When the Average Lies: Equilibrium Statistics in Non-Reciprocal Complex Systems
— When the "mean" is hijacked by rare worlds, universality collapses
> Paper: *Typical number and stability of equilibria in large non-reciprocal complex systems* > Authors: B. Lacroix-A-Chez-Toine, P. Le Doussal, V. Ros (arXiv:2611.08324, cond-mat.stat-mech)
1. Foxes and rabbits: an asymmetry ignored for decades
Ecological interactions are asymmetric: foxes eat rabbits, and the effect of foxes on rabbits is not the same as the effect of rabbits on foxes—predation obeys no Newtonian action–reaction symmetry. Synapses in neural networks, chemical reaction networks, and self-propelled particles in active matter all behave this way. Physicists call this asymmetry non-reciprocity.
Mathematically, reciprocal systems correspond to gradient flows—everything rolls downhill toward a potential minimum, like a marble in a bowl. Non-reciprocal systems correspond to non-gradient flows—no potential exists, and the system drifts in a random force field, like a marble in swirling water.
A central question for such systems: how many equilibria does this high-dimensional space contain, and how many are stable? Equilibria determine fate—where the system stalls, how it stalls, and whether a perturbation can kick it away.
2. Two statistics: millionaires vs. ordinary people
The standard tool for counting equilibria of a high-dimensional random landscape is the Kac–Rice formula. But there are two very different ways to count:
- Annealed: average over the disorder first, then count equilibria—this gives the mean number of equilibria;
- Quenched: count equilibria of a specific system, then average (the logarithm) over disorder—this gives the typical number.
- In the fully non-reciprocal limit, annealed and quenched counts agree—the 2016 picture survives intact;
- But for any finite reciprocal component (even one tending to zero), "a typical realization contains exponentially fewer equilibria than the annealed count predicts."
An income analogy: the annealed complexity is like "mean income," inflated by a few billionaires; the quenched complexity is like "median income," reflecting ordinary reality. When the distribution is mild, the two agree; when rare events grow exponentially extreme, the average lies.
In 2016, Fyodorov and Khoruzhenko computed the famous topological trivialization transition for non-gradient random flows: as nonlinearity strengthens, the (annealed) number of equilibria vanishes abruptly—the high-dimensional landscape goes from "pockmarked everywhere" to "perfectly flat" overnight. Remarkably, the result exhibited universality: the answer depended only on coarse variance statistics, not on details of the field covariance.
Ten years later, this new paper asks a stricter question: is what the average says what a typical system does?
3. The answer: any trace of reciprocity makes the average lie
The authors compute index-resolved quenched complexities for a Gaussian non-gradient force field with a tunable degree of reciprocity. The conclusions are sharp:
More sobering is the collapse of universality: the quenched complexity "explicitly depends on the covariance of the force field and the degree of non-reciprocity"—no single one-size-fits-all formula remains. Every detail of the covariance is written into the answer. The beautiful 2016 world where variance alone suffices turns out to hold only at the fully non-reciprocal extreme.
Even the meaning of the transition needs revision: annealed calculations can still locate where topological trivialization occurs, but they "miss how equilibria are typically organized at the transition, and miss their stability indices"—the real terrain near the transition differs sharply from the averaged picture.
4. Intuition: why reciprocity is so lethal
Why does a pinch of reciprocity open an exponential gap? The intuition lies in the geometry of the two field types. The reciprocal component is a gradient field—its streamlines have no vortices, and the landscape is a genuine "potential topographic map." Fitting a huge number of equilibria onto such a map (maxima, minima, and saddles appearing in pairs) requires many exquisitely balanced peaks and valleys—an event that is exponentially rare in a random realization, but exponentially abundant once it occurs. The non-reciprocal component is a rotational field—its vortices allow a potential-free system to organize equilibria through rotational structures, without relying on topographic coincidences.
So the average count is polluted by rare "topographic-map" realizations full of equilibria; the larger the reciprocal component, the more weight these carry and the deeper the average–typical gap. Non-reciprocity is, ironically, the democratic case—it makes realizations look alike, and the average becomes honest again.
5. Echoes: from May's ecological puzzle to neural network loss landscapes
This work continues the famous question Robert May posed in 1972: are large randomly connected ecosystems stable? For half a century, theorists worked mainly in the mathematical comfort zone of reciprocal (symmetric) systems, sheltered by Wigner's theorem and the semicircle law. The non-reciprocal world offered only the sparse tools of non-Hermitian random matrices. The new result is a warning: in non-reciprocal systems, intuition based on averages is not just quantitatively off—it is qualitatively misleading. Even the most basic question, "how many stable states are there?", cannot be answered trustworthily by averaging.
The same lesson applies to neural networks: the deep learning loss landscape is also a high-dimensional non-gradient field (with SGD noise and asymmetric connections). Are analytical results on the "average loss landscape" likewise being hijacked by rare realizations? No answer exists today—but this paper at least provides grounds for suspicion.
The lesson from Le Doussal, a master of quenched averaging techniques, and his collaborators can be condensed into one sentence for anyone working on high-dimensional complex systems:
> In a high-dimensional world, the "average" is not an approximation of the "typical"—it is often the name of a different world.
---
Further reading
1. Lacroix-A-Chez-Toine, Le Doussal, Ros, *Typical number and stability of equilibria in large non-reciprocal complex systems*, arXiv:2611.08324 2. Fyodorov, Khoruzhenko, *Topology trivialization transition in random non-gradient autonomous ODEs on a sphere*, J. Stat. Mech. 124003 (2016) (the annealed precursor) 3. May, *Will a large complex system be stable?*, Nature 238, 413 (1972) 4. Fyodorov, Sommers, *Counting equilibria of large complex systems by instability index* (the classical Kac–Rice framework) 5. Ros, Biroli, Cammarota, *Complexity in mean-field spin glasses* (a review of high-dimensional landscapes)
*This is part 4 of round 5 of zhichai.net's "Mathematical Foundations and Frontier Intersections" series.*