Statistical Mechanics of Random Matrices

Abstract: These lecture notes are based on the lectures on \emph{Statistical Mechanics of Random Matrices} delivered at the Spring College on the Physics of Complex Systems, held at the Abdus Salam International Centre for Theoretical Physics, Trieste, Italy, from 19 February to 15 March 2024. Their aim is to present a statistical-mechanics route to the spectral theory of sparse and diluted random matrices, with emphasis on cavity and replica methods, resolvent techniques, population dynamics, typical spectral densities, spectral-count fluctuations and large deviations, conditioned spectra, and non-Hermitian extensions. The written form of the notes has been deliberately expanded beyond the material actually covered during the lectures. This is partly because a set of lecture notes can afford a more systematic development than a sequence of blackboard lectures, and partly because several natural continuations of the material become clearer once the central methods have been introduced. Consequently, not every topic discussed here was presented during the College. The additional material is included to give a more coherent account of the subject and to indicate directions that, hopefully, can be covered in greater detail in future lectures or schools. Since these are lecture notes rather than a state-of-the-art review, the choice of topics is necessarily selective and is naturally tilted towards the author's own work and collaborations on this subject. I have nevertheless tried, within the limits of this format, to place the material in contact with the broader literature and to represent the surrounding state of the art as fairly as possible. Inevitably, some relevant contributions may be missing or treated too briefly; such omissions are unintentional and reflect the pedagogical scope of the notes rather than a judgement on their importance.

• The paper demonstrates how replica and cavity methods compute spectral properties of sparse random matrices with graph-based structures.
• It details algorithmic strategies, including belief propagation and population dynamics, to efficiently approximate spectral densities in locally tree-like graphs.
• The analysis shows that topology and local structure, beyond average degree, critically determine spectral features and break universality.

Statistical Mechanics of Random Matrices: An Expert Analysis

Introduction and Scope

"Statistical Mechanics of Random Matrices" (2606.08706) delivers a comprehensive treatise on the spectral theory of sparse and diluted random matrices from the vantage of statistical mechanics. The lecture notes emphasize theoretical tools—replica and cavity methods, resolvent techniques, and population dynamics—applied to ensembles with local tree-like structure, highlighting sparse symmetry, covariance, non-Hermiticity, large deviations, and conditioning. Rather than a broad review, the focus is on the interplay between graphical/disorder structure and spectral properties, especially as seen through the replica and cavity approaches often employed in disordered systems.

Random Matrix Ensembles and Graph-Theoretic Structure

The notes foreground a dichotomy between dense random matrices (Wigner/Dyson/Marčenko–Pastur universality) and truly sparse matrices, in which the number of nonzero entries grows linearly with size. In the latter, matrices are naturally associated with graphs (undirected, bipartite, directed), and the adjacency/connectivity structure leaves a lasting imprint on spectral observables. Key ensembles discussed include:

• Erdős–Rényi (Poissonian) graphs: Edge probability $p = c/N$ yields Poisson degree statistics, local tree-likeness, and strong sensitivity of spectral density to degree and motif fluctuations.
• Random regular graphs: All degrees fixed ($c$-regular), leading to the Kesten–McKay law for the limiting spectral density.
• Configuration models and correlated/topological ensembles: Prescribed degree sequences, degree–degree correlations, and block/community structure generalize the underlying graphical support.
• Bipartite and non-Hermitian/directed extensions: Rectangular supports for (diluted) Wishart-type covariance matrices and directed graphs for non-Hermitian generalizations.

The lecture notes demonstrate in detail how these structural ingredients control resolvent recursions, manifestly distinguishing sparse from dense spectral properties.

Spectral Observables and Gaussian Integral Representations

Spectral density is treated through resolvents, empirical measures, and associated integral transforms. For Hermitian matrices:

$$\rho_N(\lambda) = \frac{1}{N}\sum_j \delta(\lambda - \lambda_j)$$

with the resolvent (Green function):

$g_A(z) = \frac{1}{N}\text{Tr}[(zI - A)^{-1}]$

In sparse regimes, this is elevated to include local diagonal entries, since non-uniform graph topology causes $G_{ii}(z)$ to vary with the vertex's neighborhood. Non-Hermitian spectra are handled via Girko's Hermitization, expressing the density through Laplacians of logarithmic potentials on the complex plane.

The Edwards–Jones representation is central, allowing resolvents and related quantities to be expressed as logarithmic derivatives of disorder-dependent Gaussian partition functions. The critical technical challenge is the necessity of disorder averaging logarithms (quenched averages), leading to the application of the replica and cavity methods developed in spin glass theory.

The Replica and Cavity Framework

A principal contribution of these notes is laying out, in depth, the theoretical apparatus of replicas and cavity methods as applied to sparse random matrices. The methods proceed as follows:

• Replica method: Discrete copies (replicas) circumvent the calculation of averages of logarithms via analytic continuation $n\to 0$ of integer moments, producing functional order parameters when finite connectivity or graphical randomness persists.
• Cavity method: Direct exploitation of local tree-likeness. Recursions for local resolvent quantities (e.g., "messages" $G_{i \to j}$ or inverse variances $\omega_{i \to j}$) can be written exactly on trees and become asymptotically exact for locally tree-like (sparse) graphs.

An explicit connection is established: at the replica-symmetric level (ignoring RSB effects), both approaches deliver the same distributional self-consistency equations for local spectral observables.

Belief Propagation, Population Dynamics, and Algorithmic Realization

Single-instance (finite $N$) spectral densities can be efficiently approximated using belief propagation (BP), which iteratively solves for message passing recursions for local Green functions on the specific graph. For ensemble observables (thermodynamic limit), these recursions are cast as fixed-point equations for the distribution of local messages—solved numerically via population dynamics, a type of functional Monte Carlo.

Particularly strong algorithmic clarity is provided through:

• Detailed schematic for both fixed-instance BP and ensemble-level population dynamics, highlighting the difference between the site and cavity message laws.
• Demonstration that for sufficiently symmetric cases (e.g., regular graphs), these rigorous procedures reduce to classical algebraic spectral equations (e.g., the Kesten–McKay law).

Major Results: Sparse Spectrum, Topology, and Universality

Sparse Symmetric Case

In symmetric, locally tree-like ensembles, the cavity equations encode the distribution of local Green functions, connecting to the global density through an integral (population) over this distribution. This strictly generalizes the self-averaging of the dense limit (Wigner semicircle law), and instead reveals spectrum dependence on degree fluctuations, motif abundances (e.g., triangles), and graph inhomogeneities.

Key findings:

• Kesten–McKay law’s emergence as the spectral density for $c$-regular graphs.
• Spectral atoms and extended singular features: Isolated nodes/finite clusters produce discrete spectral weight; high-degree vertices produce localized outliers—absent from mean-field (effective-medium) or dense limits.
• Failure of universality: There is no single limiting law; spectral densities encode topological statistics beyond average degree.

Extensions: Correlations and Constraints

Imposing nontrivial topological constraints (degree correlations, community structure, generalized degrees) changes the form of the distributional equations for the local resolvents. The notes develop a hierarchical formalism, treating such cases as resulting in vector- or type-conditional message laws, and showing the explicit impact on spectrum (e.g., block models give rise to isolated eigenvalues corresponding to community structure, while heavy-tailed degrees manifest as spectral tails).

Covariance/Wishart and Bipartite Models

For sparse covariance (Wishart-type) matrices, a bipartite cavity formalism is derived, naturally leveraging the rectangular graph support of the data matrix $c$0. The resolvent equations again reduce to recursive equations on the bipartite graph, with population dynamics employing two coupled populations: variable-to-factor and factor-to-variable messages. This formalism elegantly reproduces the Marčenko–Pastur law as the dense limit, but at finite dilution, zero modes and non-smooth features result from graph-induced rank deficiency and isolated component structure.

Non-Hermitian and Generalized Ensembles

The treatment of non-Hermitian sparse random matrices (e.g., directed graphs) is handled using Hermitization, leading to matrix-valued messages in cavity recursions. The paper references an extensive literature extending these approaches to products of matrices, cross-correlation ensembles, and even large deviations (rare events in spectral counting), demonstrating the method's generality.

Strong Numerical and Conceptual Claims

• Non-universality of sparse spectra: Unlike dense invariant ensembles, spectral densities in sparse graphs are not determined by mean degree alone, but are directly and nontrivially a function of entire local structure (degree distribution, motifs, correlations).
• Direct computation of typical and atypical observables: The methods described permit the calculation of not only mean densities, but spectral-count large deviations, conditioned spectral densities, and fluctuations in both Hermitian and non-Hermitian settings.
• Consistency with rigorous results: While relying on physics-based methods, several results (e.g., limiting densities, localized outlier phenomena) have been corroborated by rigorous probabilistic techniques.

Implications and Future Directions

Practical: These methods underpin an accurate understanding of spectral behavior in sparse models for complex systems—ranging from network science (community detection and detectability thresholds), machine learning (spectra of sparse embeddings/covariances), to statistical physics (localization on Bethe lattices/graphs).

Theoretical: The lecture notes suggest future work in:

• Extending analytic control to ensembles with higher-order motif constraints, short loops, or more general distributions on weights and entries.
• Exploring further the large-deviation and conditioned-spectra regime, potentially linking to questions of outlier eigenvalues, localization transitions, and spectral statistics beyond the mean.
• Deepening the connection between replica-symmetry-breaking transitions and breaking of local spectral uniformity in highly inhomogeneous or constrained ensembles.

Conclusion

"Statistical Mechanics of Random Matrices" (2606.08706) provides a detailed and authoritative account of how statistical-mechanical tools compute spectral properties of sparse and structurally complex random matrix ensembles. Through rigorous derivation, explicit algorithmic procedure, and careful connection to both classical and modern results, the notes show that—beyond the universality of dense invariant models—sparse matrices exhibit a rich, topology-dependent spectral behavior accessible only via a local, message-passing (cavity/replica/stat-known) perspective. These frameworks are poised to remain central as applications and theory of sparse random operators continue to expand in physics, data science, and network theory.