Fisher Information Metric

Updated 1 March 2026

Fisher Information Metric is a canonical Riemannian metric defined on families of probability distributions, quantifying infinitesimal distinguishability.
It underlies exponential families, where its Hessian structure relates directly to the log-partition function and connects with Kähler geometry in complex manifolds.
Applications span estimation theory, quantum information, signal processing, and emergent gravity, highlighting its role in both classical and quantum frameworks.

The Fisher information metric (Fisher–Rao metric) is a canonical Riemannian metric defined on smoothly parameterized families of probability distributions. It quantifies the infinitesimal distinguishability between nearby densities in a statistical model and plays central roles in estimation theory, information geometry, quantum information, and the theory of emergent geometry.

1. Definition and Fundamental Properties

Let $\{p(x;\theta): \theta = (\theta^1, \ldots, \theta^d) \in \Theta \subset \mathbb{R}^d\}$ be a family of probability densities on a sample space $X$ with $p(x;\theta) > 0$ and $\int_X p(x; \theta) dx = 1$ . The Fisher information metric $g^F$ on parameter space $\Theta$ is defined by

$g^F_{ij}(\theta) = \mathbb{E}_{x \sim p(\cdot; \theta)} \left[ \partial_i \log p(x;\theta)\, \partial_j \log p(x;\theta) \right] = \int_X \left( \partial_i \log p(x;\theta) \right)\, \left( \partial_j \log p(x;\theta) \right)\, p(x;\theta) dx$

and, under mild regularity, equivalently as

$g^F_{ij}(\theta) = - \mathbb{E}_{x \sim p}[\partial_i \partial_j \log p(x;\theta)].$

This metric, first systematically studied in the works of Fisher, Rao, and later C̆encov and Amari, enjoys two essential invariance properties:

Monotonicity under statistics: For any sufficient statistic or measurable mapping, Fisher information cannot increase.
Invariance under sufficient statistics: The metric is preserved under mappings that retain all information relevant to parameter inference (Lê, 2013).

The uniqueness theorem (C̆encov–Morozova–Amari) asserts that, up to an overall positive scale, the Fisher information metric is the only continuous, monotone, locally statistical Riemannian metric on models over arbitrary sample spaces, given strong continuity in the appropriate mixed topology (Lê, 2013).

2. Information Geometry, Kähler Structure, and Exponential Families

In information geometry, the Fisher metric engenders a natural differential geometric structure on the statistical manifold. For exponential families,

$p(x; \theta) = \exp \left( \langle \theta, T(x) \rangle - F(\theta) \right),$

the metric admits a Hessian structure: $g_{ij}(\theta) = \partial_i \partial_j F(\theta),$ where $F$ is the log-partition (cumulant-generating) function.

A central result is that any real-analytic Kähler metric is, locally, a Fisher information metric of an exponential family: for a complex manifold with real-analytic Kähler metric, there exist holomorphic coordinates in which the Kähler metric is precisely the Fisher metric arising from an appropriate exponential family, with the Kähler potential serving as the log-partition function (Gnandi, 2024). The Calabi diastasis function associated with the Kähler metric realizes the local Kullback–Leibler divergence between distributions, providing a direct bridge between complex geometry and statistical distance.

Table: Correspondence in Exponential Family/Kähler Setting

Concept	Statistical interpretation	Geometric role
Kähler potential $\Phi$	Log-partition/cumulant function $F(\theta)$	Potential for metric
Fisher metric	Hessian of $F(\theta)$	Kähler metric
Diastasis $D^g$	Local KL divergence	Kähler-geometric distance
Kähler form $\omega$	Information-geometric symplectic form	Compatible 2-form

This realization establishes a deep connection to complex geometry and underlies the geometric structure of quantum pure states, where the Fubini–Study metric arises as a Fisher metric (Gnandi, 2024, Facchi et al., 2010).

3. Quantum Fisher Information and the Quantum Metric Tensor

For a parameterized family of quantum states $\rho(\theta)$ , the quantum Fisher information (QFI) generalizes the classical Fisher metric. For a one-parameter family, the QFI is

$F(\theta) = \mathrm{Tr}\left[ \rho(\theta) L(\theta)^2 \right]$

where $L(\theta)$ is the symmetric logarithmic derivative (SLD) defined via $\partial_\theta \rho(\theta) = \frac{1}{2} [L(\theta) \rho(\theta) + \rho(\theta) L(\theta)]$ (Zhang et al., 2022, Scandi et al., 2023). For pure states $|\psi(\theta)\rangle$ ,

$F(\theta) = 4\left( \langle \partial_\theta\psi | \partial_\theta\psi \rangle - |\langle \psi | \partial_\theta\psi \rangle|^2 \right),$

which is four times the quantum metric tensor, i.e., the real part of the quantum geometric tensor on projective Hilbert space (Facchi et al., 2010).

The QFI sets the fundamental bound on estimation precision (quantum Cramér–Rao bound), appears in quantum hypothesis testing, and quantifies response in quantum thermodynamics and criticality (Zhang et al., 2022, Scandi et al., 2023). In quantum information geometry, a large class of monotone metrics (Petz metrics) generalize QFI, characterized by operator monotone functions (Scandi et al., 2023).

4. Applications: Estimation, Signal Processing, Adversarial Analysis, and Physics

Statistical Estimation: The Fisher (and quantum Fisher) metric yields the Cramér–Rao bound, giving a lower bound on the variance of any unbiased estimator: $\mathrm{Var}(\hat\theta^i) \geq [g^F(\theta)^{-1}]^{ii}.$

Signal Processing: For additive noise channels, the Fisher information of the noise distribution upper bounds the signal-to-noise ratio gain, the asymptotic relative efficiency for signal detection, and the cross-correlation gain for transmission. The minimal value is unity (Gaussian), and non-Gaussian distributions yield strictly higher Fisher information, sometimes to the extent of unbounded (dichotomous noise) (Duan et al., 2011).

Adversarial Machine Learning: Considering the Fisher metric on data-induced output distributions of a neural network enables spectral analysis of adversarial vulnerability. The largest eigenvalues of the Fisher information matrix quantify susceptibility: adversarial perturbations aligned to the principal eigenvectors induce maximal changes in output distribution. This fact informs both construction of optimal attacks and principled detection heuristics (Zhao et al., 2018).

Optical Metrology: Fisher information quantifies the ultimate parameter estimation sensitivity in structured optical fields. For beam displacement, the Fisher information displays explicit scaling laws with mode order (Hermite–Gaussian, Laguerre–Gaussian, Bessel–Gauss), reflecting nodal complexity and enabling beam engineering for optimal sensing (Sumaya-Martinez et al., 29 Dec 2025).

Physical Geometry and Emergent Gravity: Interpreting the moduli of classical field-theoretic solutions as parameters of probability densities, the Fisher metric can reproduce familiar geometric structures. For relativistic sigma models, it gives flat or AdS spacetime metrics on moduli space depending on the structure of the underlying solutions (Miyamoto et al., 2012). The Fisher metric, when taken as the fundamental geometric object, can be used to recast the Einstein–Hilbert action and yields nontrivial RG flows and obstacles when attempting to quantize "gravity" in this formalism (Takeuchi, 2018, Matsueda, 2013).

5. Structural Uniqueness, Co-metric, and Generalizations

Uniqueness: Chentsov's theorem (and its extensions to general measure spaces) states that the Fisher information metric is, up to a constant, the unique Riemannian metric on statistical models that is monotone under statistics and continuous in the model topology (Lê, 2013). Contractivity under both classical (Markov) and quantum (completely positive trace-preserving) channels is satisfied only by the Fisher/Petz family of metrics (Scandi et al., 2023).

Fisher Co-metric: The Fisher co-metric, acting on the cotangent bundle, equates to the variance-covariance inner product on random variables modulo constants: $g^*_p(\alpha_p, \beta_p) = \operatorname{Cov}_p(A, B)$ for differentials $d(A)_p, d(B)_p$ . The Cramér–Rao bound is immediate in co-metric form, and invariance under Markov maps trivially determines the variance/covariance structure up to scale (cotangent Čencov theorem) (Nagaoka, 2023).

Extension to Quantum and Kähler Geometry: The quantum counterpart, for pure states, arises as the real part of the Hermitian (Kähler) tensor on projective Hilbert space; for exponential families and Kähler manifolds, the Fisher metric is the Hessian of the cumulant-generating function, and the associated geometry is naturally Kähler (Gnandi, 2024, Facchi et al., 2010). This unifies quantum and classical distinguishability metrics, with the quantum metric upper bounding the classical Fisher metric in general (with equality for phase-constant states) (Facchi et al., 2010).

Operations on PDFs and the Inverse Problem: The sum rule for spatially disjoint products of probability densities yields additive Fisher metrics. Arbitrary Riemannian metrics can be inverted to families of PDFs via Nash embedding followed by construction using translation-invariant (e.g., Gaussian) kernels; the mapping is highly non-unique (Clingman et al., 2015).

6. Geometry, Torsion, and Conformal Transformations

Conformal rescalings of the Fisher information metric generally break invariants like scalar curvature, unless the affine connection is extended to include torsion. Introducing a torsionful (metric-compatible) connection allows the formation of new invariants (torsion scalar) that can distinguish between PDFs with the same Fisher metric but different normalization factors (e.g., Gaussian vs. Cauchy). In thermodynamic geometry, the torsion scalar displays distinct behaviors, such as diverging along spinodal curves—providing physically sensitive differentiators invisible to scalar curvature (Pal et al., 2022).

7. Holography, Field Theory, and General Relativity

Holographic Fisher Information Metric: In quantum field theories and their holographic (AdS/CFT or Schrödinger) duals, the quantum Fisher metric computed via fidelity or two-point functions matches the coefficient of relative entropy between nearby states. In the gravitational dual, this quadratic information metric equates to canonical energy in the Rindler wedge of AdS, and the agreement of leading divergences in the bulk/boundary computations is a precision check of the duality (Dimov et al., 2020, Lashkari et al., 2015). Subleading divergences correspond to multi-trace/contact data, highlighting the sensitivity of the bulk Fisher metric to higher-point correlations.

Emergent Gravity from Information Geometry: Treating Fisher geometry as the foundational spacetime structure, one can derive the Einstein tensor entirely from the Fisher metric constructed from a statistical family, with the entropy or spectrum density functioning as a matter (scalar) field. The resulting system behaves as a classical field theory of geometry sourced by coarse-grained information fields. Notably, the emergence of curvature and dynamical equations is sourced by statistical rather than fundamental fields, aligning with perspectives on gravity as an emergent or entropic force (Matsueda, 2013).

References:

(Gnandi, 2024) "Any Kähler metric is a Fisher information metric"
(Lê, 2013) "The uniqueness of the Fisher metric as information metric"
(Facchi et al., 2010) "Classical and Quantum Fisher Information in the Geometrical Formulation of Quantum Mechanics"
(Zhang et al., 2022) "Direct measurement of quantum Fisher information"
(Nagaoka, 2023) "The Fisher metric as a metric on the cotangent bundle"
(Clingman et al., 2015) "Probability Density Functions from the Fisher Information Metric"
(Miyamoto et al., 2012) "Information metric from a linear sigma model"
(Takeuchi, 2018) "Coarse-graining of the Einstein-Hilbert Action rewritten by the Fisher information metric"
(Zhao et al., 2018) "The Adversarial Attack and Detection under the Fisher Information Metric"
(Duan et al., 2011) "Fisher information as a performance metric for locally optimum processing"
(Sumaya-Martinez et al., 29 Dec 2025) "Fisher Information as an Operational Metric for Structured Optical Beams"
(Pal et al., 2022) "Conformal Fisher information metric with torsion"
(Matsueda, 2013) "Emergent General Relativity from Fisher Information Metric"
(Lashkari et al., 2015) "Canonical Energy is Quantum Fisher Information"
(Dimov et al., 2020) "Holographic Fisher Information Metric in Schrödinger Spacetime"
(Scandi et al., 2023) "Quantum Fisher Information and its dynamical nature"
(Mondal, 2015) "Generalized Fubini-Study Metric and Fisher Information Metric"
(Cirilo-Lombardo et al., 2012) "Information metric from Riemannian superspaces"

The Fisher information metric thus stands as the uniquely natural, geometric, and physically rich structure underlying statistical inference, quantum estimation, and the analysis of emergent spacetime and field theories. It subsumes notions of distinguishability, curvature, physical observables in estimation and sensing, and the dynamical constraints of both classical and quantum channels.