Skip to main content
computer-science economics engineering-and-technology linguistics

​
Feedback loop

​
Description

Output influences future input through a returning signal — the system’s own behavior feeds back to change what happens next. The polarity decides the character: positive feedback amplifies (small perturbations grow; runaway behavior; vicious or virtuous cycles); negative feedback damps (perturbations get pushed back toward equilibrium; thermostats, governors, immune response). Most stable systems combine both at different timescales. The diagnostic question — what is the signal-and-receiver pair? — separates real feedback from coincidence. If you cannot name what got measured, what got sent back, and what changed in response, you do not have a loop; you have correlation.

​
Triggers

User-initiated: User describes a system where something “spiraled,” “compounded,” or “got out of control,” or where stability arose unexpectedly. Vocabulary cues: “vicious cycle,” “virtuous cycle,” “snowball,” “runaway,” “self-reinforcing,” “cascade,” “compound.” Agent-initiated: Agent notices a system’s behavior depends not just on current inputs but on its own recent outputs being fed back. Candidate inference: “what is the polarity — is this amplifying or damping?” Situation-shape signals: Phenomena that are surprising under no-feedback assumptions (sudden equilibrium, sudden runaway, oscillation, hysteresis-like path-dependence). Long-running processes where you would expect monotonic behavior but observe oscillation or threshold crossings.

​
Exclusions

  • Open-loop control systems — output set without measuring outcome (e.g., a microwave timer doesn’t measure food temperature; it just runs for N seconds). The concept doesn’t fire; the system is by-construction non-adaptive.
  • Pure forcing / external drivers — sun-driven daily temperature isn’t a feedback loop; it’s a driven oscillation with no return signal.
  • Coincidence vs causation — correlated movements between two variables aren’t a feedback loop unless you can identify the actual signal-and-receiver. Naming “this is a feedback loop” without naming the channel is a label-without-mapping failure.

​
Structure

Internal structure of feedback-loop: a table of its component slots and the concepts that fill them.

​
Relationships

Relationship neighborhood of feedback-loop: a graph of the concepts it connects to and the concepts it is a part of.
  • backpressure — backpressure is feedback-loop applied to a flow; the loop signals upstream when downstream lags. Backpressure is the higher-order concept (flow + feedback-loop); feedback-loop is the primitive it builds on.
  • cadence — stable cadences in feedback-controlled systems emerge from the loop’s natural period (PID-tuned settling time, biological circadian rhythms, code-review cycle times).
  • substrate-surface-amplifier — the amplifier tier explicitly is a feedback-loop: accumulated outcomes feed back as progressively better-targeted retrievals.
  • loop-completion — loop-completion is the diagnostic for “is this loop actually closing?”; useful when a feedback channel exists in principle but the signal isn’t reaching the actor.

​
Examples

Thermostat / governor · engineering-and-technology

canonical negative-feedback example; measured temperature signals back to the heater’s on-off control.

Stock-market bubbles · economics

positive feedback: rising prices attract buyers; buyers raise prices; cycle continues until exhaustion.
review feedback informs next commit; rapid review establishes a tight loop, slow review unbounds drift.
positive feedback on the reinvestment loop; small advantage compounds over many cycles.
Control theory is the feedback loop made quantitative — specifically the negative, self-regulating kind. The founding text is Maxwell’s 1868 “On Governors,” which took Watt’s centrifugal flyball governor (a device that throttles a steam engine’s fuel as the engine speeds up, slowing it back toward a setpoint) and asked the question that turns engineering intuition into a science: when does the correcting loop settle, and when does it oscillate? As engines grew more powerful, governors began “hunting” — overshooting, correcting too hard, overshooting the other way. Maxwell linearized the governing differential equations around equilibrium and proved that stability depends on a precise condition: every root of the system’s characteristic polynomial must have a negative real part. That is the mathematical statement of “the loop converges instead of running away.”This is the same returning-signal structure the feedback-loop concept names, with the sign and the dynamics made explicit. Output (engine speed) is measured, compared to a setpoint, and fed back as a correction; whether the loop stabilizes or amplifies is now a checkable property of the equations, not a hope. The later apparatus elaborates the same question: the PID controller decomposes the correction into proportional (respond to current error), integral (accumulate past error to kill steady-state offset — exactly the “droop” Maxwell flagged), and derivative (anticipate via rate of change, adding damping); the root-locus method and Nyquist criterion are tools for predicting from the open-loop response whether closing the loop will be stable.Inference: When a system has a returning correction signal, the load-bearing question is not whether feedback exists but whether the loop’s gain and timing put it in the convergent regime or the oscillatory/divergent one. Control theory supplies the diagnostic: too much corrective gain or too much lag and a stabilizing loop starts hunting (overshoot, oscillation) or goes unstable — the same failure mode that shows up in supply chains, thermostats, monetary policy, and any other negative-feedback system tuned with the wrong gain.
Donella Meadows, Thinking in Systems — accessible systems-thinking treatment; popularized “leverage points” framing where feedback-loop changes have outsized effect.
In the image-schema tradition associated with George Lakoff and Mark Johnson (and extended by Leonard Talmy’s force-dynamics work), feedback isn’t an abstract control-theoretic notion bolted onto English vocabulary — it’s encoded directly in everyday language as a force-dynamic interaction. “Push back,” “give and take,” “build pressure,” “let off steam,” “fight the tide” all describe agents and counter-agents whose outputs change each other’s future inputs.That folk-vocabulary encoding is evidence that the feedback-loop schema is a primitive cognitive structure humans reason from across domains. The same schema that handles “the thermostat pushes back against rising temperature” also handles “the market pushes back against overpriced offerings” and “the team pushes back against unrealistic deadlines.” The image-schema lineage grounds why the concept transfers so freely.
Norbert Wiener, Cybernetics: Or Control and Communication in the Animal and the Machine (1948) — the discipline-founding text; named the concept.
reward signal feeds back into policy updates; the agent’s own outputs determine the next training signal.