{
  "id": "negative-feedback",
  "title": "Negative Feedback",
  "lang": "en",
  "source_lang": "en",
  "summary": "Negative feedback counteracts changes in a system by returning information about its behavior to influence the processes producing that behavior.",
  "category": "technology",
  "subcategory": "engineering",
  "type": "concept",
  "aliases": [
    "Degenerative feedback",
    "Balancing feedback"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/negative-feedback/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/negative-feedback/",
      "json": "https://aiwiki.page/zh/negative-feedback.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Negative feedback is a process in which a change in a system produces effects that tend to oppose that change. In engineered systems, it commonly involves measuring an output, comparing it with a desired value, and adjusting an input to reduce the difference. In natural systems, the same principle can operate through interacting physical or biological processes without a separate measuring device or explicitly represented target.\n\nNegative feedback is fundamental to control theory, electronic amplification, industrial automation, and physiological regulation. It can make behavior less sensitive to disturbances and uncertain components, improve accuracy, and maintain operating conditions within a limited range. Its effects depend on the strength, timing, and structure of the feedback loop: an opposing response can arrive too late, overshoot, or produce sustained oscillation.\n\nThe concept therefore describes a causal relationship rather than a guarantee of stability. Its mathematical treatment connects the dynamics of individual components with the behavior of the complete loop, while applications range from voltage regulation and motor control to hormone secretion and the response of the climate system to warming."
    },
    {
      "heading": "Causal structure and distinguishing features",
      "text": "A feedback loop exists when the consequences of an action return, directly or indirectly, to influence the process that produced them. A typical engineered loop contains a controlled process, often called the plant; a sensor that measures a relevant variable; a controller that determines a corrective action; and an actuator that applies that action. A room-heating system illustrates the arrangement: temperature is measured, compared with a setting, and used to change heat delivery. The room, heater, and heat losses together determine how rapidly the measured temperature responds.\n\nThe defining feature of negative feedback is the direction of the returned influence. If room temperature rises, the controller reduces heating; if it falls, the controller increases heating. By contrast, positive feedback reinforces an initial change. “Negative” and “positive” refer to these causal signs, not to whether an outcome is desirable. Negative feedback can maintain an unwanted condition, while positive feedback can be useful in switching circuits and regenerative devices.\n\nThe loop sign is determined by the combined effects around the entire causal path. A pathway containing one inhibitory interaction and otherwise activating interactions is negative; two inhibitory interactions can make a loop positive. In a simple monotonic causal model, this is equivalent to multiplying the signs of the individual interactions. More complicated systems may have gains that change with operating conditions, multiple overlapping loops, or interactions whose effects depend on frequency. Their behavior cannot always be classified adequately from a single diagram of positive and negative arrows.\n\nFeedback also differs from feedforward action. A feedforward controller responds to a command or a measured disturbance before the resulting output error develops. Feedback instead uses information about the actual behavior of the controlled system. The two are often combined: a motor controller may calculate an expected torque from a motion command and then use measured position to correct the remaining error. Feedback can compensate for unanticipated disturbances, whereas feedforward compensation depends more directly on the accuracy of a model.\n\nA recognizable negative-feedback mechanism does not necessarily require a fixed set point. Local feedback in an amplifier may establish a proportional relationship between input and output, while a biological interaction may limit the rate of a chemical process. Nor does an observed return toward an average prove that feedback occurred. Establishing feedback requires evidence for a closed causal pathway, not merely a pattern of statistical recovery."
    },
    {
      "heading": "Mathematical description and sensitivity",
      "text": "For a single-input, single-output linear time-invariant system, negative feedback is commonly represented using transfer functions. Let \\(r(s)\\) be a reference input, \\(y(s)\\) the output, \\(G(s)\\) the forward-path transfer function, and \\(H(s)\\) the measurement or return-path transfer function. With zero initial conditions, the equations are\n\n\\[e(s)=r(s)-H(s)y(s), \\qquad y(s)=G(s)e(s).\\]\n\nHere \\(s\\) is the complex frequency variable of the Laplace transform. Solving for the relationship between reference and output gives\n\n\\[\\frac{y(s)}{r(s)}=\\frac{G(s)}{1+G(s)H(s)}.\\]\n\nThe product \\(L(s)=G(s)H(s)\\) is the loop gain. It includes the dynamic behavior of the complete path around the loop, not simply the gain of the amplifier or controller. When \\(|L|\\) is large and the closed loop is stable, the reference-to-output response approaches \\(1/H\\). The overall behavior can consequently depend more on a deliberately chosen feedback network than on an uncertain forward-path gain.\n\nTwo related functions express central performance trade-offs:\n\n\\[S(s)=\\frac{1}{1+L(s)}, \\qquad T(s)=\\frac{L(s)}{1+L(s)}.\\]\n\nThey are called sensitivity and complementary sensitivity, respectively, and satisfy \\(S+T=1\\). For unity feedback, with the forward path divided into controller \\(C\\) and plant \\(P\\), the reference-to-output response is \\(T\\). A disturbance added directly at the output is transmitted through \\(S\\), while a disturbance added at the plant input is transmitted through \\(PS\\). Measurement noise added to the feedback signal reaches the output through \\(-T\\). These distinctions explain why disturbance rejection and noise rejection require attention to where each signal enters.\n\nSensitivity also quantifies reduced dependence on uncertain components. With \\(H\\) fixed, a small fractional change in forward gain produces approximately\n\n\\[\\frac{\\Delta(G/(1+GH))}{G/(1+GH)} \\approx S\\frac{\\Delta G}{G}.\\]\n\nAt a frequency where \\(L=99\\) is real and positive, \\(S=0.01\\): a sufficiently small fractional gain change is reduced by approximately a factor of 100 in the closed-loop gain. This local result does not mean that arbitrary component failures or large nonlinear changes are suppressed by the same factor.\n\nThe algebra is compact, but its assumptions matter. Transfer functions describe linear behavior; real systems may require linearization about an operating point. Multivariable systems use transfer matrices and can exhibit coupling between controlled variables. Initial conditions, actuator limits, and internal dynamics must also be considered separately when assessing whether the complete physical system behaves acceptably."
    },
    {
      "heading": "Stability, delay, and oscillation",
      "text": "Negative feedback opposes a change only in a static or sufficiently slow sense. Dynamic components introduce phase lag: a heater continues transferring stored heat after its power is reduced, a motor retains momentum, and a communication channel delays information. If a correction reaches the system after the original deviation has reversed, it can reinforce the new deviation rather than suppress it. This is the central reason that increasing feedback strength does not always improve performance.\n\nFor a finite-dimensional continuous-time linear model, asymptotic internal stability requires the relevant closed-loop poles to lie strictly in the left half of the complex plane. The characteristic equation is commonly written\n\n\\[1+L(s)=0.\\]\n\nA pole with positive real part produces an exponentially growing component; a complex-conjugate pair can produce growing or decaying oscillations. Poles on the imaginary axis require special treatment and do not establish asymptotic stability. Checking only a simplified input–output expression can be misleading if cancellation conceals an unstable internal mode.\n\nThe Nyquist stability criterion relates closed-loop stability to the frequency response of the open loop and the number of its unstable poles. Its critical point is \\(-1\\), where \\(1+L=0\\). A Bode plot displays loop magnitude and phase against frequency and provides common measures of stability margin. Phase margin measures the additional phase lag needed to reach the critical condition at a unity-gain crossover; gain margin measures the corresponding allowable gain change at a relevant phase crossover. Multiple crossovers and unusual open-loop dynamics require interpretation beyond simple rules of thumb.\n\nA pure time delay \\(\\tau\\) contributes the factor \\(e^{-s\\tau}\\). At angular frequency \\(\\omega\\), its magnitude is one and its phase lag is \\(-\\omega\\tau\\) radians. It therefore reduces phase margin without appearing as an attenuation in the magnitude response. Long pipelines, remote control links, biological transport processes, and digital computation all introduce delays of this kind. A controller that works well on a fast local process may oscillate when applied to an otherwise similar delayed process.\n\nLinear stability analysis is not a complete description of nonlinear behavior. Saturation, friction, dead zones, and switching can generate limit cycles even when a local linear approximation is stable. Conversely, a system may remain bounded despite a locally unstable equilibrium. Practical analysis therefore distinguishes local stability, global behavior, transient overshoot, and robustness to uncertainty rather than treating “stable” as a single comprehensive performance rating."
    },
    {
      "heading": "Electronic amplifiers and precision circuits",
      "text": "Negative feedback became especially important in electrical engineering because electronic amplifiers can provide high gain but have characteristics that vary with manufacturing, temperature, aging, and signal level. Returning a controlled fraction of the output in opposition to the input allows a circuit to exchange some of that gain for more predictable operation. The resulting amplifier can have a gain set primarily by passive components rather than by the poorly controlled gain of individual active devices.\n\nThe operational amplifier is a standard example. In an ideal non-inverting circuit, a feedback divider made from resistances \\(R_f\\) and \\(R_g\\) produces\n\n\\[A_{\\mathrm{cl}}=1+\\frac{R_f}{R_g}.\\]\n\nThe approximation assumes sufficiently large open-loop gain, negligible input current, and operation within voltage, current, and frequency limits. With \\(R_f=9\\,\\mathrm{k\\Omega}\\) and \\(R_g=1\\,\\mathrm{k\\Omega}\\), the nominal closed-loop gain is 10. The amplifier drives its output so that the difference between its two input voltages becomes small; feedback does not physically short the inputs together.\n\nFeedback can also reduce distortion generated inside the loop. If an output stage introduces a small unwanted signal, the loop responds to its effect on the measured output and creates a partly opposing correction. The reduction depends on the loop gain at the distortion frequencies and on the location of the nonlinear element. Distortion introduced after the feedback sampling point is not corrected in the same way, and clipping cannot generally be undone by demanding still more output from an already saturated amplifier.\n\nBandwidth and impedance can also change. For a simple dominant-pole amplifier, lowering closed-loop gain commonly increases useful bandwidth, with the gain–bandwidth product approximately constant over a limited range. Feedback topology determines its effect on terminal impedances: a voltage-series arrangement typically raises input impedance and lowers output impedance. These results are not universal properties of every feedback connection; current sensing and alternative mixing arrangements lead to different impedance transformations.\n\nReal circuits require frequency compensation because transistor stages, wiring, and loads introduce additional poles and zeros. A capacitive load can add enough phase lag to cause ringing or oscillation. Feedback also does not eliminate all electrical noise: resistor noise, input-referred amplifier noise, and reference noise remain subject to their own transfer paths. The relation between feedback, bandwidth, and signal-to-noise ratio therefore depends on the complete circuit rather than on loop gain alone."
    },
    {
      "heading": "Industrial regulation and motion control",
      "text": "Industrial feedback systems regulate quantities such as temperature, pressure, flow, liquid level, speed, and position. Their plants differ substantially: a tank accumulates incoming liquid, a furnace stores thermal energy, and a flexible mechanical assembly may contain several resonant modes. Controller design consequently depends on the dynamics and constraints of the particular process rather than on a universal correction strength.\n\nA widely used controller is the proportional–integral–derivative controller, described in ideal continuous form by\n\n\\[u(t)=K_p e(t)+K_i\\int_0^t e(\\xi)\\,d\\xi +K_d\\frac{de(t)}{dt}.\\]\n\nThe proportional term responds to present error, the integral term accumulates persistent error, and the derivative term responds to its rate of change. In suitable stable systems, integral action eliminates steady error under constant references and disturbances. Derivative action can improve damping, but it amplifies high-frequency measurement noise unless filtered. Many industrial applications use only proportional–integral control.\n\nA motor-speed loop provides a concrete example. An increased mechanical load initially reduces speed. The measured speed deviation causes the controller to increase motor torque, usually through an inner current-control loop. Proportional control alone generally requires a residual speed error to sustain the additional torque command. Integral action can retain the necessary command after the speed error returns to zero. The achievable response remains limited by motor current, available voltage, inertia, and thermal constraints.\n\nLarge installations often use nested or cascade loops. In a process-heating system, an outer temperature controller may command a faster inner steam-flow controller. The inner loop rejects steam-supply disturbances before they produce a large temperature change. Motion systems similarly combine current, velocity, and position loops, with inner loops normally designed to respond faster than outer ones. Such structures are implemented through programmable logic controllers, dedicated drives, or a distributed control system.\n\nPhysical limits create important departures from the ideal equations. If a valve is fully open or a drive has reached its current limit, increasing the controller command cannot produce the expected action. An integral term may nevertheless continue accumulating, a condition called windup, delaying recovery when the error reverses. Anti-windup mechanisms limit or revise that accumulation. In robotics, feedback must additionally account for joint coupling, friction, flexible structures, and changing loads; accurate sensing and a mathematically stable controller cannot remove the consequences of insufficient actuator authority."
    },
    {
      "heading": "Digital implementation, uncertainty, and performance limits",
      "text": "Digital controllers operate on sampled measurements rather than continuously available signals. A typical cycle acquires a sensor value, calculates a control command, and holds that command until the next update. Sampling, computation, and communication introduce dynamics that must be included in the model. For a discrete-time linear system, asymptotic stability requires its closed-loop poles to lie inside the unit circle in the complex \\(z\\)-plane, rather than in the left half of the continuous-time \\(s\\)-plane.\n\nThe sampling frequency must be high enough to represent relevant behavior, but the sampling theorem alone does not establish a satisfactory control rate. A controller needs adequate timing resolution near its operating bandwidth, and unmeasured high-frequency motion can contaminate lower-frequency measurements through aliasing. Analog filtering before conversion can reduce this problem. Quantization introduces finite measurement and command steps, while scheduling jitter makes the effective delay vary between updates.\n\nNetworked systems add packet loss, congestion, and irregular transmission intervals. These issues arise in remote machinery, distributed energy systems, and connected vehicles. A controller may use timestamps, prediction, or local fallback loops, but these techniques depend on assumptions about timing and plant behavior. An average communication delay does not fully characterize the risk if rare delays are much larger than typical ones.\n\nUncertainty places another limit on useful feedback strength. A model may accurately describe slow motion but omit structural resonances, sensor dynamics, or fast electrical effects. Raising loop gain into that poorly modeled frequency range can amplify the omitted behavior and undermine stability. Robust control addresses this issue by specifying classes of model uncertainty and requiring acceptable stability and performance across those classes, rather than only for one nominal plant.\n\nFundamental constraints also prevent arbitrary suppression of error at all frequencies. Under specified assumptions, sensitivity-integral results show that reducing sensitivity in one frequency range must be balanced by increased sensitivity elsewhere. Unstable poles, certain zeros, and time delays impose additional restrictions. This is often called the “waterbed effect.” The relevant integral relation depends on the system’s mathematical properties; it is not a claim that every real controller obeys one identical numerical trade-off.\n\nFeedback design is therefore a joint allocation of bandwidth, noise amplification, disturbance rejection, actuator effort, and stability margin. More precise sensors or faster computation can improve particular limits, but cannot remove all of them simultaneously. A practical closed loop is defined as much by its measurement path and physical constraints as by its control algorithm."
    },
    {
      "heading": "Physiological and cellular regulation",
      "text": "In biology, negative feedback is a major mechanism of homeostasis: the maintenance of internal conditions within ranges compatible with normal function. Biological regulation is usually distributed among receptors, signaling pathways, tissues, and organs rather than concentrated in a single controller. A regulated variable can also change with time of day, developmental stage, or physiological state, so homeostasis should not be equated with an absolutely fixed value.\n\nBlood-glucose regulation illustrates an interacting hormonal network. Increased glucose availability promotes insulin secretion from pancreatic beta cells. Insulin promotes glucose uptake in tissues including skeletal muscle and adipose tissue and suppresses hepatic glucose production, helping reduce the original rise. During fasting, glucagon contributes to increased hepatic glucose output. These processes interact with nutrient signals, neural inputs, and other hormones; they are not adequately represented by a single switch controlling glucose concentration.\n\nThe endocrine system contains prominent multilevel feedback loops. In the hypothalamic–pituitary–thyroid axis, hypothalamic thyrotropin-releasing hormone promotes pituitary thyroid-stimulating hormone secretion, which stimulates thyroid hormone production. Thyroid hormones inhibit upstream secretion and synthesis at the pituitary and hypothalamic levels. Delays in hormone synthesis, transport, and clearance influence the dynamics. Other endocrine axes combine negative feedback with pulsatile release and circadian timing, producing organized fluctuations rather than constant concentrations.\n\nThe baroreflex provides a faster neural example. Stretch-sensitive receptors in the carotid sinus and aortic arch respond to changes in arterial pressure. Their signals alter autonomic output through brainstem circuits, changing heart rate, cardiac contractility, and vascular tone. A pressure rise normally elicits responses that oppose that rise. This reflex is especially important for short-term pressure regulation; longer-term control also involves renal fluid balance, hormonal systems, and adaptation of the receptors themselves.\n\nNegative feedback also operates within cells. In feedback inhibition, a downstream metabolic product reduces the activity of an enzyme earlier in its production pathway. The mechanism can restrict unnecessary synthesis when the product is abundant. Regulatory proteins can likewise reduce their own production by influencing transcription, translation, or degradation. Such interactions connect feedback to gene expression and metabolism.\n\nBiological feedback does not imply a formal comparison with an explicit numerical reference. Molecular binding, transport, and reaction kinetics can produce the same opposing causal structure. Nor does it imply absence of variability: feedback loops can transmit noise, generate oscillations when delayed, and interact with positive-feedback mechanisms to produce switching or adaptation."
    },
    {
      "heading": "Environmental and climate feedbacks",
      "text": "In environmental science, negative feedback describes an induced process that reduces the effect of an initial perturbation. It need not return a system to its original condition, and it does not imply that the total system response is small. An ecosystem or the climate system may contain many negative and positive feedbacks operating on different timescales. Their combined effect depends on their strengths, interactions, and changes in background conditions.\n\nThe central stabilizing response of Earth’s energy balance is the Planck feedback. As the planet warms, it generally emits more thermal radiation to space. Other factors held fixed, that increase opposes the original energy gain. A simplified global energy-balance model can be written\n\n\\[C\\frac{dT}{dt}=F-\\lambda T,\\]\n\nwhere \\(T\\) is a temperature anomaly, \\(C\\) is effective heat capacity per unit area, \\(F\\) is an imposed radiative forcing, and \\(\\lambda>0\\) is a restoring coefficient. For constant forcing in this model, the equilibrium response is \\(T=F/\\lambda\\). The return influence limits warming but does not require the equilibrium anomaly to vanish.\n\nReal climate feedbacks modify that restoring response. Increasing atmospheric water vapor generally amplifies warming because water vapor is a greenhouse gas. Loss of reflective snow and ice can also amplify warming by increasing absorbed sunlight. Cloud changes and changes in the vertical temperature structure contribute additional effects. Climate literature uses more than one sign convention for feedback parameters, so a “negative feedback parameter” may denote a stabilizing radiative response even when the restoring coefficient in the equation above is defined as positive.\n\nThe distinction between forcing and feedback depends on the causal question. An externally imposed increase in atmospheric carbon dioxide can be treated as a forcing, while temperature-driven changes in carbon storage can act as feedbacks. Ocean uptake of heat slows surface warming during adjustment but should not automatically be classified as a radiative feedback. Heat storage, energy loss to space, and carbon exchange are physically different processes even when all influence the observed temperature trajectory.\n\nOn geological timescales, silicate weathering participates in a potential stabilizing carbon-cycle loop. Under suitable conditions, warmer and wetter environments can accelerate reactions that consume carbon dioxide; transport of reaction products and eventual carbonate burial contribute to long-term carbon removal. The strength of this response depends on rock exposure, erosion, runoff, and biological activity. Its long adjustment timescale distinguishes it from rapid atmospheric radiative responses and limits its relevance as an offset to abrupt climate change."
    },
    {
      "heading": "Historical development and unification",
      "text": "Devices exhibiting feedback existed long before a general mathematical theory. Float mechanisms maintained liquid levels, while mechanical governors regulated speed by linking machine motion to the supply of driving energy. During the Industrial Revolution, the centrifugal governor became closely associated with the steam engine. In 1788, James Watt applied this mechanism to steam-engine regulation: rising speed moved rotating masses outward and altered a linkage that reduced steam admission.\n\nThe governor also exposed a theoretical difficulty. A mechanism intended to correct speed could produce “hunting,” or persistent speed fluctuations, when its mechanical dynamics and correction strength were poorly matched. James Clerk Maxwell’s 1868 paper “On Governors” analyzed such behavior with differential equations and connected stable regulation to the mathematical properties of those equations. Edward Routh and Adolf Hurwitz subsequently developed algebraic criteria that became important tools for determining stability.\n\nElectronic communication created a different motivation. Long-distance telephone systems required amplifiers whose gain and distortion remained controlled despite variations in vacuum tubes and operating conditions. Harold Stephen Black conceived the negative-feedback amplifier at Bell Telephone Laboratories in 1927. His paper “Stabilized Feed-Back Amplifiers,” published in 1934, explained how returning part of the output in opposition to the input could improve amplifier behavior. The development helped establish feedback as a deliberate design method rather than merely a property of mechanical regulators.\n\nHarry Nyquist’s 1932 paper “Regeneration Theory” provided a frequency-domain stability criterion suited to feedback amplifiers. Hendrik Bode developed influential methods for analyzing gain, phase, and network design, consolidated in his 1945 book Network Analysis and Feedback Amplifier Design. These methods connected measurable frequency response with closed-loop stability and performance, making them useful well beyond telephone electronics.\n\nDuring World War II, work on servomechanisms, gun directors, radar, and automatic tracking further brought electrical and mechanical feedback problems together. Norbert Wiener’s 1948 book Cybernetics helped extend the language of feedback to communication, organisms, and behavior. The resulting field of cybernetics emphasized similarities among systems with very different physical components.\n\nLater developments in state-space analysis, digital computation, robust control, and nonlinear dynamics expanded the treatment beyond simple single-loop models. These approaches retained the central insight of earlier feedback engineering while making internal states, multiple interacting variables, uncertain models, and computational implementation explicit parts of the analysis."
    }
  ],
  "links": [
    "control-theory",
    "automation",
    "sensor",
    "actuator",
    "positive-feedback",
    "linear-time-invariant-system",
    "transfer-function",
    "bode-plot",
    "electrical-engineering",
    "signal-to-noise-ratio",
    "temperature",
    "pressure",
    "pid-controller",
    "programmable-logic-controller",
    "distributed-control-system",
    "robotics",
    "aliasing",
    "algorithm",
    "homeostasis",
    "endocrine-system",
    "enzyme",
    "protein",
    "gene-expression",
    "metabolism",
    "ecosystem",
    "climate",
    "greenhouse-gas",
    "carbon-dioxide",
    "climate-change",
    "steam-engine",
    "james-watt",
    "world-war-ii",
    "radar",
    "cybernetics"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}