ρ_{sc} (E)$ be that density, and rescale so the mean spacing is one: $x = E + \frac{u}{n ρ _{sc} ( E )}$ , $y = E + \frac{v}{n ρ _{sc} ( E )}$ . Then $$ \frac{1}{n\rho_{\mathrm{sc}}(E)}, K_n!\Big(E + \tfrac{u}{n\rho_{\mathrm{sc}}(E)},, E + \tfrac{v}{n\rho_{\mathrm{sc}}(E)}\Big) \longrightarrow K_{\sin}(u,v) = \frac{\sin\pi(u - v)}{\pi(u - v)}. $$ The gap probability — the chance that a Borel set $B$ contains no points — is the Fredholm determinant of the kernel restricted to $B$ : $$ \mathbb{P}\big(N(B) = 0\big) = \det\big(I - K|B\big) = \sum{k=0}^{\infty}\frac{(-1)^k}{k!}\int_{B^k}\det\big[K(x_i,x_j)\big]_{i,j=1}^k, dx_1\cdots dx_k. $$

Counterexamples to common slips Intermediate+

Determinantal $\neq =$ Gaussian-distributed. The kernel entries are not random; $K$ is a fixed function and the randomness is entirely in which points appear. A determinantal process is a recipe for the correlation functions, not a normal distribution.
The kernel is not unique. Conjugating $K (x, y) \mapsto \frac{g ( x )}{g ( y )} K (x, y)$ by any nonvanishing $g$ leaves every determinant $det [K (x_{i}, x_{j})]$ unchanged, so it leaves the process unchanged. Two different-looking kernels can describe the same process.
The diagonal carries the density. $ρ_{1} (x) = K (x, x)$ is the one-point intensity. For the Hermite kernel $K_{n} (x, x)$ is the GUE spectral density; it is not constant, and forgetting this conflates the local sine limit (constant intensity one after rescaling) with the global semicircle profile.
Bulk and edge limits differ. Rescaling at a fixed bulk point gives the sine kernel; rescaling at the spectral edge $\pm 2 n$ on the finer scale $n^{- 1/6}$ gives the Airy kernel, not the sine kernel. The scaling exponent and the limit object both change at the edge.

Key theorem with proof Intermediate+

Theorem (sine-kernel bulk limit of the GUE). Let $K_{n}$ be the Hermite kernel of the $n \times n$ GUE, and fix a bulk point $E$ with $E / 2 n \to e \in (- 1, 1)$ , so the rescaled semicircle density at $E$ is $ϱ := ρ_{sc} (e) = \frac{1}{π} 1 - e^{2}$ in the units where the spectrum fills $[- 2 n, 2 n]$ . Then, with the mean-spacing rescaling $x = E + u / (n ϱ)$ , $y = E + v / (n ϱ)$ , $$ \lim_{n\to\infty}\frac{1}{n\varrho},K_n!\big(x, y\big) = \frac{\sin\pi(u - v)}{\pi(u - v)} = K_{\sin}(u, v), $$ uniformly for $u, v$ in compact sets. Consequently every rescaled $k$ -point correlation function converges to $det [K_{s i n} (u_{i}, u_{j})]$ .

Proof. By Christoffel-Darboux the kernel is controlled by the two top Hermite functions $ϕ_{n}, ϕ_{n - 1}$ , so the limit reduces to their Plancherel-Rotach asymptotics in the oscillatory (bulk) regime. For $x = 2 n cos θ$ with $θ$ in a compact subset of $(0, π)$ , $$ \phi_n(x) = \Big(\frac{2}{\pi}\Big)^{1/2}(2n)^{-1/4}(\sin\theta)^{-1/2}\Big[\cos\big(n\psi(\theta) - \tfrac{\theta}{2} + \tfrac{\pi}{4}\big) + O(n^{-1})\Big], $$ where $ψ (θ) = θ - \frac{1}{2} sin 2 θ$ has derivative $ψ^{'} (θ) = 2 sin^{2} θ$ , and the phase advances by $ρ_{sc}$ per unit length: differentiating the local phase $n ψ (θ)$ in $x$ gives exactly $π n ϱ$ per unit $x$ at the centre point $E$ . Write the phase of $ϕ_{n}$ at the point $x = E + u / (n ϱ)$ as $Θ_{n} + π u + o (1)$ , where $Θ_{n}$ is the phase at $E$ ; the shift by $u / (n ϱ)$ multiplies the slowly varying amplitude by $1 + o (1)$ and advances the fast phase by $π u$ .

Abbreviate the common amplitude $a_{n} = (2/ π)^{1/2} (2 n)^{- 1/4} (sin θ)^{- 1/2}$ , evaluated at $E$ up to $1 + o (1)$ . Then $$ \phi_n(x) = a_n\cos(\Theta_n + \pi u) + o(a_n),\qquad \phi_{n-1}(x) = a_n\cos(\Theta_n + \pi u - \chi_n) + o(a_n), $$ where $χ_{n}$ is the phase lag of $ϕ_{n - 1}$ relative to $ϕ_{n}$ ; the recurrence relation for Hermite functions forces $χ_{n} \to π /2$ in the bulk (the two top functions are asymptotically in quadrature). Insert these into the Christoffel-Darboux numerator. With $x = E + u / (n ϱ)$ , $y = E + v / (n ϱ)$ , the denominator is $x - y = (u - v) / (n ϱ)$ , and $$ \phi_n(x)\phi_{n-1}(y) - \phi_{n-1}(x)\phi_n(y) = a_n^2\big[\cos(\alpha + \pi u)\sin(\alpha + \pi v) - \sin(\alpha + \pi u)\cos(\alpha + \pi v)\big] + o(a_n^2), $$ using $cos (\cdot - π /2) = sin (\cdot)$ with $α = Θ_{n}$ . The bracket is $sin ((α + π v) - (α + π u)) = sin π (v - u) = - sin π (u - v)$ , an exact angle-subtraction identity that erases the unknown phase $α$ .

Collect constants. Christoffel-Darboux carries the factor $n$ , the amplitude squared is $a_{n}^{2} = \frac{2}{π} (2 n)^{- 1/2} (sin θ)^{- 1}$ , and the diagonal density satisfies $K_{n} (E, E) = n ϱ (1 + o (1))$ , which fixes $n a_{n}^{2} \cdot (n ϱ) = - (sign) \frac{1}{π}$ after matching the one-point intensity to $n ϱ$ . Putting the pieces together, $$ \frac{1}{n\varrho}K_n(x,y) = \frac{1}{n\varrho}\cdot\sqrt n,\frac{a_n^2,(-\sin\pi(u-v))}{(u - v)/(n\varrho)} + o(1) = \frac{\sin\pi(u-v)}{\pi(u-v)} + o(1), $$ the normalisation absorbing every constant because the diagonal limit $\frac{1}{n ϱ} K_{n} (E, E) \to 1 = K_{s i n} (u, u)$ pins it. Uniformity on compacts follows from the uniform Plancherel-Rotach error $O (n^{- 1})$ . Since each $ρ_{k} = det [K_{n} (x_{i}, x_{j})]$ and the determinant is continuous in its entries, the rescaled $k$ -point functions converge to $det [K_{s i n} (u_{i}, u_{j})]$ . $□$

Bridge. This computation builds toward the universal local statistics of all Hermitian random matrices, and it appears again in the gap-probability and level-spacing analysis where the Fredholm determinant $det (I - K_{s i n})$ is evaluated. The foundational reason a single trigonometric kernel survives is that Christoffel-Darboux reduces the entire $n$ -term sum to the top two oscillating Hermite functions, and an angle-subtraction identity annihilates their common unknown phase — this is exactly why the limit is phase-independent and hence universal across bulk locations. The determinantal structure generalises the resolvent self-consistency of 37.08.02 from one-point density to all-order correlations: there one analytic equation pins the global density, here one kernel pins every local correlation. Putting these together, the bridge is that the projection kernel onto the lowest $n$ Hermite modes, rescaled at the mean spacing, converges to the projection onto the half-line of frequencies $[- \frac{1}{2}, \frac{1}{2}]$ — the sine kernel is the Fourier projection — and this is exactly the central insight that the same limit appears for every $β = 2$ ensemble and, by the universality theorems, for every Wigner matrix with finite moments.

Exercises Intermediate+

Exercise 3 (medium, symbolic).

Prove the Christoffel-Darboux formula $\sum_{k = 0}^{n - 1} p_{k} (x) p_{k} (y) = \frac{κ _{n - 1}}{κ _{n}} \frac{p _{n} ( x ) p _{n - 1} ( y ) - p _{n - 1} ( x ) p _{n} ( y )}{x - y}$ for any family of orthonormal polynomials with three-term recurrence $x p_{k} (x) = a_{k} p_{k + 1} (x) + b_{k} p_{k} (x) + a_{k - 1} p_{k - 1} (x)$ .

Hint

Form $(x - y) p_{k} (x) p_{k} (y)$ using the recurrence and telescope the sum over $k$ .

Answer

From the symmetric three-term recurrence, $x p_{k} (x) = a_{k} p_{k + 1} (x) + b_{k} p_{k} (x) + a_{k - 1} p_{k - 1} (x)$ and likewise in $y$ . Multiply the first by $p_{k} (y)$ , the second by $p_{k} (x)$ , and subtract: $$ (x - y)p_k(x)p_k(y) = a_k\big[p_{k+1}(x)p_k(y) - p_k(x)p_{k+1}(y)\big] - a_{k-1}\big[p_k(x)p_{k-1}(y) - p_{k-1}(x)p_k(y)\big]. $$ Writing $D_{k} = a_{k} (p_{k + 1} (x) p_{k} (y) - p_{k} (x) p_{k + 1} (y))$ , the right side is $D_{k} - D_{k - 1}$ . Summing $k = 0, \dots, n - 1$ telescopes to $D_{n - 1} - D_{- 1} = a_{n - 1} (p_{n} (x) p_{n - 1} (y) - p_{n - 1} (x) p_{n} (y))$ , since $D_{- 1} = 0$ . Dividing by $x - y$ gives the formula with $a_{n - 1} = κ_{n - 1} / κ_{n}$ , the ratio of leading coefficients. For the Hermite functions in this unit's normalisation $a_{n - 1} = n$ .

Exercise 4 (medium, symbolic).

Show that the sine kernel is the Fourier projection onto frequencies in $[- \frac{1}{2}, \frac{1}{2}]$ : prove $K_{s i n} (u, v) = \int_{- 1/2}^{1/2} e^{2 π i ξ (u - v)} d ξ$ , and deduce that $K_{s i n}$ , as an integral operator, is an orthogonal projection.

Hint

Integrate the exponential in $ξ$ ; then check $K_{s i n}^{2} = K_{s i n}$ using that the indicator of a set squares to itself in Fourier space.

Answer

$\int_{- 1/2}^{1/2} e^{2 π i ξ t} d ξ = \frac{e ^{π i t} - e ^{- π i t}}{2 π i t} = \frac{s i n π t}{π t}$ with $t = u - v$ , which is $K_{s i n} (u, v)$ . As an operator on $L^{2} (R)$ , $K_{s i n}$ is the conjugate by the Fourier transform of multiplication by $1_{[- 1/2, 1/2]} (ξ)$ . Since $1_{[- 1/2, 1/2]}^{2} = 1_{[- 1/2, 1/2]}$ and the indicator is real, the multiplier is an orthogonal projection, hence so is its Fourier conjugate $K_{s i n}$ . This is the continuum analogue of the Hermite kernel $K_{n}$ being the projection onto the first $n$ Hermite modes: under the bulk rescaling the band of occupied frequencies converges to $[- \frac{1}{2}, \frac{1}{2}]$ .

Exercise 5 (medium, numeric).

The expected number of points of the sine process in an interval of length $L$ is $\int_{0}^{L} ρ_{1} = L$ . Compute the variance of the count, given the determinantal variance formula $Var N (B) = \int_{B} K (x, x) d x - \int_{B \times B} K (x, y)^{2} d x d y$ , in the regime of large $L$ where it grows like $\frac{1}{π ^{2}} lo g L$ .

Hint

$Var N (B) = \int_{B} K_{s i n} (x, x) d x - \iint_{B^{2}} K_{s i n} (x, y)^{2} d x d y$ . The double integral of $(\frac{s i n π t}{π t})^{2}$ over a long interval grows like $L - \frac{1}{π ^{2}} lo g L$ .

Answer

With $B = [0, L]$ , $\int_{B} K_{s i n} (x, x) d x = L$ , and $\iint_{B^{2}} K_{s i n} (x, y)^{2} d x d y = \int_{- L}^{L} (L - ∣ t ∣) (\frac{s i n π t}{π t})^{2} d t$ . Using $\int_{R} (\frac{s i n π t}{π t})^{2} d t = 1$ and the slow logarithmic divergence $\int_{1}^{L} \frac{1}{π ^{2} t ^{2}} \cdot t d t = \frac{1}{π ^{2}} lo g L$ for the $∣ t ∣$ -weighted tail, one finds $\iint = L - \frac{1}{π ^{2}} lo g L + O (1)$ . Hence $Var N ([0, L]) = L - (L - \frac{1}{π ^{2}} lo g L) + O (1) = \frac{1}{π ^{2}} lo g L + O (1)$ . The variance grows only logarithmically, not linearly as for independent (Poisson) points whose variance equals the mean $L$ — a quantitative signature of rigidity from repulsion.

Exercise 6 (hard, short-answer).

Show that the gap probability $P (N (B) = 0) = det (I - K ∣_{B})$ follows from the inclusion-exclusion expansion of the correlation functions, and identify the Fredholm-determinant series.

Hint

The void probability is $E \prod (1 - 1_{B} (x_{i}))$ over points; expand and recognise the integrals against $ρ_{k}$ .

Answer

For a simple point process the probability of no point in $B$ is the generating functional $E [\prod_{x \in config} (1 - 1_{B} (x))]$ . Expanding the product over the configuration and taking expectations termwise against the correlation functions gives $$ \mathbb{P}(N(B) = 0) = \sum_{k=0}^{\infty}\frac{(-1)^k}{k!}\int_{B^k}\rho_k(x_1,\dots,x_k),dx_1\cdots dx_k, $$ the $1/ k!$ accounting for unordered $k$ -subsets. Substituting the determinantal form $ρ_{k} = det [K (x_{i}, x_{j})]$ yields exactly the Fredholm-determinant series $\sum_{k} \frac{( - 1 ) ^{k}}{k !} \int_{B^{k}} det [K (x_{i}, x_{j})] d x = det (I - K ∣_{B})$ , where $K ∣_{B}$ is the operator $f \mapsto \int_{B} K (\cdot, y) f (y) d y$ . Convergence is guaranteed because $K ∣_{B}$ is trace-class on the bounded set $B$ (Hadamard's bound controls the $k \times k$ determinants). Thus every gap probability of a determinantal process is a Fredholm determinant of its kernel.

Exercise 7 (hard, short-answer).

Explain why the sine kernel is the universal bulk limit — what is fixed and what is free — and contrast the bulk sine limit with the edge Airy limit, including the differing scaling exponents.

Hint

Track which feature of the matrix the local correlations depend on, and recall the square-root vanishing of the semicircle density at the edge.

Answer

After rescaling to unit mean spacing at a fixed bulk point, the limiting correlation kernel is $K_{s i n}$ for every Hermitian ensemble with a smooth bulk density: the universality theorems of Erdős-Schlein-Yau and Tao-Vu show that all Wigner matrices with matching second moment (and finite higher moments) share this local limit, so the entry distribution is free while the symmetry class ( $β = 2$ , complex Hermitian) is fixed. The sine kernel depends on the matrix only through the local density $ρ_{sc} (E)$ , which the rescaling normalises away. At the spectral edge the density vanishes like $(2 - x)^{1/2}$ 37.08.01, so the local spacing is not $n^{- 1}$ but $n^{- 2/3}$ ; rescaling there by $n^{1/6}$ at the edge $2 n$ gives the Airy kernel $K_{Ai} (u, v) = \frac{Ai ( u ) Ai ^{'} ( v ) - Ai ^{'} ( u ) Ai ( v )}{u - v}$ , whose gap probability is the Tracy-Widom distribution. Same Christoffel-Darboux machinery, different Plancherel-Rotach regime (turning-point Airy asymptotics rather than oscillatory cosine), hence a different universal kernel and exponent.

Advanced results Master

The determinantal structure of the GUE is exact at every finite $n$ , not merely asymptotic. The squared Vandermonde $\prod_{i < j} (λ_{i} - λ_{j})^{2}$ factors as $det [λ_{j}^{i - 1}]^{2}$ , and row-reducing the monomials into the orthonormal Hermite functions converts the joint density into $\frac{1}{n !} det [K_{n} (λ_{i}, λ_{j})]_{i, j = 1}^{n}$ . Integrating out $n - k$ variables, using the reproducing identity $\int K_{n} (x, z) K_{n} (z, y) d z = K_{n} (x, y)$ (because $K_{n}$ is a projection) and $\int K_{n} (x, x) d x = n$ , the Mehta-Gaudin-Dyson integration lemma collapses each integral and yields $ρ_{k} (λ_{1}, \dots, λ_{k}) = det [K_{n} (λ_{i}, λ_{j})]_{i, j = 1}^{k}$ for all $k \leq n$ . This is the finite- $n$ determinantal identity from which the sine-kernel limit is extracted.

The gap probability and level spacing are governed by the Fredholm determinant of the sine kernel on an interval. Writing $E_{2} (s) = det (I - K_{s i n}) ∣_{(- s /2, s /2)}$ for the probability that an interval of length $s$ (in mean-spacing units) is empty, the nearest-neighbour spacing density is $p (s) = \frac{d ^{2}}{d s ^{2}} E_{2} (s)$ . Gaudin and Mehta computed $E_{2}$ through the eigenvalues of the sine-kernel integral operator (prolate spheroidal functions); Jimbo-Miwa-Môri-Sato later identified $\frac{d}{d s} lo g E_{2} (s)$ with a Painlevé V transcendent, the bulk analogue of the Painlevé II governing the Tracy-Widom edge law. The small- $s$ behaviour $p (s) \sim \frac{π ^{2}}{3} s$ encodes linear level repulsion at $β = 2$ ; this linear vanishing (rather than the $s^{0}$ of Poisson or the $s^{β}$ of general $β$ ) is the determinantal fingerprint.

Soshnikov's theorem characterises which point processes are determinantal: a simple point process whose correlation functions are nonnegative and satisfy the determinantal identity for some Hermitian locally-trace-class kernel exists if and only if the operator $K$ satisfies $0 \leq K \leq I$ . Projection kernels ( $K^{2} = K$ ) such as the Hermite and sine kernels are the extreme case, and they describe processes with exactly the rigidity seen in eigenvalues. The general $0 \leq K \leq I$ condition is the determinantal analogue of a probability lying in $[0, 1]$ , and it makes the determinantal class closed under the operations — restriction, conditioning, complementation $K \mapsto I - K$ — that the eigenvalue applications require.

Bulk universality is the statement that the sine kernel is the entry-distribution-independent local limit. For Wigner matrices the proof proceeds by the three-step strategy: (i) a local semicircle law 37.08.02 controlling the density down to scale $n^{- 1 + ε}$ ; (ii) the relaxation of Dyson Brownian motion 37.08.03 to local equilibrium in time $t \sim n^{- 1 + ε}$ , which produces sine-kernel statistics for the Gaussian-divisible ensemble $M + t GUE$ ; and (iii) a Green's-function comparison (the four-moment theorem of Tao-Vu, or the Erdős-Yau continuity estimate) removing the added Gaussian component because local statistics depend only on the first four moments. The same architecture, with the Airy kernel replacing the sine kernel, gives edge universality and the Tracy-Widom law 37.08.01. The sine kernel is therefore not a feature of the Gaussian weight but a universal attractor for the local correlations of Hermitian spectra.

Synthesis. The foundational reason a single kernel organises all the local statistics is that the GUE eigenvalue process is the determinantal projection onto the lowest $n$ Hermite modes, and Christoffel-Darboux compresses this projection into the top two oscillating functions whose common phase the bulk scaling annihilates — this is exactly why the limit is the Fourier projection $K_{s i n}$ onto the band $[- \frac{1}{2}, \frac{1}{2}]$ and why it is universal. Putting these together, the moment method 37.08.01, the resolvent self-consistent equation 37.08.02, and the log-gas correlation functions 37.08.03 are three views of one spectral object: the first fixes the global semicircle, the second pins it analytically and reaches the edge, and the determinantal kernel of this unit refines both to all-order local correlations, so the global density is the diagonal $K_{n} (x, x) / n$ of the very kernel whose bulk limit is the sine process. This generalises the central-limit principle of classical probability 37.03.01 — there a universal Gaussian governs sums, here a universal sine kernel governs local eigenvalue correlations — and it is dual to the free-probability reading in which the semicircle is the free Gaussian. The central insight is that universality is a statement about which projection survives a scaling limit, and the bridge to the frontier is that the three-step Dyson-Brownian-motion strategy turns this exact-GUE computation into a theorem for every Wigner matrix, making the sine kernel the genuine local limit of Hermitian spectra.

Full proof set Master

The Christoffel-Darboux reduction and the sine-kernel limit are proved above. The remaining Master claims are recorded here.

Proposition (the GUE process is determinantal with the Hermite kernel). The eigenvalue density $p_{n} \propto \prod_{i < j} (λ_{i} - λ_{j})^{2} e^{- \sum λ_{i}^{2} /2}$ has $k$ -point correlation functions $ρ_{k} (λ_{1}, \dots, λ_{k}) = det [K_{n} (λ_{i}, λ_{j})]_{i, j = 1}^{k}$ , where $K_{n} (x, y) = \sum_{m = 0}^{n - 1} ϕ_{m} (x) ϕ_{m} (y)$ .

Proof. The Vandermonde determinant $\prod_{i < j} (λ_{j} - λ_{i}) = det [λ_{j}^{i - 1}]_{i, j}$ is unchanged, up to the constant of the leading coefficients, by replacing the monomials $λ^{i - 1}$ with the orthonormal Hermite polynomials $p_{i - 1} (λ)$ (each is a triangular linear combination of monomials). Hence $\prod_{i < j} (λ_{i} - λ_{j})^{2} = c det [p_{i - 1} (λ_{j})]^{2}$ . Folding $e^{- \sum λ_{i}^{2} /2} = \prod_{i} e^{- λ_{i}^{2} /2}$ into the rows gives $det [ϕ_{i - 1} (λ_{j})]^{2} = det [ϕ_{i - 1} (λ_{j})] det [ϕ_{i - 1} (λ_{ℓ})]$ , and the Cauchy-Binet / Andréief identity rewrites the product of the two determinants as $det [\sum_{m = 0}^{n - 1} ϕ_{m} (λ_{i}) ϕ_{m} (λ_{j})] = det [K_{n} (λ_{i}, λ_{j})]$ . So $p_{n} = \frac{1}{n !} det [K_{n} (λ_{i}, λ_{j})]_{i, j = 1}^{n}$ . The integration lemma — using $\int K_{n} (x, x) d x = n$ and the reproducing property $\int K_{n} (x, z) K_{n} (z, y) d z = K_{n} (x, y)$ valid because $K_{n}$ projects onto the span of $ϕ_{0}, \dots, ϕ_{n - 1}$ — integrates out the last $n - k$ variables one at a time, each integration reducing a size- $m$ determinant to a size- $(m - 1)$ determinant times the scalar $(n - m + 1)$ , and the factorials cancel to leave $ρ_{k} = det [K_{n} (λ_{i}, λ_{j})]_{i, j = 1}^{k}$ . $□$

Proposition (reproducing / projection property of $K_{n}$ ). $K_{n}$ is the kernel of the orthogonal projection of $L^{2} (R, d x)$ onto $V_{n} = span {ϕ_{0}, \dots, ϕ_{n - 1}}$ ; in particular $\int K_{n} (x, z) K_{n} (z, y) d z = K_{n} (x, y)$ and $\int K_{n} (x, x) d x = n$ .

Proof. By orthonormality $\int ϕ_{m} (z) ϕ_{m^{'}} (z) d z = δ_{m m^{'}}$ , so $\int K_{n} (x, z) K_{n} (z, y) d z = \sum_{m, m^{'} = 0}^{n - 1} ϕ_{m} (x) ϕ_{m^{'}} (y) \int ϕ_{m} (z) ϕ_{m^{'}} (z) d z = \sum_{m = 0}^{n - 1} ϕ_{m} (x) ϕ_{m} (y) = K_{n} (x, y)$ , the idempotence $K_{n}^{2} = K_{n}$ . Symmetry and reality of $ϕ_{m}$ give self-adjointness, so $K_{n}$ is an orthogonal projection; its range is $V_{n}$ because $K_{n} ϕ_{m} = ϕ_{m}$ for $m < n$ and $K_{n} ϕ_{m} = 0$ for $m \geq n$ . The trace is $\int K_{n} (x, x) d x = \sum_{m = 0}^{n - 1} \int ϕ_{m} (x)^{2} d x = n$ . $□$

Proposition (sine kernel is a projection; level repulsion). $K_{s i n}$ is the kernel of the orthogonal projection of $L^{2} (R)$ onto the band-limited subspace of functions with Fourier support in $[- \frac{1}{2}, \frac{1}{2}]$ , and the two-point function of the sine process is $ρ_{2} (u, v) = 1 - K_{s i n} (u, v)^{2}$ , vanishing quadratically as $u \to v$ .

Proof. From Exercise 4, $K_{s i n} (u, v) = \int_{- 1/2}^{1/2} e^{2 π i ξ (u - v)} d ξ$ is the Fourier conjugate of multiplication by $1_{[- 1/2, 1/2]}$ , an orthogonal projection since the indicator is real and idempotent; hence $K_{s i n}^{2} = K_{s i n}$ as an operator. The determinantal two-point function is $ρ_{2} (u, v) = K_{s i n} (u, u) K_{s i n} (v, v) - K_{s i n} (u, v) K_{s i n} (v, u) = 1 - K_{s i n} (u, v)^{2}$ , using $K_{s i n} (u, u) = 1$ and symmetry. Near coincidence, $K_{s i n} (u, v) = 1 - \frac{π ^{2}}{6} (u - v)^{2} + O ((u - v)^{4})$ from the Taylor expansion of $\frac{s i n π t}{π t}$ , so $ρ_{2} (u, v) = \frac{π ^{2}}{3} (u - v)^{2} + O ((u - v)^{4})$ , the quadratic vanishing that is linear level repulsion at the level of spacings. $□$

Proposition (gap probability is a Fredholm determinant). For a determinantal process with locally-trace-class kernel $K$ and bounded Borel $B$ , $P (N (B) = 0) = det (I - K ∣_{B})$ .

Proof. The void probability equals the generating functional $E [\prod_{x} (1 - 1_{B} (x))]$ , whose expansion against the correlation functions is $\sum_{k \geq 0} \frac{( - 1 ) ^{k}}{k !} \int_{B^{k}} ρ_{k} d x$ (Exercise 6). Substituting $ρ_{k} = det [K (x_{i}, x_{j})]$ gives the series $\sum_{k} \frac{( - 1 ) ^{k}}{k !} \int_{B^{k}} det [K (x_{i}, x_{j})]_{i, j = 1}^{k} d x$ , which is by definition the Fredholm determinant $det (I - K ∣_{B})$ of the trace-class operator $K ∣_{B} : f \mapsto \int_{B} K (\cdot, y) f (y) d y$ . Trace-class-ness on the bounded set $B$ and Hadamard's inequality $∣ det [K (x_{i}, x_{j})] ∣ \leq \prod_{i} ∥ K (x_{i}, \cdot) ∥$ make the series absolutely convergent, so the interchange of sum and expectation is justified. $□$

Connections Master

The Wigner semicircle law and the moment method 37.08.01 supplies the global density that the determinantal kernel refines. The diagonal of the Hermite kernel, $K_{n} (x, x) / n$ , is exactly the GUE spectral density whose large- $n$ limit is the semicircle; the local rescaling at a bulk point divides this density out to expose the sine kernel, and the square-root vanishing of the semicircle at the edge is what changes the local scaling from $n^{- 1}$ to $n^{- 2/3}$ and the sine kernel to the Airy kernel. This unit is the all-order local-correlation companion of that unit's one-point global law.

The Stieltjes transform and the resolvent 37.08.02 is the analytic route to the same density and the technical input to universality. The local semicircle law proved there — control of $s_{n} (z)$ down to imaginary part $n^{- 1 + ε}$ — is the first of the three steps that upgrade the exact-GUE sine-kernel computation of this unit to bulk universality for all Wigner matrices, since it places eigenvalues precisely enough that the Dyson-Brownian-motion relaxation can act on the correct local scale.

The GUE joint density and the log-gas 37.08.03 is the source of the determinantal structure used throughout this unit. The squared Vandermonde of the $β = 2$ log-gas is what factors into a single determinant of the Hermite kernel; the $β = 1$ and $β = 4$ log-gases are only Pfaffian, not determinantal, which is why the clean sine-kernel determinant is special to the unitary symmetry class, and the Dyson-Brownian-motion dynamics of that unit drives the relaxation step of the universality proof.

The QFT large- $N$ matrix model and topological expansion 08.14.06 meets this unit at the level of correlations: the connected two-point eigenvalue correlation of a one-matrix model has a universal bulk form — the sine kernel after local rescaling — independent of the matrix potential, the same universality the orthogonal-polynomial method makes exact here, and the loop-equation / Riemann-Hilbert analysis of that model is the field-theoretic counterpart of the Plancherel-Rotach asymptotics used in the proof.

Historical & philosophical context Master

The correlation-function structure of the Gaussian ensembles was worked out by Freeman Dyson and Madan Lal Mehta around 1960-1963. Mehta and Gaudin computed the level-spacing distribution as a Fredholm determinant of an integral operator in 1960 ^{[Gaudin 1960]}, reducing it to the spectrum of the sine-kernel operator via prolate spheroidal functions; Dyson's 1962 series on the statistical theory of energy levels ^{[Dyson 1962]} introduced the correlation functions of the circular ensembles and identified the sine kernel as their bulk limit, alongside his threefold ( $β = 1, 2, 4$ ) classification by time-reversal and spin. The determinantal identity $ρ_{k} = det [K (x_{i}, x_{j})]$ for the unitary ensemble and the orthogonal-polynomial method are systematised in Mehta's monograph ^{[Mehta 2004]}, long the standard reference.

The abstraction to general determinantal point processes was pursued by Odile Macchi, who introduced them in 1975 to model fermionic systems, and the systematic theory — including the $0 \leq K \leq I$ existence criterion — was given by Alexander Soshnikov in 2000 ^{[Soshnikov 2000]}. The connection to the moment problem and to integrable systems deepened when the gap probabilities were tied to Painlevé transcendents: Jimbo, Miwa, Môri and Sato linked the sine-kernel Fredholm determinant to Painlevé V in 1980, and Tracy and Widom recast both the sine (bulk) and Airy (edge) level-spacing laws in their 1994 Fredholm-determinant framework ^{[Tracy 1994]}. The proof that the sine kernel is universal — the local limit for all Wigner matrices, not merely the Gaussian ones — was completed in the late 2000s by the Erdős-Schlein-Yau program and independently by Tao and Vu, vindicating Wigner's hypothesis at the level of fine spacing statistics, and the Riemann-Hilbert asymptotic analysis of Deift and collaborators supplied the orthogonal-polynomial route to the same universality.

Bibliography Master

@article{gaudinmehta1960,
  author  = {Mehta, Madan Lal and Gaudin, Michel},
  title   = {On the density of eigenvalues of a random matrix},
  journal = {Nuclear Physics},
  volume  = {18},
  pages   = {420--427},
  year    = {1960}
}

@article{dyson1962,
  author  = {Dyson, Freeman J.},
  title   = {Statistical theory of the energy levels of complex systems. III},
  journal = {Journal of Mathematical Physics},
  volume  = {3},
  number  = {1},
  pages   = {166--175},
  year    = {1962}
}

@book{mehta2004,
  author    = {Mehta, Madan Lal},
  title     = {Random Matrices},
  edition   = {3rd},
  publisher = {Elsevier/Academic Press, Amsterdam},
  year      = {2004}
}

@article{soshnikov2000,
  author  = {Soshnikov, Alexander},
  title   = {Determinantal random point fields},
  journal = {Russian Mathematical Surveys},
  volume  = {55},
  number  = {5},
  pages   = {923--975},
  year    = {2000}
}

@article{tracywidom1994,
  author  = {Tracy, Craig A. and Widom, Harold},
  title   = {Level-spacing distributions and the Airy kernel},
  journal = {Communications in Mathematical Physics},
  volume  = {159},
  number  = {1},
  pages   = {151--174},
  year    = {1994}
}

@book{deift1999,
  author    = {Deift, Percy},
  title     = {Orthogonal Polynomials and Random Matrices: A Riemann-Hilbert Approach},
  series    = {Courant Lecture Notes},
  volume    = {3},
  publisher = {American Mathematical Society},
  year      = {1999}
}

@book{agz2010,
  author    = {Anderson, Greg W. and Guionnet, Alice and Zeitouni, Ofer},
  title     = {An Introduction to Random Matrices},
  series    = {Cambridge Studies in Advanced Mathematics},
  volume    = {118},
  publisher = {Cambridge University Press},
  year      = {2010}
}

Prerequisites

37.08.01

Tier anchors

beginner: Tao, Topics in Random Matrix Theory §2.6 (eigenvalue repulsion, the picture of points that avoid each other); the physical image of identical fermions filling levels and refusing to coincide; Mehta, Random Matrices 3e §5 (level-spacing histograms)
intermediate: Anderson-Guionnet-Zeitouni, An Introduction to Random Matrices §3.1-3.5 (orthogonal polynomials, Christoffel-Darboux, determinantal correlation functions, the sine kernel); Mehta, Random Matrices 3e Ch. 5-6; Tao §2.6
master: Anderson-Guionnet-Zeitouni, An Introduction to Random Matrices (Cambridge, 2010) Ch. 3-4; Mehta, Random Matrices 3e (Elsevier, 2004) Ch. 5-7; Deift, Orthogonal Polynomials and Random Matrices: A Riemann-Hilbert Approach (AMS, 1999); Tracy-Widom, Commun. Math. Phys. 159 (1994); Soshnikov, Russian Math. Surveys 55 (2000) (determinantal random point fields)

References

Mehta — Random Matrices · Elsevier/Academic Press, 3rd ed., 2004, Ch. 5-7 (Gaussian ensembles, orthogonal polynomials, n-point correlation functions, sine and Airy kernels)
Anderson, Guionnet, Zeitouni — An Introduction to Random Matrices · Cambridge University Press, 2010, Ch. 3-4 (determinantal structure of the GUE, Christoffel-Darboux, bulk scaling limit)
Dyson — Statistical theory of the energy levels of complex systems III · Journal of Mathematical Physics 3 (1962), 166-175 (correlation functions, the sine kernel for the circular ensembles)
Gaudin, Mehta — On the density of eigenvalues of a random matrix · Nuclear Physics 18 (1960), 420-427 (gap probability as a Fredholm determinant)
Soshnikov — Determinantal random point fields · Russian Mathematical Surveys 55 (2000), 923-975 (general theory of determinantal point processes)
Tracy, Widom — Level-spacing distributions and the Airy kernel · Communications in Mathematical Physics 159 (1994), 151-174 (Fredholm-determinant level-spacing, sine and Airy kernels)

Estimated time

beginner: 20m
intermediate: 55m
master: 95m