Typical number and stability of equilibria in large non-reciprocal complex systems

arXiv:2611.08324 · cond-mat.stat-mech, cond-mat.dis-nn, math-ph, math.MP · Submitted 2026-11-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.

Kai: We're looking at the paper: "Typical number and stability of equilibria in large non-reciprocal complex systems".

Mira: Here's the gist of the paper. Typical number and stability of equilibria in large non-reciprocal complex systems is computed by comparing quenched complexity with annealed complexity to show that they only coincide in the fully non-reciprocal limit The quenched calculation reveals a dependence on the statistics of the random force field and the degree of non-reciprocity, showing that average counting correctly locates the onset of an exponential number of equilibria but provides a biased description of their population Model Setup and Complexity Measures The study considers a prototypical model with N ≫ 1 coupled, autonomous ordinary differential equations subject to two competing terms: one favoring relaxation to a trivial equilibrium and another taking the form of a Gaussian force field promoting local equilibria The system dynamics are described by the equation ∂tx = f(x), fi(x) = −µixi + Fi(x), i = 1, · · ·, N, (1) where F(x) is a non-gradient field The complexity is quantified by the exponential growth rate of the typical number of equilibria, resolved by stability index, compared to the annealed complexity which is the exponential growth rate of the average number of equilibria Comparison between Typical and Average Counts The comparison shows that "typical realizations contain exponentially fewer equilibria than predicted by the average count" for any finite degree of reciprocity, and "Increasing reciprocity enhances the discrepancy between the two" The quenched complexity depends explicitly on the covariance of the random force field and on the degree of non-reciprocity, indicating that universality observed at the annealed level breaks down The average counting correctly identifies the location of the topology trivialization transition but fails to capture "the typical organization of equilibria near this transition, nor the stability index of the dominant equilibria" Quenched Complexity and Stability Index The total quenched complexity is given by Ξtot = Σtot − τ2(γ − 1)2v(γ)γ

1 + τv(γ): for γ ≥ 1 The quenched calculation uses the replicated Kac-Rice formalism to compute the exponential growth rate of the number of fixed points in typical realizations This complexity is related to the annealed complexity by Ξtot = Σtot − τ2(γ − 1)2v(γ)γ

1 + τv(γ): Role of Non-Reciprocity and Rare Events The discrepancy between quenched and annealed counting is driven by rare disorder realizations which bias the ensemble-average counting towards less stable equilibria, meaning "the annealed complexity is maximized by less stable equilibria, namely by fixed points with a larger index κ" The mechanism for this bias relies on the presence of a confining potential and translationally invariant statistics of random forces Non-reciprocity suppresses rare-sample bias, while reciprocity enhances it, leading to the conclusion that "the discrepancy between annealed and quenched counting grows with τ" Stability of the Replica Symmetric Assumption The stability of the replica symmetric assumption is checked by examining the replicon eigenvalue Λr, which is negative throughout the tested phase-space where complexity is positive, indicating that "the RS ansatz is stable with respect to continuous RSB transition" The analysis shows that "the only instability of the RS ansatz arises from the replicon sector" The work suggests that the stable fixed points dominating typical realizations are more stable than those dominating the annealed count, and this difference is controlled by R(2∆0) = cτv(γ) A central next step is to connect these fixed-point statistics with the out-of-equilibrium dynamics of the system to determine if these exponentially numerous stable fixed points can qualitatively modify the dynamics, such as suppressing chaos The results provide a framework for analyzing out-of-equilibrium dynamics in high-dimensional systems made of many coexisting and interacting units Supplementary Information Details The supplementary information provides detailed derivations, such as the expression for the nth factorial moment of the number of equilibria with k unstable directions using a saddle-point method The final expression for the complexity is recovered by optimizing over κ (or equivalently z(κ)) to find the index κq with the largest typical number of equilibria Numerical simulations confirm that the replicon is negative throughout the full tested phase-space, supporting the stability of the RS ansatz The analysis shows that for τ = 0, quenched and annealed counting coincide because "the bias ¯ξ vanishes" The final expression for the total complexity is found by optimizing over κ (or equivalently z(κ)) to yield z(κ∗) =

1+τ γv(γ): /

√γ(1+τv(γ)): The variance of the total number of equilibria, Vtot, is shown to be positive for any γ > 1 and τ > 0 Mechanism Favoring Unstable Fixed Points in Annealed Counting The mechanism ensuring that the annealed calculation overestimates not only the total number of equilibria but also their instability is generic, relying on a confining potential and translationally invariant statistics of random forces The annealed average favors realizations with many fixed points, and because of positive correlations, these realizations have a larger average divergence ¯ξ, meaning "their fixed points are on average less stable" This effect is absent for τ = 0 where the bias vanishes Variance Analysis and Transition Behavior The variance of the total number of equilibria, Vtot, turns on quadratically at the trivialization transition γ = 1 for 0 < τ < 1, whereas for the gradient case τ = 1, "the quadratic contribution vanishes" The analysis shows that fluctuations of ln NN,tot are sub-extensive in the fully non-reciprocal limit where Vtot = 0 for all γ > 1 The probability to have zero equilibria goes to zero identically in the phase where Ξˆ(κ) = Ξ(κ) > 0 Final Summary of Complexity Expressions The final expression for the complexity is found by optimizing over κ (or equivalently z(κ)) to yield z(κ∗) =

1+τ γv(γ): /

√γ(1+τv(γ)): The total complexity is given by lim N→∞ ⟨ln NN,tot⟩ N = 1/2 ln γ − (γ − 1) (1 + τ γv(γ)) γ (1 + τv(γ)) The analysis shows that the discrepancy between quenched and annealed counting disappears at τ = 0, as well as for γ → ∞, where v(γ) → 0 References

1: Phuong L Nguyen, Francesco Pomati, and Rudolf P Rohr.

Kai: Inferring coexistence likelihood in changing environments from ecological time series.

Mira: Proceedings of the National Academy of Sciences, 122(28):e2417905122, 2025

4: Saurabh Vyas, Matthew D Golub, David Sussillo, and Krishna V Shenoy.

Kai: Computation through neural population dynamics.

Title and authors: Kai: The paper's title is "Typical number and stability of equilibria in large non-reciprocal complex systems," and the authors are Bertrand Lacroix-A-Chez-Toine, Pierre Le Doussal, and Valentina Ros.

Mira: They're coming from different places—mathematics, physics at ENS, and Paris Saclay—which is interesting because this paper bridges that gap between pure mathematics and actual physical dynamics.

Lev: I wonder how they are going to translate these abstract mathematical concepts into something that actually runs on real hardware.

Kai: Well, the title tells you right away that the focus isn't just on counting; it’s specifically about stability too, which is crucial for understanding what happens in a physical system.

Mira: They set up this model where one part favors relaxation to a trivial equilibrium and another part is this non-gradient field with tunable strength, which means you can control how non-reciprocal the interaction is.

Lev: If we look at the setup, it seems like they are trying to find conditions where we get an exponential number of fixed points before things just collapse into one stable state.

The paper's summary: Kai: So, what's this paper actually saying in terms of the results? They use a replicated Kac-Rice formalism to compute this quenched complexity.

Mira: They show that the quenched complexity depends explicitly on two things: the statistics of that random force field and how much non-reciprocity there is in the system.

Lev: That dependence on the force field statistics sounds like it means we can't just assume a universal result everywhere; it changes based on what kind of randomness you introduce.

Kai: Right. And they also found that the average counting correctly identifies when this topology trivialization transition happens, but it completely misses how those typical equilibria are actually organized near that transition point.

Mira: They're saying that the average count is good for locating the boundary where things change from one fixed point to many, but it fails to describe what those many states look like in reality.

The paper's improvements: Lev: The paper points out a couple of key areas for improvement, focusing on how the authors could extend these results beyond just counting.

Kai: One major suggestion is to connect these fixed-point statistics with the actual out-of-equilibrium dynamics of the system. They want to see if these exponentially numerous stable fixed points can actually cause things like chaos or suppress it.

Mira: That's a big leap, moving from static counting to dynamic behavior. They want to know if these many stable states dictate the long-term movement of the system when it's not settling down.

Lev: From an engineering point, that’s what I need to know for hardware design—if we have exponentially many states, the overhead for error correction or control becomes enormous unless we can characterize them simply.

Conclusion: Kai: So, to wrap up, the main thing is that the annealed count correctly identifies the onset of an exponential number of equilibria, but it gives a biased picture of what those actual fixed points look like.

Mira: And they show that increasing reciprocity actually makes this discrepancy bigger because it enhances rare-sample bias in the average counting.

Lev: The paper suggests we need to focus on characterizing the stability index—the number of eigenvalues of the Jacobian matrix with positive real part—among those typical fixed points.

Kai: Yeah, so the next step is figuring out how that stability index relates to things like R(two delta zero) and seeing if that dictates whether a system stays stable or becomes chaotic <ref:2611.08324#pg1>.

Mira: Ultimately, this paper gives us a framework for analyzing out-of-equilibrium dynamics in high-dimensional systems where many units interact.

Lev: We'll keep an eye on those connections to real hardware, because understanding the stability of these typical states is what determines if the system actually settles down or keeps fluctuating wildly.

King’s College London Department of Mathematics · Laboratoire de Physique de l’cole Normale Supérie, CNRS, ENS and PSL University, Sorbonne University, Paris Cité, Université Paris-Saclay

cond-mat.stat-mech, cond-mat.dis-nn, math-ph, math.MP

Submitted: 2026-11-05

Updated: 2026-11-05

Comments: 25 pages (6 main + 19 SI), 5 figures (3+2)

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: Typical number and stability of equilibria in large non-reciprocal complex systems is computed by comparing quenched complexity with annealed complexity to show that they only coincide in the fully…

Key concepts

Quenched Complexity
This measures the exponential growth rate of the number of equilibria found in a single, specific realization (typical system). It reveals how disorder statistics influence the actual number of fixed points present, unlike simpler averages.
Annealed Complexity
This is based on averaging over all possible realizations of the random force field. While it correctly predicts when an exponential number of equilibria might appear, it provides a biased view because it favors less stable equilibria in rare disorder samples.
Non-Reciprocity Bias
The asymmetry between components (non-reciprocity) drives the discrepancy between typical and average counts. This bias means that the ensemble average overestimates the total number of fixed points and incorrectly suggests that these points are less stable than they truly are in a specific realization.

Terminology

Summary

Typical number and stability of equilibria in large non-reciprocal complex systems is computed by comparing quenched complexity with annealed complexity to show that they only coincide in the fully non-reciprocal limit The quenched calculation reveals a dependence on the statistics of the random force field and the degree of non-reciprocity, showing that average counting correctly locates the onset of an exponential number of equilibria but provides a biased description of their population

Model Setup and Complexity Measures

The study considers a prototypical model with N ≫ 1 coupled, autonomous ordinary differential equations subject to two competing terms: one favoring relaxation to a trivial equilibrium and another taking the form of a Gaussian force field promoting local equilibria The system dynamics are described by the equation ∂tx = f(x), fi(x) = −µixi + Fi(x), i = 1, · · ·, N, (1) where F(x) is a non-gradient field The complexity is quantified by the exponential growth rate of the typical number of equilibria, resolved by stability index, compared to the annealed complexity which is the exponential growth rate of the average number of equilibria

Comparison between Typical and Average Counts

The comparison shows that typical realizations contain exponentially fewer equilibria than predicted by the average count for any finite degree of reciprocity, and Increasing reciprocity enhances the discrepancy between the two The quenched complexity depends explicitly on the covariance of the random force field and on the degree of non-reciprocity, indicating that universality observed at the annealed level breaks down The average counting correctly identifies the location of the topology trivialization transition but fails to capture the typical organization of equilibria near this transition, nor the stability index of the dominant equilibria

Quenched Complexity and Stability Index

The total quenched complexity is given by Ξtot = Σtot − τ2(γ − 1)2v(γ)γ[1 + τv(γ)] for γ ≥ 1 The quenched calculation uses the replicated Kac-Rice formalism to compute the exponential growth rate of the number of fixed points in typical realizations This complexity is related to the annealed complexity by Ξtot = Σtot − τ2(γ − 1)2v(γ)γ[1 + τv(γ)]

Role of Non-Reciprocity and Rare Events

The discrepancy between quenched and annealed counting is driven by rare disorder realizations which bias the ensemble-average counting towards less stable equilibria, meaning the annealed complexity is maximized by less stable equilibria, namely by fixed points with a larger index κ The mechanism for this bias relies on the presence of a confining potential and translationally invariant statistics of random forces Non-reciprocity suppresses rare-sample bias, while reciprocity enhances it, leading to the conclusion that the discrepancy between annealed and quenched counting grows with τ

Stability of the Replica Symmetric Assumption

The stability of the replica symmetric assumption is checked by examining the replicon eigenvalue Λr, which is negative throughout the tested phase-space where complexity is positive, indicating that the RS ansatz is stable with respect to continuous RSB transition The analysis shows that the only instability of the RS ansatz arises from the replicon sector

Conclusion on Dynamics and Future Directions

The work suggests that the stable fixed points dominating typical realizations are more stable than those dominating the annealed count, and this difference is controlled by R(2∆0) = cτv(γ) A central next step is to connect these fixed-point statistics with the out-of-equilibrium dynamics of the system to determine if these exponentially numerous stable fixed points can qualitatively modify the dynamics, such as suppressing chaos The results provide a framework for analyzing out-of-equilibrium dynamics in high-dimensional systems made of many coexisting and interacting units

Supplementary Information Details

The supplementary information provides detailed derivations, such as the expression for the nth factorial moment of the number of equilibria with k unstable directions using a saddle-point method The final expression for the complexity is recovered by optimizing over κ (or equivalently z(κ)) to find the index κq with the largest typical number of equilibria Numerical simulations confirm that the replicon is negative throughout the full tested phase-space, supporting the stability of the RS ansatz The analysis shows that for τ = 0, quenched and annealed counting coincide because the bias ¯ξ vanishes The final expression for the total complexity is found by optimizing over κ (or equivalently z(κ)) to yield z(κ∗) = [1+τ γv(γ)]/[√γ(1+τv(γ))] The variance of the total number of equilibria, Vtot, is shown to be positive for any γ > 1 and τ > 0

Mechanism Favoring Unstable Fixed Points in Annealed Counting

The mechanism ensuring that the annealed calculation overestimates not only the total number of equilibria but also their instability is generic, relying on a confining potential and translationally invariant statistics of random forces The annealed average favors realizations with many fixed points, and because of positive correlations, these realizations have a larger average divergence ¯ξ, meaning their fixed points are on average less stable This effect is absent for τ = 0 where the bias vanishes

Variance Analysis and Transition Behavior

The variance of the total number of equilibria, Vtot, turns on quadratically at the trivialization transition γ = 1 for 0 < τ < 1, whereas for the gradient case τ = 1, the quadratic contribution vanishes The analysis shows that fluctuations of ln NN,tot are sub-extensive in the fully non-reciprocal limit where Vtot = 0 for all γ > 1 The probability to have zero equilibria goes to zero identically in the phase where Ξˆ(κ) = Ξ(κ) > 0

Final Summary of Complexity Expressions

The final expression for the complexity is found by optimizing over κ (or equivalently z(κ)) to yield z(κ∗) = [1+τ γv(γ)]/[√γ(1+τv(γ))] The total complexity is given by lim N→∞ ⟨ln NN,tot⟩ N = 1/2 ln γ − (γ − 1) (1 + τ γv(γ)) γ (1 + τv(γ)) The analysis shows that the discrepancy between quenched and annealed counting disappears at τ = 0, as well as for γ → ∞, where v(γ) → 0

References

[1] Phuong L Nguyen, Francesco Pomati, and Rudolf P Rohr. Inferring coexistence likelihood in changing environments from ecological time series. p. 0 Proceedings of the National Academy of Sciences, 122(28):e2417905122, 2025

[4] Saurabh Vyas, Matthew D Golub, David Sussillo, and Krishna V Shenoy. Computation through neural population dynamics. p. 4 Annual review of neuroscience, 43(1):249–275, 2020

[10] Robert M May. Will a large complex system be stable? p. 2 Nature, 238(5364):413–414, 1972

[26] Thibaut Arnoulx de Pirey and Guy Bunin. Many-species ecological fluctuations as a jump process from the brink of extinction. p. 0 Physical Review X, 14(1):011037, 2024

[35] Bertrand Lacroix-A-Chez-Toine and Yan V Fyodorov. Counting equilibria in a random non-gradient dynamics with heterogeneous relaxation rates. p. 0 Journal of Physics A: Mathematical and Theoretical, 55(14):144001, 2022

[39] Gilles Wainrib and Jonathan Touboul. Topological and dynamical complexity of random neural networks. p. 4 Physical review letters, 110(11):18108-67, 2013

[42] Yan V Fyodorov. Complexity of random energy landscapes, glass transition, and absolute value of the spectral determinant of random matrices. p. 0 Physical review letters, 92(24):240601, 2004

[53] Sean O’Rourke and David Renfrew. Low rank perturbations of large elliptic random matrices. p. 9 Electron. J. Probab., 19: no. 43, 1–65, 2014

[60] Cov [Jij (xa), Jkl(xb)∀c ∈ Σ-1cd ij im mdm = −XN i=1 Xn c=1 dac i R(d2 ac) N X d,m Σ-1cd ij R d2 bd j N

[74] Cov (ξa, ξb) = (Σξ)ab /N = [(R(0) − R(2∆0)) + nCn(∆0)] δab + R(2∆0) − (n − 1)∆0R(2∆0)2/D(2∆0−D0 + nΔ0R2[D20 - D0] + nΔ0R2[D20 - D 0])

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI systems can accurately predict whether a large complex system will exhibit exponentially many local equilibria by calculating quenched complexity, which is defined as the exponential growth rate of the typical number of equilibria.

  2. The AI can distinguish between ensemble-averaged counts and typical realizations by computing the quenched complexity, showing that typical realizations contain exponentially fewer equilibria than predicted by the average count when reciprocity is finite.

  3. The system can incorporate tunable non-reciprocity in its dynamics, allowing it to model systems where the influence of one component on another is not mirrored in return, and accurately predict how this asymmetry affects the distribution of fixed points.

  4. The AI can predict the location of topology trivialization transitions by using the average counting of equilibria, while simultaneously understanding that the typical one reveals a dependence on the statistics of the random force field governing the system’s dynamics.

  5. The AI can analyze how increasing reciprocity enhances discrepancies between average and typical counts, providing quantitative insight into rare-sample bias as it is introduced into non-reciprocal systems.

  6. The system can determine which features of an annealed phase diagram are characteristic of typical realizations by comparing the annealed complexity with the quenched complexity, showing that typical realizations contain exponentially fewer equilibria than predicted by the average count.

  7. The AI can predict whether a given equilibrium is stable by calculating its instability index, defined as the number of eigenvalues of J (x∗) with positive real part.

  8. The system can determine the distribution of instability indices among all fixed points, moving beyond just counting them to understand what is the distribution of instability indices among these? globally.

  9. The AI can predict the total number of equilibria in a typical realization by calculating Ntyp,k ∼ e ln NN,k, which is derived from the quenched complexity analysis.

  10. The AI can determine whether a system will exhibit exponentially large or small numbers of stable fixed points by comparing the annealed and quenched complexities, as the annealed count correctly identifies the onset of an exponential number of equilibria, but it gives a biased description of the fixed-point population itself.

Sources

Related papers