{
  "id": "state-space-model",
  "title": "State-Space Model",
  "lang": "en",
  "source_lang": "en",
  "summary": "A state-space model describes a dynamic system through internal states, their evolution, and their relationship to observable measurements.",
  "category": "mathematics",
  "subcategory": "applied",
  "type": "concept",
  "aliases": [
    "State-space representation",
    "state-space formulation",
    "state-space system"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/state-space-model/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/state-space-model/",
      "json": "https://aiwiki.page/zh/state-space-model.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A state-space model is a mathematical representation of a dynamical system in which an internal state evolves over time and generates observable outputs. The state contains the information needed to determine the system’s subsequent evolution, given future inputs and, in a stochastic model, the relevant probability laws. Its components may represent directly meaningful quantities, such as position and velocity, or unobserved quantities, such as an underlying economic trend. State-space models accommodate deterministic and random dynamics, continuous and discrete time, and linear and nonlinear relationships.\n\nThe framework connects differential equations, control engineering, time-series analysis, and statistical inference. Engineers use it to design feedback controllers and reconstruct unmeasured variables; statisticians use it to separate latent signals from noisy observations and to produce forecasts with quantified uncertainty. Its central distinction is between the evolution of the system itself and the process by which that system is measured. This distinction makes it possible to represent measurement error, incomplete observations, external interventions, and changing dynamics within a common mathematical structure."
    },
    {
      "heading": "States, transitions, and observations",
      "text": "A deterministic continuous-time state-space model typically takes the form\n\n\\[\\dot{x}(t)=f\\bigl(x(t),u(t),t\\bigr),\\qquad y(t)=h\\bigl(x(t),u(t),t\\bigr),\\]\n\nwhere \\(x(t)\\) is the state, \\(u(t)\\) is an externally specified input, and \\(y(t)\\) is the output. The dot denotes differentiation with respect to time. In a finite-dimensional model, these quantities are usually vectors, with \\(x(t)\\in\\mathbb R^n\\), \\(u(t)\\in\\mathbb R^m\\), and \\(y(t)\\in\\mathbb R^p\\). The functions \\(f\\) and \\(h\\) respectively specify the dynamics and the observation relationship. An initial state \\(x(t_0)\\), together with the input trajectory and appropriate existence and uniqueness conditions, determines the subsequent trajectory.\n\nThe defining feature of a state is its role as a complete description of the system’s relevant memory. For a mechanical particle governed by a second-order equation, position alone generally does not determine future motion; position and velocity together do. Introducing velocity as a separate state converts the second-order equation into two first-order equations. More generally, a differential equation of order \\(n\\) can often be rewritten as a first-order system with \\(n\\) state components. States need not coincide with measured outputs, and their choice is not unique.\n\nA discrete-time stochastic model is commonly specified through transition and observation distributions:\n\n\\[x_0\\sim p(x_0),\\qquad x_t\\mid x_{t-1},u_{t-1}\\sim p(x_t\\mid x_{t-1},u_{t-1}),\\]\n\n\\[y_t\\mid x_t,u_t\\sim p(y_t\\mid x_t,u_t).\\]\n\nThe transition equation expresses the Markov property: conditional on the current state and specified inputs, earlier states provide no additional information about the next state. The observation assumption expresses conditional independence: once the current state is known, the observation does not additionally depend on other states or observations. These are modeling assumptions, not automatic properties of every physical measurement process.\n\nWith inputs suppressed, the resulting joint distribution factorizes as\n\n\\[p(x_{0:T},y_{1:T}) =p(x_0)\\prod_{t=1}^{T} p(x_t\\mid x_{t-1})p(y_t\\mid x_t).\\]\n\nHere \\(x_{0:T}\\) denotes the sequence \\(x_0,\\ldots,x_T\\). This factorization makes a stochastic state-space model a particular probabilistic graphical model. If disturbances or observation errors have temporal dependence, additional variables can sometimes be included in the state to restore this structure. Such state augmentation is powerful, but it increases dimension and may complicate estimation."
    },
    {
      "heading": "Linear dynamics and continuous-to-discrete conversion",
      "text": "The most widely studied finite-dimensional form is the linear model\n\n\\[x_{t+1}=A_tx_t+B_tu_t+w_t,\\qquad y_t=C_tx_t+D_tu_t+v_t.\\]\n\nThe matrices \\(A_t,B_t,C_t,D_t\\) describe state transition, input effects, measurement, and direct input-to-output transmission. The disturbances \\(w_t\\) and \\(v_t\\) represent process noise and observation noise. A common specification gives them zero means and covariance matrices \\(Q_t\\) and \\(R_t\\), respectively, with independence across time and between the two noise sequences. Correlated disturbances are possible, but require correspondingly modified inference formulas. A linear model is Gaussian only when its initial state and disturbances have the necessary joint Gaussian distribution.\n\nFor constant matrices and no disturbances, repeated substitution gives\n\n\\[x_t=A^t x_0+ \\sum_{j=0}^{t-1}A^{t-1-j}Bu_j.\\]\n\nThis separates the response to the initial condition from the accumulated effects of inputs. The eigenvalues of \\(A\\) determine the asymptotic behavior of the unforced system. All initial states converge to zero precisely when every eigenvalue lies strictly inside the complex unit circle. In continuous time, for \\(\\dot x=Ax\\), asymptotic stability instead requires strictly negative real parts. Eigenvalues on the boundary require additional analysis, including the sizes of associated Jordan blocks; they cannot be classified solely by their absolute values or real parts.\n\nThe continuous-time linear system\n\n\\[\\dot{x}(t)=Ax(t)+Bu(t)\\]\n\nhas the solution\n\n\\[x(t)=e^{A(t-t_0)}x(t_0) +\\int_{t_0}^{t}e^{A(t-s)}Bu(s)\\,ds,\\]\n\nwhere \\(e^{At}\\) is the matrix exponential. If inputs are held constant over sampling intervals of length \\(\\Delta\\), exact discretization gives\n\n\\[A_d=e^{A\\Delta},\\qquad B_d=\\int_0^\\Delta e^{A\\tau}B\\,d\\tau.\\]\n\nThese formulas avoid treating a continuous-time derivative as merely a finite difference. The frequently used approximation \\(A_d\\approx I+A\\Delta\\) is a first-order numerical approximation and can distort stability when the sampling interval is too large.\n\nNoise must also be discretized consistently. For a linear stochastic differential equation\n\n\\[dx=Ax\\,dt+Bu\\,dt+L\\,dW_t,\\]\n\nwhere \\(W_t\\) is a standard vector Wiener process, the discrete process-noise covariance is\n\n\\[Q_d=\\int_0^\\Delta e^{A\\tau}LL^{\\mathsf T}e^{A^{\\mathsf T}\\tau}\\,d\\tau.\\]\n\nThus sampling changes both deterministic dynamics and uncertainty accumulation. This matters in tracking and sensor fusion, where observations may arrive at unequal intervals and each interval requires its own transition matrix and process covariance."
    },
    {
      "heading": "Observability, controllability, and equivalent representations",
      "text": "Two structural questions distinguish what can be inferred from what can be influenced. Observability asks whether the internal state can be uniquely recovered from ideal, noise-free outputs over a suitable interval when the inputs are known. For a discrete-time, linear time-invariant system with \\(n\\) states, the observability matrix is\n\n\\[\\mathcal O= \\begin{bmatrix} C\\\\CA\\\\ \\vdots\\\\CA^{n-1} \\end{bmatrix}.\\]\n\nThe pair \\((A,C)\\) is observable exactly when \\(\\mathcal O\\) has rank \\(n\\). If the rank is smaller, some distinct initial states produce identical output sequences. Adding more measurements of the same uninformative combination does not eliminate this structural ambiguity.\n\nControllability, in its standard finite-dimensional linear sense, concerns the ability to drive the state between arbitrary values using admissible inputs over a finite interval. For an unconstrained linear time-invariant system, the familiar rank test uses\n\n\\[\\mathcal C= \\begin{bmatrix} B&AB&\\cdots&A^{n-1}B \\end{bmatrix}.\\]\n\nFull row rank establishes reachability from the zero state and, under the usual assumptions, complete state controllability. Input magnitude limits and state constraints introduce additional questions: a rank condition alone does not establish that a maneuver is possible within a specified time or under physical actuator limits.\n\nState coordinates are not intrinsic. If \\(z=Tx\\), where \\(T\\) is invertible, then\n\n\\[A'=TAT^{-1},\\quad B'=TB,\\quad C'=CT^{-1},\\quad D'=D\\]\n\ndescribes the same input-output behavior. This similarity transformation changes the interpretation of individual state components without changing the underlying dynamics. Consequently, an estimated state vector may have no unique component-by-component physical meaning unless the model imposes an interpretable coordinate system.\n\nThe connection to classical frequency-domain analysis is supplied by the transfer function. For continuous time, zero initial state, and suitable transform conditions,\n\n\\[G(s)=C(sI-A)^{-1}B+D.\\]\n\nA single transfer function generally has many state-space realizations. A realization of the smallest possible dimension is called minimal; for ordinary finite-dimensional linear systems, it is minimal precisely when it is both controllable and observable. Nonminimal models can contain internal modes that never affect measured outputs or cannot be excited by the specified inputs.\n\nPractical estimation also depends on conditioning, not just exact rank. A theoretically observable mode may have such a weak output signature that measurement noise makes its reconstruction unreliable. Related weaker properties are detectability, which permits unobservable modes provided they are stable, and stabilizability, which permits uncontrollable modes provided they are stable. These properties are central to the existence of stabilizing estimators and controllers, even when complete observability or controllability is absent."
    },
    {
      "heading": "Filtering and recursive state estimation",
      "text": "Filtering estimates the current state using observations available up to the current time. In probabilistic notation, its target is \\(p(x_t\\mid y_{1:t})\\). The general Bayesian recursion has a prediction step,\n\n\\[p(x_t\\mid y_{1:t-1}) =\\int p(x_t\\mid x_{t-1}) p(x_{t-1}\\mid y_{1:t-1})\\,dx_{t-1},\\]\n\nfollowed by an update,\n\n\\[p(x_t\\mid y_{1:t}) \\propto p(y_t\\mid x_t) p(x_t\\mid y_{1:t-1}).\\]\n\nThe first propagates uncertainty through the dynamics; the second reweights the prediction by the measurement likelihood. These equations are exact, although evaluating their integrals may be computationally difficult.\n\nFor linear-Gaussian models, the Kalman filter carries out this recursion using only a state mean and covariance. With known inputs omitted for clarity, its prediction equations are\n\n\\[\\hat x_{t|t-1}=A_{t-1}\\hat x_{t-1|t-1},\\]\n\n\\[P_{t|t-1} =A_{t-1}P_{t-1|t-1}A_{t-1}^{\\mathsf T}+Q_{t-1}.\\]\n\nThe difference between the actual observation and its prediction,\n\n\\[e_t=y_t-C_t\\hat x_{t|t-1},\\]\n\nis the innovation. Its covariance and the Kalman gain are\n\n\\[S_t=C_tP_{t|t-1}C_t^{\\mathsf T}+R_t,\\qquad K_t=P_{t|t-1}C_t^{\\mathsf T}S_t^{-1}.\\]\n\nThe updated state mean is\n\n\\[\\hat x_{t|t}=\\hat x_{t|t-1}+K_te_t.\\]\n\nThe gain balances predicted state uncertainty against measurement uncertainty. A highly uncertain state and a precise measurement generally produce a larger correction in the directions measured by the sensor. Conversely, an unreliable measurement receives less weight. Under the stated Gaussian assumptions, the updated mean is the exact conditional expectation and minimizes conditional mean squared error. With appropriate second-moment assumptions but without Gaussianity, the same recursion gives the best linear estimator, not necessarily the best estimator among all nonlinear functions of the data.\n\nFinite-precision implementation requires care. Although the covariance update can be written compactly as \\(P_{t|t}=(I-K_tC_t)P_{t|t-1}\\), the Joseph form\n\n\\[P_{t|t}=(I-K_tC_t)P_{t|t-1}(I-K_tC_t)^{\\mathsf T} +K_tR_tK_t^{\\mathsf T}\\]\n\nbetter preserves symmetry and nonnegative definiteness under rounding. Square-root filters instead propagate matrix factors, often using Cholesky decomposition or orthogonal transformations.\n\nMissing observations have a direct interpretation: prediction continues while the unavailable measurement update is omitted. If only some measurement components are missing, the update uses the observed rows of the measurement system and their associated covariance submatrix. This handles incomplete sampling without manufacturing observations, although informative missingness may require an additional model for the observation process itself."
    },
    {
      "heading": "Smoothing, prediction, and uncertainty",
      "text": "Filtering, smoothing, and forecasting differ in the information they condition on. Filtering estimates \\(x_t\\) from observations through time \\(t\\). Smoothing estimates a past state using observations extending beyond it, such as \\(p(x_t\\mid y_{1:T})\\) with \\(T>t\\). Forecasting concerns future states or outputs, such as \\(p(y_{T+h}\\mid y_{1:T})\\). These tasks can use the same model but answer different questions. A historical estimate revised after later measurements is a smoothed estimate, not a real-time filtered estimate.\n\nFor linear-Gaussian models, the Rauch–Tung–Striebel smoother follows a forward Kalman-filter pass with a backward recursion. Define\n\n\\[J_t=P_{t|t}A_t^{\\mathsf T}P_{t+1|t}^{-1}.\\]\n\nThen\n\n\\[\\hat x_{t|T} =\\hat x_{t|t} +J_t\\bigl(\\hat x_{t+1|T}-\\hat x_{t+1|t}\\bigr),\\]\n\n\\[P_{t|T} =P_{t|t} +J_t\\bigl(P_{t+1|T}-P_{t+1|t}\\bigr)J_t^{\\mathsf T}.\\]\n\nThe backward correction transmits information from future observations to earlier states. Under fixed, correctly specified parameters, the smoothed covariance cannot exceed the filtered covariance in the positive-semidefinite ordering. This is an uncertainty statement, not a claim that every realized smoothed point estimate must be closer to the true state.\n\nFor a time-invariant model with no future inputs, the \\(h\\)-step state forecast is\n\n\\[\\hat x_{T+h|T}=A^h\\hat x_{T|T},\\]\n\nwith covariance\n\n\\[P_{T+h|T} =A^hP_{T|T}(A^{\\mathsf T})^h +\\sum_{j=0}^{h-1}A^jQ(A^{\\mathsf T})^j.\\]\n\nThe first term propagates uncertainty about the present state; the second accumulates future process noise. For the observation equation \\(y=Cx+v\\), the forecast covariance is \\(CP_{T+h|T}C^{\\mathsf T}+R\\). Omitting \\(R\\) gives uncertainty about the latent signal rather than about a future noisy observation.\n\nFor stable \\(A\\), the accumulated process uncertainty can converge to a finite limit satisfying\n\n\\[P_\\infty=AP_\\infty A^{\\mathsf T}+Q.\\]\n\nFor a random-walk state, uncertainty instead grows with the forecast horizon. This difference explains why a model with an evolving trend can produce widening long-range intervals even when its recent observations are highly precise.\n\nStandard filtering and smoothing covariances condition on the model parameters. If those parameters have been estimated from a short record, plug-in intervals can understate total uncertainty. Bayesian inference accounts for parameter uncertainty by integrating over a posterior distribution, while simulation-based frequentist procedures can approximate its effect through repeated fitting. The distinction is especially consequential when estimating weakly identified trend variances, persistent dynamics, or long-horizon forecasts."
    },
    {
      "heading": "Parameter estimation and system identification",
      "text": "The state trajectory and the model parameters are different unknowns. A Kalman filter estimates states conditional on matrices and noise covariances; estimating those matrices and covariances is a separate problem. In engineering, this broader task is called system identification. Parameters can be derived from physical laws, estimated from observed inputs and outputs, or determined through a combination of physical constraints and statistical fitting. A physically specified model may still contain unknown masses, damping coefficients, sensor biases, or disturbance intensities.\n\nFor Gaussian models, the innovations provide an efficient likelihood. If the parameter vector is \\(\\theta\\), the log-likelihood is\n\n\\[\\ell(\\theta) =-\\frac12\\sum_{t=1}^{T} \\left[ p_t\\log(2\\pi)+\\log\\det S_t +e_t^{\\mathsf T}S_t^{-1}e_t \\right],\\]\n\nwhere \\(p_t\\) is the number of observed measurement components at time \\(t\\), and both \\(e_t\\) and \\(S_t\\) depend on \\(\\theta\\). Maximum likelihood estimation therefore combines a numerical optimizer with repeated filter evaluations. Covariance parameters are often represented through triangular factors or transformed scales so that optimization respects their required constraints.\n\nThe expectation–maximization algorithm treats the unobserved states as missing data. Its expectation step uses smoothing to compute conditional moments such as\n\n\\[E[x_t\\mid y_{1:T}],\\quad E[x_tx_t^{\\mathsf T}\\mid y_{1:T}],\\quad E[x_tx_{t-1}^{\\mathsf T}\\mid y_{1:T}].\\]\n\nIts maximization step updates parameters using these quantities. In unrestricted linear-Gaussian models, many updates resemble multivariate least-squares calculations. EM has a nondecreasing likelihood under exact implementation, but it may converge slowly and does not guarantee the global maximum of a nonconcave likelihood.\n\nIdentification places limits on what fitting can recover. Similarity-related state realizations are observationally equivalent unless constraints select particular coordinates. Process noise and measurement noise may also be difficult to distinguish from short or weakly informative records. When inputs are used to identify dynamics, their variation must sufficiently excite the relevant modes. Data collected under feedback can introduce dependencies between inputs and disturbances that invalidate naive regression approaches.\n\nModel assessment examines more than fitted likelihood. Under a correctly specified Gaussian model, appropriately standardized innovations should have the assumed distribution and no remaining serial dependence. Residual patterns can reveal omitted seasonality, incorrect transition dynamics, or a mischaracterized observation process. State dimension and structural complexity can be compared using likelihood penalties and temporally ordered cross-validation. Ordinary random splitting can leak future information into training or destroy the forecasting task’s temporal structure. Near-zero variance estimates and nearly redundant states require particular attention because they can indicate boundary solutions or unnecessary model components."
    },
    {
      "heading": "Structural time-series models and latent components",
      "text": "State-space models provide a direct way to express observations as combinations of evolving components. A basic local-level model is\n\n\\[y_t=\\mu_t+\\varepsilon_t,\\qquad \\mu_{t+1}=\\mu_t+\\eta_t,\\]\n\nwhere the independent disturbances have variances \\(r\\) and \\(q\\). The latent level \\(\\mu_t\\) follows a random walk, while measurement noise represents short-lived irregular variation. If \\(q=0\\), the level is constant; as \\(q/r\\) increases, the estimated level becomes more responsive to recent observations. The distinction between a changing signal and noisy measurement is therefore encoded through separate stochastic mechanisms.\n\nA local linear trend adds a changing slope:\n\n\\[\\mu_{t+1}=\\mu_t+\\beta_t+\\eta_t,\\qquad \\beta_{t+1}=\\beta_t+\\zeta_t,\\]\n\n\\[y_t=\\mu_t+\\varepsilon_t.\\]\n\nThis does not force the entire series to follow a single straight line. Instead, the slope can evolve gradually, with its variability controlled by the disturbance variance of \\(\\zeta_t\\). Damped-trend variants replace the unit coefficient on \\(\\beta_t\\) by a parameter of magnitude below one. Such a change alters long-range extrapolation substantially, because recent growth is no longer projected indefinitely at an unchanged rate.\n\nSeasonal components can be represented through a state vector of seasonal effects subject to a sum-to-zero structure, or through pairs of states that rotate by seasonal frequencies. A deterministic harmonic pair uses a transition matrix\n\n\\[\\begin{bmatrix} \\cos\\omega&\\sin\\omega\\\\ -\\sin\\omega&\\cos\\omega \\end{bmatrix}.\\]\n\nAdding process disturbances allows seasonal amplitude and phase to evolve. Several harmonic pairs can represent complex seasonal patterns, linking this construction to Fourier series without requiring seasonal behavior to remain fixed throughout the record.\n\nAutoregressive and autoregressive moving-average models also admit state-space representations. For example,\n\n\\[y_t=\\phi_1y_{t-1}+\\phi_2y_{t-2}+\\varepsilon_t\\]\n\ncan use the state \\(x_t=(y_t,y_{t-1})^{\\mathsf T}\\), a companion transition matrix, and a disturbance entering only the first coordinate. This representation has a singular process covariance, which is legitimate. Additional measurement error can be included when the recorded series is itself noisy, producing a distinction absent from the basic autoregression.\n\nDynamic regression treats regression coefficients as states:\n\n\\[y_t=z_t^{\\mathsf T}\\beta_t+\\varepsilon_t,\\qquad \\beta_{t+1}=\\beta_t+\\eta_t.\\]\n\nHere \\(z_t\\) contains observed predictors, and the coefficient vector can change over time. Dynamic-factor models similarly use a low-dimensional latent process to explain many observed series. Applications in econometrics include trend-cycle decomposition and forecasting from indicators released at different times. The model does not, by itself, give latent components a causal interpretation: a mathematically well-defined “trend” remains dependent on the chosen transition and disturbance assumptions."
    },
    {
      "heading": "Nonlinear, non-Gaussian, and switching models",
      "text": "Many systems require nonlinear transitions or observation mechanisms. A general formulation is\n\n\\[x_{t+1}=f(x_t,u_t)+w_t,\\qquad y_t=h(x_t)+v_t,\\]\n\nbut additive noise is not essential; transition and observation distributions can be specified directly. A Poisson observation with mean \\(\\exp(x_t)\\), for example, can describe counts driven by a changing latent log-intensity. A Bernoulli observation with a logistic transformation of the state can represent a time-varying event probability. These models retain the state-space conditional-independence structure while replacing Gaussian measurement assumptions.\n\nThe extended Kalman filter linearizes nonlinear functions near the current estimate. Its covariance calculations use the Jacobian matrices of the transition and observation functions. This can work well when uncertainty is concentrated in a region where the functions are approximately linear. Strong curvature, ambiguous measurements, or poor initialization can cause substantial errors or divergence. The unscented Kalman filter instead propagates a carefully selected set of representative points through the nonlinear functions to approximate transformed means and covariances. Neither method generally reproduces a multimodal posterior faithfully.\n\nA particle filter represents the filtering distribution with weighted samples,\n\n\\[p(x_t\\mid y_{1:t}) \\approx\\sum_{i=1}^{N} w_t^{(i)}\\delta_{x_t^{(i)}}.\\]\n\nParticles are propagated through a proposal distribution, reweighted using observations, and commonly resampled to concentrate computation in plausible regions. In the bootstrap particle filter, the state transition itself supplies the proposal. Particle methods can handle non-Gaussian and multimodal distributions, but poor proposals or highly informative measurements can concentrate almost all weight on very few particles. In high-dimensional states, this degeneracy can make straightforward implementations prohibitively expensive.\n\nA finite-state hidden Markov model is a closely related construction in which the latent state takes values in a finite set. Terminology varies: some fields reserve “state-space model” for continuous-valued states, while others include hidden Markov models within it. A switching state-space model combines a discrete regime with continuous states, allowing different transition matrices or noise levels in different regimes. Exact inference can involve a rapidly growing mixture of Gaussian components, motivating approximation, pruning, or simulation.\n\nVariational inference, Laplace approximations, and Markov chain Monte Carlo offer further approaches to latent trajectories and parameter uncertainty. Their suitability depends on state dimension, record length, smoothness, and whether inference must operate online. In models that are conditionally linear-Gaussian, some components can be integrated out analytically while particles represent the remaining variables. This Rao–Blackwellization reduces sampling variability by exploiting structure rather than simulating every unknown quantity."
    },
    {
      "heading": "Feedback control and scientific applications",
      "text": "In feedback control, the state-space model connects estimation to action. If the state is observed, a linear feedback law\n\n\\[u_t=-Kx_t\\]\n\nchanges the closed-loop transition matrix from \\(A\\) to \\(A-BK\\). Under controllability assumptions, pole-placement methods choose \\(K\\) to assign desired closed-loop eigenvalues. Other methods choose feedback to balance competing objectives. The linear-quadratic regulator minimizes a quadratic cost penalizing state deviations and input effort, with feedback gains obtained through a Riccati equation. The weighting matrices specify the trade-off rather than emerging from the state-space representation alone.\n\nWhen the state is not fully measured, a controller can use an estimated state. In the standard linear-quadratic-Gaussian problem, appropriate assumptions yield the separation principle: an optimal state estimator and an optimal full-state feedback controller can be designed separately and combined. This is a special result, not a general property of uncertain nonlinear control. In nonlinear or partially observed problems, an action may change both the physical state and the information available about it, so estimation and control can become inseparable.\n\nModel predictive control uses a state-space model to predict trajectories over a finite horizon, solves a constrained optimization problem, applies the first selected input, and repeats after receiving new measurements. Unlike unconstrained linear-quadratic regulation, it can explicitly accommodate actuator saturation, operating limits, and coupled input constraints. Its performance depends on model accuracy, state estimation, optimization speed, and the treatment of disturbances. The framework is widely used in industrial process control, where temperature, flow, and concentration may evolve on different timescales.\n\nTracking illustrates how physical reasoning determines states and noise. A one-dimensional constant-velocity model with interval \\(\\Delta\\) uses\n\n\\[x_t= \\begin{bmatrix}p_t\\\\v_t\\end{bmatrix}, \\qquad A= \\begin{bmatrix} 1&\\Delta\\\\0&1 \\end{bmatrix}.\\]\n\nA position sensor has measurement matrix \\(C=[1\\;\\;0]\\). Velocity is not measured directly but can be inferred from successive positions. If acceleration is modeled as continuous white noise with spectral intensity \\(q\\), its exact discrete covariance is\n\n\\[Q=q \\begin{bmatrix} \\Delta^3/3&\\Delta^2/2\\\\ \\Delta^2/2&\\Delta \\end{bmatrix}.\\]\n\nThe off-diagonal terms encode the fact that acceleration uncertainty affects position and velocity together.\n\nScientific uses extend beyond engineered control. Ecological models distinguish latent population abundance from imperfect detection; geophysical data assimilation combines evolving physical models with observations; and physiological models represent hidden processes measured indirectly by sensors. In each case, the observation equation is scientifically consequential. An apparent change in measured abundance, for example, may arise from a change in detection rather than population dynamics. State-space modeling makes those alternatives explicit, but separating them requires informative data and defensible structural assumptions."
    },
    {
      "heading": "Historical development and modern computational architectures",
      "text": "State-based descriptions have roots in classical mechanics, where positions and momenta specify a system’s instantaneous condition. The twentieth-century state-space approach developed alongside electronic computation and multivariable control, offering a time-domain alternative to methods centered on transfer functions. Rudolf E. Kálmán’s 1960 paper on linear filtering introduced the recursive discrete-time estimator now bearing his name. Kálmán and Richard S. Bucy developed the corresponding continuous-time filtering theory in 1961. Their work connected dynamic models, uncertainty propagation, and estimation in a form suited to sequential computation.\n\nThe approach became influential in aerospace guidance and navigation, where measurements arrive sequentially and essential quantities cannot all be sensed directly. Observability, controllability, and minimal-realization theory established a systematic language for internal dynamics and input-output equivalence. In statistics, state-space methods provided a common representation for latent-component time series and recursive forecasting. The 1965 Rauch–Tung–Striebel smoothing method and later computational developments expanded the framework from real-time estimation to retrospective trajectory reconstruction and likelihood-based analysis.\n\nGeneral nonlinear filtering remained more difficult because probability distributions no longer closed under prediction and updating. Sequential Monte Carlo methods became especially influential after the 1993 bootstrap-filter work of Neil Gordon, David Salmond, and Adrian Smith. Their practical appeal grew with computing power, while approximate Gaussian methods remained important in applications requiring low latency. Modern implementations combine statistical algorithms with sparse computations, matrix factorizations, and automatic differentiation to fit large models or embed known physical dynamics in trainable systems.\n\nIn machine learning, the phrase “state-space model” also describes sequence-processing architectures built around recurrences such as\n\n\\[h_t=\\bar A h_{t-1}+\\bar B u_t,\\qquad z_t=Ch_t+Du_t.\\]\n\nFor a time-invariant linear recurrence with zero initial state, this is equivalent to a causal convolution whose kernel is\n\n\\[k_j=C\\bar A^j\\bar B.\\]\n\nThis equivalence allows different computational strategies: recurrent updates for streaming inference and convolution-based evaluation for full sequences. The hidden state in such an architecture need not represent a calibrated probability distribution, and its training objective may be predictive rather than generative.\n\nStructured State Space sequence models, commonly called S4, were introduced in 2021 by Albert Gu, Karan Goel, and Christopher Ré, using structured parameterizations to support long-range sequence computation. Selective state-space models developed this direction further; Mamba, introduced by Gu and Tri Dao in 2023, made parts of the recurrence depend on the current input. Such input dependence generally removes equivalence to one fixed convolution kernel while enabling content-dependent state updates. These architectures share state-transition mathematics with classical control and statistical models, but they do not automatically provide Bayesian filtering, physical interpretability, or uncertainty estimates. Their computational performance depends on the parameterization, state dimension, hardware implementation, and sequence-processing task."
    }
  ],
  "links": [
    "dynamical-system",
    "differential-equation",
    "time-series",
    "markov-property",
    "conditional-independence",
    "probabilistic-graphical-model",
    "matrix",
    "covariance-matrix",
    "eigenvalues-and-eigenvectors",
    "exponential-function",
    "stochastic-differential-equation",
    "wiener-process",
    "observability",
    "matrix-rank",
    "controllability",
    "matrix-similarity",
    "transfer-function",
    "kalman-filter",
    "conditional-expectation",
    "mean-squared-error",
    "cholesky-decomposition",
    "bayesian-inference",
    "posterior-distribution",
    "likelihood-function",
    "maximum-likelihood-estimation",
    "expectation-maximization-algorithm",
    "statistical-identification",
    "cross-validation",
    "random-walk",
    "fourier-series",
    "econometrics",
    "poisson-distribution",
    "jacobian-matrix",
    "hidden-markov-model",
    "variational-inference",
    "markov-chain-monte-carlo",
    "model-predictive-control",
    "automatic-differentiation",
    "machine-learning",
    "convolution"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}