Abstract: We compute the correlation functions of the eigenvalues in the Gaussian unitary ensemble using the fermionic replica method. We show that non-trivial saddle points, which break replica symmetry, must be included in the calculation in order to reproduce correctly the exact results for the correlation functions at large distance.
Introduction
For over fifteen years, a significant divide existed in the field of Random Matrix Theory (RMT). While RMT is indispensable for describing the energy levels of complex nuclei and the transport properties of disordered metallic grains, the mathematical tools used to analyze it were seen as unequal. The Supersymmetry (SUSY) method was widely regarded as the gold standard for its mathematical elegance and accuracy in capturing fine-grained statistics. Conversely, the replica method—a technique involving the creation of $n$ copies of a system and taking the limit as $n \to 0$—was largely considered a failure for these specific applications.
Previous attempts to apply the replica method to the Gaussian Unitary Ensemble (GUE) failed to capture the characteristic oscillatory "level-spacing" behavior, known as Wigner-Dyson statistics. This led many researchers to believe the replica method was fundamentally "ill" or inapplicable to RMT. In this work, we demonstrate that the replica method is not broken; rather, it was previously incomplete. By accounting for non-trivial saddle points that break replica symmetry, we reconcile the replica formalism with known exact results.
Methodology
We employ the fermionic replica method to calculate the correlation functions of eigenvalues within the GUE. The derivation follows a field-theoretic approach, structured into three primary stages:
- Mapping to a $\sigma$-model: The random matrix problem is first transformed into a field-theoretic "sigma model" via a Hubbard-Stratonovich transformation. This allows us to treat the matrix elements as fields.
- Itzykson-Zuber Integration: To simplify the resulting functional integrals, we perform an exact integration over the "angular" degrees of freedom (rotations). This step isolates the eigenvalues of the field as the primary variables of interest.
- Saddle Point Analysis: This represents the core innovation of our study. Traditional replica calculations typically only consider the "replica symmetric" saddle point, where all $n$ replicas are treated identically. We systematically explore non-trivial saddle points—configurations where the symmetry between replicas is spontaneously broken.
Key Insights: The Necessity of Symmetry Breaking
Our analysis reveals that the perceived failure of the replica method was rooted in the assumption of replica symmetry. We find that for the two-point correlation function, non-trivial saddle points contribute to the leading-order result.
Vector Replica Symmetry Breaking (RSB)
The pattern of symmetry breaking identified here is classified as "vector" RSB. This is notably simpler than the hierarchical RSB structures found in spin glass theory, yet it is mathematically essential. It arises from the existence of multiple distinct "ground states" within the effective action of the sigma model.
Oscillatory Corrections
While the trivial (replica symmetric) saddle point successfully yields the famous Wigner semi-circle law for the average density of states (DOS), it cannot account for the fine-grained oscillations. We demonstrate that a specific non-trivial saddle point (denoted as $p=1$) is responsible for the $1/N$ oscillatory corrections that characterize the local fluctuations of the eigenvalues.
Results and Comparison
By incorporating these non-trivial saddle points, we successfully derive the exact GUE two-point correlation function:
$$S_2(\omega) = 1 - \frac{\sin^2 \omega}{\omega^2}$$
This result is expressed as the sum of the perturbative (replica symmetric) and non-trivial saddle point contributions. Furthermore, our method derives the average Density of States, including the $1/N$ terms that match exact results from the method of orthogonal polynomials.
This derivation provides a unified physical picture by identifying the replica analogs of the "Andreev-Altshuler" saddle points used in the SUSY method. It also justifies the "universality" of these statistics for large matrices, as the final results depend on the underlying saddle point structure rather than the specific distribution of the matrix elements.
Discussion and Future Directions
While this work resolves the long-standing discrepancy between the replica and SUSY methods for the GUE, several avenues for future research remain. The limit $n \to 0$ (the "replica limit") currently relies on heuristic arguments regarding the summation of series; a more rigorous mathematical proof of this analytic continuation is needed.
Additionally, while we focused on the Unitary ensemble, the approach should be extended to the Orthogonal (GOE) and Symplectic (GSE) ensembles. Perhaps most significantly, because the replica method can handle non-quadratic actions more flexibly than SUSY, this understanding of RSB saddle points may prove to be a powerful tool for investigating disordered systems with interacting electrons, where traditional methods often struggle.
Summary updated