Brown Measure: Non-Normal Spectral Analysis

• Brown measure is the canonical probability measure on ℂ associated with non-normal operators, linking logarithmic potential with spectral distributions.
• It provides a noncommutative analogue of eigenvalue distributions, coinciding with empirical spectra for finite matrices and usual spectral measures for normal operators.
• Analytic techniques like Hermitization, operator-valued subordination, and PDE methods are central to computing Brown measures in free probability and random matrix theory.

Searching arXiv for recent and foundational Brown measure papers to ground the article. Brown measure is the canonical probability measure on $\mathbb C$ associated with a possibly non-normal operator in a tracial von Neumann algebra. It is defined through the logarithmic potential induced by the Fuglede–Kadison determinant and serves as the noncommutative analogue of eigenvalue distribution. For finite matrices it coincides with the empirical spectral distribution, while for normal operators it agrees with the usual spectral measure. In free probability and non-Hermitian random matrix theory, Brown measure is the natural object linking operator-theoretic limits to limiting eigenvalue distributions of matrix ensembles (Adhikari et al., 2017).

1. Definition and basic characterization

Let $(\mathcal A,\varphi)$ be a tracial non-commutative probability space, equivalently a tracial von Neumann algebra with faithful normal trace $\varphi$. For $a\in\mathcal A$, the Fuglede–Kadison determinant is defined for invertible $a$ by

$$\Delta(a)=\exp\Big(\tfrac12\,\varphi(\log(aa^*))\Big),$$

and extended by regularization through

$$\Delta(a):=\lim_{\varepsilon\downarrow 0}\Delta_\varepsilon(a),\qquad \Delta_\varepsilon(a):=\exp\Big(\tfrac12\,\varphi(\log(aa^*+\varepsilon^2))\Big).$$

The logarithmic potential of $a$ is then

$$U_a(\lambda):=\log\Delta(a-\lambda),\qquad \lambda\in\mathbb C.$$

This function is subharmonic on $\mathbb C$, and the Brown measure $(\mathcal A,\varphi)$0 is defined by its distributional Laplacian: $(\mathcal A,\varphi)$1 Equivalently,

$(\mathcal A,\varphi)$2

This is the potential-theoretic formulation that underlies most analytic and random-matrix applications (Adhikari et al., 2017).

A closely related regularized form uses

$(\mathcal A,\varphi)$3

and

$(\mathcal A,\varphi)$4

with $(\mathcal A,\varphi)$5. This formulation is standard in PDE- and Hermitization-based computations (Ho, 2020).

For finite-dimensional matrices the construction reduces exactly to eigenvalue counting. If $(\mathcal A,\varphi)$6 is an $(\mathcal A,\varphi)$7 matrix, then

$(\mathcal A,\varphi)$8

Thus Brown measure is literally the empirical spectral distribution in finite dimensions (Adhikari et al., 2017). This is why Brown measure is the natural candidate for the large-$(\mathcal A,\varphi)$9 eigenvalue limit of non-Hermitian random matrices converging in $\varphi$0-distribution to an operator in a tracial von Neumann algebra (Driver et al., 2019).

2. Relation to spectral measure and non-normality

If $\varphi$1 is normal, the Brown measure agrees with the usual spectral measure computed by functional calculus (Adhikari et al., 2017). The distinction becomes substantive only for non-normal operators, where there is no direct scalar-valued spectral theorem on $\varphi$2 of the same form. Brown measure replaces spectral projection data by a logarithmic-potential description based on $\varphi$3 (Zhou, 2024).

Its support is contained in the spectrum, but it can be strictly smaller than the spectrum (Ho, 2020). This gap is one of the characteristic features of non-normal operator theory. A plausible implication is that Brown measure captures the “eigenvalue-distribution aspect” of a non-normal operator without encoding the full pseudospectral or invariant-subspace geometry.

The Brown measure is also the correct object in many asymptotic matrix problems, but not in all $\varphi$4-distribution limits automatically. The literature emphasizes that convergence in $\varphi$5-moments alone does not force convergence of empirical eigenvalue distributions to Brown measure; appropriate regularity or perturbative hypotheses are needed (Ho, 2020). This point is central in later work on polynomial images of Ginibre matrices, where least singular value estimates are used precisely to upgrade $\varphi$6-moment convergence to Brown-measure convergence (Han, 1 Jun 2026).

3. Analytic machinery: Hermitization, regularization, and subordination

A standard route to Brown measure is Hermitization. For non-normal $\varphi$7, one studies $\varphi$8 rather than $\varphi$9 directly, often through the regularized potential

$a\in\mathcal A$0

The Brown measure is then recovered by taking the Laplacian in $a\in\mathcal A$1 and sending $a\in\mathcal A$2 (Driver et al., 2019).

A rigorous operator-valued version of Hermitian reduction embeds $a\in\mathcal A$3 into

$a\in\mathcal A$4

so that the scalar regularized Cauchy transform of $a\in\mathcal A$5 becomes an entry of an operator-valued Cauchy transform of a selfadjoint block operator. This makes available operator-valued free convolution, linearization, and subordination techniques (Belinschi et al., 2015). That framework is particularly effective for polynomials in free variables and underlies algorithmic computations of Brown measures for non-Hermitian polynomial models (Belinschi et al., 2015).

For sums of free variables, especially when one summand is $a\in\mathcal A$6-diagonal, the problem can be reduced to selfadjoint free additive convolution of symmetrized singular-value laws. The resulting subordination functions yield explicit logarithmic potentials and, on suitable open sets, real-analytic Brown measure densities (Bercovici et al., 2022). This is a general structural mechanism: singular values are easier to access than eigenvalues, and Brown measure is recovered from their logarithmic data.

PDE and Hamilton–Jacobi methods form another major line. For free multiplicative Brownian motion and related diffusions, one derives nonlinear first-order PDEs for the regularized logarithmic potential, solves them by characteristics, and identifies the Brown measure from the limiting Laplacian (Driver et al., 2019). This approach has since been adapted to other non-normal free diffusions and deformations (2002.04585).

4. Core explicit examples

Several families of operators have explicit Brown measures and serve as templates for the general theory.

For a circular element, the Brown measure is the circular law, uniform on a disk (Driver et al., 2019). For an elliptic element, the Brown measure is the elliptic law, uniform on an ellipse, matching the limiting spectral distribution of elliptic random matrices (Adhikari et al., 2017). These are the foundational non-normal analogues of semicircular and Wigner-type laws.

For $a\in\mathcal A$7-diagonal elements, the Brown measure is rotationally invariant and is determined by the distribution of $a\in\mathcal A$8 through the $a\in\mathcal A$9-transform. This is the Haagerup–Larsen paradigm revisited in operator-valued Hermitization language (Belinschi et al., 2015). The support is typically a disk or annulus, and the radial cumulative distribution is expressed by the inverse $a$0-transform (Bercovici et al., 2022).

For the free multiplicative Brownian motion $a$1, the Brown measure is supported on the closure of a domain $a$2, with a continuous density $a$3 on $a$4, strictly positive and real analytic on $a$5. In polar coordinates,

$a$6

where $a$7 is determined by the geometry of $a$8 (Driver et al., 2019). The support theorem for $a$9 was first established separately, together with the general inclusion of Brown measure support in the $$\Delta(a)=\exp\Big(\tfrac12\,\varphi(\log(aa^*))\Big),$$0-spectrum (Hall et al., 2018).

For sums of a self-adjoint element and a circular or elliptic perturbation, one gets bounded planar domains with densities constant along vertical lines. In the case $$\Delta(a)=\exp\Big(\tfrac12\,\varphi(\log(aa^*))\Big),$$1, the Brown measure is supported on

$$\Delta(a)=\exp\Big(\tfrac12\,\varphi(\log(aa^*))\Big),$$2

and has an explicit absolutely continuous density depending only on the real coordinate $$\Delta(a)=\exp\Big(\tfrac12\,\varphi(\log(aa^*))\Big),$$3 (Ho, 2020). The case $$\Delta(a)=\exp\Big(\tfrac12\,\varphi(\log(aa^*))\Big),$$4 exhibits the same vertical-constancy phenomenon and a support region $$\Delta(a)=\exp\Big(\tfrac12\,\varphi(\log(aa^*))\Big),$$5 obtained by a Hamilton–Jacobi analysis (Hall et al., 2020).

For atomic selfadjoint data, new geometric phenomena appear. The Brown measure of $$\Delta(a)=\exp\Big(\tfrac12\,\varphi(\log(aa^*))\Big),$$6, with $$\Delta(a)=\exp\Big(\tfrac12\,\varphi(\log(aa^*))\Big),$$7 free projections or more generally free finitely atomic selfadjoints, is supported on algebraic curves; in the two-atom case it lies on hyperbolas inside a rectangle, with explicit atomic and continuous parts (Zhou, 2024). A related quaternionic Green’s function analysis shows that, under the corresponding heuristic, the boundary is algebraic in the finitely atomic case (Zhou, 2024).

5. Brown measure and random matrix limits

Brown measure is central to the free-probabilistic interpretation of non-Hermitian random matrix asymptotics. In the finite-dimensional setting it is exactly the empirical eigenvalue distribution, so the question is when the empirical eigenvalue law of a random matrix model converges to the Brown measure of its free limit.

A major example is the product of independent elliptic matrices. Independent elliptic matrices converge in $$\Delta(a)=\exp\Big(\tfrac12\,\varphi(\log(aa^*))\Big),$$8-distribution to freely independent elliptic elements, and for $$\Delta(a)=\exp\Big(\tfrac12\,\varphi(\log(aa^*))\Big),$$9 the Brown measure of the product $$\Delta(a):=\lim_{\varepsilon\downarrow 0}\Delta_\varepsilon(a),\qquad \Delta_\varepsilon(a):=\exp\Big(\tfrac12\,\varphi(\log(aa^*+\varepsilon^2))\Big).$$0 is rotationally invariant with radial distribution

$$\Delta(a):=\lim_{\varepsilon\downarrow 0}\Delta_\varepsilon(a),\qquad \Delta_\varepsilon(a):=\exp\Big(\tfrac12\,\varphi(\log(aa^*+\varepsilon^2))\Big).$$1

This coincides with the limiting spectral distribution of the corresponding random matrix product (Adhikari et al., 2017). The structural reason is that products of free elliptic elements are $$\Delta(a):=\lim_{\varepsilon\downarrow 0}\Delta_\varepsilon(a),\qquad \Delta_\varepsilon(a):=\exp\Big(\tfrac12\,\varphi(\log(aa^*+\varepsilon^2))\Big).$$2-diagonal when $$\Delta(a):=\lim_{\varepsilon\downarrow 0}\Delta_\varepsilon(a),\qquad \Delta_\varepsilon(a):=\exp\Big(\tfrac12\,\varphi(\log(aa^*+\varepsilon^2))\Big).$$3, and their $$\Delta(a):=\lim_{\varepsilon\downarrow 0}\Delta_\varepsilon(a),\qquad \Delta_\varepsilon(a):=\exp\Big(\tfrac12\,\varphi(\log(aa^*+\varepsilon^2))\Big).$$4-distribution matches that of products of circular elements (Adhikari et al., 2017).

Another prominent example is the free multiplicative Brownian motion, the $$\Delta(a):=\lim_{\varepsilon\downarrow 0}\Delta_\varepsilon(a),\qquad \Delta_\varepsilon(a):=\exp\Big(\tfrac12\,\varphi(\log(aa^*+\varepsilon^2))\Big).$$5-distribution limit of Brownian motion on $$\Delta(a):=\lim_{\varepsilon\downarrow 0}\Delta_\varepsilon(a),\qquad \Delta_\varepsilon(a):=\exp\Big(\tfrac12\,\varphi(\log(aa^*+\varepsilon^2))\Big).$$6. Its Brown measure is the natural large-$$\Delta(a):=\lim_{\varepsilon\downarrow 0}\Delta_\varepsilon(a),\qquad \Delta_\varepsilon(a):=\exp\Big(\tfrac12\,\varphi(\log(aa^*+\varepsilon^2))\Big).$$7 candidate for the empirical eigenvalue distribution of the matrix diffusion, and the support/density computations provide a detailed prediction for the macroscopic spectrum (Driver et al., 2019).

The most general current convergence result in the provided literature concerns polynomials in Ginibre matrices. For any non-commutative polynomial $$\Delta(a):=\lim_{\varepsilon\downarrow 0}\Delta_\varepsilon(a),\qquad \Delta_\varepsilon(a):=\exp\Big(\tfrac12\,\varphi(\log(aa^*+\varepsilon^2))\Big).$$8, if $$\Delta(a):=\lim_{\varepsilon\downarrow 0}\Delta_\varepsilon(a),\qquad \Delta_\varepsilon(a):=\exp\Big(\tfrac12\,\varphi(\log(aa^*+\varepsilon^2))\Big).$$9 with independent Ginibre matrices, then the empirical spectral distribution of $a$0 converges weakly in probability to the Brown measure of $a$1, where $a$2 are free circular variables (Han, 1 Jun 2026). The proof relies on least singular value lower bounds for $a$3, showing that Brown measure is not merely a heuristic candidate but the actual limiting spectral law in this full polynomial Ginibre setting (Han, 1 Jun 2026).

The same Brown measure limit persists beyond Gaussian entries under mean-zero, variance-one, bounded-density, finite-moment assumptions (Han, 1 Jun 2026). This suggests a robust universality principle for non-Hermitian polynomial ensembles.

6. Structural themes, edge behavior, and current directions

One recurring theme is that Brown measure often depends on less microscopic information than the full operator model. For products of free elliptic elements, the Brown measure depends only on the number of factors $a$4, not on the individual elliptic parameters $a$5 (Adhikari et al., 2017). This suggests a coarse universality phenomenon in multiplicative free models.

Another theme is the geometry of support and edge singularities. For free circular Brownian motion $a$6, under the condition

$a$7

the Brown measure density near the boundary of its support has exactly two possible local behaviors: either a sharp cut, meaning a jump discontinuity across the boundary, or quadratic decay at critical boundary points where $a$8 (Erdős et al., 2023). This is the non-Hermitian analogue of square-root edges and cubic cusps in free semicircular Brownian motion (Erdős et al., 2023).

Support questions remain an active area. For products such as $a$9, with $$U_a(\lambda):=\log\Delta(a-\lambda),\qquad \lambda\in\mathbb C.$$0 free unitary Brownian motion and $$U_a(\lambda):=\log\Delta(a-\lambda),\qquad \lambda\in\mathbb C.$$1 a free projection, the support can be described explicitly via Hamiltonian characteristics and analytic maps $$U_a(\lambda):=\log\Delta(a-\lambda),\qquad \lambda\in\mathbb C.$$2, but the full density is not yet given in that work (2002.04585). Likewise, for $$U_a(\lambda):=\log\Delta(a-\lambda),\qquad \lambda\in\mathbb C.$$3, where $$U_a(\lambda):=\log\Delta(a-\lambda),\qquad \lambda\in\mathbb C.$$4 is non-negative and free from a free multiplicative Brownian motion $$U_a(\lambda):=\log\Delta(a-\lambda),\qquad \lambda\in\mathbb C.$$5, the support is identified through domains $$U_a(\lambda):=\log\Delta(a-\lambda),\qquad \lambda\in\mathbb C.$$6 and $$U_a(\lambda):=\log\Delta(a-\lambda),\qquad \lambda\in\mathbb C.$$7, while the full Brown measure remains open (Hall et al., 25 Mar 2025).

Recent work also continues to develop computational methods for sums of non-normal free variables. When one summand is $$U_a(\lambda):=\log\Delta(a-\lambda),\qquad \lambda\in\mathbb C.$$8-diagonal, subordination methods yield a real-analytic density on an open set $$U_a(\lambda):=\log\Delta(a-\lambda),\qquad \lambda\in\mathbb C.$$9 whose closure contains the support (Bercovici et al., 2022). For $\mathbb C$0 with $\mathbb C$1 selfadjoint and $\mathbb C$2 free Poisson, matrix-valued subordination with an explicit left inverse gives a reparametrized density formula for the absolutely continuous part of the Brown measure on a domain $\mathbb C$3 (Lehner et al., 29 Dec 2025). This suggests that reparametrization by subordination coordinates is a productive general technique.

A recurring misconception is that Brown measure should always be rotationally invariant or always coincide with random matrix limits once $\mathbb C$4-distribution convergence is known. The first is false except in structured classes such as $\mathbb C$5-diagonal operators; the second requires extra control, typically through singular-value estimates or explicit model structure (Adhikari et al., 2017, Han, 1 Jun 2026). Another misconception is that support should determine density. The atomic-operator examples show that boundaries may be algebraic while the measure can mix atoms and continuous parts in a delicate way (Zhou, 2024, Zhou, 2024).

In summary, Brown measure occupies the role that spectral measure cannot play for non-normal operators: it is the canonical planar distribution extracted from the logarithmic potential $\mathbb C$6, compatible with finite matrices, with normal spectral theory, and with a large and expanding class of random matrix limits. Its modern theory combines potential theory, free convolution, operator-valued subordination, Hermitization, Hamilton–Jacobi PDE, and singular-value analysis, and it has become a central organizing concept in non-Hermitian free probability and random matrix theory (Belinschi et al., 2015, Driver et al., 2019, Han, 1 Jun 2026).