Feedback & Control Loops

The unifying principle

Every system here is a loop: an output is sensed, compared to a reference, and the error is fed back into the input.

The sign of feedback decides behavior:

• Negative feedback ($e$ opposes deviation): the $+1$ in $1+L$ shrinks the error; the system self-corrects and stabilizes.
• Positive feedback ($e$ reinforces deviation): effectively $1 - L$; small perturbations amplify, giving switches, bistability, or blow-up.

Stability is set by the poles — the roots of $1 + L = 0$. If any pole has positive real part, oscillation grows. A simple first-order corrector behaves like

$\dot{e} = -k\,e \implies e(t) = e_0\,e^{-kt},$

decaying for $k>0$ (stabilizing) and exploding for $k<0$ (runaway). Delay turns even negative feedback into oscillation: correcting based on stale information overshoots.

How it shows up in each field

Hardware — the thermostat / PID controller

The canonical negative-feedback loop. A thermostat senses temperature $y$, compares to setpoint $r$, and switches a heater. Continuous controllers use PID: $u(t) = K_p\,e + K_i\!\int e\,dt + K_d\,\dot{e}$

• $K_p$: proportional pushback on current error
• $K_i$: eliminates steady-state offset
• $K_d$: damps overshoot by reacting to error rate

Biology — homeostasis

Bodies hold internal variables near setpoints via negative feedback; positive feedback drives commitment/switch events.

• Negative: blood glucose rises → pancreas releases insulin → cells absorb glucose → glucose falls. Error $e = [\text{glucose}] - r$ is driven to zero. Thermoregulation (sweating/shivering) is the biological thermostat.
• Positive: oxytocin in childbirth — contractions release oxytocin, which strengthens contractions, an intentional runaway to completion. Also action-potential Na⁺ channel opening.

AI-ML — reinforcement learning & training loops

The agent's loop is control: action → environment → reward/error → policy update.

• Negative feedback: gradient descent minimizes loss $J$ by stepping down the error gradient: $\theta_{t+1} = \theta_t - \eta\,\nabla_\theta J$ This is proportional negative feedback on prediction error; the learning rate $\eta$ is loop gain — too high → divergent oscillation, too low → sluggish.
• Positive feedback (pathological): reward hacking, model collapse from training on own outputs, or divergent value estimates in bootstrapped Q-learning where overestimates reinforce themselves.

Stock-Market — price correction & momentum

Prices are a feedback loop between valuation and trading.

• Negative feedback (mean reversion): price above fair value → sellers enter → price falls. Arbitrageurs act as the corrector driving $e = \text{price} - \text{value} \to 0$.
• Positive feedback (bubbles/crashes): rising prices attract momentum buyers → further rises; margin calls in a crash force selling that drives prices lower still. This is $1 - L$ behavior — self-amplifying until a nonlinearity (liquidity, circuit breakers) intervenes.

Why this bridge matters

The same three failure/success modes recur, so intuition transfers directly:

• Stability ↔ loop gain & delay. Control theory's rule "too much gain + lag = oscillation" explains drone wobble, glucose overshoot in diabetes, exploding learning rates, and boom-bust market cycles. If you can tune a PID, you can reason about all four.
• Negative vs positive is a design choice, not a value judgment. Biology uses positive feedback for decisive switches; markets suffer it as bubbles; ML exploits negative feedback for learning but must guard against positive-feedback collapse.
• The error term is universal. TD-error, glucose deviation, thermostat error, and mispricing are all $e = r - y$. Recognizing this lets a biologist read RL, or a trader read control theory, and immediately ask the right question: what's the setpoint, what's the gain, and how stale is the sensor?

Connections

• 01-Hardware-PID-Controllers
• 02-Hardware-Sensors-and-Actuators
• 03-Biology-Homeostasis
• 04-Biology-Signaling-Cascades
• 05-AI-ML-Reinforcement-Learning
• 06-AI-ML-Gradient-Descent
• 07-Stock-Market-Mean-Reversion
• 08-Stock-Market-Momentum-and-Bubbles
• 09-Control-Theory-Stability-and-Poles

#bridge