Certified feedback laws that run on fixed-weight spikes alone

What the work claims

This is a theory paper with a simulation study; there is no hardware and no biological experiment, and the claims are theorems about idealized systems, not measurements. The setting is a network game: several agents, each a linear dynamical system with an unknown model and an exogenous disturbance, each minimizing a local cost that depends on its own output and its neighbors' outputs, communicating over a directed graph. The objective is the Nash equilibrium, the joint state where no agent can improve its cost by unilateral change, here relaxed to epsilon-equilibrium: everyone within epsilon of optimal.¹

The claim has three stacked parts. First, the feedback gains that steer the agents to equilibrium can be computed directly from noisy local input-state trajectories, using robust linear matrix inequalities, without ever identifying the agent dynamics. Second, those gains can be realized using only fixed-weight spikes: the control commands are encoded in the firing times and channel selections of leaky integrate-and-fire units, not in continuously varying amplitudes. Third, the spiking implementation degrades the ideal guarantee gracefully: a continuous auxiliary coordinate exposes the impulsive closed loop as the stable analogue system plus a bounded implementation error, yielding forward completeness, Zeno-freeness, and an ultimate bound on the pseudo-gradient that implies eventual epsilon-Nash equilibrium for every epsilon above a finite threshold.

How it works

The trick that makes equilibrium seeking look like a regulation problem is classical: for the quadratic games studied, the pseudo-gradient, the vector of each agent's marginal costs, is affine in the outputs, and its unique zero coincides with the Nash equilibrium. Drive the pseudo-gradient to zero and you have solved the game. Treating that gradient as a regulated error connects game theory to the internal model principle of control: the agents' controllers embed a local copy of the disturbance generator, a known exosystem, so persistent disturbances such as orbital perturbations are rejected exactly rather than merely attenuated.

The data-driven step replaces system identification with a consistency argument. Each agent collects one offline trajectory in isolation, an 8-second run in the numerical study, applying zero-order-held probing inputs and recording 80 samples of its own state. From this single noisy trajectory, a robust linear matrix inequality certifies a stabilizing analogue feedback gain for every model in the consistent set, which is why the dynamics never need to be identified.

The spiking step is where the paper earns its novelty. The analogue gain is encoded into populations of leaky integrate-and-fire units whose weights are fixed once and never adapted. The authors construct an exact rate-equivalent encoding: the spiking command reproduces the analogue feedback through firing times and channel selection, and a continuous auxiliary coordinate, built by stacking the agent state with the neuronal states so that spike-induced jumps cancel exactly, rewrites the impulsive closed loop as the stable analogue system driven by a bounded implementation residual. That rewriting is what buys the guarantees. Two realizations are given: one with non-interacting units, where each LIF neuron handles one component, and one with reset coupling, where spikes redistribute state among a smaller set of generally nonorthogonal units. Zeno-freeness, the absence of infinitely many events in finite time, is proved for both, and the connected architecture carries explicit event-processing conditions under which everything holds.

The demonstration is a formation reconfiguration of seven spacecraft with distinct masses, 10.5 to 13.5 kilograms, in low Earth orbit, modeled by the Clohessy-Wiltshire-Hill equations at an orbital mean motion of 1.1068 times ten to the minus three per second. Each spacecraft's cost rewards reaching a prescribed relative position while staying near its neighbors; sinusoidal acceleration disturbances at the orbital frequency test the internal model. Both spiking realizations converge to the same prescribed formation as their analogue reference, with the normalized output error and pseudo-gradient norm decaying over time as the theory predicts.

Where a skeptic should push

The load-bearing assumption is linearity. Every certificate in the paper, the data-based LMI, the exosystem internal model, the rate-equivalent encoding, the ultimate equilibrium bound, is constructed for linear agents with quadratic costs and disturbances generated by known linear exosystems. Real systems, and living neural tissue above all, are none of these. The paper is explicit that its guarantees live inside this class; the risk is entirely in how it gets quoted outside it. An epsilon-Nash theorem for linear spacecraft says nothing certifiable about stimulating a cortical culture, and the distance between those settings is not a detail to be bridged by enthusiasm.

Second, epsilon is doing quiet work. The theorem promises equilibrium only for epsilon above a finite threshold, and the threshold is determined by the implementation error bound, which in turn depends on design constants that the numerical study selects rather than derives from physical constraints. The spacecraft example shows convergence of normalized errors, but the paper's own framing is that these are practical equilibria, not exact ones. Anyone importing this machinery should budget for that residual and ask what epsilon their application can tolerate.

Third, the probing step. Certifying the gain requires an offline excitation trajectory with probing inputs, and in the study those inputs span minus 50 to plus 50 newtons on 10-kilogram spacecraft. Applied to tissue, the equivalent is a deliberate, amplitude-bounded but aggressive stimulation protocol run precisely to elicit dynamics, before any closed-loop control begins. That is a real ethical and practical constraint on the data-driven half of the method, and the paper offers no substitute for it. Finally, Zeno-freeness and forward completeness are proved, but the inter-event-time bounds are finite-horizon, and the connected realization's event-processing conditions are stated as explicit assumptions; how easily they are met in hardware or in tissue is untested.

Fixed-weight spikes already compute

For organoid intelligence the non-obvious implication sits on the stimulation side of the loop, where the field's attention is thinnest. Most of the intellectual effort in biological computing goes to reading the tissue out and training it. This paper asks the neglected complementary question and answers it with a theorem: how much certified computation can you push into the spike encoding itself, before any synapse in the target changes? The answer is a surprising amount. A bank of fixed-weight integrate-and-fire units, which is to say neurons with no learning at all, can carry an arbitrary linear feedback law with an exact rate-equivalent encoding, bounded implementation error, and an equilibrium guarantee that degrades gracefully rather than collapsing.

The opportunity is a design discipline for closed-loop tissue experiments that currently does not exist. The framework supplies vocabulary the field badly needs: an explicit bound on what spiking jitter costs, in the form of the epsilon threshold; a proof that event storms cannot occur, which is the formal content of a safe stimulation rate; and a data-driven route to controller gains from short probing trajectories when the system resists modeling, which describes cultured neural tissue almost exactly. None of this transfers as a certificate, the plants are wrong. But it transfers as a specification of what a stimulation encoder must satisfy to be called well engineered, and that is a higher bar than the field currently sets for itself.

The threat is calibration failure in the opposite direction. Control theorists will read this paper as what it is, a clean result about linear spacecraft. Biologists will be tempted to read it as permission: fixed-weight spikes compute, therefore the stimulation side is solved, therefore only the biology remains. That inference is wrong twice over. The guarantees assume the receiving plant is linear, low-dimensional, and disturbed by known exosystems; tissue is nonlinear, high-dimensional, and noisy in ways nobody can write as an exosystem. And the one step that cannot be imported is the one the certificate depends on, deliberate bounded-amplitude probing to elicit dynamics, whose tissue equivalent collides directly with welfare constraints on living substrates. The honest transfer is narrower and more useful: treat the stimulation encoder as a designed, analyzable component with its own error budget, and refuse to let a game-theoretic theorem for spacecraft stand in for evidence on living tissue.

The bottom line

Established: for linear agents in quadratic network games with exosystem disturbances, feedback gains computed from a single noisy offline trajectory via robust linear matrix inequalities can be realized exactly in the rate-equivalent sense by fixed-weight spiking neurons, and the resulting impulsive closed loop is complete, Zeno-free, and converges to an epsilon-Nash equilibrium for epsilon above an implementation-error-dependent threshold; the seven-spacecraft study illustrates each stage. Not established: anything nonlinear, any biological substrate, any hardware validation, or any epsilon tight enough to matter for a specified application. What would extend the claim is a nonlinear or hybrid version with the same auxiliary-coordinate device; what would break it is a demonstration that realistic event timing destroys the rate-equivalence in practice. For biological computing the durable content is a standard: stimulation encoders deserve the same error budgets and safety proofs this paper gives spacecraft.