Fisher information: Wikis

Note: Many of our articles have direct quotes from sources you can cite, within the Wikipedia article! This article doesn't yet, but we're working on it! See more info or our list of citable articles.

Encyclopedia

In mathematical statistics and information theory, the Fisher information (sometimes simply called information[1]) is the variance of the score. Its role in the asymptotic theory of maximum-likelihood estimation was emphasized by the statistician R.A. Fisher (following some initial results by F. Y. Edgeworth). In Bayesian statistics, the Fisher information is used in the calculation of the Jeffreys prior.

An important use of the Fisher information matrix in statistical analyses is its contribution to the calculation of the covariance matrices of estimates of parameters fitted by maximium likelihood. It can also be used in the formulation of test statistics, such as the Wald test.

History

The Fisher information was discussed by several early statisticians, notably F. Y. Edgeworth.[2] For example, Savage[3] says: "In it [Fisher information], he [Fisher] was to some extent anticipated (Egeworth 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[4] and a number of reviews of this early work.[5][6][7]

Definition

The Fisher information is a way of measuring the amount of information that an observable random variable X carries about an unknown parameter θ upon which the likelihood function of θ, L(θ) = f(X;θ), depends. The likelihood function is the joint probability of the data, the Xs, conditional on the value of θ, as a function of θ. Since the expectation of the score is zero, the variance is simply the second moment of the score, the derivative of the log of the likelihood function with respect to θ. Hence the Fisher information can be written as:

$\mathcal{I}(\theta) = \mathrm{E} \left\{\left. \left[ \frac{\partial}{\partial\theta} \ln f(X;\theta) \right]^2 \right|\theta\right\},$

which implies $0 \leq \mathcal{I}(\theta) < \infty$. The Fisher information is thus the expectation with respect to $P(X \mid \theta)$ of the squared score. A random variable carrying high Fisher information implies that the absolute value of the score is often high.

The Fisher information is not a function of a particular observation, as the random variable X has been averaged out. The concept of information is useful when comparing two methods of observing a given random process.

If lnf(x;θ) is twice differentiable with respect to θ, and if the regularity condition

$\int \frac{\partial^2}{\partial \theta^2} f(x ; \theta ) \, dx = 0$

holds, then the Fisher information may also be written as[8]

$\mathcal{I}(\theta) = - \mathrm{E} \left[\left. \frac{\partial^2}{\partial\theta^2} \ln f(X;\theta)\right|\theta \right].$

Thus Fisher information is the negative of the expectation under $P(X \mid \theta)$ of the second derivative of the log of f with respect to θ. Information may thus be seen to be a measure of the "sharpness" of the support curve near the maximum likelihood estimate of θ. A "blunt" support curve (one with a shallow maximum) would have a low negative expected second derivative, and thus low information; while a sharp one would have a high negative expected second derivative and thus high information.

Information is additive, in that the information yielded by two independent experiments is the sum of the information from each experiment separately:

$\mathcal{I}_{X,Y}(\theta) = \mathcal{I}_X(\theta) + \mathcal{I}_Y(\theta).$

This result follows from the elementary fact that if random variables are independent, the variance of their sum is the sum of their variances. Hence the information in a random sample of size n is n times that in a sample of size 1 (if observations are independent).

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. See sufficient statistic for a more detailed explanation. The equality of information then follows from the following fact:

$\frac{\partial}{\partial\theta} \ln \left[f(X ;\theta)\right] = \frac{\partial}{\partial\theta} \ln \left[g(T(X);\theta)\right]$

which follows from the definition of Fisher information, and the independence of h(X) from θ. More generally, if T = t(X) is a statistic, then

$\mathcal{I}_T(\theta) \leq \mathcal{I}_X(\theta)$

with equality if and only if T is a sufficient statistic.

The Cramér-Rao inequality states that the inverse of the Fisher information is a lower bound on the variance of any unbiased estimator of θ.

Informal derivation

Van Trees (1968) and Frieden (2004) provide the following method of deriving the Cramér-Rao bound, a result which describes use of the Fisher information, informally:

Consider an unbiased estimator $\hat\theta(X)$. Mathematically, we write

$\mathrm{E}\left[ \hat\theta(X) - \theta \right] = \int \left[ \hat\theta(X) - \theta \right] \cdot f(X ;\theta) \, dx = 0.$

The likelihood function f(X;θ) describes the probability that we observe a given sample x given a known value of θ. If f is sharply peaked, it is easy to intuit the "correct" value of θ given the data, and hence the data contains a lot of information about the parameter. If the likelihood f is flat and spread-out, then it would take many, many samples of X to estimate the actual "true" value of θ. Therefore, we would intuit that the data contain much less information about the parameter.

Now, we differentiate the unbiased-ness condition above to get

$\frac{\partial}{\partial\theta} \int \left[ \hat\theta(X) - \theta \right] \cdot f(X ;\theta) \, dx = \int \left(\hat\theta-\theta\right) \frac{\partial f}{\partial\theta} \, dx - \int f \, dx = 0.$

We now make use of two facts. The first is that the likelihood f is just the probability of the data given the parameter. Since it is a probability, it must be normalized, implying that

$\int f \, dx = 1.$

Second, we know from basic calculus that

$\frac{\partial f}{\partial\theta} = f \, \frac{\partial \ln f}{\partial\theta}.$

Using these two facts in the above let us write

$\int \left(\hat\theta-\theta\right) f \, \frac{\partial \ln f}{\partial\theta} \, dx = 1.$

Factoring the integrand gives

$\int \left(\left(\hat\theta-\theta\right) \sqrt{f} \right) \left( \sqrt{f} \, \frac{\partial \ln f}{\partial\theta} \right) \, dx = 1.$

If we square the equation, the Cauchy-Schwarz inequality lets us write

$\left[ \int \left(\hat\theta - \theta\right)^2 f \, dx \right] \cdot \left[ \int \left( \frac{\partial \ln f}{\partial\theta} \right)^2 f \, dx \right] \geq 1.$

The right-most factor is defined to be the Fisher Information

$\mathcal{I}\left(\theta\right) = \int \left( \frac{\partial \ln f}{\partial\theta} \right)^2 f \, dx.$

The left-most factor is the expected mean-squared error of the estimator θ, since

$\mathrm{E}\left[ \left( \hat\theta\left(X\right) - \theta \right)^2 \right] = \int \left(\hat\theta - \theta\right)^2 f \, dx.$

Notice that the inequality tells us that, fundamentally,

$\mbox{Var}\left[\hat\theta\right] \, \geq \, {1} / {\mathcal{I}\left(\theta\right)}.$

In other words, the precision to which we can estimate θ is fundamentally limited by the Fisher Information of likelihood function.

Single-parameter Bernoulli experiment

A Bernoulli trial is a random variable with two possible outcomes, "success" and "failure", with "success" having a probability of θ. The outcome can be thought of as determined by a coin toss, with the probability of obtaining a "head" being θ and the probability of obtaining a "tail" being 1 − θ.

The Fisher information contained in n independent Bernoulli trials may be calculated as follows. In the following, A represents the number of successes, B the number of failures, and n = A + B is the total number of trials.

$\mathcal{I}(\theta) = -\mathrm{E} \left[ \frac{\partial^2}{\partial\theta^2} \ln(f(A;\theta)) \right] \qquad (1)$
$= -\mathrm{E} \left[ \frac{\partial^2}{\partial\theta^2} \ln \left[ \theta^A(1-\theta)^B\frac{(A+B)!}{A!B!} \right] \right] \qquad (2)$
$= -\mathrm{E} \left[ \frac{\partial^2}{\partial\theta^2} \left[ A \ln (\theta) + B \ln(1-\theta) \right] \right] \qquad (3)$
$= -\mathrm{E} \left[ \frac{\partial}{\partial\theta} \left[ \frac{A}{\theta} - \frac{B}{1-\theta} \right] \right]$ (on differentiating ln x, see logarithm) $\qquad (4)$
$= +\mathrm{E} \left[ \frac{A}{\theta^2} + \frac{B}{(1-\theta)^2} \right] \qquad (5)$
$= \frac{n\theta}{\theta^2} + \frac{n(1-\theta)}{(1-\theta)^2}$ (as the expected value of A = nθ, etc.) $\qquad (6)$
$= \frac{n}{\theta(1-\theta)} \qquad (7)$

(1) defines Fisher information. (2) invokes the fact that the information in a sufficient statistic is the same as that of the sample itself. (3) expands the log term and drops a constant. (4) and (5) differentiate with respect to θ. (6) replaces A and B with their expectations. (7) is algebra.

The end result, namely,

$\mathcal{I}(\theta) = \frac{n}{\theta(1-\theta)},$

is the reciprocal of the variance of the mean number of successes in n Bernoulli trials, as expected (see last sentence of the preceding section).

Matrix form

When there are N parameters, so that θ is a Nx1 vector $\theta = \begin{bmatrix} \theta_{1}, \theta_{2}, \dots , \theta_{N} \end{bmatrix},$ then the Fisher information takes the form of an NxN matrix, the Fisher Information Matrix (FIM), with typical element:

${\left(\mathcal{I} \left(\theta \right) \right)}_{i, j} = \mathrm{E} \left[\left. \frac{\partial}{\partial\theta_i} \ln f(X;\theta) \frac{\partial}{\partial\theta_j} \ln f(X;\theta) \right|\theta\right].$

The FIM is a NxN positive semidefinite symmetric matrix, defining a Riemannian metric on the N-dimensional parameter space, thus connecting Fisher information to differential geometry. In that context, this metric is known as the Fisher information metric, and the topic is called information geometry.

If the following regularity condition is met:

$\int \frac{\partial^2}{\partial \theta_i \partial \theta_j} f(X ; \theta ) \, dx = 0,$

then the Fisher Information Matrix may also be written as:

${\left(\mathcal{I} \left(\theta \right) \right)}_{i, j} = - \mathrm{E} \left[\left. \frac{\partial^2}{\partial\theta_i \partial\theta_j} \ln f(X;\theta) \right|\theta\right].$

Orthogonal parameters

We say that two parameters θi and θj are orthogonal if the element of the i-th row and j-th column of the Fisher Information Matrix is zero. Orthogonal parameters are easy to deal with in the sense that their maximum likelihood estimates are independent and can be calculated separately. When dealing with research problems, it is very common for the researcher to invest some time searching for an orthogonal parametrization of the densities involved in the problem.

Multivariate normal distribution

The FIM for a N-variate multivariate normal distribution has a special form. Let $\mu(\theta) = \begin{bmatrix} \mu_{1}(\theta), \mu_{2}(\theta), \dots , \mu_{N}(\theta) \end{bmatrix}^\top,$ and let Σ(θ) be the covariance matrix. Then the typical element $\mathcal{I}_{m,n}$, 0 ≤ m, n < N, of the FIM for $X \sim N(\mu(\theta), \Sigma(\theta))$ is:

$\mathcal{I}_{m,n} = \frac{\partial \mu^\top}{\partial \theta_m} \Sigma^{-1} \frac{\partial \mu}{\partial \theta_n} + \frac{1}{2} \mathrm{tr} \left( \Sigma^{-1} \frac{\partial \Sigma}{\partial \theta_m} \Sigma^{-1} \frac{\partial \Sigma}{\partial \theta_n} \right),$

where $(..)^\top$ denotes the transpose of a vector, tr(..) denotes the trace of a square matrix, and:

• $\frac{\partial \mu}{\partial \theta_m} = \begin{bmatrix} \frac{\partial \mu_1}{\partial \theta_m} & \frac{\partial \mu_2}{\partial \theta_m} & \cdots & \frac{\partial \mu_N}{\partial \theta_m} & \end{bmatrix}^\top;$
• $\frac{\partial \Sigma}{\partial \theta_m} = \begin{bmatrix} \frac{\partial \Sigma_{1,1}}{\partial \theta_m} & \frac{\partial \Sigma_{1,2}}{\partial \theta_m} & \cdots & \frac{\partial \Sigma_{1,N}}{\partial \theta_m} \\ \ \frac{\partial \Sigma_{2,1}}{\partial \theta_m} & \frac{\partial \Sigma_{2,2}}{\partial \theta_m} & \cdots & \frac{\partial \Sigma_{2,N}}{\partial \theta_m} \\ \ \vdots & \vdots & \ddots & \vdots \\ \ \frac{\partial \Sigma_{N,1}}{\partial \theta_m} & \frac{\partial \Sigma_{N,2}}{\partial \theta_m} & \cdots & \frac{\partial \Sigma_{N,N}}{\partial \theta_m} \end{bmatrix}.$

Note that a special, but very common case is the one where Σ(θ) = Σ, a constant. Then

$\mathcal{I}_{m,n} = \frac{\partial \mu^\top}{\partial \theta_m} \Sigma^{-1} \frac{\partial \mu}{\partial \theta_n}.\$

In this case the Fisher information matrix may be identified with the coefficient matrix of the normal equations of least squares estimation theory.

Properties

Reparametrization

The Fisher information depends on the parametrization of the problem. If θ and η are two scalar parametrizations of an estimation problem, such that θ = h(η) and h is a differentiable function, then

${\mathcal I}_\eta(\eta) = {\mathcal I}_\theta(h(\eta)) \left( h'(\eta) \right)^2$

where ${\mathcal I}_\eta$ and ${\mathcal I}_\theta$ are the Fisher information measures of η and θ, respectively.[9]

In the vector case, suppose ${\boldsymbol \theta}$ and ${\boldsymbol \eta}$ are k-vectors which parametrize an estimation problem, and suppose that ${\boldsymbol \theta} = {\boldsymbol h}({\boldsymbol \eta})$ where ${\boldsymbol h}: {\mathbb R}^k \rightarrow {\mathbb R}^k$ is continuously differentiable. Then,[10]

${\mathcal I}_{\boldsymbol \eta}({\boldsymbol \eta}) = {\boldsymbol J} {\mathcal I}_{\boldsymbol \theta} ({\boldsymbol h}({\boldsymbol \eta})) {\boldsymbol J}^*$

where the (i,j)th element of the Jacobian matrix $\boldsymbol J$ is defined by

$J_{ij} = \frac{\partial h_j}{\partial \theta_i}.$

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 parametrization.

Applications

Optimal design of experiments

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 matrix 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 a mean-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-matrix addition, under matrix-inversion, and under the multiplication of positive real-numbers and matrices. An exposition of matrix theory and the Loewner-order appears in Pukelsheim.

The traditional optimality-criteria are the information-matrix's invariants; algebraically, the traditional optimality-criteria are functionals of the eigenvalues of the (Fisher) information matrix: see optimal design.

Jeffreys prior in Bayesian statistics

In Bayesian statistics, the Fisher information is used to calculate the Jeffreys prior, which is a standard, non-informative prior for continuous distribution parameters.[11]

Other measures employed in information theory:

Notes

1. ^ Lehmann and Casella, p. 115
2. ^ Savage (1976)
3. ^ Savage(1976), page 156
4. ^ Edgeworth (Sept. 1980, Dec. 1908)
5. ^ Pratt(1976)
6. ^ Stigler (1978,1986,1999)
7. ^ Hald (1998,1999)
8. ^ Lehmann and Casella, eq. (2.5.16).
9. ^ Lehmann and Casella, eq. (2.5.11).
10. ^ Lehmann and Casella, eq. (2.6.16)
11. ^ Bayesian theory, Jose M. Bernardo and Adrian FM. Smith, John Wiley & Sons, 1994

References

• Schervish, Mark J. (1995). "Section 2.3.1". Theory of Statistics. New York: Springer. ISBN 0387945466.
• Stephen Stigler (1978). The History of Statistics: The Measurement of Uncertainty before 1900. Harvard University Press. ISBN 0-674-40340-1.
• Stephen Stigler (1999). Statistics on the Table: The History of Statistical Concepts and Methods. Harvard University Press. ISBN 0-674-83601-4.
• Anders Hald (May 1999). "On the History of Maximum Likelihood in Relation to Inverse Probability and Least Squares". Statistical Science 14 (2).   Stable URL: http://www.jstor.org/stable/2676741
• Hald, A. (1998). A History of Mathematical Statistics from 1750 to 1930. New York: Wiley.
• Van Trees, H. L. (1968). Detection, Estimation, and Modulation Theory, Part I. New York: Wiley. ISBN 0471095176.
• Lehmann, E. L.; Casella, G. (1998). Theory of Point Estimation. Springer. pp. 2nd ed. ISBN 0-387-98502-6.