Feedback Loops: Stability, Runaway and Oscillation

1. Introduction

A thermostat maintains a temperature. The human body regulates blood sugar. A microphone held too close to a speaker produces a piercing squeal. These three phenomena share a common structure: a consequence returns to influence its own cause. That is the definition of a feedback loop.

Feedback is one of the most cross-disciplinary concepts in science. It structures cybernetics, control theory, systems biology, ecology, climatology and economics. Understanding its precise mechanisms — not just its metaphor — allows us to distinguish what stabilizes, what oscillates and what runs away.

This article builds this vocabulary progressively, from the formal definition through phase planes, delays, limit cycles and causal diagrams. It also clarifies when feedback produces a spiral trajectory — and when it does not.

2. Defining Feedback

In an open-loop system, a cause produces an effect without that effect returning to modify the cause. In a closed-loop system, the effect is measured, compared to a reference, and the difference — the error — is used to adjust the cause. This closed loop is the fundamental structure of feedback.

Two polarities are classically distinguished. In negative feedback, the effect opposes the deviation: if temperature rises above the set point, the system cools. In positive feedback, the effect amplifies the deviation: if temperature rises, the system heats further. The term "positive" does not mean beneficial; it simply indicates that the sign of the loop matches that of the initial perturbation.

3. Negative Feedback: Correction and Stabilization

Negative feedback is the basic mechanism of regulation. It compares the current state to a reference value — the set point — and generates a corrective action proportional to the deviation. A thermostat, a car cruise control, the pupil of the eye dilating in darkness: all use this principle.

In its simplest form, a negative feedback system can be described by a first-order differential equation. If x(t) is the state variable and x* the set point, the dynamics are dx/dt = −k(x − x*), where k > 0 is the gain. The solution is an exponential convergence toward x* with time constant τ = 1/k. The larger k, the faster the convergence.

Blood glucose regulation illustrates a complex biological negative feedback. When blood sugar rises after a meal, pancreatic beta cells secrete insulin, which promotes glucose uptake by muscle and fat cells. When blood sugar falls, alpha cells secrete glucagon, which stimulates glucose release from the liver. These two opposing loops maintain blood glucose within a narrow range.

Body temperature regulation mobilizes several parallel mechanisms: cutaneous vasodilation or vasoconstriction, sweating, shivering, changes in basal metabolism. Each mechanism has its own time constant and activation threshold. Together they form a multi-layer feedback system.

4. Positive Feedback: Reinforcement and Runaway

Positive feedback amplifies deviations. It can be useful — the depolarization of a neuron is a positive feedback process that produces an all-or-nothing action potential — or dangerous, as in thermal runaway or a financial bubble.

The most cited climate example is the ice-albedo feedback. Ice and snow reflect about 80–90% of incident solar radiation (high albedo). When ice melts, it exposes dark water or soil, which absorbs 90–95% of radiation. The increased absorption warms the surface further, melting more ice. This positive loop amplifies the initial warming.

Positive feedback does not necessarily produce infinite growth. In most real systems, it is limited by physical, biological or economic constraints. Population growth is initially exponential (positive feedback: more individuals → more births), but is slowed by available resources (negative feedback: overpopulation → increased mortality). The logistic model captures this interaction.

Hysteresis is an important consequence of strong positive feedbacks. When a system crosses a threshold in one direction, it may not return to the same point when the cause is reversed. One must go further in the opposite direction to trigger the return. This phenomenon is observable in magnetic materials, ecosystems and certain financial markets.

5. Delays, Oscillations and Limit Cycles

Negative feedback does not guarantee smooth convergence. If the corrective response is delayed, the system may overshoot the set point before the correction arrives. It then corrects in the other direction, overshoots again, and oscillates. This behavior is universal: it appears in poorly tuned industrial regulators, commodity markets, predator-prey populations and electronic circuits.

The dynamics of a two-variable coupled system can be described by a system of ordinary differential equations. Consider dx/dt = f(x, y) and dy/dt = g(x, y). Around an equilibrium point, the system can be linearized to obtain a Jacobian matrix J. The eigenvalues of J determine the nature of the equilibrium.

If the eigenvalues are real and negative, trajectories converge directly to equilibrium (stable node). If they are complex with negative real part, trajectories spiral toward equilibrium (stable focus). If the real part is positive, trajectories spiral away (unstable focus). If the real part is zero, trajectories form closed ellipses (center).

A limit cycle is an isolated closed trajectory in phase space. Unlike centers in the linear model, a limit cycle is robust: neighboring trajectories converge toward it (stable limit cycle) or diverge from it (unstable limit cycle). Biological oscillations — heartbeat, circadian rhythm, Belousov-Zhabotinsky oscillator — are often modeled as limit cycles.

The Poincaré-Bendixson theorem establishes that in a plane (two-variable system), if a trajectory remains bounded and does not converge to a fixed point, it must converge to a limit cycle. This result limits the possible complexity in planar systems. In three or more dimensions, chaotic behavior becomes possible.

6. Phase Plane and Eigenvalues

The phase plane represents the state of a two-variable system as a point in a two-dimensional space. The evolution of the system traces a trajectory in this space. This representation allows all possible trajectories to be visualized simultaneously and attractors, repellers and saddles to be identified.

Four types of local behavior are possible around an equilibrium point. A stable node corresponds to real negative eigenvalues: trajectories converge directly, without oscillation. A stable focus corresponds to complex eigenvalues with negative real part: trajectories spiral toward equilibrium. An unstable focus corresponds to complex eigenvalues with positive real part: trajectories spiral away. A saddle point corresponds to real eigenvalues of opposite signs: some trajectories converge, others diverge.

It is essential to understand that the spiral in the phase plane is a geometric property of the trajectory, not a physical property of the system. It indicates that two variables oscillate with a phase shift. It does not imply that the system "rotates" in physical space. A spiral in the (blood glucose, insulin) plane does not mean the body is spinning.

7. Causal Loop Diagrams

A causal loop diagram (CLD) represents relationships between variables using directed arrows. Each arrow carries a sign: positive if an increase in the cause produces an increase in the effect, negative otherwise. A loop is positive if the product of the signs of its arrows is positive, negative otherwise.

CLDs are useful for formulating hypotheses, communicating causal structure and identifying dominant loops. They are used in system dynamics, management and public policy. Their main limitation is that they do not give intensities, delays, nonlinearities or thresholds. Two systems with the same CLD can have very different dynamic behaviors.

To predict the behavior of a system, the CLD must be converted into a quantitative model: differential equations, difference equations or agent-based simulation. This step requires specifying functional forms, parameters and initial conditions. Calibration against real data is often difficult.

8. Examples: Biology, Climate, Economics

In molecular biology, gene regulatory networks combine positive and negative feedback loops. A repressor inhibits its own production (negative feedback, stabilization). An activator stimulates its own transcription (positive feedback, bistability). The combination of the two can produce a bistable switch, an oscillator or a low-pass filter depending on the network topology.

In climatology, feedbacks are classified by their sign and magnitude. Positive feedbacks amplify the initial forcing: ice-albedo, water vapor (the strongest), methane release from permafrost. Negative feedbacks attenuate it: blackbody radiation (the most robust), low clouds in some regions. The balance of these feedbacks determines climate sensitivity.

In economics, commodity markets illustrate delay-induced oscillations. A high price encourages producers to increase supply, but production takes time. When the additional supply arrives, the price collapses. Producers then reduce production, and the cycle repeats. This phenomenon, known as the hog cycle or cobweb theorem, is a classic example of delay-induced oscillation in a negative feedback loop.

PID (proportional-integral-derivative) control is the most widespread engineering solution for managing delays and oscillations. The proportional term corrects the present error; the integral term eliminates accumulated residual error; the derivative term anticipates future variations. Correct tuning of the three gains achieves fast convergence without excessive oscillation. This principle applies to industrial regulators, drones and temperature control systems.

9. Limits and Caveats

Linear eigenvalue analysis is only valid near an equilibrium point. Far from this equilibrium, nonlinearities can radically change behavior. A linearly stable system can be globally unstable if a sufficiently large perturbation moves it outside the region where the approximation is valid.

Real systems often have distributed delays, nonlinearities, stochastic noise and time-varying parameters. Simple two-variable models capture essential mechanisms but must not be confused with precise quantitative predictions. Calibration against real data is essential before any application.

Finally, feedback does not automatically produce a spiral. Simple negative feedback produces monotone exponential convergence. A spiral in the phase plane requires two coupled variables with an appropriate phase shift. The spiral metaphor is useful for certain specific systems; it must not be generalized to all feedback dynamics.