{
  "id": "markov-random-field",
  "title": "Markov random field",
  "lang": "en",
  "source_lang": "en",
  "summary": "A Markov random field models joint dependence through an undirected graph whose missing edges encode conditional independences.",
  "category": "mathematics",
  "subcategory": "probability-statistics",
  "type": "concept",
  "aliases": [
    "Markov network",
    "undirected graphical model",
    "undirected Markov model"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/markov-random-field/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/markov-random-field/",
      "json": "https://aiwiki.page/zh/markov-random-field.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A Markov random field (MRF) is a collection of random variables whose conditional dependence structure is described by an undirected graph. Each vertex represents a variable, and the graph specifies which groups of variables become independent when others are known. The essential principle is local: conditional on an appropriate neighborhood, a variable does not depend on variables farther away. Unlike a Markov chain, a random field need not be indexed by time or arranged along a line; its variables may occupy a lattice, a geographical map, a molecular structure, or an arbitrary network.\n\nMarkov random fields connect probability, graph theory, and statistical mechanics. They provide models for interacting physical systems, spatial observations, image pixels, and structured outputs in machine learning. Their mathematical importance lies in the relationship between local conditional independences and a global distribution assembled from local interaction terms. Their practical difficulty is that simple local specifications can produce complex global behavior, making normalization, inference, and parameter estimation computationally demanding."
    },
    {
      "heading": "Graphs and conditional independence",
      "text": "Let \\(G=(V,E)\\) be an undirected graph, and let \\(X=(X_v)_{v\\in V}\\) be variables indexed by its vertices. For \\(A\\subseteq V\\), write \\(X_A=(X_v)_{v\\in A}\\), and let \\(N(v)\\) denote the neighbors of \\(v\\). A joint probability distribution for \\(X\\) is Markov with respect to \\(G\\) when it satisfies suitable graphical conditional independence statements. Three standard formulations distinguish pairwise, local, and global properties.\n\nThe pairwise Markov property states that nonadjacent variables are conditionally independent given all remaining variables:\n\n\\[X_u\\perp X_v\\mid X_{V\\setminus\\{u,v\\}} \\qquad\\text{if }\\{u,v\\}\\notin E.\\]\n\nThe local Markov property states that the neighbors of a vertex shield it from the rest of the graph:\n\n\\[X_v\\perp X_{V\\setminus(\\{v\\}\\cup N(v))} \\mid X_{N(v)}.\\]\n\nThus, the full conditional distribution of \\(X_v\\), given every other variable, depends only on its neighbors. In an image grid with four-neighbor adjacency, for example, a pixel variable is conditionally independent of all other pixels once its immediate horizontal and vertical neighbors are known.\n\nThe global Markov property uses graph separation. For disjoint vertex sets \\(A,B,S\\), the set \\(S\\) separates \\(A\\) from \\(B\\) if every path from \\(A\\) to \\(B\\) meets \\(S\\). The property requires\n\n\\[X_A\\perp X_B\\mid X_S\\]\n\nwhenever this separation holds. Global Markov implies local Markov, which implies pairwise Markov. The converse implications require additional distributional assumptions; they hold, in particular, for strictly positive distributions on finite product state spaces. Positivity is important because conditional independence can behave differently when some configurations have probability zero.\n\nAn edge does not necessarily imply a nonzero marginal correlation, nor does a missing edge imply marginal independence. On the chain \\(A-B-C\\), the variables \\(A\\) and \\(C\\) are separated by \\(B\\), so \\(A\\perp C\\mid B\\), but they can still be strongly dependent before conditioning. Conversely, a graph may include an unnecessary edge whose associated interaction is actually absent. A graph is an independence map when its separation statements are valid for the distribution; it is a perfect map only when it captures all relevant conditional independences.\n\nThe word “Markov” therefore describes a conditional dependence structure rather than a temporal evolution. A finite random field can be defined on any undirected graph without assuming spatial coordinates, stationarity, or identical distributions across vertices. The graph is also not, by itself, a causal diagram: an edge records a possible conditional association, not a direction of influence or the result of an intervention."
    },
    {
      "heading": "Clique factorization and Gibbs distributions",
      "text": "A clique is a set of vertices every pair of which is connected by an edge. A distribution factorizes over an undirected graph if it can be written as\n\n\\[p(x)=\\frac{1}{Z}\\prod_{C\\in\\mathcal C}\\psi_C(x_C),\\]\n\nwhere \\(\\mathcal C\\) is a collection of cliques, \\(\\psi_C\\) are nonnegative potential functions, and \\(Z\\) normalizes the distribution. Factors over smaller cliques can be absorbed into factors over maximal cliques, so either convention is possible. A potential is not generally a probability distribution: it need not sum to one and does not have to describe a conditional probability.\n\nFor discrete variables, the partition function is\n\n\\[Z=\\sum_x\\prod_{C\\in\\mathcal C}\\psi_C(x_C).\\]\n\nFor continuous variables, the sum is replaced by an integral with respect to a specified product reference measure. A model defines a probability distribution only when \\(0<Z<\\infty\\). Finite discrete state spaces make finiteness straightforward, but continuous models may be improper even when every individual factor has an apparently reasonable form.\n\nThe Hammersley–Clifford theorem supplies the central bridge between graphical independence and factorization. In its standard finite discrete form, a strictly positive distribution is Markov with respect to an undirected graph if and only if it factorizes over that graph’s cliques. Strict positivity means \\(p(x)>0\\) for every configuration in the product state space. Without it, graphical Markov properties need not imply clique factorization. Conversely, a well-defined clique-factorized distribution satisfies the global Markov property even when some factors are zero.\n\nFor positive potentials, define clique energies \\(U_C=-\\log\\psi_C\\). The same model becomes\n\n\\[p(x)=\\frac{1}{Z}\\exp[-E(x)], \\qquad E(x)=\\sum_{C\\in\\mathcal C}U_C(x_C).\\]\n\nThis is a Gibbs distribution. In physical notation, an inverse temperature \\(\\beta\\) is often displayed explicitly as \\(p(x)\\propto\\exp[-\\beta H(x)]\\). Statistical models commonly absorb \\(\\beta\\) into their parameters. Adding a constant to the total energy does not change the normalized distribution, and interaction terms can sometimes be redistributed among factors. Consequently, potentials are not uniquely determined by the joint distribution.\n\nFactorization explains why local conditional distributions are comparatively simple. Holding all other variables fixed gives\n\n\\[p(x_v\\mid x_{V\\setminus\\{v\\}}) \\propto \\prod_{C:\\,v\\in C}\\psi_C(x_C).\\]\n\nEvery factor not involving \\(v\\) cancels. The remaining factors involve only \\(v\\) and its neighbors, producing the local Markov property. This cancellation also underlies local simulation and estimation methods. However, constructing a model from local conditional distributions requires care: an arbitrary collection of conditional formulas need not be compatible with any common joint distribution."
    },
    {
      "heading": "Discrete interaction models",
      "text": "The best-known discrete example is the Ising model. Each variable is a spin \\(x_i\\in\\{-1,+1\\}\\), with distribution\n\n\\[p(x)=\\frac{1}{Z} \\exp\\left( \\sum_{\\{i,j\\}\\in E}J_{ij}x_ix_j +\\sum_i h_ix_i \\right).\\]\n\nThe parameter \\(J_{ij}\\) controls the interaction between neighboring spins, while \\(h_i\\) is a local field. With this sign convention, positive \\(J_{ij}\\) favors agreement and negative \\(J_{ij}\\) favors disagreement. The variables need not represent physical spins: they may instead denote binary labels, occupied sites, or two-state decisions in a statistical model.\n\nThe one-site conditional distribution follows directly from the energy:\n\n\\[\\Pr(X_i=+1\\mid x_{-i}) = \\frac{1}{ 1+\\exp\\left[-2\\left(h_i+\\sum_{j\\in N(i)}J_{ij}x_j\\right)\\right] }.\\]\n\nIt therefore has a logistic form. A network of pairwise interactions produces local conditional models resembling logistic regression, but with compatibility constraints: in the joint Ising model, the interaction coefficient associated with an undirected edge is shared between its two endpoints. Independently fitting arbitrary logistic regressions does not automatically produce a consistent joint model.\n\nThe Potts model extends the agreement interaction to \\(q\\) labels. A common statistical form is\n\n\\[p(x)\\propto \\exp\\left( \\sum_{\\{i,j\\}\\in E}\\beta_{ij}\\mathbf 1\\{x_i=x_j\\} +\\sum_i a_i(x_i) \\right).\\]\n\nHere \\(\\mathbf 1\\{\\cdot\\}\\) is an indicator function, and \\(a_i(k)\\) supplies a site-specific preference for label \\(k\\). Positive \\(\\beta_{ij}\\) rewards equal neighboring labels. This makes the model useful for spatial classification, where adjacent locations are expected to share categories but local observations provide competing evidence. For two labels, a suitable recoding relates the Potts and Ising parameterizations.\n\nNot every discrete MRF is pairwise. If a graph contains a three-vertex clique, its potential can include an interaction that cannot be expressed as a sum of unary and pairwise terms. Conversely, on a triangle-free graph, every clique has size at most two, so a positive distribution Markov with respect to that graph has a pairwise factorization. Adding auxiliary variables can sometimes replace higher-order interactions with pairwise ones on an enlarged graph, but marginalizing those variables can restore higher-order dependence among the original variables.\n\nZero-valued potentials express hard constraints, such as forbidding adjacent occupied vertices. They can make the model’s support disconnected under single-site changes. This distinction matters computationally: a sampler that changes one variable at a time may be unable to move between valid configurations even though a normalized joint distribution exists. Strong but finite penalties instead retain positive probability for every configuration, although movement between competing high-probability regions may still be extremely slow."
    },
    {
      "heading": "Gaussian fields and sparse precision matrices",
      "text": "A Gaussian Markov random field is a random vector with a multivariate normal distribution whose conditional independences are represented by a graph. For a nonsingular covariance matrix \\(\\Sigma\\), define the precision matrix \\(Q=\\Sigma^{-1}\\). Its density is\n\n\\[p(x)= \\frac{\\det(Q)^{1/2}}{(2\\pi)^{n/2}} \\exp\\left[ -\\frac12(x-\\mu)^\\mathsf TQ(x-\\mu) \\right].\\]\n\nThe graph is encoded by zeros in \\(Q\\), rather than by zeros in the covariance matrix. Specifically,\n\n\\[X_i\\perp X_j\\mid X_{V\\setminus\\{i,j\\}} \\quad\\Longleftrightarrow\\quad Q_{ij}=0.\\]\n\nThis exact equivalence is a distinctive advantage of Gaussian models.\n\nThe full conditional distribution is univariate normal, with\n\n\\[\\operatorname E[X_i\\mid X_{-i}] = \\mu_i-\\frac{1}{Q_{ii}} \\sum_{j\\ne i}Q_{ij}(X_j-\\mu_j), \\qquad \\operatorname{Var}(X_i\\mid X_{-i})=\\frac{1}{Q_{ii}}.\\]\n\nOnly entries corresponding to neighbors contribute to the conditional mean. The conditional association can also be expressed by the partial correlation\n\n\\[\\rho_{ij\\cdot V\\setminus\\{i,j\\}} = -\\frac{Q_{ij}}{\\sqrt{Q_{ii}Q_{jj}}}.\\]\n\nA sparse precision matrix can coexist with a dense covariance matrix: local direct dependence can generate long-range marginal dependence throughout a connected graph.\n\nConditional specifications must respect symmetry and positive definiteness. Suppose a spatial model specifies conditional means \\(\\sum_j b_{ij}x_j\\) and conditional variances \\(\\tau_i^2\\), with zero overall mean. Compatibility requires\n\n\\[Q_{ii}=\\tau_i^{-2}, \\qquad Q_{ij}=-b_{ij}\\tau_i^{-2},\\]\n\nso \\(b_{ij}/\\tau_i^2=b_{ji}/\\tau_j^2\\). A proper nonsingular joint Gaussian distribution additionally requires \\(Q\\) to be a positive-definite matrix. Plausible-looking neighborhood regressions can violate either condition.\n\nA common construction uses an adjacency matrix \\(W\\) and the diagonal degree matrix \\(D\\). A precision proportional to \\(D-\\rho W\\) yields local spatial dependence when the parameter range makes it positive definite. At the boundary \\(\\rho=1\\), the graph Laplacian \\(D-W\\) is singular: constant vectors have zero quadratic penalty on each connected component. The corresponding intrinsic Gaussian field is not a normalized density on the full Euclidean space. It can be treated through constraints, such as a zero-sum condition, or as an improper component within a model whose posterior is proper.\n\nThe principal computational benefit is sparsity. Cholesky decomposition of \\(Q\\) supports Gaussian simulation, likelihood evaluation, and solutions of linear systems without explicitly forming a dense inverse. Elimination can nevertheless create additional nonzero entries, called fill-in, so computational cost depends on graph structure and vertex ordering. Gaussian fields are therefore both probabilistic models and a major application of sparse numerical linear algebra."
    },
    {
      "heading": "Infinite lattices, boundary conditions, and phase transitions",
      "text": "On an infinite graph, a global expression \\(p(x)=Z^{-1}e^{-E(x)}\\) is generally not a literal probability density obtained by summing over all infinite configurations. The total energy may diverge, and there is no uniform counting probability measure on the configuration space. Infinite-volume models are instead specified through conditional distributions on finite regions, given the configuration outside each region.\n\nFor a finite vertex set \\(\\Lambda\\), let \\(\\eta\\) denote an exterior configuration. A local Gibbs specification has the form\n\n\\[\\gamma_\\Lambda(x_\\Lambda\\mid\\eta) = \\frac{1}{Z_\\Lambda(\\eta)} \\exp[-\\beta H_\\Lambda(x_\\Lambda\\mid\\eta)],\\]\n\nwhere \\(H_\\Lambda\\) includes the interactions involving the region and its boundary. An infinite-volume Gibbs measure is a probability measure whose finite-region conditional distributions agree with these kernels. These consistency requirements are known as the Dobrushin–Lanford–Ruelle equations. Their formulation uses conditional probabilities rather than a single finite global normalizer.\n\nBoundary conditions can affect the interior even when the boundary is moved arbitrarily far away. Consider the ferromagnetic Ising model with no external field. One may take increasingly large boxes and fix all exterior spins to \\(+1\\), or instead to \\(-1\\). At sufficiently high temperature, the limiting distributions coincide in standard lattice settings. At sufficiently low temperature in dimensions at least two, the two constructions can yield distinct infinite-volume measures. The same local interaction rule then supports different macroscopic phases.\n\nFor the nearest-neighbor Ising model on the infinite square lattice, with uniform coupling energy \\(J>0\\), the critical inverse temperature satisfies\n\n\\[\\beta_cJ=\\frac12\\log(1+\\sqrt2).\\]\n\nAbove this value of \\(\\beta J\\), corresponding to lower temperature, plus and minus phases exist at zero external field. The result highlights the difference between a local specification and a uniquely determined global law. On the one-dimensional nearest-neighbor chain, by contrast, there is no phase transition at positive temperature for finite coupling.\n\nA finite-state model on a finite graph has a partition function that is a finite sum of positive analytic terms in finite real-valued parameters. It therefore does not exhibit the same thermodynamic nonanalyticity as an infinite-volume phase transition. It may nevertheless show sharp crossovers, multimodality, and very slow simulation. Near criticality, correlations extend over long distances, so local algorithms can require many updates to move information across the system.\n\nThese issues also clarify the meaning of locality. A local Markov property does not guarantee weak long-range dependence, rapid decay of correlations, or uniqueness of an infinite-volume distribution. Strong interactions can transmit dependence through arbitrarily long chains of intermediate vertices. Conditions ensuring uniqueness or decay depend on interaction strength, graph connectivity, state spaces, and the precise family of potentials."
    },
    {
      "heading": "Exact and approximate inference",
      "text": "Inference asks questions about an MRF after its structure and parameters have been specified. Common tasks include computing marginal probabilities, evaluating the partition function, drawing samples, and finding a most probable configuration. These are distinct problems. Maximum a posteriori estimation seeks\n\n\\[x^\\star=\\arg\\max_x p(x)=\\arg\\min_x E(x),\\]\n\nwhereas marginal inference sums or integrates over configurations. The most probable joint configuration need not equal the vector obtained by choosing each variable’s most probable marginal state.\n\nOn a tree, belief propagation computes exact marginals through local messages. For a pairwise discrete model, a message from \\(i\\) to \\(j\\) is\n\n\\[m_{i\\to j}(x_j) = \\sum_{x_i} \\psi_i(x_i)\\psi_{ij}(x_i,x_j) \\prod_{k\\in N(i)\\setminus\\{j\\}}m_{k\\to i}(x_i).\\]\n\nAfter messages have passed in both directions, multiplying the local potential by incoming messages gives each vertex marginal up to normalization. With \\(q\\) states per vertex, the arithmetic cost is typically \\(O(|E|q^2)\\). Related recursions recover the partition function or a maximizing configuration.\n\nFor graphs with cycles, variable elimination remains exact but creates intermediate factors involving multiple variables. Its cost is governed by the size of these factors, closely related to treewidth. A junction-tree construction groups variables into larger clusters and performs exact message passing between them. For bounded state size, the worst-case complexity is exponential in treewidth rather than necessarily in the total number of vertices. Sparse graphs can still have large treewidth, so having few edges does not by itself ensure easy inference.\n\nMarkov chain Monte Carlo provides a different approach. Gibbs sampling repeatedly draws variables or blocks from their full conditional distributions. Under suitable irreducibility and recurrence conditions, its stationary distribution is the target MRF. Local conditional calculations avoid the global partition function, but strong dependence, hard constraints, and separated modes can make mixing slow. Cluster updates can greatly improve sampling for some spin systems, while methods such as tempering address barriers by considering distributions at different temperatures.\n\nApproximate deterministic methods trade computational cost against accuracy. Variational inference searches within a tractable family, often minimizing a Kullback–Leibler divergence. Mean-field approximations use product distributions and can miss strong dependencies. Loopy belief propagation applies tree-style updates on graphs with cycles; it may work well, but convergence and correctness are not guaranteed in general. For MAP inference, special binary pairwise energies satisfying a submodularity condition can be minimized exactly by graph cuts. General discrete MRF inference remains computationally intractable in the worst case."
    },
    {
      "heading": "Parameter estimation and graph learning",
      "text": "Learning an MRF can mean estimating interaction strengths on a known graph, identifying which edges should be present, or learning both structure and parameters. In an exponential-family parameterization,\n\n\\[p_\\theta(x)= \\exp\\{\\theta^\\mathsf T T(x)-A(\\theta)\\}, \\qquad A(\\theta)=\\log Z(\\theta),\\]\n\nwhere \\(T(x)\\) collects interaction statistics. For \\(m\\) independent observations, the log likelihood is\n\n\\[\\ell(\\theta)= \\sum_{r=1}^{m}\\theta^\\mathsf TT(x^{(r)}) -mA(\\theta).\\]\n\nDifferentiation gives the moment-matching equation\n\n\\[\\nabla\\ell(\\theta) = \\sum_{r=1}^{m}T(x^{(r)}) -m\\,\\operatorname E_\\theta[T(X)].\\]\n\nThus maximum likelihood estimation repeatedly requires expectations under the model. The second derivative of \\(A\\) is the covariance matrix of the sufficient statistics, making \\(A\\) convex. In an identifiable exponential family, the likelihood has favorable optimization geometry, yet each gradient evaluation may require difficult inference. Monte Carlo likelihood methods approximate model expectations by simulation; variational methods replace them with tractable approximations, potentially changing the resulting estimates.\n\nPseudolikelihood replaces the joint likelihood with a product of full conditional probabilities:\n\n\\[\\operatorname{PL}(\\theta;x) = \\prod_{v\\in V} p_\\theta(x_v\\mid x_{V\\setminus\\{v\\}}).\\]\n\nBecause global normalization cancels from each conditional, this objective is often much cheaper to evaluate. For an Ising model, its terms are logistic likelihood contributions. Pseudolikelihood is not the true joint likelihood, but it can yield consistent estimators under suitable sampling and regularity conditions. Statistical uncertainty must account for its composite nature rather than treating all conditional terms as independent observations.\n\nGraph learning introduces additional difficulties. In high-dimensional settings, there may be more potential edges than observations can reliably identify. Nodewise regression with sparsity penalties estimates each variable’s neighborhood; separately estimated neighborhoods must then be reconciled, for example by taking their union or intersection. Recovery guarantees require assumptions about signal strength, dependence, sample size, and the structure of the underlying graph. A fitted missing edge is therefore a model-based conditional independence claim, not an unconditional empirical fact.\n\nFor Gaussian data, a prominent method is the graphical lasso. Given a sample covariance matrix \\(S\\), it estimates a sparse precision matrix by solving\n\n\\[\\widehat Q = \\arg\\min_{Q\\succ0} \\left[ -\\log\\det Q+\\operatorname{tr}(SQ) +\\lambda\\sum_{i\\ne j}|Q_{ij}| \\right].\\]\n\nThe penalty encourages off-diagonal zeros and hence a sparse graph. This is a convex optimization problem, but the selected graph depends on the penalty, the sampling assumptions, and preprocessing. Latent variables can induce additional observed dependencies: marginalizing an unobserved vertex often creates edges among its neighbors. Consequently, a learned graph over measured variables need not reproduce the graph of a larger underlying system."
    },
    {
      "heading": "Applications, related models, and historical development",
      "text": "Statistical mechanics supplied the original setting for many random-field models. Ernst Ising’s 1925 analysis established the absence of a positive-temperature transition in the one-dimensional nearest-neighbor model, while Lars Onsager’s 1944 solution of the zero-field two-dimensional model demonstrated the richness of lattice interactions. The name “Markov random field” later emphasized conditional locality rather than a particular physical interpretation. The Markov–Gibbs connection associated with John Hammersley and Peter Clifford dates to their 1971 work, and Julian Besag’s 1974 paper on spatial interaction helped establish these models in spatial statistics.\n\nImage analysis became a major application during the 1980s. Pixels or image regions can serve as vertices, with potentials representing local appearance and compatibility between neighboring labels. Stuart and Donald Geman’s 1984 work connected image restoration, Gibbs distributions, and stochastic relaxation. In a simple denoising model, latent clean pixels have an MRF prior distribution, while observed noisy pixels contribute likelihood factors. The resulting posterior distribution combines data fit with spatial regularity. Pairwise smoothing can suppress isolated errors, although overly strong agreement terms can also remove small structures and blur boundaries.\n\nSpatial statistics uses related constructions for observations on administrative regions, ecological sites, and irregular maps. Gaussian conditional autoregressive models describe residual dependence through geographic adjacency; discrete models represent spatially varying categories or counts through appropriate local specifications. In hierarchical models, observed data can be conditionally independent given a latent random field even though they are dependent marginally. Applications to physical fields also exploit the relationship between sparse precision operators and discretized differential operators. Certain Gaussian processes can be approximated by Gaussian Markov fields derived through stochastic partial differential equations and finite-element discretization.\n\nWithin probabilistic graphical models, MRFs are contrasted with a Bayesian network, whose directed acyclic graph factorizes a distribution into conditional probabilities. Undirected models naturally represent symmetric compatibility constraints; directed models naturally express ordered conditional generation. Their graphical independence languages are not identical. For example, an unshielded directed collider can represent independences that no undirected graph represents exactly. Moralizing a directed graph produces an undirected representation useful for inference, but can lose some independence information.\n\nA conditional random field directly models a structured output \\(Y\\) given observations \\(X=x\\):\n\n\\[p_\\theta(y\\mid x) = \\frac{1}{Z_\\theta(x)} \\prod_C\\psi_C(y_C,x).\\]\n\nIts normalizer depends on the observed input, and its graphical Markov assumptions concern the output variables conditional on that input. Introduced under that name by John Lafferty, Andrew McCallum, and Fernando Pereira in 2001, conditional random fields became important in sequence labeling and other structured prediction tasks. A linear-chain version permits dynamic programming much like a hidden Markov model, while allowing potentials to depend on extensive observed context without specifying a generative distribution for the observations."
    }
  ],
  "links": [
    "random-variable",
    "markov-chain",
    "probability",
    "graph-theory",
    "statistical-mechanics",
    "machine-learning",
    "joint-probability-distribution",
    "conditional-independence",
    "correlation",
    "statistical-independence",
    "partition-function",
    "integral",
    "boltzmann-distribution",
    "ising-model",
    "logistic-function",
    "logistic-regression",
    "indicator-function",
    "multivariate-normal-distribution",
    "covariance-matrix",
    "positive-definite-matrix",
    "sparse-matrix",
    "cholesky-decomposition",
    "maximum-a-posteriori-estimation",
    "belief-propagation",
    "markov-chain-monte-carlo",
    "gibbs-sampling",
    "variational-inference",
    "kullback-leibler-divergence",
    "maximum-likelihood-estimation",
    "convex-optimization",
    "statistics",
    "prior-distribution",
    "posterior-distribution",
    "gaussian-process",
    "finite-element-method",
    "probabilistic-graphical-model",
    "bayesian-network",
    "conditional-random-field",
    "hidden-markov-model"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}