Deformed Wigner Random Matrices

• Deformed Wigner random matrices are additive perturbations of classical Wigner ensembles that incorporate structured modifications, such as full-rank diagonal and finite-rank perturbations, to reveal novel macroscopic and microscopic spectral features.
• They exhibit deterministic limits governed by free additive convolution and matrix Dyson equations while maintaining GOE/GUE universality in the regular one‐cut regime, even as local statistics adapt to the deformation.
• Finite-rank deformations introduce outlier eigenvalues and eigenvector localization phenomena, contrasting with full-rank deformations where the bulk spectrum remains semicircular and deterministic corrections dominate.

Searching arXiv for recent and foundational papers on deformed Wigner matrices. Deformed Wigner random matrices are additive perturbations of Wigner ensembles by deterministic or random structured operators, most prominently diagonal full-rank perturbations $H=W+\lambda V$ and finite-rank perturbations $\widehat H=H+VDV^*$. They form a central class in random matrix theory because they preserve the mean-field character of Wigner matrices while introducing nontrivial macroscopic structure, outlier mechanisms, and eigenvector effects. In the regular one-cut regime, their empirical law is governed by free additive convolution and their local bulk and edge statistics remain GOE/GUE-universal; in other regimes, especially finite-rank spiking or convex-edge full-rank diagonal deformation, the extremal spectrum becomes model-dependent and may exhibit Gaussian, Tracy–Widom, Weibull, or deformed Fréchet behavior, together with partial localization of extremal eigenvectors (Lee et al., 2014, Lee et al., 2014, Knowles et al., 2011, Knowles et al., 2012).

1. Canonical formulations and normalization

The term usually refers to several closely related additive models. In the full-rank diagonal setting, one studies

$H=\lambda_0 V+W,$

where $W$ is a real symmetric or complex Hermitian Wigner matrix with centered entries of variance $N^{-1}$, and $V=\operatorname{diag}(v_1,\dots,v_N)$ is a real diagonal matrix, deterministic or random, independent of $W$, with empirical measure converging to a compactly supported law $\nu$ (Lee et al., 2014). A broader deterministic-deformation form is

$H=W+B,$

with $B=B^*$ bounded in operator norm, and an associated monoparametric family

$\widehat H=H+VDV^*$0

where $\widehat H=H+VDV^*$1 is deterministic Hermitian and $\widehat H=H+VDV^*$2 is a scalar random parameter used to probe quenched universality (Cipolloni et al., 2021).

A second major class is finite-rank additive deformation. Here one starts from a Wigner matrix $\widehat H=H+VDV^*$3 and adds

$\widehat H=H+VDV^*$4

with fixed rank $\widehat H=H+VDV^*$5, orthonormal spike directions $\widehat H=H+VDV^*$6, and bounded spike strengths $\widehat H=H+VDV^*$7. This is the additive spiked-Wigner model underlying BBP-type outlier transitions (Knowles et al., 2011, Knowles et al., 2012). A related formulation,

$\widehat H=H+VDV^*$8

is used in finite-rank fluctuation theory under weak moment hypotheses (Pizzo et al., 2011).

The same terminology also covers dynamical or Hamiltonian variants such as

$\widehat H=H+VDV^*$9

with deterministic $H=\lambda_0 V+W,$0 and small random perturbation $H=\lambda_0 V+W,$1. In that context the deformed matrix is viewed simultaneously as a random matrix and as a quantum Hamiltonian, and resolvent technology is used to derive kinetic-time relaxation and prethermalization (Erdős et al., 2023). A common feature across these formulations is the Wigner normalization $H=\lambda_0 V+W,$2 on matrix entries, which puts the undeformed spectrum on an $H=\lambda_0 V+W,$3 scale and makes additive perturbations of operator norm $H=\lambda_0 V+W,$4 genuinely nonperturbative.

A recurrent misconception is that “deformed Wigner” means only low-rank spiking. In the literature, full-rank diagonal deformations $H=\lambda_0 V+W,$5 are equally standard, and many of the deepest results concern precisely this non-spiked, macroscopic regime (Lee et al., 2014, Lee et al., 2014).

2. Deterministic limits and self-consistent equations

For full-rank diagonal deformation, the global spectral law is the deformed semicircle law, i.e. the free additive convolution of the semicircle law with the limiting diagonal measure. Its Stieltjes transform $H=\lambda_0 V+W,$6 is characterized by

$H=\lambda_0 V+W,$7

and the density is recovered by

$H=\lambda_0 V+W,$8

Under a convexity-type condition on $H=\lambda_0 V+W,$9, the support is a single interval $W$0, the density is positive and analytic in the interior, the edges are square-root, and asymptotic outliers are excluded (Lee et al., 2014). The same free-convolution picture underlies bulk universality for $W$1, where $W$2 replaces the semicircle law as the macroscopic density (Lee et al., 2014).

For general deterministic additive perturbations, the correct deterministic object is the matrix Dyson equation. For $W$3, the deterministic resolvent approximation $W$4 solves

$W$5

with self-consistent density of states

$W$6

In the deformed Wigner case considered there, $W$7, so the MDE collapses to a scalar self-consistency along the identity direction, but it still provides the natural language for bulk quantiles, local laws, and quenched universality (Cipolloni et al., 2021). The same structure reappears in the dynamical model $W$8, where $W$9 satisfies

$N^{-1}$0

which is the MDE for free convolution of the density of states of $N^{-1}$1 with a semicircle of variance $N^{-1}$2 (Erdős et al., 2023).

Finite-rank deformation is different at the macroscopic level: since the perturbation modifies only finitely many eigenvalues, the empirical spectral measure remains asymptotically semicircular. The deformation acts only through outliers and edge rearrangements, not through a new bulk law (Knowles et al., 2012). This clean separation between global semicircle behavior and finite outlier structure is one of the key reasons finite-rank spiking is analytically tractable.

The one-cut square-root regime is not exhaustive. For diagonal $N^{-1}$3 with Jacobi-type density

$N^{-1}$4

the right edge can undergo a transition at a critical $N^{-1}$5. If $N^{-1}$6 and $N^{-1}$7, then near the upper edge one has

$N^{-1}$8

whereas for $N^{-1}$9 the usual square-root law persists (Lee et al., 2013). The same fast-decaying edge profile is used in spherical SK analysis, where the free-convolution density has power-law edge vanishing rather than Airy-type regularity (Lee et al., 2021). This distinction is decisive for extremal statistics.

3. Local laws, rigidity, and universality

A fundamental step in the subject is replacing global free-convolution control by local resolvent control. For undeformed Wigner matrices, the isotropic local semicircle law estimates

$V=\operatorname{diag}(v_1,\dots,v_N)$0

for arbitrary deterministic $V=\operatorname{diag}(v_1,\dots,v_N)$1, not just matrix entries or traces. This isotropic form is the key input for analyzing additive deformations, because spike directions need not align with coordinate axes (Knowles et al., 2011). It allows one to treat arbitrary deterministic perturbation eigenvectors and to derive outlier equations directly from $V=\operatorname{diag}(v_1,\dots,v_N)$2.

For full-rank diagonal deformation, the local deformed semicircle law asserts that the random Stieltjes transform follows the self-consistent one down to spectral scales matching the local spacing. Near the edge, after a normalization adapted to the deformed law, one has

$V=\operatorname{diag}(v_1,\dots,v_N)$3

with corresponding entrywise control on $V=\operatorname{diag}(v_1,\dots,v_N)$4, and edge rigidity

$V=\operatorname{diag}(v_1,\dots,v_N)$5

for the ordered eigenvalues $V=\operatorname{diag}(v_1,\dots,v_N)$6 and their classical locations $V=\operatorname{diag}(v_1,\dots,v_N)$7 (Lee et al., 2014). In the bulk, the analogous strong local law yields rigidity relative to quantiles of $V=\operatorname{diag}(v_1,\dots,v_N)$8, which is the starting point for DBM comparison and local equilibrium arguments (Lee et al., 2014).

In the regular one-cut regime, universality survives deformation. In the bulk, after unfolding by $V=\operatorname{diag}(v_1,\dots,v_N)$9, local correlation functions and single-gap statistics agree with those of GOE/GUE, first in annealed energy-averaged form and then at the level of individual gaps (Lee et al., 2014). At the edge, if the deformation preserves a square-root edge and prevents outliers, the largest eigenvalue satisfies

$W$0

with $W$1 the GOE Tracy–Widom law; the same method extends to joint top-$W$2 edge statistics and, in the complex Hermitian case, to $W$3 (Lee et al., 2014). Thus deformation changes centering and local scale through $W$4 and $W$5, but not the microscopic edge universality class.

A later development is quenched bulk universality. For a fixed typical deformed Wigner matrix $W$6, one can recover GOE/GUE gap statistics by sampling along the spectrum of $W$7 itself or by varying a single scalar parameter in $W$8. The proof combines MDE local laws, DBM on coupled families $W$9, a two-resolvent local law, and asymptotic orthogonality of eigenvectors for separated parameters or energies (Cipolloni et al., 2021). This result sharpens the standard annealed picture: universality can be observed with high probability in a single realization.

Universality is therefore conditional, not absolute. It is tied to one-cut regularity, square-root edges, and the absence of macroscopic outliers. Once those conditions fail, the extremal law can exit the Tracy–Widom class.

4. Outliers and finite-rank deformations

In the finite-rank additive model $\nu$0, the outlier threshold is $\nu$1 in Wigner normalization. If $\nu$2, the classical outlier location is

$\nu$3

and an eigenvalue separates from the bulk whenever $\nu$4 lies beyond the $\nu$5 edge scale. Equivalently, the transition occurs on the spike scale $\nu$6 (Knowles et al., 2011). Away from this critical window, the paper establishes both the law of the outliers and the sticking behavior of the non-outliers near the spectral edge, the latter remaining universal and governed by the corresponding Gaussian ensemble (Knowles et al., 2011).

The full outlier problem is substantially more intricate when several spikes are present. If several $\nu$7 are close enough that their classical locations differ by at most the fluctuation scale, they form an overlapping group. The joint asymptotic distribution of all outliers is then described by the eigenvalues of an explicit finite random matrix whose deterministic and covariance terms depend on higher moments of the Wigner entries and on the geometry of the spike eigenvectors. In particular, outlier fluctuations are generally nonuniversal, and distinct outlier groups can remain strongly correlated even when their spectral locations are far apart (Knowles et al., 2012).

Under weaker moment assumptions, earlier finite-rank work already identified a dichotomy between localized and delocalized perturbation eigenvectors. For localized spike directions, the outlier fluctuations retain dependence on fourth cumulants and on the finite-dimensional geometry of the perturbation; for delocalized spike directions, the limit becomes GOE/GUE after the natural $\nu$8 scaling (Pizzo et al., 2011). This picture was sharpened further: in the delocalized case, the outlier fluctuations are asymptotically those of a small GOE/GUE matrix plus an explicit deterministic shift driven by third moments, with no need for sub-Gaussian tails or identical entry distributions (Renfrew et al., 2012).

Full-rank diagonal deformation produces a different transition. For Jacobi-type $\nu$9 with $H=W+B,$0, once $H=W+B,$1, the upper edge is no longer regular square-root. Instead,

$H=W+B,$2

and the top eigenvalues are asymptotically governed by the order statistics of the diagonal entries: $H=W+B,$3 Consequently the largest eigenvalue has a Weibull law, and the top point process becomes asymptotically Poissonian rather than Airy (Lee et al., 2013). This regime is not a BBP outlier phase in the finite-rank sense; it is an edge-dominated phase in which the diagonal extremes determine the extremal spectrum.

5. Eigenvectors, localization, and mesoscopic structure

The eigenvector theory of deformed Wigner matrices is as regime-sensitive as the eigenvalue theory. In the convex-edge full-rank diagonal regime $H=W+B,$4, the eigenvector $H=W+B,$5 associated with the $H=W+B,$6-th largest eigenvalue is partially localized on the $H=W+B,$7-th largest diagonal site. More precisely,

$H=W+B,$8

while the remaining mass is distributed across the other coordinates with the bound

$H=W+B,$9

up to small powers of $B=B^*$0. For $B=B^*$1, by contrast, the extremal eigenvectors remain completely delocalized (Lee et al., 2013). The same parameter that governs the edge-law transition therefore also governs a delocalization-to-partial-localization transition at the spectral edge.

Mesoscopic deformations of diagonal matrices produce another canonical picture. In the generalized Rosenzweig–Porter model

$B=B^*$2

with $B=B^*$3, eigenvector entries in a bulk window are asymptotically Gaussian with an explicit variance profile, and for regular $B=B^*$4 this profile is universally Cauchy-shaped. The resulting eigenvectors are extended over $B=B^*$5 sites, yielding a partially delocalized, nonergodic regime. For smooth-entry Wigner noise, one also obtains a strong form of quantum unique ergodicity on mesoscopic scales (Benigni, 2017). This model interpolates between diagonal localization and mean-field delocalization and is often viewed as a deformed Wigner laboratory for nonergodic metallic behavior.

Rank-one deformation exhibits yet another mechanism. For

$B=B^*$6

below the BBP threshold $B=B^*$7, each individual eigenvector is asymptotically uninformative about $B=B^*$8, but the overlaps are not structureless. If $B=B^*$9 denotes the associated bulk eigenvalue, then

$\widehat H=H+VDV^*$00

an explicit profile increasing with $\widehat H=H+VDV^*$01, equivalently the generating function of Chebyshev polynomials of the second kind (Haddadi et al., 2018). This identifies a coherent subcritical overlap structure invisible to fixed-index asymptotics alone.

A broad lesson is that deformation can localize eigenvectors without necessarily creating finite-rank outliers, and can also produce intermediate phases between $\widehat H=H+VDV^*$02-delocalization and site localization. The eigenvector problem is therefore not reducible to the eigenvalue problem.

6. Extensions: spectral statistics, dynamics, and generalized ensembles

Beyond local universality and outliers, deformed Wigner matrices support a rich fluctuation theory for global observables. For $\widehat H=H+VDV^*$03 with deterministic diagonal $\widehat H=H+VDV^*$04, linear spectral statistics

$\widehat H=H+VDV^*$05

admit Gaussian fluctuations around a deterministic equivalent expressed through the finite-$\widehat H=H+VDV^*$06 free convolution $\widehat H=H+VDV^*$07. The CLT holds for $\widehat H=H+VDV^*$08 when centered by the mean, and the $\widehat H=H+VDV^*$09-bias relative to $\widehat H=H+VDV^*$10 converges for $\widehat H=H+VDV^*$11; both the variance and bias depend explicitly on $\widehat H=H+VDV^*$12, $\widehat H=H+VDV^*$13, $\widehat H=H+VDV^*$14, and the subordination map of the free convolution (Dallaporta et al., 2019).

Large deviations of the top eigenvalue are also understood in additive deformation settings. For

$\widehat H=H+VDV^*$15

there is an LDP at speed $\widehat H=H+VDV^*$16 for $\widehat H=H+VDV^*$17 in the Gaussian case, and a restricted LDP in the sharp sub-Gaussian diagonal-deformation case. The good rate function is expressed through spherical integrals and the free-convolution edge $\widehat H=H+VDV^*$18, linking rare-event asymptotics to the same self-consistent objects that govern typical behavior (McKenna, 2019).

In dynamics, the two-resolvent theory of deformed Wigner matrices yields quantitative relaxation formulas. For $\widehat H=H+VDV^*$19, the perturbed Heisenberg evolution relaxes to a thermal state through an intermediate prethermal state with lifetime of order $\widehat H=H+VDV^*$20, and the derivation rests on a two-resolvent isotropic global law for the deformed ensemble (Erdős et al., 2023). In a different direction, deformed Wigner input with fast-decaying free-convolution edges implies a sharp Gaussian-to-Weibull transition for the free energy of the spherical SK model, paralleling the underlying spectral transition between bulk-dominated and edge-dominated regimes (Lee et al., 2021).

Recent work has also pushed the subject beyond mean-field and finite-variance universality classes. In the tail crossover regime defined by

$\widehat H=H+VDV^*$21

dense, sparse, banded, and regular-graph Wigner-type matrices exhibit a deformed Fréchet law for the largest eigenvalue, together with a deformed Poisson point process for the top eigenvalues and localization of outlying eigenvectors (Han, 2024). For low-rank deformations of inhomogeneous Wigner-type matrices with variance profile, the BBP law of large numbers survives—the supercritical outlier remains at $\widehat H=H+VDV^*$22—but the fluctuation law becomes strongly nonuniversal, depending on sparsity, geometry, the variance profile, and the spike eigenvectors (Geng et al., 2024).

Taken together, these developments show that deformed Wigner random matrices are not a single model but a unifying framework. Free convolution and MDEs govern their deterministic limits; isotropic and deformed local laws govern their local structure; DBM, resolvent comparison, and ribbon-graph methods govern universality and nonuniversality; and the deformation itself selects among sharply different edge, outlier, and eigenvector regimes.