Regularized Bulk Universality versus Bounded-Disorder Nonuniversality for Annealed Complexity of Spherical $p$-Spin Landscapes

Technical Summary: Regularized Bulk Universality Versus Bounded-Disorder Nonuniversality for Annealed Complexity of Spherical $p$-Spin Landscapes

1. Problem Statement
The paper investigates the universality of the annealed complexity (the exponential rate of the expected number of critical points) for the spherical $p$-spin Hamiltonian ($p \ge 3$) when the disorder tensor entries are non-Gaussian. Specifically, it addresses whether matching a finite number of moments of the disorder distribution to the standard Gaussian distribution is sufficient to guarantee that the annealed complexity matches the Gaussian limit.

The central tension identified is between two regimes:

1. Regularized Bulk Universality: The behavior of the system when the critical-point count is "regularized" (smoothed via mollifiers and restricted to "incoherent" frames where coordinates are delocalized).
2. Unregularized Nonuniversality: The behavior of the raw, unregularized count of critical points in a compact energy window, which may be dominated by rare, localized events (spikes or coherent blocks) that are invisible to finite moment matching.

2. Methodology
The author employs a combination of probabilistic invariance principles, large deviation theory, and geometric analysis of spherical landscapes.

• Multiplicative Lindeberg Principle: Unlike standard additive invariance principles, the paper develops a multiplicative comparison for the Kac–Rice partition function. This is necessary because the partition function itself is exponential in $N$. The method uses a product-mixture path to interpolate between non-Gaussian and Gaussian disorder, preserving the aggregate influence budget of the tensor coordinates.
• Incoherent Frames: The analysis restricts the Kac–Rice functional to frames where the basis vectors have small entries (incoherent frames). This ensures that no single tensor coordinate exerts a dominant influence on the gradient or Hessian, allowing for universality proofs under moment matching.
• De-regularization and Defects: The paper defines "de-regularization defects" to quantify the gap between the regularized (smoothed) pressure and the exact, unregularized critical-point count. These defects measure the failure of the regularized functional to capture the exact count as the smoothing parameters vanish.
• Counterexample Construction: To prove nonuniversality, the author constructs specific smooth, compactly supported probability laws that match Gaussian moments up to order $2p$ (or any finite order) but possess specific tail behaviors or "bump" structures.
• Localized Obstructions: The paper analyzes two mechanisms that break universality in the unregularized setting:
  • One-spike branch: A single large entry ($O(\sqrt{N})$) creating a local maximum.
  • Mesoscopic coherent block: A block of $O(N^{1/p})$ coordinates aligning to create a rank-one energy profile.

3. Key Contributions and Results

A. Regularized Bulk Universality (Unconditional)
The paper establishes that for the regularized Kac–Rice functional on incoherent frames, universality holds under relatively weak conditions:

• Moment Matching: If the disorder coordinates match the first $m$ Gaussian moments and satisfy a uniform subexponential tail bound, the regularized annealed pressure converges to the Gaussian limit for all $p \ge 3$ and $m \ge 2$.
• Ratio Convergence: If $(m-1)(p-2) > 2$, the ratio of the non-Gaussian to Gaussian expectations tends to 1. Notably, for $p \ge 5$, matching only the mean and variance ($m=2$) is sufficient for ratio universality.
• Robustness: These results hold for both ordered-tensor models and standard symmetric-coordinate models, and extend to Weibull-tail distributions under specific conditions.
• Variational Limit: The Gaussian limit is identified explicitly via a one-dimensional variational formula involving the semicircle law and the Kac–Rice complexity potential $\theta_p(u)$.

B. Unregularized Nonuniversality (Counterexamples)
The paper proves that finite moment matching is insufficient for the unregularized annealed count:

• Bounded-Disorder Coexistence: There exist symmetric, compactly supported, smooth laws matching Gaussian moments through order $2p$ (and any finite order) such that:
  • The regularized expectation ratio converges to 1 (universality holds for the regularized observable).
  • The unregularized expected critical-point count in a compact high-energy window has a lower exponential rate strictly greater than the Gaussian limit.
• Mechanisms of Failure:
  • One-spike: A single entry of size $a\sqrt{N}$ can create a local maximum with an exponential cost lower than the Gaussian cost if the tail of the disorder law is sufficiently heavy (e.g., wider Gaussian tails or specific bumps).
  • Coherent Block: A block of $O(N^{1/p})$ coordinates falling into a specific bounded interval can create a critical point with a speed-$N$ probability cost that exceeds the Gaussian rate. This obstruction persists even if the disorder is bounded (i.e., entries are uniformly bounded by $N^{1/2-\epsilon}$).
• Implication: No prescribed finite collection of moments forces the Gaussian energy-window rate for the spatially unrestricted, unregularized annealed count.

C. Conditional Reduction to Exact Universality
The paper does not claim exact universality for general non-Gaussian disorder. Instead, it provides a quantitative reduction:

• Defect Analysis: Exact universality is equivalent to the vanishing of two one-sided "de-regularization defects" (measuring the error in approximating the hard Kac–Rice slice by the regularized functional) and the control of a "localized complement" (critical points outside the incoherent region).
• Necessary and Sufficient Conditions: If the defects vanish, the exact complexity is determined by the maximum of the bulk pressure and the localized branch pressure. The paper shows that for the constructed counterexamples, the localized branch dominates, preventing universality.

4. Significance and Claims
The paper clarifies the distinct roles of moment matching in different regimes of random matrix and spin glass theory:

• Distinction from Thermodynamics: While finite $2p$ moments are known to be sharp for thermodynamic observables (free energy, ground state) in spherical models, this paper demonstrates they are not sufficient for annealed complexity. The complexity is sensitive to "mesoscopic" coherent deviations and quadratic-scale tails that do not affect the free energy.
• Regularization vs. Reality: The results highlight a dichotomy: the "bulk" of the landscape (delocalized, regularized) is universal, but the "edges" or rare localized events (unregularized) retain a dependence on the specific disorder law.
• Limitations: The author explicitly states that the paper does not prove a quenched universality theorem. The counterexamples rely on rare events (probability $e^{-O(N)}$) that dominate the annealed expectation but may not affect the typical (quenched) behavior. The paper also notes that the existence of a non-Gaussian limit for the annealed complexity is not asserted; only a strict separation of lower rates is proven.
• Open Problems: The paper identifies the control of the "mixed-profile" branch (points with both localized and delocalized mass) and the vanishing of de-regularization defects for general non-Gaussian disorder as open problems required for a full exact universality theorem.

In summary, the work establishes a rigorous separation between the universality of regularized bulk observables and the nonuniversality of unregularized critical-point counts, demonstrating that finite moment matching fails to control the annealed complexity due to localized, coherent obstructions.