"__Optimal control of excitable systems near criticality __ why is this paper's title so powerful " tldr this paper, go into it, and explain why it is so powerful. SUMMARIZE IT

Thought for 52s

That title hits like it knows it’s right: “Optimal control of excitable systems near criticality.” It’s short, it’s ambitious, and it welds together three big, high-status ideas (control theory, excitable dynamics, critical phenomena) into one promise: there’s a principled sweet spot where you can steer noisy, twitchy networks best. Humans love “there’s a sweet spot” almost as much as they love pretending they’ll actually use it.

TL;DR of the paper

They study a simple stochastic network model of “binary neurons” (active/inactive) and ask: if you use feedback control to hold the population firing rate near a target, when is that easiest? Answer: near criticality. Even though critical dynamics are the noisiest, they’re also the most controllable across the widest range of desired activity levels.

The setup (what they actually do)

The dynamical system

• A network of NNN excitable nodes, each either 0 or 1 each time step.
• Directed, weighted, signed adjacency matrix AAA with both excitation and inhibition (they assume 20% inhibitory nodes).
• Activation is probabilistic via a piecewise-linear transfer function σ(x)\sigma(x)σ(x) (0 below 0, linear between 0 and 1, saturates at 1).
• The key “distance to criticality” knob is the largest eigenvalue λ\lambdaλ of AAA:
  • λ<1\lambda<1λ<1: activity dies out
  • λ>1\lambda>1λ>1: activity blows up
  • λ=1\lambda=1λ=1: critical boundary

The control task (important: it’s macroscopic)

They’re not doing full state controllability (“force neuron-by-neuron trajectories”). They do something much more biologically/plausibly relevant: keep a global activity variable close to a target.

They define a global activity S=∑bmsmS = \sum b_m s_mS=∑bmsm and apply proportional feedback using the error (S^−S)(\hat S - S)(S^−S), injected back into node activation probabilities with gains μn\mu_nμn.

This is meant to resemble things like optogenetic or closed-loop stimulation controlling population rate, not micromanaging every neuron.

The core result (why “near criticality” wins)

Linear-regime punchline: λ=1\lambda=1λ=1 kills steady-state error

They derive an analytic expression for expected relative error RRR and show something very clean: