Fisher information¶

The expected squared score, or negative expected log-likelihood curvature under regularity conditions, measuring local sensitivity of a probability model to its parameter.

Version: v2 · 2026-09-19 · History
Domain-specific #: 4548
Domain group: Formal Sciences
Origin domain: Experimental Design & Statistics
Subdomain: Mathematical Statistics and Information Geometry → Experimental Design & Statistics

Core Idea¶

Fisher information determines Cramer-Rao bounds, asymptotic estimator covariance, experimental design criteria, Jeffreys priors, local distinguishability, and a Riemannian metric while transforming covariantly under reparameterization. Differentiate log likelihood with respect to the parameter; averaging the score outer product under the model yields information, and independent observations add their contributions because scores sum and cross terms vanish under regularity. The abstraction is therefore identified by a declared carrier, a transformation or constraint over that carrier, and an invariant that tells an analyst whether the named structure is genuinely present.

How would you explain it like I'm…

Clue Strength Score

Imagine a hidden dial, and every clue you collect helps you guess where it is set. Some clues change a lot when the dial moves even a tiny bit, so they help you tell nearby settings apart. Fisher information is a score for how much a clue does that. More clues add up to more help.

How Much Data Pins Down

Scientists often want to learn a hidden number, like how often a coin lands heads, from data. Fisher information measures how strongly the chances of seeing your data depend on that hidden number. If a small change in the number would make your data look very different, the data carries a lot of information, and you can pin the number down precisely. Independent measurements each add their own information to the total. This helps scientists plan experiments that teach them the most.

Likelihood Sensitivity Measure

Fisher information comes from a statistical model: a rule giving the probability of the data for each value of an unknown parameter. Take the log of that probability and see how steeply it changes as you nudge the parameter; that slope is called the score. Fisher information is the average squared score, averaged over the data the model expects. High information means the likelihood is sharply peaked, so the parameter can be estimated precisely, and it sets a limit on how small an unbiased estimator's variance can be (the Cramer-Rao bound). Independent observations add their information. It differs from 'observed information', which uses the curvature for the actual data you got rather than the model average.

For a model p(x|θ) with respect to a dominating measure, the score is the gradient of the log-likelihood in θ. Under regularity conditions (you can differentiate under the integral, and the support does not depend on θ), the score has mean zero, and Fisher information is its covariance: the expected outer product of the score, or equivalently minus the expected Hessian of the log-likelihood. For n independent observations the information is additive because the scores add and their cross terms have zero expectation. The inverse information gives the Cramer-Rao lower bound for unbiased estimators and the asymptotic covariance of the maximum-likelihood estimator. It also drives optimal experimental design criteria, defines the Jeffreys prior through the square root of its determinant, and serves as a Riemannian metric measuring how distinguishable nearby models are. Under reparameterization it transforms covariantly through the Jacobian. Being precise means stating the model, the coordinates, the observation unit and sample size, whether information is expected or observed, scalar or matrix, and how singular cases are handled.

Scope of Application¶

Fisher information belongs to mathematical statistics and information geometry and is useful where the analyst can specify the typed mathematical statistics and information geometry carrier, defining objects and relations, parameters, conventions, evidence, boundary cases, and comparison targets, then evaluate the statistical model and dominating measure, parameter space and coordinates, observation unit and sample size, score definition, differentiation and support regularity, expected versus observed information, scalar or matrix form, singular cases, reparameterization, and inferential use are explicit.

Clarity¶

The abstraction clarifies a crowded vocabulary by making the statistical model and dominating measure, parameter space and coordinates, observation unit and sample size, score definition, differentiation and support regularity, expected versus observed information, scalar or matrix form, singular cases, reparameterization, and inferential use are explicit the center of the account. A claim should name the carrier, the governing operation or relation, the applicable assumptions, and the recognition test.

Manages Complexity¶

Without the abstraction, an analyst must reason directly over many local details: the carrier roles, admissibility assumptions, competing conventions, derived invariants, boundary cases, and proof or validation obligations specific to Fisher information. Fisher information compresses them into the roles in the structural signature. That compression permits comparison across instances without erasing the variables that determine validity. It also exposes which details may be varied safely and which are constitutive.

Abstract Reasoning¶

Identify the carrier. State what the elements, states, objects, or observations are: the typed mathematical statistics and information geometry carrier, defining objects and relations, parameters, conventions, evidence, boundary cases, and comparison targets. Reject examples whose alleged carrier belongs to a different problem. 2.

Knowledge Transfer¶

Knowledge transfers strongly among subfields of mathematical statistics and information geometry because they reuse the typed mathematical statistics and information geometry carrier, defining objects and relations, parameters, conventions, evidence, boundary cases, and comparison targets, Differentiate log likelihood with respect to the parameter; averaging the score outer product under the model yields information, and independent observations add their contributions because scores sum and cross terms vanish under regularity., and type the carrier, state every parameter and convention in the definition, test that the statistical model and dominating measure, parameter space and coordinates, observation unit and sample size, score definition, differentiation and support regularity, expected versus observed information, scalar or matrix form, singular cases, reparameterization, and inferential use are explicit, compare the nearest accepted identity, and report counterexamples, uncertainty, and limiting cases.

Relationships to Other Abstractions¶

Current abstraction Fisher information Domain-specific

Parents (1) — outgoing typed relations

Fisher information is a kind of Measurement Prime

It remains its own entry because its identity is fixed by the statistical model and dominating measure, parameter space and coordinates, observation unit and sample size, score definition, differentiation and support regularity, expected versus observed information, scalar or matrix form, singular cases, reparameterization, and inferential use.

Children (1) — abstractions linked here by typed relations

Cramér–Rao Estimator Efficiency Domain-specific presupposes Fisher information

The score's regular-model benchmark is inverse Fisher information.

Hierarchy path (1) — routes to 1 parentless root

Fisher information → is a kind of Measurement

Neighborhood in Abstraction Space¶

Fisher information sits in a crowded region of the domain-specific corpus (6^th percentile for distinctiveness): several abstractions share nearly its structure, so a description that fits it tends to fit its neighbors too.

Family — Probability Distributions & Second-Moment Structures (32 abstractions)

Nearest neighbors

Statistical manifold — 0.94
Maximum likelihood estimation — 0.92
Widely applicable information criterion — 0.92
Control variates — 0.92
Studentization — 0.92

Computed from structural-signature embeddings · 2026-10-25