Criticality Is a Control Loop

01The evidence

• Critical-state dynamics of avalanches and oscillations emerge jointly from balanced excitation and inhibition in neuronal networks, which means the critical state can be produced by a simple homeostatic balance rather than by fine tuning for a computational purpose.Journal of Neuroscience
• The same result is reproduced in network models where the critical state appears as the fixed point of a balance rule, supporting the view that criticality is where the control loop settles rather than where the system is designed to be.PMC full text

02The tension

Criticality is routinely reported in cortex, in ecosystems, and in gene networks, and is routinely interpreted as functional: the system is critical because criticality maximises information transmission or dynamic range. The interpretation has not been tested by pushing a system to criticality and showing that function improves.

03The theory

A homeostatic controller that adjusts gain to keep activity in a usable band will naturally land near the point where response is most sensitive, because that is where the controller has the most leverage. The critical point is the fixed point of that control law, not a design target.

If this is right, criticality is a side effect of control, and the exponent is a readout of the controller's gain. Change the gain and the exponent moves. That makes criticality an instrument for measuring control, not a goal in itself.

The strong prediction is counterintuitive: interventions that move the system away from criticality should not necessarily impair function, and may improve dynamic range, because dynamic range is what the controller is actually optimising. Conversely, forcing a system to criticality should not improve function unless the controller was already trying to get there.

This dissolves a puzzle. If criticality were functional, evolution should have stabilised it precisely, and yet the exponent varies across states and regions. If it is a byproduct of a controller optimising something else, variation is expected.

04What this is not

05Predictions

• P1Changing the homeostatic setpoint should shift the avalanche exponent away from the critical value while increasing measured dynamic range.
• P2The exponent should track the control gain across states, asleep versus awake, and across individuals.
• P3In food webs, the same relation should hold between a control parameter and the reported critical exponent.

06What would refute it

A demonstration that driving a system to its critical point improves function, in a controlled manipulation, would support the functional interpretation instead.

07What it would change

It would reframe a large literature, and give experimenters a reason to measure the controller rather than the exponent.