In mathematical statistics, the Fisher information (sometimes simply called information[1]) is a way of measuring the amount of information that an observable random variable X carries about an unknown parameter θ of a distribution that models X. Formally, it is the variance of the score, or the expected value of the observed information.
The role of the Fisher information in the asymptotic theory of maximum-likelihood estimation was emphasized and explored by the statistician Sir Ronald Fisher (following some initial results by Francis Ysidro Edgeworth). The Fisher information matrix is used to calculate the covariance matrices associated with maximum-likelihood estimates. It can also be used in the formulation of test statistics, such as the Wald test.
In Bayesian statistics, the Fisher information plays a role in the derivation of non-informative prior distributions according to Jeffreys' rule.[2] It also appears as the large-sample covariance of the posterior distribution, provided that the prior is sufficiently smooth (a result known as Bernstein–von Mises theorem, which was anticipated by Laplace for exponential families).[3] The same result is used when approximating the posterior with Laplace's approximation, where the Fisher information appears as the covariance of the fitted Gaussian.[4]
Statistical systems of a scientific nature (physical, biological, etc.) whose likelihood functions obey shift invariance have been shown to obey maximum Fisher information.[5] The level of the maximum depends upon the nature of the system constraints.
X
\theta
X
f(X;\theta)
X
\theta
X
\theta
f
\theta
\theta
X
\theta
f
X
\theta
\theta
Formally, the partial derivative with respect to
\theta
\theta
X
f(X;\theta)
\theta
\begin{align} \operatorname{E}\left[\left.
\partial | |
\partial\theta |
logf(X;\theta)\right|\theta\right]={}&\intR
| |||||
f(x;\theta) |
f(x;\theta)dx\\[6pt] ={}&
\partial | |
\partial\theta |
\intRf(x;\theta)dx\\[6pt] ={}&
\partial | |
\partial\theta |
1\\[6pt] ={}&0. \end{align}
The Fisher information is defined to be the variance of the score:[7]
l{I}(\theta)=\operatorname{E}\left[\left.\left(
\partial | |
\partial\theta |
logf(X;\theta)\right)2\right|\theta\right]=\intR\left(
\partial | |
\partial\theta |
logf(x;\theta)\right)2f(x;\theta)dx,
Note that
l{I}(\theta)\geq0
If is twice differentiable with respect to θ, and under certain regularity conditions, then the Fisher information may also be written as[8]
l{I}(\theta)=-\operatorname{E}\left[\left.
\partial2 | |
\partial\theta2 |
logf(X;\theta)\right|\theta\right],
\partial2 | |
\partial\theta2 |
logf(X;\theta)=
| |||||
f(X;\theta) |
-\left(
| |||||
f(X;\theta) |
\right)2 =
| |||||
f(X;\theta) |
-\left(
\partial | |
\partial\theta |
logf(X;\theta)\right)2
\operatorname{E}\left[\left.
| |||||
f(X;\theta) |
\right|\theta\right]=
\partial2 | |
\partial\theta2 |
\intRf(x;\theta)dx=0.
The regularity conditions are as follows:[9]
If θ is a vector then the regularity conditions must hold for every component of θ. It is easy to find an example of a density that does not satisfy the regularity conditions: The density of a Uniform(0, θ) variable fails to satisfy conditions 1 and 3. In this case, even though the Fisher information can be computed from the definition, it will not have the properties it is typically assumed to have.
Because the likelihood of θ given X is always proportional to the probability f(X; θ), their logarithms necessarily differ by a constant that is independent of θ, and the derivatives of these logarithms with respect to θ are necessarily equal. Thus one can substitute in a log-likelihood l(θ; X) instead of in the definitions of Fisher Information.
The value X can represent a single sample drawn from a single distribution or can represent a collection of samples drawn from a collection of distributions. If there are n samples and the corresponding n distributions are statistically independent then the Fisher information will necessarily be the sum of the single-sample Fisher information values, one for each single sample from its distribution. In particular, if the n distributions are independent and identically distributed then the Fisher information will necessarily be n times the Fisher information of a single sample from the common distribution. Stated in other words, the Fisher Information of i.i.d. observations of a sample of size n from a population is equal to the product of n and the Fisher Information of a single observation from the same population.
The Cramér–Rao bound[10] [11] states that the inverse of the Fisher information is a lower bound on the variance of any unbiased estimator of θ. H.L. Van Trees (1968) and B. Roy Frieden (2004) provide the following method of deriving the Cramér–Rao bound, a result which describes use of the Fisher information.
\hat\theta(X)
\operatorname{E}\left[\left.\hat\theta(X)-\theta\right|\theta\right] =\int\left(\hat\theta(x)-\theta\right)f(x;\theta)dx=0regardlessofthevalueof\theta.
This expression is zero independent of θ, so its partial derivative with respect to θ must also be zero. By the product rule, this partial derivative is also equal to
0=
\partial | |
\partial\theta |
\int\left(\hat\theta(x)-\theta\right)f(x;\theta)dx =\int\left(\hat\theta(x)-\theta\right)
\partialf | |
\partial\theta |
dx-\intfdx.
For each θ, the likelihood function is a probability density function, and therefore
\intfdx=1
logf
f(x;\theta)
\partialf | |
\partial\theta |
=f
\partiallogf | |
\partial\theta |
.
Using these two facts in the above, we get
\int\left(\hat\theta-\theta\right)f
\partiallogf | |
\partial\theta |
dx=1.
Factoring the integrand gives
\int\left(\left(\hat\theta-\theta\right)\sqrt{f}\right)\left(\sqrt{f}
\partiallogf | |
\partial\theta |
\right)dx=1.
Squaring the expression in the integral, the Cauchy–Schwarz inequality yields
1= l(\int\left[\left(\hat\theta-\theta\right)\sqrt{f}\right] ⋅ \left[\sqrt{f}
\partiallogf | |
\partial\theta |
\right]dxr)2 \le \left[\int\left(\hat\theta-\theta\right)2fdx\right] ⋅ \left[\int\left(
\partiallogf | |
\partial\theta |
\right)2fdx\right].
The second bracketed factor is defined to be the Fisher Information, while the first bracketed factor is the expected mean-squared error of the estimator
\hat\theta
\operatorname{Var}\left(\hat\theta\right)\geq
1 | |
l{I |
\left(\theta\right)}.
In other words, the precision to which we can estimate θ is fundamentally limited by the Fisher information of the likelihood function.
Alternatively, the same conclusion can be obtained directly from the Cauchy–Schwarz inequality for random variables,
|\operatorname{Cov}(A,B)|2\le\operatorname{Var}(A)\operatorname{Var}(B)
\hat\theta(X)
\partial\thetalogf(X;\theta)
A Bernoulli trial is a random variable with two possible outcomes, 0 and 1, with 1 having a probability of θ. The outcome can be thought of as determined by the toss of a biased coin, with the probability of heads (1) being θ and the probability of tails (0) being .
Let X be a Bernoulli trial of one sample from the distribution. The Fisher information contained in X may be calculated to be:
\begin{align} l{I}(\theta) &=-\operatorname{E}\left[\left.
\partial2 | |
\partial\theta2 |
log\left(\thetaX(1-\theta)1\right)\right|\theta\right]\\[5pt] &=-\operatorname{E}\left[\left.
\partial2 | |
\partial\theta2 |
\left(Xlog\theta+(1-X)log(1-\theta)\right)\right|\theta\right]\\[5pt] &=\operatorname{E}\left[\left.
X | |
\theta2 |
+
1-X | |
(1-\theta)2 |
\right|\theta\right]\\[5pt] &=
\theta | |
\theta2 |
+
1-\theta | |
(1-\theta)2 |
\\[5pt] &=
1 | |
\theta(1-\theta) |
. \end{align}
Because Fisher information is additive, the Fisher information contained in n independent Bernoulli trials is therefore
l{I}(\theta)=
n | |
\theta(1-\theta) |
.
If
xi
2n
xij
xi
p(xi,\theta)=\prod
n | |
j=0 |
xij | |
\theta |
xij | |
(1-\theta) |
\mui=
n | |
(1/n)\sum | |
j=1 |
xij
E(\mu)=\sum | |
xi |
\muip(xi,\theta)=\theta
2n
2)=\sum | |
E(\mu | |
xi |
2 | |
\mu | |
i |
p(xi,\theta)=
(1+(n-1)\theta)\theta | |
n |
E(\mu2)-E(\mu)2=(1/n)\theta(1-\theta)
It is seen that the Fisher information is the reciprocal of the variance of the mean number of successes in n Bernoulli trials. This is generally true. In this case, the Cramér–Rao bound is an equality.
\theta
X\sim\operatorname{Bern}(\sqrt\theta)
X
p0=1-\sqrt\theta
p1=\sqrt\theta
\theta\in[0,1]
\theta
X
The Fisher information reads in this caseThis expression can also be derived directly from the change of reparametrization formula given below. More generally, for any sufficiently regular function
f
f(\theta)\in[0,1]
\theta
X\sim\operatorname{Bern}(f(\theta))
\theta=\begin{bmatrix}\theta1&\theta2&...&
sf{T}, | |
\theta | |
N\end{bmatrix} |
l[l{I}(\theta)r]i,= \operatorname{E}\left[\left. \left(
\partial | |
\partial\thetai |
logf(X;\theta)\right) \left(
\partial | |
\partial\thetaj |
logf(X;\theta)\right) \right|\theta\right].
The FIM is a positive semidefinite matrix. If it is positive definite, then it defines a Riemannian metric[12] on the N-dimensional parameter space. The topic information geometry uses this to connect Fisher information to differential geometry, and in that context, this metric is known as the Fisher information metric.
Under certain regularity conditions, the Fisher information matrix may also be written as
l[l{I}(\theta)r]i,= -\operatorname{E}\left[\left.
\partial2 | |
\partial\thetai\partial\thetaj |
logf(X;\theta) \right|\theta\right].
The result is interesting in several ways:
We say that two parameter component vectors θ1 and θ2 are information orthogonal if the Fisher information matrix is block diagonal, with these components in separate blocks.[17] Orthogonal parameters are easy to deal with in the sense that their maximum likelihood estimates are asymptotically uncorrelated. When considering how to analyse a statistical model, the modeller is advised to invest some time searching for an orthogonal parametrization of the model, in particular when the parameter of interest is one-dimensional, but the nuisance parameter can have any dimension.[18]
If the Fisher information matrix is positive definite for all, then the corresponding statistical model is said to be regular; otherwise, the statistical model is said to be singular.[19] Examples of singular statistical models include the following: normal mixtures, binomial mixtures, multinomial mixtures, Bayesian networks, neural networks, radial basis functions, hidden Markov models, stochastic context-free grammars, reduced rank regressions, Boltzmann machines.
In machine learning, if a statistical model is devised so that it extracts hidden structure from a random phenomenon, then it naturally becomes singular.[20]
The FIM for a N-variate multivariate normal distribution,
X\simN\left(\mu(\theta),\Sigma(\theta)\right)
\theta=\begin{bmatrix}\theta1&...&\thetaK\end{bmatrix}sf{T}
X=\begin{bmatrix}X1&...&XN\end{bmatrix}sf{T}
\mu(\theta)=\begin{bmatrix}\mu1(\theta)&...&\muN(\theta)\end{bmatrix}sf{T}
\Sigma(\theta)
1\lem,n\leK
l{I}m,n=
\partial\musf{T | |
where
( ⋅ )sf{T}
\operatorname{tr}( ⋅ )
\begin{align}
\partial\mu | |
\partial\thetam |
&= \begin{bmatrix} \dfrac{\partial\mu1}{\partial\thetam}& \dfrac{\partial\mu2}{\partial\thetam}& … & \dfrac{\partial\muN}{\partial\thetam} \end{bmatrix}sf{T};\\[8pt] \dfrac{\partial\Sigma}{\partial\thetam}&= \begin{bmatrix} \dfrac{\partial\Sigma1,1
Note that a special, but very common, case is the one where
\Sigma(\theta)=\Sigma
l{I}m,n=
\partial\musf{T | |
In this case the Fisher information matrix may be identified with the coefficient matrix of the normal equations of least squares estimation theory.
Another special case occurs when the mean and covariance depend on two different vector parameters, say, β and θ. This is especially popular in the analysis of spatial data, which often uses a linear model with correlated residuals. In this case,[22]
l{I}(\beta,\theta)=\operatorname{diag}\left(l{I}(\beta),l{I}(\theta)\right)
where
\begin{align} l{I}{(\beta)m,n
Similar to the entropy or mutual information, the Fisher information also possesses a chain rule decomposition. In particular, if X and Y are jointly distributed random variables, it follows that:[23]
l{I}X,Y(\theta)=l{I}X(\theta)+l{I}Y\mid(\theta),
where
l{I}Y\mid(\theta)=\operatorname{E}X\left[l{I}Y\mid(\theta)\right]
l{I}Y\mid(\theta)
\theta
As a special case, if the two random variables are independent, the information yielded by the two random variables is the sum of the information from each random variable separately:
l{I}X,Y(\theta)=l{I}X(\theta)+l{I}Y(\theta).
Consequently, the information in a random sample of n independent and identically distributed observations is n times the information in a sample of size 1.
Given a convex function
f:[0,infty)\to(-infty,infty]
f(x)
x>0
f(1)=0
f(0)=\lim | |
t\to0+ |
f(t)
Df
f
1
\theta\in\Theta
P\theta
\theta
f(x;\theta)
In this form, it is clear that the Fisher information matrix is a Riemannian metric, and varies correctly under a change of variables. (see section on Reparameterization.)
The information provided by a sufficient statistic is the same as that of the sample X. This may be seen by using Neyman's factorization criterion for a sufficient statistic. If T(X) is sufficient for θ, then
f(X;\theta)=g(T(X),\theta)h(X)
for some functions g and h. The independence of h(X) from θ implies
\partial | |
\partial\theta |
log\left[f(X;\theta)\right]=
\partial | |
\partial\theta |
log\left[g(T(X);\theta)\right],
and the equality of information then follows from the definition of Fisher information. More generally, if is a statistic, then
l{I}T(\theta)\leql{I}X(\theta)
with equality if and only if T is a sufficient statistic.[25]
The Fisher information depends on the parametrization of the problem. If θ and η are two scalar parametrizations of an estimation problem, and θ is a continuously differentiable function of η, then
{lI}η(η)={lI}\theta(\theta(η))\left(
d\theta | |
dη |
\right)2
where
{lI}η
{lI}\theta
In the vector case, suppose
{\boldsymbol\theta}
{\boldsymbolη}
{\boldsymbol\theta}
{\boldsymbolη}
{lI}\boldsymbol({\boldsymbolη})={\boldsymbol
sf{T}{lI} | |
J} | |
\boldsymbol\theta |
({\boldsymbol\theta}({\boldsymbolη})){\boldsymbolJ}
\boldsymbolJ
Jij=
\partial\thetai | |
\partialηj |
,
{\boldsymbolJ}sf{T}
{\boldsymbolJ}.
In information geometry, this is seen as a change of coordinates on a Riemannian manifold, and the intrinsic properties of curvature are unchanged under different parametrizations. In general, the Fisher information matrix provides a Riemannian metric (more precisely, the Fisher–Rao metric) for the manifold of thermodynamic states, and can be used as an information-geometric complexity measure for a classification of phase transitions, e.g., the scalar curvature of the thermodynamic metric tensor diverges at (and only at) a phase transition point.[28]
In the thermodynamic context, the Fisher information matrix is directly related to the rate of change in the corresponding order parameters.[29] In particular, such relations identify second-order phase transitions via divergences of individual elements of the Fisher information matrix.
The Fisher information matrix plays a role in an inequality like the isoperimetric inequality.[30] Of all probability distributions with a given entropy, the one whose Fisher information matrix has the smallest trace is the Gaussian distribution. This is like how, of all bounded sets with a given volume, the sphere has the smallest surface area.
The proof involves taking a multivariate random variable
X
f
\{f(x-\theta)\mid\theta\inRn\}
X
S(X)=\lim\varepsilon
| |||||||
\varepsilon |
Z\varepsilon
\varepsilonI
eH(X)
S(X)
S(X)
Fisher information is widely used in optimal experimental design. Because of the reciprocity of estimator-variance and Fisher information, minimizing the variance corresponds to maximizing the information.
When the linear (or linearized) statistical model has several parameters, the mean of the parameter estimator is a vector and its variance is a matrix. The inverse of the variance matrix is called the "information matrix". Because the variance of the estimator of a parameter vector is a matrix, the problem of "minimizing the variance" is complicated. Using statistical theory, statisticians compress the information-matrix using real-valued summary statistics; being real-valued functions, these "information criteria" can be maximized.
Traditionally, statisticians have evaluated estimators and designs by considering some summary statistic of the covariance matrix (of an unbiased estimator), usually with positive real values (like the determinant or matrix trace). Working with positive real numbers brings several advantages: If the estimator of a single parameter has a positive variance, then the variance and the Fisher information are both positive real numbers; hence they are members of the convex cone of nonnegative real numbers (whose nonzero members have reciprocals in this same cone).
For several parameters, the covariance matrices and information matrices are elements of the convex cone of nonnegative-definite symmetric matrices in a partially ordered vector space, under the Loewner (Löwner) order. This cone is closed under matrix addition and inversion, as well as under the multiplication of positive real numbers and matrices. An exposition of matrix theory and Loewner order appears in Pukelsheim.[32]
The traditional optimality criteria are the information matrix's invariants, in the sense of invariant theory; algebraically, the traditional optimality criteria are functionals of the eigenvalues of the (Fisher) information matrix (see optimal design).
In Bayesian statistics, the Fisher information is used to calculate the Jeffreys prior, which is a standard, non-informative prior for continuous distribution parameters.[33]
The Fisher information has been used to find bounds on the accuracy of neural codes. In that case, X is typically the joint responses of many neurons representing a low dimensional variable θ (such as a stimulus parameter). In particular the role of correlations in the noise of the neural responses has been studied.[34]
Fisher information was used to study how informative different data sources are for estimation of the reproduction number of SARS-CoV-2.[35]
Fisher information plays a central role in a controversial principle put forward by Frieden as the basis of physical laws, a claim that has been disputed.[36]
The Fisher information is used in machine learning techniques such as elastic weight consolidation,[37] which reduces catastrophic forgetting in artificial neural networks.
Fisher information can be used as an alternative to the Hessian of the loss function in second-order gradient descent network training.[38]
Using a Fisher information metric, da Fonseca et. al [39] investigated the degree to which MacAdam ellipses (color discrimination ellipses) can be derived from the response functions of the retinal photoreceptors.
See also: Fisher information metric. Fisher information is related to relative entropy.[40] The relative entropy, or Kullback–Leibler divergence, between two distributions
p
q
KL(p:q)=\intp(x)log
p(x) | |
q(x) |
dx.
Now, consider a family of probability distributions
f(x;\theta)
\theta\in\Theta
D(\theta,\theta')=KL(p({} ⋅ {};\theta):p({} ⋅ {};\theta'))=\intf(x;\theta)log
f(x;\theta) | |
f(x;\theta') |
dx.
If
\theta
\theta'=\theta
\theta'
\theta
D(\theta,\theta')=
1 | |
2 |
(\theta'-
| ||||
\theta) |
D(\theta,\theta')\right)\theta'=\theta(\theta'-\theta)+o\left((\theta'-\theta)2\right)
But the second order derivative can be written as
\left( | \partial2 |
\partial\theta'i\partial\theta'j |
D(\theta,\theta')\right)\theta'=\theta=-\intf(x;\theta)\left(
\partial2 | |
\partial\theta'i\partial\theta'j |
log(f(x;\theta'))\right)\theta'=\thetadx=[l{I}(\theta)]i,j.
Thus the Fisher information represents the curvature of the relative entropy of a conditional distribution with respect to its parameters.
The Fisher information was discussed by several early statisticians, notably F. Y. Edgeworth.[41] For example, Savage[42] says: "In it [Fisher information], he [Fisher] was to some extent anticipated (Edgeworth 1908–9 esp. 502, 507–8, 662, 677–8, 82–5 and references he [Edgeworth] cites including Pearson and Filon 1898 [. . .])." There are a number of early historical sources[43] and a number of reviews of this early work.[44] [45] [46]
Other measures employed in information theory:
. Lucien Le Cam . 1986 . Asymptotic Methods in Statistical Decision Theory . New York . Springer . 618–621 . 0-387-96307-3 .