Statistic¶

A measurable function of the observed sample alone, with no dependence on unknown population parameters, used to summarize data or support estimation and testing.

Version: v1 · 2026-08-30 · History
Domain-specific #: 2845
Origin domain: statistics
Subdomain: mathematical statistics
Aliases: Sample statistic, Statistical quantity

Core Idea¶

In mathematical statistics, a statistic is a measurable function of a sample that does not depend on unknown population parameters. If the sample is a random vector \(X=(X_1,\ldots,X_n)\) on sample space \(\mathcal X\), a statistic is a measurable mapping

\[ T:\mathcal X\to\mathcal T, \]

and \(T(X)\) is a random variable or random element before observation. Once the realized sample is \(x\), the observed statistic is the value \(T(x)\). Casella and Berger emphasize the parameter-free sample-function requirement because an expression containing an unknown \(\theta\) cannot be calculated from the data alone and is not a statistic in that model.^[1]

Statistics can summarize location, spread, order, association, or shape; estimate parameters; define rejection regions; or reduce data while retaining selected information. The identity is deliberately narrower than ordinary “statistics” meaning numerical facts or the discipline as a whole. It is also broader than aggregation: an order statistic, empirical distribution function, or full sample identity map can be a statistic even when it does not collapse everything to one scalar.

Structural Signature¶

Sampling model: a family \(\{P_\theta:\theta\in\Theta\}\) on a declared sample space.
Random sample: \(X\in\mathcal X\), possibly scalar, vector, matrix, path, or other measurable object.
Mapping rule: one fixed function \(T\) applied to every admissible sample.
Measurability: \(T(X)\) is a legitimate random element with a distribution under each \(P_\theta\).
Parameter independence: the formula for \(T\) contains no unknown population parameter.
Sampling distribution: \(P_\theta\circ T^{-1}\) describes variation across repeated samples and may depend on \(\theta\).
Observed value: \(T(x)\) is computed after data realization.
Declared purpose: description, estimation, testing, diagnostics, or information reduction.
Information boundary: different samples can map to the same value unless \(T\) is injective.

The statistic may have a scalar, vector, function, set, or graph as its value. “Numerical summary” is common but not defining.

What It Is Not¶

It is not a population parameter. The sample mean \(\bar X\) is a statistic; the population mean \(\mu\) is a parameter. It is not automatically an estimator: a statistic becomes an estimator when assigned the role of estimating a parameter. It is not the observed number alone, because the sampling behavior belongs to the mapping \(T(X)\), not only its realized value.

It is not necessarily sufficient, unbiased, consistent, complete, ancillary, robust, or efficient. Those are properties relative to a model and purpose. A pivotal quantity can fail to be a statistic if it contains an unknown parameter, even though its distribution is parameter-free. A test statistic is a statistic used within a hypothesis-testing rule, not a separate mathematical species.

Scope of Application¶

Statistics mediate between raw samples and inference. Descriptive statistics report features of observed data. Estimators use statistics to infer unknown parameters. Test procedures order evidence through a statistic and compare its observed value with a null distribution. Confidence procedures often invert distributions of statistics or pivots.

Data reduction is especially important. A sufficient statistic retains all model-relevant information about a parameter in the factorization or conditional-distribution sense, while an ancillary statistic has a distribution independent of that parameter. These labels cannot be assigned from the formula alone; they require a model.^[1]

The abstraction covers non-scalar statistics: the empirical distribution function \(F_n(t)\), an entire vector of order statistics, a contingency table, or a covariance matrix.^[2] It also covers statistics from dependent data when the sample space and distribution are declared. It does not turn arbitrary database metrics into formal statistics unless their sampling interpretation and rule are specified.

Clarity¶

For real observations \(X_1,\ldots,X_n\), the sample mean

\[ \bar X=\frac1n\sum_{i=1}^nX_i \]

is a statistic. Its formula uses \(n\) and observed sample values but no unknown population mean. Its distribution can still depend on \(\mu\) and \(\sigma^2\); parameter independence concerns the mapping rule, not its law.

The sample variance

\[ S^2=\frac1{n-1}\sum_{i=1}^n(X_i-\bar X)^2 \]

is also a statistic. By contrast,

\[ \frac{\bar X-\mu}{S/\sqrt n} \]

is not a statistic when \(\mu\) is unknown, although under a normal model it is a pivot expression used to derive procedures. Replacing \(\mu\) by a specified null value \(\mu_0\) produces a computable test statistic for that null.

Manages Complexity¶

A sample may be high-dimensional, unordered, noisy, and inconvenient for inference. A statistic selects a function of that sample that exposes a feature or retains information relevant to a task. The mapping makes repeated-sampling behavior analyzable and can greatly reduce storage or computation.

Reduction entails possible information loss. Two samples with the same mean can have different spreads, tails, and dependence. Calling the mean “the data” commits a category error. Sufficiency theory asks when a reduction loses no information about a parameter under a model; robustness asks how sensitive a statistic is to deviations or outliers. The abstraction makes these questions possible without predetermining their answers.

Abstract Reasoning¶

The sample-to-statistic map pushes each model distribution \(P_\theta\) forward to a sampling distribution \(P_\theta^T=P_\theta\circ T^{-1}\). Statistical reasoning can therefore occur on the smaller space \(\mathcal T\) while remaining connected to the original experiment.

This perspective separates three layers: the unknown data-generating parameter, the random sample, and the statistic. Conflating them leads to errors such as treating an estimator as fixed before sampling or treating a parameter-dependent pivot as directly observable.

Knowledge Transfer¶

The mapping concept transfers to machine-learning features, signal summaries, quality-control indices, and survey estimators. Transfer is legitimate when the input sample, rule, and sampling model are explicit. A “dashboard statistic” may merely be a deterministic database aggregation with no stated repeated-sampling claim; it shares form but not the full inferential frame.

Vector- and function-valued statistics also show that transfer should preserve codomain. Reducing every statistic to a number erases empirical processes, sufficient sigma-fields, and structured summaries. The correct abstraction is measurable sample function, not scalar aggregation.

Examples¶

Sample maximum: \(X_{(n)}=\max_iX_i\) is an order statistic useful for endpoint models.
Count: \(T=\sum_i\mathbf1\{X_i\in A\}\) is a binomial count under independent Bernoulli membership assumptions.
Empirical distribution: \(F_n(t)=n^{-1}\sum_i\mathbf1\{X_i\le t\}\) is a function-valued statistic.
Sufficient statistic: for independent Poisson observations, \(\sum_iX_i\) is sufficient for the rate under the standard model.
Not a statistic: \((\bar X-\mu)/\sigma\) is not computable when \(\mu,\sigma\) are unknown.
Observed value: reporting \(S^2=4.2\) identifies a realization; uncertainty calculations concern the random variable \(S^2\).

Structural Tensions¶

Function vs. realization. \(T(X)\) varies before observation while \(T(x)\) is fixed afterward. Diagnostic: state whether discussing the statistic or its observed value.
Parameter-free rule vs. parameter-dependent distribution. A statistic's law generally depends on \(\theta\). Diagnostic: inspect the formula separately from the sampling distribution.
Reduction vs. information loss. Compact summaries can discard task-relevant structure. Diagnostic: justify reduction through sufficiency, risk, or the declared descriptive purpose.
Description vs. inference. The same statistic can serve different roles. Diagnostic: name the estimator, test, or descriptive use rather than infer it from the formula.
Formal vs. ordinary language. “A statistic” can mean a published number colloquially. Diagnostic: require a declared sample-function identity in this node.
Autonomous statistic vs. generic mapping or aggregation. Every statistic is a sample-space mapping, and many statistics aggregate observations, but neither neighbor requires a parameter-free measurable rule evaluated on a random sample with a sampling distribution. Diagnostic: retain this node only when the object is defined as (T(X)), with the rule independent of unknown parameters and the random-before-observed distinction load-bearing; otherwise route the case to Function Mapping or Aggregation.

Structural–Framed Character¶

The structural core is a single-valued mapping from sample to derived object. The statistics frame supplies sampling models, measurability, unknown parameters, sampling distributions, and inferential roles. Without that frame one has Function Mapping or Aggregation, not the technical statistic identity.

The candidate is domain-specific. Its parameter-independence condition and random-before-observed distinction are not generic requirements of every function.

Structural Core vs. Domain Accent¶

Structural core: declared input space, fixed measurable rule, output space, repeated applications, and information selection.

Domain accent: random sample, parameterized distributions, observed values, estimators, tests, sufficiency, ancillary structure, and sampling uncertainty.

Statistic is a strict specialization of Function (Mapping): it has a sample-space domain, statistic-space codomain, and a single-valued measurable rule, with the additional restriction that the rule not depend on unknown parameters. Aggregation is a frequent but nonuniversal case; Statistical Inference is a downstream use. Function Mapping is the minimal parent that remains literal for scalar, vector, function-valued, injective, and reductive statistics alike.

Relationships to Other Abstractions¶

Current abstraction Statistic Domain-specific

Parents (1) — more general patterns this builds on

Statistic is a kind of Function (Mapping) Prime

Statistic is a strict specialization of Function (Mapping): it has a sample-space domain, statistic-space codomain, and a single-valued measurable rule, with the additional restriction that the rule not depend on unknown parameters.

Hierarchy path (1) — routes to 1 parentless root

Statistic → Function (Mapping)

Neighborhood in Abstraction Space¶

Statistic sits in a sparse region of the domain-specific corpus (90^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Statistical Tests & Distribution Calibration (7 abstractions)

Nearest neighbors

Fisher Consistency — 0.80
Independent and Identically Distributed Random Variables — 0.80
Empirical Measure — 0.79
Learnable Function Class — 0.78
Universal Hypothesis Testing — 0.78

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Parameter: fixed but unknown feature of the population or model.
Estimator: statistic assigned to estimate a parameter.
Estimate: observed numerical value of an estimator.
Test statistic: statistic used to order evidence in a testing procedure.
Pivotal quantity: expression with parameter-free distribution that may itself contain unknown parameters.
Sufficient statistic: statistic satisfying an information-preservation criterion relative to a model.
Statistics discipline: field studying data and uncertainty, not one sample function.

References¶

[1] George Casella and Roger L. Berger, Statistical Inference, 2^nd ed., Duxbury, 2002, ISBN 978-0-534-24312-8. registry ↩a ↩b

[2] Larry Wasserman, All of Statistics: A Concise Course in Statistical Inference, Springer, 2004, DOI: 10.1007/978-0-387-21736-9. registry ↩