🔬 condensed matter

Quenched complexity of marginal states in the Sherrington--Kirkpatrick spin glass

This paper resolves the long-standing problem of computing the quenched complexity of marginal metastable states in the Sherrington-Kirkpatrick spin glass model using a one-step replica symmetry breaking Ansatz, revealing that the complexity vanishes near the full-RSB equilibrium free energy and suggesting that the lowest marginal states correspond to the equilibrium states themselves.

Original authors: Tiziana De Chirico, Luca Leuzzi

Published 2026-09-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tiziana De Chirico, Luca Leuzzi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast, abstract landscape of modern physics, there is a field dedicated to understanding how disorder shapes reality. Imagine a system made of countless tiny magnets, each trying to align with its neighbors while being pulled in conflicting directions by random, invisible forces. This is the world of spin glasses, a class of materials that behave like frozen chaos. For decades, scientists have used these systems as a testing ground for understanding not just magnets, but also the complex networks inside our brains and the algorithms that power artificial intelligence. The central question in this field is about the "energy landscape" of these systems. Think of this landscape as a mountain range where the height represents energy. A system naturally wants to settle into the deepest valley, the state of lowest energy, which is its stable equilibrium. However, because the forces are random, the landscape is not smooth; it is riddled with millions of tiny valleys, ridges, and plateaus. These are the metastable states: places where a system can get stuck, thinking it has found the bottom, when it is actually just resting in a shallow dip. Counting how many of these dips exist and how deep they go is crucial, because it tells us whether an algorithm or a physical process can actually find the true bottom or if it will remain trapped in a false one.

For a specific, famous model of this disorder known as the Sherrington-Kirkpatrick model, a long-standing mystery has persisted for forty-five years. Scientists knew that the vast majority of these metastable states were "marginal." In plain terms, this means they are perched on the very edge of stability, like a ball balanced on a razor's edge. They are so delicate that standard mathematical tools, which work well for stable valleys, fail to count them. For forty years, researchers could only count these states using a rough approximation that worked well for high-energy, unstable regions but broke down completely when trying to find the deepest, most stable states. This approximation predicted that stable states could exist at energy levels lower than the true equilibrium of the system, a physical impossibility that suggested the math was missing something fundamental. It was as if a map showed a valley deeper than the ocean floor, signaling that the cartography itself was flawed.

In a recent study, researchers Tiziana de Chirico and Luca Leuzzi have finally solved this problem by computing the count of these marginal states using a more rigorous, "quenched" method. Instead of averaging the number of states across all possible random configurations, they calculated the count for a specific, fixed configuration, which is the correct way to describe real physical systems. They applied a sophisticated mathematical technique that breaks the symmetry of the problem into two groups, allowing them to track the delicate relationship between the magnetization of the states and their soft, unstable modes. The result is a new, accurate curve that maps out the complexity, or the number of available states, across different energy levels.

The findings reveal that the old, rough approximation was indeed wrong about the deepest states. The new calculation shows that the number of available states drops to zero at an energy level that is much higher than the flawed approximation suggested. In fact, the point where the count of states vanishes moves significantly closer to the true equilibrium energy of the system. This shift is profound: it implies that the lowest-energy marginal states are not separate, exotic entities, but are likely the equilibrium states themselves. The researchers found that as the energy decreases toward this limit, the angle between the magnetization of a state and its unstable direction grows until they are perfectly perpendicular. This geometric change suggests that the deepest states are indeed the true equilibrium, resolving a decades-old inconsistency where the math predicted states that could not physically exist.

The study also clarifies the structure of the energy landscape at low temperatures. As the system moves toward lower energies, the states do not remain isolated; they begin to cluster together in increasingly correlated groups. This clustering prefigures the complex, hierarchical organization that characterizes the equilibrium phase of the system. The researchers' work confirms that the "gap" between the unstable approximation and the true physical reality is not a feature of nature, but an artifact of the mathematical method used. By correcting the method, they have unified the counting of these excited states with the established laws of thermodynamics.

While the full mathematical picture requires even more complex layers of symmetry breaking to be perfectly precise, the current results are a major step forward. They show that the lowest-lying states are not a separate category of "marginal" objects, but are the very ground states the system seeks. This reconciliation means that the tools used to count these states can now be trusted to describe the true behavior of the system, from the highest energy peaks down to the deepest valleys. The study does not claim to have solved every aspect of the problem, particularly the full continuous hierarchy of states, but it firmly establishes that the previous predictions of unphysical, ultra-low energy states were incorrect. The landscape, it turns out, is consistent with the laws of physics, and the deepest valleys are exactly where the theory says they should be.

Technical Summary: Quenched Complexity of Marginal States in the Sherrington–Kirkpatrick Spin Glass

Problem Statement
The Sherrington–Kirkpatrick (SK) model exhibits an exponentially large number of metastable states below the spin-glass transition temperature. Unlike models with a one-step replica symmetry breaking (1RSB) equilibrium phase (e.g., the spherical $p$ -spin model), where metastable states are genuine minima of the Thouless–Anderson–Palmer (TAP) free energy, the overwhelming majority of TAP states in the SK model are marginal. These states possess a soft mode in the free-energy Hessian, rendering them unstable to the standard supersymmetric Becchi–Rouet–Stora–Tyutin (BRST) counting methods.

Historically, the complexity (configurational entropy) of these marginal states in the SK model has been computed only within the annealed approximation (averaging the number of states rather than the logarithm of the number). This annealed solution, derived via the Bray-Moore two-group replica Ansatz, is exact only in a limited free-energy window ( $f^* < f < f_{\max}$ ). Below the stability threshold $f^*$ , the annealed solution becomes unstable (indicated by a negative replicon eigenvalue) and predicts a zero-complexity energy $f_0^{\text{ann}}$ that lies strictly below the equilibrium free energy $f_0^{\text{eq}}$ of the Parisi full replica symmetry breaking (FRSB) solution. This unphysical prediction implies the existence of states with energies lower than the thermodynamic ground state. A quenched computation was previously attempted but remained within the BRST-symmetric framework, which is unstable at all free energy levels except the equilibrium point. Consequently, the quenched complexity of the dominant, marginal, BRST-breaking states has remained an open problem for forty-five years.

Methodology
The authors compute the quenched complexity of the SK model's marginal states using a two-group replica Ansatz combined with a one-step replica symmetry breaking (1RSB) structure in the space of the marginal states.

Two-Group Formalism: The approach utilizes the formalism introduced by Bray and Moore, where the replica overlap matrix $Q$ is partitioned into two groups. This structure describes the coalescence of a minimum and a rank-one saddle into a single marginal state. The order parameters include:
- $Q_{ab}$ : Overlap between local magnetizations of states $a$ and $b$ .
- $A_{ab}$ : Overlap between the magnetization of state $a$ and the soft-mode direction of state $b$ .
- $C_{ab}$ : Soft-mode–soft-mode overlap.
  The angle $\gamma$ between the magnetization and the soft mode is defined via $\cos \gamma = A_{aa} / \sqrt{Q_{aa}C_{aa}}$ .
Quenched 1RSB Computation: The authors introduce a 1RSB structure among the marginal states themselves. The matrices $Q, A, C$ are parameterized with diagonal values, inner-block values (size $m_1$ ), and outer values.
- In zero external field, the outer overlaps vanish ( $Q_0 = A_0 = C_0 = 0$ ).
- The saddle-point equations reveal that the inner-block parameters $A_1$ and $C_1$ are effectively zero ( $A_1 \simeq C_1 \simeq 0$ ) across all temperatures and $m$ values explored. This simplifies the problem significantly, implying that while magnetizations within a cluster are correlated ( $Q_1 \neq 0$ ), the soft-mode directions of different marginal states within the same cluster are uncorrelated.
- The complexity $\Sigma(f)$ is obtained via the Legendre transform of the generalized two-group free energy functional.
Zero-Temperature Limit: The scheme is analyzed in the $T \to 0$ limit (with fixed $y = \beta m$ ), collapsing the system into two closed equations for the parameters $k_A = \lim_{T\to 0} A/T$ and $l_{Q_1} = \lim_{T\to 0} Q_1$ , allowing direct evaluation of the ground-state complexity $\Sigma(e)$ .

Key Results
The paper presents the quenched complexity curve $\Sigma(f)$ for various temperatures ( $T=0, 0.05, 0.1, 0.2$ ) and compares it with the annealed approximation and the equilibrium FRSB solution.

Correction of the Zero-Complexity Point: The quenched 1RSB complexity coincides with the annealed curve at high free energies (where the annealed solution is stable). However, at the replicon instability point $f^*$ , the quenched curve branches off. Crucially, the quenched complexity vanishes at a free energy $f_0^{\text{1RSB}}$ that is significantly higher than the unphysical annealed value $f_0^{\text{ann}}$ .
- At $T=0$ , $f_0^{\text{ann}} \approx -0.79067$ , while the equilibrium FRSB energy is $f_0^{\text{eq}} \approx -0.76321$ .
- The 1RSB quenched result yields $f_0^{\text{1RSB}} \approx -0.76522$ , drastically reducing the gap between the zero-complexity point and the equilibrium energy.
Behavior of Observables:
- Overlap Correlation: As free energy decreases, the difference between diagonal and off-diagonal overlaps ( $Q - Q_1$ ) decreases, indicating that low-energy states reside in increasingly correlated clusters, prefiguring the continuous ultrametric organization of the equilibrium phase.
- Soft Mode Orientation: The angle $\gamma$ between the magnetization and the soft mode increases as $f \to f_0^{\text{1RSB}}$ , approaching $\pi/2$ . This suggests that in the lowest marginal states, the soft mode becomes orthogonal to the magnetization, a property expected for the exact equilibrium FRSB states.
Convergence Pattern: The results support a systematic shift where $f_0^{\text{ann}} < f_0^{\text{1RSB}} \lesssim f_0^{\text{eq}}$ . The residual mismatch between $f_0^{\text{1RSB}}$ and $f_0^{\text{eq}}$ is of the same relative order as the known error of the 1RSB approximation to the equilibrium free energy itself.

Significance and Claims
The paper claims to solve a forty-five-year-old open problem by providing the first quenched computation of the complexity for the dominant marginal states in the SK model. The primary significance lies in reconciling the counting of TAP excited states with Parisi thermodynamics:

Resolution of Unphysical Predictions: The work demonstrates that the unphysical prediction of states existing below the equilibrium free energy is an artifact of the annealed approximation. The quenched calculation pushes the zero-complexity point back toward the equilibrium energy.
Conjecture on FRSB Limit: The authors conjecture that in the limit of full replica symmetry breaking (FRSB) within the two-group theory, the complexity of marginal states will vanish exactly at the equilibrium free energy ( $f_0^{\text{FRSB}} = f_0^{\text{eq}}$ ). This implies that the lowest-lying marginal TAP states are, in fact, the equilibrium states of the Parisi solution, and there is no interval of free energies above $f_0^{\text{eq}}$ devoid of states.
Methodological Advancement: The paper establishes a numerically stable formulation for the inner-block saddle-point equations, which can be extended to external fields and models with mixed 1RSB–FRSB phases. It also provides a framework to compare with future quenched extensions of the superspace Kac–Rice formulation.

The authors remain modest regarding the 1RSB approximation, noting that it is an approximation to the stable 2G FRSB Ansatz and that the zero-complexity endpoint does not yet coincide exactly with the equilibrium solution, but the trend strongly supports the convergence to the exact result in the full-RSB limit.

Technical Summary: Quenched Complexity of Marginal States in the Sherrington–Kirkpatrick Spin Glass

More like this