---
title: 'Fisher Metric: Definition and Applications'
url: https://www.emergentmind.com/topics/fisher-metric
type: topic
---

# Fisher Metric: Definition and Applications

The **Fisher metric**, or **Fisher–Rao metric**, is a Riemannian metric on a statistical model that quantifies the local distinguishability of neighboring probability distributions. For a smooth family of normalized densities $p(x;\theta)$, it is defined by
$$
g_{ij}(\theta)
=
\int p(x;\theta)\,
\frac{\partial\log p(x;\theta)}{\partial\theta^i}
\frac{\partial\log p(x;\theta)}{\partial\theta^j}\,d\mu(x),
$$
or equivalently,
$$
g_{ij}(\theta)
=
\int
\frac{\partial_i p(x;\theta)\,\partial_j p(x;\theta)}
{p(x;\theta)}\,d\mu(x).
$$
Under the usual regularity conditions it is also the negative expected Hessian of the log-likelihood,
$$
g_{ij}(\theta)
=
-\mathbb E_\theta
\left[
\frac{\partial^2\log p(X;\theta)}
{\partial\theta^i\partial\theta^j}
\right].
$$
The metric induces geodesics, distances, curvature, and natural notions of statistical speed and information loss. Its square-root representation identifies probability densities with a submanifold of an $L^2$ sphere, while its monotonicity under probabilistic maps underlies its characterization by Čencov-type theorems. In quantum theory, the corresponding Hermitian tensor combines a Riemannian Fisher or Fubini–Study metric with a symplectic component; in geometric analysis, analogous constructions measure the distinguishability of heat-kernel measures.

## 1. Definition and fundamental structures

Let $\mathcal P$ be a statistical model consisting of probability distributions $p_\theta$ indexed by $\theta=(\theta^1,\ldots,\theta^d)$. A tangent vector corresponds to an infinitesimal perturbation of the density. In a finite sample space, if $X$ and $Y$ have coordinate representations $X(\omega)$ and $Y(\omega)$ with zero total sum, the Fisher metric is
$$
g_p(X,Y)
=
\sum_{\omega}
\frac{X(\omega)Y(\omega)}{p(\omega)}.
$$
Writing the score, or $e$-representation, as
$$
L_X(\omega)=\frac{X(\omega)}{p(\omega)},
$$
gives
$$
g_p(X,Y)=\mathbb E_p[L_XL_Y].
$$
Because score functions have zero expectation, this is also their covariance.

The metric is positive semidefinite:
$$
g_p(X,X)=\mathbb E_p[L_X^2]\geq 0.
$$
It is positive definite when the model is identifiable in the relevant tangent directions. In non-identifiable models, distinct tangent vectors can produce the same score and the metric can be degenerate.

For a parameter curve $\theta(t)$, the Fisher length is
$$
L(\theta)
=
\int
\sqrt{\dot\theta(t)^{\mathsf T}G(\theta(t))\dot\theta(t)}\,dt,
$$
where $G(\theta)=[g_{ij}(\theta)]$. The Fisher–Rao distance is the infimum of such lengths over curves joining two distributions. Unlike the Kullback–Leibler divergence, it is symmetric, satisfies the triangle inequality, and separates points under the usual regularity assumptions. Locally,
$$
D_{\mathrm{KL}}(p_{\theta_1}\|p_{\theta_2})
=
\frac12d_{\mathrm{FR}}^2(p_{\theta_1},p_{\theta_2})
+
o\!\left(d_{\mathrm{FR}}^2\right),
$$
so relative entropy is locally a squared Fisher distance rather than a distance itself [2304.14885].

The Fisher metric also has a square-root form:
$$
g_{ij}
=
4\int
\partial_i\sqrt{p_\theta}\,
\partial_j\sqrt{p_\theta}\,d\mu.
$$
Consequently, the map
$$
p_\theta\longmapsto 2\sqrt{p_\theta}
$$
pulls back the $L^2$ metric to the Fisher metric. For categorical distributions, this map embeds the simplex into the positive orthant of a sphere of radius $2$. The resulting geometry makes the Fisher–Rao distance a spherical arc length rather than the Euclidean chord associated with the Hellinger distance [2304.14885].

## 2. Distances, geodesics, and geometric models

The Fisher distance generally requires solving a geodesic boundary-value problem. In one dimension it is obtained by direct integration:
$$
d_{\mathrm{FR}}(\theta_1,\theta_2)
=
\left|
\int_{\theta_1}^{\theta_2}
\sqrt{g_{11}(\theta)}\,d\theta
\right|.
$$

For categorical distributions $p=(p_1,\ldots,p_n)$ and $q=(q_1,\ldots,q_n)$,
$$
d_{\mathrm{FR}}(p,q)
=
2\arccos\left(\sum_{i=1}^n\sqrt{p_iq_i}\right).
$$
For a multinomial model with $m$ trials, the distance is multiplied by $\sqrt m$:
$$
d_{\mathrm{FR}}(p,q)
=
2\sqrt m\,
\arccos\left(\sum_{i=1}^n\sqrt{p_iq_i}\right).
$$

Many two-parameter location-scale models have Fisher metrics of the form
$$
G(x,y)
=
\begin{pmatrix}
a/y^2&0\\
0&b/y^2
\end{pmatrix}.
$$
After rescaling coordinates, this becomes a constant multiple of the Poincaré half-plane metric. The corresponding distance is hyperbolic:
$$
d_{H^2}((x_1,y_1),(x_2,y_2))
=
2\operatorname{arctanh}
\sqrt{
\frac{(x_1-x_2)^2+(y_1-y_2)^2}
{(x_1-x_2)^2+(y_1+y_2)^2}
}.
$$
Gaussian, Laplace, Cauchy, Student, logistic, generalized Gaussian, inverse Gaussian, Gumbel, Fréchet, Weibull, Pareto, and power-function models provide examples in which parameter transformations reduce the metric to spherical, hyperbolic, logarithmic, or matrix geometries. Wishart models yield the affine-invariant metric on the cone of positive-definite matrices:
$$
d_{\mathrm{FR}}(\Sigma_1,\Sigma_2)
=
\sqrt{\frac n2}\,
\left\|
\log\left(
\Sigma_1^{-1/2}\Sigma_2\Sigma_1^{-1/2}
\right)
\right\|_F.
$$

For product distributions,
$$
p(x_1,\ldots,x_m;\theta_1,\ldots,\theta_m)
=
\prod_{i=1}^m p_i(x_i;\theta_i),
$$
the Fisher metric is block diagonal and the squared distance is additive:
$$
d_{\mathrm{FR}}^2
=
\sum_{i=1}^m d_{\mathrm{FR},i}^2.
$$
Closed-form distances are exceptional. Gamma and Beta families generally have nonconstant negative curvature and do not usually admit elementary distance formulas [2304.14885].

## 3. Invariance, information loss, and uniqueness

A measurable statistic or Markov kernel maps one statistical model to another. If $\kappa:\Omega_1\to\Omega_2$ is a statistic, the processed distribution is $\kappa_*p$. The Fisher metric is monotone:
$$
g^{\Omega_2}(V,V)
\leq
g^{\Omega_1}(V,V).
$$
Thus data processing cannot increase Fisher information. The proof is based on the $L^2$ contraction of conditional expectation:
$$
\|\kappa_*f\|_{L^2(\Omega_2,\kappa_*\mu)}
\leq
\|f\|_{L^2(\Omega_1,\mu)}.
$$

If the statistic is sufficient, no information is lost and equality holds. In this sense, sufficient statistics act isometrically on the statistical model. The same principle applies to probabilistic mappings between more general, possibly singular models. Diffeological statistical models replace an ordinary parameter manifold with a family of plots and can accommodate boundaries, corners, self-intersections, varying local dimensions, and other singularities. For an almost $2$-integrable diffeological model, tangent scores lie in $L^2$, and the Fisher quadratic form is
$$
g_\xi(v,w)
=
\int
\frac{dv}{d\xi}\frac{dw}{d\xi}\,d\xi.
$$
Probabilistic mappings preserve almost $2$-integrability, and Fisher monotonicity continues to hold [1912.02090].

Čencov’s theorem characterizes the Fisher metric, up to a positive multiplicative constant, by invariance under suitable Markov embeddings on finite probability simplices. For general separable metrizable sample spaces, the uniqueness result requires monotonicity together with strong continuity in a mixed topology on the space of measures and their $L^2$ tangent fields. Under these assumptions,
$$
F=c\,g_F
$$
for a universal constant $c\geq0$. Monotonicity and strong continuity also imply invariance under sufficient statistics [1306.1465].

On conditional probability polytopes, the precise normalization depends on the selected structural principle. For a row-stochastic matrix $K$, the three principal metrics are
$$
g_K^{\mathrm{inv}}(u,v)
=
\frac{C}{k}
\sum_{a,b}
\frac{u_{ab}v_{ab}}{K_{ab}},
$$
the unscaled product metric
$$
g_K^{\mathrm{prod}}(u,v)
=
\sum_{a,b}
\frac{u_{ab}v_{ab}}{K_{ab}},
$$
and the marginal-weighted metric
$$
g_K^{\rho}(u,v)
=
\sum_{a,b}
\rho_a
\frac{u_{ab}v_{ab}}{K_{ab}}.
$$
The first arises from invariance under homogeneous conditional embeddings, the second from exponential-family embeddings of Cartesian products, and the third from embedding conditional distributions into joint distributions with fixed marginal $\rho$ [1404.0198].

## 4. Dual, generalized, and continuous formulations

The inverse Fisher metric is naturally a metric on the cotangent bundle. For a finite sample space, every cotangent vector is the differential of an expectation:
$$
\delta_p(A)=d(\mathbb E_p[A]).
$$
Random variables represent cotangent vectors modulo constants because constants have zero differential. Covariance therefore defines the Fisher co-metric:
$$
g_p(\delta_p(A),\delta_p(B))
=
\operatorname{Cov}_p(A,B).
$$
In particular,
$$
\|\delta_p(A)\|_p^2=\operatorname{Var}_p(A).
$$
The inverse Fisher information matrix is the coordinate form of this cotangent metric. The Cramér–Rao inequality becomes a minimum-norm extension property and can be written
$$
V_p(A)\geq G_M(p)^{-1}.
$$
Under Markov maps, the co-metric is monotone in the dual direction:
$$
\|\Phi^*\alpha\|_p
\leq
\|\alpha\|_{\Phi(p)}.
$$
For random variables this is conditional-expectation contraction of variance [2310.13237].

The Fisher metric is also the Hessian of relative entropy. For the Kullback–Leibler divergence,
$$
D_{\mathrm{KL}}(p(\theta)\|p(\theta_0))
=
\frac12
g_{ij}(\theta_0)
\Delta\theta^i\Delta\theta^j
+
O(\Delta\theta^3).
$$
More generally, if a relative entropy group is generated by a function $G$, its Hessian at coincidence is
$$
(g_{ij})_G
=
\left[G'(0)+G''(0)\right]g_{ij}^{\mathrm F}.
$$
The Boltzmann–Gibbs and Kaniadakis classes reproduce the ordinary Fisher metric, while the Tsallis class gives
$$
(g_{ij})_q=(2-q)g_{ij}^{\mathrm F}.
$$
The Levi–Civita connection and Ricci tensor are unchanged by this constant rescaling, whereas scalar curvature is divided by the same factor [1805.11157].

For smooth positive densities on a compact manifold, the Fisher–Rao metric is the unique diffeomorphism-invariant metric on probability densities up to a positive constant. On the larger space of all positive densities, diffeomorphism-invariant metrics have the form
$$
G_\mu(\alpha,\beta)
=
C_1(\mu(M))
\int_M
\frac{\alpha}{\mu}\frac{\beta}{\mu}\,\mu
+
C_2(\mu(M))
\left(\int_M\alpha\right)
\left(\int_M\beta\right).
$$
The first term governs fixed-volume variations and the second governs total-mass variations. Under the square-root map, these metrics become radial warped metrics on an $L^2$ cone. On probability densities, the radial direction disappears and the geometry is a constant-curvature sphere [1607.04550].

## 5. Quantum, geometric, and holographic interpretations

For a normalized amplitude
$$
\psi(x;\theta)=p(x;\theta)^{1/2}e^{i\alpha(x;\theta)},
$$
the probability density contains amplitude information while $\alpha$ contains phase information. Pulling back the Hermitian tensor of projective Hilbert space gives
$$
h_\Theta
=
\frac14 E_p[(d\ln p)^2]
+
E_p[(d\alpha)^2]
-
[E_p(d\alpha)]^2
-
iE_p[d\ln p\wedge d\alpha].
$$
Its real part is
$$
g
=
\frac14F+\operatorname{Cov}_p(d\alpha),
$$
where
$$
F=E_p[(d\ln p)^2]
$$
is the classical Fisher metric. If $d\alpha=0$, then
$$
g=\frac14F.
$$
Thus the classical Fisher metric is the phase-independent sector of the quantum geometric metric, while parameter-dependent phase contributes additional metric and symplectic information [1009.5219].

For density matrices, the symmetric logarithmic derivative is defined by
$$
d\rho
=
\frac12(\rho\,dL_\rho+dL_\rho\,\rho),
$$
and the SLD quantum Fisher tensor is
$$
F_Q=\operatorname{Tr}[\rho(dL_\rho)^2].
$$
For pure states, $dL_\rho=2d\rho$, giving
$$
F_Q=4\,\operatorname{Tr}[\rho(d\rho)^2].
$$
The resulting tensor is proportional to the Fubini–Study metric. For a mixed two-level system with eigenvalues $k_1$ and $k_2$, the fixed-eigenvalue orbit has
$$
I=(k_1-k_2)^2g_{\mathrm{FS}}.
$$
The orbit collapses at the maximally mixed state $k_1=k_2=1/2$, while the pure-state limit recovers the ordinary projective geometry [1205.2561].

For bipartite quantum systems, the local-unitary quantum Fisher tensor has block form
$$
K=
\begin{pmatrix}
K^{(A)}&K^{(AB)}\\
K^{(BA)}&K^{(B)}
\end{pmatrix}.
$$
The off-diagonal block is the covariance matrix between observables on the two subsystems. It vanishes for product states and is generally nonzero for entangled states. In spin-network Hilbert spaces, the quantity
$$
\mathcal E
\propto
\operatorname{Tr}\left[
K^{(AB)T}K^{(AB)}
\right]
$$
serves as a geometric entanglement monotone. For fixed-spin maximally entangled link states, it is tied to the spin Casimir and hence to the area carried by the link [1703.05231].

Fisher geometry also appears in holographic constructions. If $p_n(\xi,x)$ is the probability distribution formed from squared Schmidt coefficients,
$$
p_n(\xi,x)=|\lambda_n(\xi,x)|^2,
$$
then
$$
g_{\mu\nu}
=
\sum_n p_n\,
\partial_\mu\log p_n\,
\partial_\nu\log p_n.
$$
For the entropy-derived construction,
$$
g_{\mu\nu}
=
-\partial_\mu\partial_\nu S
$$
up to normalization. In a $1+1$-dimensional conformal field theory, the special logarithmic entropy case yields the constant-time hyperbolic metric
$$
ds^2=\frac{l^2}{\xi^2}(d\xi^2+dx^2),
$$
whose geodesics are semicircles and whose regulated lengths reproduce the logarithmic form of the Ryu–Takayanagi entropy [1408.6633].

A related holographic result identifies the relative-entropy Hessian of a vacuum ball with bulk canonical energy. For a ball-shaped region $B$,
$$
\langle\delta\rho_B,\delta\rho_B\rangle_{\rho_B^{\mathrm{vac}}}
=
\mathcal E_B(\delta g,\delta g),
$$
where $\mathcal E_B$ is the canonical energy of the corresponding perturbation of the AdS Rindler wedge. Positivity of relative entropy implies
$$
\mathcal E_B(\delta g,\delta g)\geq0.
$$
The first-order version yields the linearized Einstein equations, while the second-order version constrains quadratic gravitational and matter perturbations [1508.00897].

## 6. Applications beyond finite statistical models

The Fisher metric is used whenever a parameterized family of distributions, fields, or measures is available. In functional data registration, the Fisher–Rao metric on absolutely continuous functions is
$$
\langle v_1,v_2\rangle_f
=
\frac14\int_0^1
\frac{\dot v_1(t)\dot v_2(t)}
{|\dot f(t)|}\,dt.
$$
The square-root velocity function
$$
q(t)
=
\operatorname{sign}(\dot f(t))
\sqrt{|\dot f(t)|}
$$
maps this variable metric to the standard $L^2$ metric:
$$
d_{\mathrm{FR}}(f_1,f_2)
=
\|q_1-q_2\|_{L^2}.
$$
Time warps act isometrically by
$$
q\mapsto(q\circ\gamma)\sqrt{\dot\gamma}.
$$
Quotienting by the warping group separates phase variability from amplitude variability, enables elastic distances, and supports Karcher means and registration through dynamic programming [1103.3817].

For structured optical beams, normalized measured intensity is treated as a probability density:
$$
p(x|\theta)
=
\frac{I(x;\theta)}
{\int I(x;\theta)\,dx}.
$$
For transverse displacement,
$$
\mathcal F
=
\int
\frac{[\partial_xp(x)]^2}{p(x)}\,dx
=
4\int
[\partial_x\sqrt{p(x)}]^2\,dx.
$$
The metric is sensitive to intensity gradients, nodal structure, local curvature, and spatial oscillations. Hermite–Gaussian modal order, Laguerre–Gaussian vortex charge, and the radial oscillation parameter of finite-energy Bessel–Gauss beams are all reported to increase displacement Fisher information under the specified measurement model [2512.23538].

In Ricci-flow geometry, the pointed heat-kernel probability measure
$$
d\nu_{x,t;t-\tau}(y)
=
K(x,t;y,t-\tau)\,dg_{t-\tau}(y)
$$
defines a scale-dependent Fisher metric on the source manifold:
$$
(g_\tau^F)_{(x,t)}
=
2\tau
\int
d_x\log K\otimes d_x\log K\,
d\nu_{x,t;t-\tau}.
$$
It satisfies
$$
0\leq g_\tau^F\leq g_t,
\qquad
g_\tau^F\longrightarrow g_t
\quad\text{as }\tau\downarrow0,
$$
and is nonincreasing with scale:
$$
\partial_\tau g_\tau^F\leq0.
$$
The defect $g_t-g_\tau^F$ measures the failure of heat-kernel measures to behave like translated Euclidean Gaussians. Its trace is related to pointed Nash entropy, while equality in a nonzero direction forces Euclidean splitting. Large Fisher eigenvalues generate heat-kernel splitting maps and provide curvature-regularity criteria [2608.14478].

Fisher geometry can also be placed on Lie groups. In a Souriau-type formulation, the geometric temperature $\beta$ belongs to a Lie algebra, the dual heat belongs to its dual, and a cocycle modifies the orbit metric through Lie brackets and central-extension terms. For the explicit $SO(2)$ and $SO(3)$ constructions, the generalized metric is positive on the relevant coadjoint-orbit directions; in $SO(3)$ the stabilizer direction produces a zero eigenvalue on the full Lie algebra but not on the two-dimensional orbit tangent space [2509.20910].

The Fisher metric is therefore not restricted to ordinary finite-dimensional likelihood models. It appears as a metric on probability simplices, spaces of densities, singular and diffeological models, conditional distributions, quantum-state manifolds, functional-data quotient spaces, optical intensity distributions, heat-kernel measures, and coadjoint orbits. Across these settings, its defining principle remains the same: infinitesimal geometric length is determined by the $L^2$ sensitivity of a normalized statistical object to changes in the parameters or base point.

Source: https://www.emergentmind.com/topics/fisher-metric