{
  "id": "value-of-information",
  "title": "Value of Information",
  "lang": "en",
  "source_lang": "en",
  "summary": "Value of information measures the expected improvement in a decision’s outcome from obtaining additional information before choosing an action.",
  "category": "mathematics",
  "subcategory": "applied",
  "type": "concept",
  "aliases": [
    "Information value",
    "Decision-theoretic value of information",
    "VOI"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/value-of-information/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/value-of-information/",
      "json": "https://aiwiki.page/zh/value-of-information.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Value of information is the expected improvement in the outcome of a decision that becomes possible when additional information is available before an action is chosen. In decision theory, it compares the best decision achievable with existing knowledge against the best decisions that could be made after observing new evidence. Information is valuable insofar as it helps distinguish between actions with different consequences—not simply because it reduces uncertainty. The concept underlies the evaluation of experiments, forecasts, surveys, inspections, and other information-gathering activities."
    },
    {
      "heading": "Decision-theoretic formulation",
      "text": "Consider a decision-maker choosing an action \\(a\\) from a set \\(A\\). The consequences depend on an uncertain state \\(\\theta\\), represented by a random variable with a prior distribution \\(p(\\theta)\\). Let \\(u(a,\\theta)\\) denote the utility of taking action \\(a\\) when the state is \\(\\theta\\).\n\nWithout additional information, the optimal expected utility is\n\n\\[V_0=\\max_{a\\in A}\\mathbb{E}_{\\theta}[u(a,\\theta)],\\]\n\nwhere \\(\\mathbb{E}\\) denotes expected value. The decision is an instance of mathematical optimization under uncertainty.\n\nSuppose a signal or dataset \\(Y\\) can be observed before the action is chosen. Its relationship to the state is specified by a sampling model \\(p(y\\mid\\theta)\\). After observing \\(Y=y\\), the decision-maker uses Bayesian inference to obtain the posterior distribution \\(p(\\theta\\mid y)\\) and chooses an action maximizing posterior expected utility.\n\nBefore the signal is observed, the expected value of this decision procedure is\n\n\\[V_Y= \\mathbb{E}_{Y} \\left[ \\max_{a\\in A} \\mathbb{E}[u(a,\\theta)\\mid Y] \\right].\\]\n\nThe gross expected value of the information is\n\n\\[\\operatorname{VOI}(Y)=V_Y-V_0.\\]\n\nThe outer expectation uses the predictive distribution of possible observations, obtained by combining the prior distribution and the sampling model. Thus, value of information is an ex ante quantity: it evaluates the opportunity to learn before the evidence is known.\n\nThe placement of the maximization is essential. Without the signal, one action must be chosen for all possible observations. With the signal, the action can depend on what is observed. Formally, the latter problem optimizes over decision rules \\(\\delta\\) mapping observations to actions:\n\n\\[V_Y=\\sup_{\\delta}\\mathbb{E}[u(\\delta(Y),\\theta)].\\]\n\nThese expressions assume that the expectations exist and that suitable decision rules are available. Suprema replace maxima when optimal actions are not attained."
    },
    {
      "heading": "Principal forms",
      "text": "Expected value of perfect information\n\nThe expected value of perfect information, abbreviated EVPI, measures the improvement available if the uncertain state relevant to the decision were revealed completely before an action was chosen:\n\n\\[\\operatorname{EVPI} = \\mathbb{E}_{\\theta} \\left[ \\max_{a\\in A}u(a,\\theta) \\right] - \\max_{a\\in A} \\mathbb{E}_{\\theta}[u(a,\\theta)].\\]\n\nThe first term allows a different best action for every state. The second requires a single action before the state is known.\n\nPerfect information concerns the modeled uncertainty. Learning model parameters exactly does not necessarily reveal every future random outcome. For example, knowing a machine’s failure probability does not establish whether it will fail on a particular occasion. An EVPI calculation must therefore specify what is assumed to become known.\n\nExpected value of sample information\n\nThe expected value of sample information, or EVSI, evaluates an imperfect observation, such as a sample, experiment, or noisy forecast:\n\n\\[\\operatorname{EVSI}(Y) = \\mathbb{E}_{Y} \\left[ \\max_{a\\in A} \\mathbb{E}[u(a,\\theta)\\mid Y] \\right] - V_0.\\]\n\nEVSI is therefore the general signal-based value of information applied to a specified study or observation process. It depends on both the decision problem and the information-generating mechanism.\n\nThe same dataset can have high value for one decision and little value for another. Likewise, a highly accurate forecast can have no decision value if its possible results would never change the preferred action.\n\nExpected value of partial perfect information\n\nSuppose the uncertain state consists of two components, \\(\\theta=(\\phi,\\psi)\\). The expected value of partial perfect information, commonly abbreviated EVPPI, measures the benefit of learning \\(\\phi\\) exactly while uncertainty about \\(\\psi\\) remains:\n\n\\[\\operatorname{EVPPI}(\\phi) = \\mathbb{E}_{\\phi} \\left[ \\max_{a\\in A} \\mathbb{E}_{\\psi\\mid\\phi}[u(a,\\phi,\\psi)] \\right] - V_0.\\]\n\nThe conditional distribution of \\(\\psi\\) matters: revealing \\(\\phi\\) may also change beliefs about \\(\\psi\\) when the two are dependent.\n\nEVPPI can identify which groups of uncertain parameters are worth investigating. If a proposed study depends on the full state only through \\(\\phi\\), its EVSI cannot exceed the EVPPI of \\(\\phi\\). That bound requires the study not to provide additional information about other components beyond what learning \\(\\phi\\) would provide."
    },
    {
      "heading": "Worked example",
      "text": "A firm can either undertake a project or decline it. Declining yields zero. Undertaking yields a profit of 100 monetary units if the project succeeds and a loss of 40 if it fails. The prior probability of success is \\(0.30\\).\n\nThe expected payoff from undertaking the project is\n\n\\[0.30(100)+0.70(-40)=2.\\]\n\nThe firm therefore undertakes it, giving a baseline expected payoff of \\(V_0=2\\).\n\nWith perfect information, the firm would undertake only successful projects. Its expected payoff would be\n\n\\[V_{\\mathrm{PI}}=0.30(100)+0.70(0)=30,\\]\n\nso\n\n\\[\\operatorname{EVPI}=30-2=28.\\]\n\nNow suppose an assessment produces a positive or negative signal, with\n\n\\[P(+\\mid\\text{success})=0.80, \\qquad P(+\\mid\\text{failure})=0.10.\\]\n\nUsing conditional probability, the probability of a positive signal is\n\n\\[P(+)=0.30(0.80)+0.70(0.10)=0.31.\\]\n\nBy Bayes’ theorem, the posterior probability of success after a positive signal is \\(0.24/0.31\\), approximately \\(0.774\\). After a negative signal, it is \\(0.06/0.69\\), approximately \\(0.087\\).\n\nThe firm should undertake the project after a positive signal and decline after a negative one. The expected payoff of this rule, before receiving the signal, is\n\n\\[V_Y=0.24(100)+0.07(-40)=21.2.\\]\n\nConsequently,\n\n\\[\\operatorname{EVSI}=21.2-2=19.2.\\]\n\nIf the assessment costs 5 monetary units and has no other effects, its net expected value is \\(19.2-5=14.2\\). These monetary comparisons assume that maximizing expected money is the appropriate decision objective."
    },
    {
      "heading": "Fundamental properties",
      "text": "Nonnegativity and upper bounds\n\nFor a single decision-maker who may freely ignore information,\n\n\\[\\operatorname{VOI}(Y)\\geq 0.\\]\n\nThe original action remains available after observing the signal, so optimizing with the additional option to condition on it cannot reduce expected utility.\n\nA formal argument uses conditional expectation:\n\n\\[\\mathbb{E}_{Y} \\left[ \\max_a\\mathbb{E}[u(a,\\theta)\\mid Y] \\right] \\geq \\max_a \\mathbb{E}_{Y} \\left[ \\mathbb{E}[u(a,\\theta)\\mid Y] \\right] = V_0.\\]\n\nWhen perfect information reveals the complete payoff-relevant state,\n\n\\[0\\leq\\operatorname{EVSI}(Y)\\leq\\operatorname{EVPI}.\\]\n\nNonnegativity concerns the gross value before acquisition costs. Costs, delays, or restrictions associated with information can make its net value negative.\n\nZero value despite learning\n\nInformation can change beliefs substantially without improving a decision. Its value is zero if the same action remains optimal for almost every possible signal realization.\n\nFor example, evidence about a small component of operating costs may substantially reduce uncertainty about that component but leave the preferred production method unchanged. Conversely, a modest amount of evidence can have considerable value when it frequently changes a choice between actions with sharply different consequences.\n\nValue of information therefore depends on the payoff structure, available actions, and current beliefs—not on uncertainty alone.\n\nMore informative experiments\n\nOne experiment is more informative than another in the Blackwell sense if the less informative observation can be generated from the more informative one by an additional random transformation, often called garbling.\n\nBecause a decision-maker observing the richer signal can reproduce the poorer signal and follow its decision rule, the richer experiment cannot have lower gross value in any decision problem satisfying the standard assumptions. Under suitable regularity conditions, a converse characterization also holds: comparison across all decision problems characterizes this informational ordering.\n\nThis gives a decision-based account of informativeness, distinct from ranking signals by a single statistical accuracy measure.\n\nAdditional information need not have diminishing value\n\nIf earlier evidence is already available, the value of a new signal must be measured relative to decisions using that earlier evidence.\n\nAlthough freely available additional evidence cannot lower optimal expected utility, its marginal value need not decline. Signals can be substitutes, where one makes another less useful, or complements, where neither is useful alone but together they identify an appropriate action. Diminishing returns require additional assumptions about the decision and observation models."
    },
    {
      "heading": "Relationship to statistical information",
      "text": "Value of information differs from information measures used in information theory.\n\nEntropy describes uncertainty in a probability distribution. Mutual information measures statistical dependence between variables. Neither ordinarily incorporates the actions or consequences defining a particular decision.\n\nThis distinction has several implications:\n\n- A signal can have positive mutual information about the state but zero decision value.\n- Two signals with similar predictive accuracy can have different values if they distinguish different payoff-relevant states.\n- Reducing variance is not necessarily valuable when the decision depends on another feature of the distribution.\n- Fisher information concerns the sensitivity of a sampling distribution to parameter changes, rather than directly measuring improved decisions.\n\nThere are important connections under particular objectives. With logarithmic loss and unrestricted probabilistic predictions, the expected reduction in optimal predictive loss equals mutual information. Such identities depend on the specified decision problem and do not apply to arbitrary utilities."
    },
    {
      "heading": "Loss-based formulation",
      "text": "In statistics, decisions are often expressed through a loss function \\(L(a,\\theta)\\), which is minimized rather than maximized.\n\nThe minimum prior expected loss is\n\n\\[R_0=\\min_a\\mathbb{E}[L(a,\\theta)].\\]\n\nWith a signal \\(Y\\), the expected minimum posterior loss is\n\n\\[R_Y= \\mathbb{E}_{Y} \\left[ \\min_a\\mathbb{E}[L(a,\\theta)\\mid Y] \\right].\\]\n\nThe value of information is then\n\n\\[\\operatorname{VOI}(Y)=R_0-R_Y.\\]\n\nThis is the reduction in Bayes risk. It allows information value to be expressed in monetary losses, prediction errors, classification penalties, or other consequences appropriate to the application."
    },
    {
      "heading": "Applications and computation",
      "text": "Value-of-information analysis is used in experimental design to compare studies by their expected effect on decisions rather than solely by precision or statistical power. It can help determine which uncertainties to investigate, which observations to collect, and how much sampling is worthwhile.\n\nIn operations research, applications include inspections, demand forecasting, inventory decisions, and resource exploration. In machine learning, related calculations support the selection of observations or labels whose acquisition could improve later decisions.\n\nSequential problems require accounting for both current rewards and future learning. In a partially observable Markov decision process, actions can affect both the underlying system and the information available for subsequent choices.\n\nSmall discrete problems can often be solved by enumeration. Continuous or high-dimensional models may require numerical integration, Monte Carlo methods, or approximation. A direct simulation approach samples possible datasets, computes a posterior for each, solves the corresponding decision problem, and averages the resulting optimized values.\n\nThis nested procedure can be expensive. Noisy estimates inside a maximization can also create upward bias in estimated optimized values. Reliable computation therefore requires attention to optimization error, simulation uncertainty, and approximation quality."
    },
    {
      "heading": "Costs, interpretation, and limitations",
      "text": "The relevant quantity for purchasing or collecting information is usually its net value. If utility is measured in money and acquisition has a fixed monetary cost \\(c\\),\n\n\\[\\operatorname{NetVOI}(Y)=\\operatorname{VOI}(Y)-c.\\]\n\nWith nonlinear utility for money, subtracting a cash price from a utility-valued VOI is not valid. Instead, willingness to pay is determined by incorporating payment into the outcomes and comparing expected utilities.\n\nA broader evaluation may also include delay, processing costs, implementation costs, privacy effects, or opportunity costs. If an experiment changes the system being studied, those consequences must be represented alongside its informational benefit.\n\nResults are conditional on the model. They depend on prior beliefs, the observation process, the available actions, and the utility or loss assigned to outcomes. Sensitivity analysis can examine how valuations change when these assumptions vary, but a precise calculation does not eliminate model uncertainty.\n\nThe usual nonnegativity result also assumes a single decision-maker with unchanged options and the ability to ignore evidence. In game theory, disclosing information can alter other participants’ strategies, so an individual’s payoff or aggregate welfare need not improve. Information can likewise have adverse consequences when disclosure is compulsory or changes contractual, legal, or institutional constraints. These situations require modeling how information changes the decision environment, rather than treating it only as an optional signal."
    }
  ],
  "links": [
    "decision-theory",
    "random-variable",
    "prior-distribution",
    "expected-value",
    "mathematical-optimization",
    "bayesian-inference",
    "posterior-distribution",
    "conditional-probability",
    "bayes-theorem",
    "conditional-expectation",
    "information-theory",
    "information-entropy",
    "mutual-information",
    "variance",
    "fisher-information",
    "statistics",
    "loss-function",
    "experimental-design",
    "statistical-power",
    "operations-research",
    "machine-learning",
    "partially-observable-markov-decision-process",
    "monte-carlo-method",
    "sensitivity-analysis",
    "game-theory"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}