Written and maintained by CASRAI Editorial Board
Published · Last updated
On this page
Fisher information measures how much a sample tells you about an unknown parameter. This page is for analysts who run or report a maximum likelihood analysis and need to turn a model into a standard error. By the end you can compute Fisher information for a Bernoulli or normal model by hand, read it, tell it apart from the Hessian, and get it from R, Python or Excel.
Ask CASRAI · free to tryQuestion about this guide? Ask CASRAI answers with linked sources. 2 free questions a day, no account.
What Is Fisher Information? Definition
Fisher information, written I(θ), is the expected amount of information that a random variable X carries about a parameter θ of its distribution. Formally it is the variance of the score, where the score is the derivative of the log-likelihood with respect to θ. Under standard regularity conditions it also equals the negative expected second derivative of the log-likelihood, so a sharply peaked log-likelihood means high information (Ly, Marsman, Verhagen, Grasman and Wagenmakers, 2017, Definition 1.1). In practice its reciprocal gives the variance of the maximum likelihood estimate, which is why it matters for standard errors, confidence intervals and sample-size planning.
Fisher Information Formula
For a single observation X with density or mass function f(x | θ):
- Score: s(x; θ) = ∂/∂θ log f(x | θ).
- Definition 1 (expected squared score): I(θ) = E[ (∂/∂θ log f(X | θ))² ], a sum over x for discrete data or an integral for continuous data.
- Definition 2 (curvature): I(θ) = −E[ ∂²/∂θ² log f(X | θ) ], valid under mild regularity conditions.
Symbols: X is the random variable, θ is the unknown parameter, f(x | θ) is its probability function, log is the natural logarithm, E[·] is expectation taken with the true θ. Fisher information has to exist for every parameter value in the model (Ly et al., 2017). For n independent, identically distributed observations the information adds: In(θ) = n · I(θ). The asymptotic variance of the maximum likelihood estimator is then 1 / (n · I(θ)) (Ly et al., 2017, Eq. 2.1 to 2.2).
Two closed forms worth memorising:
- Bernoulli (success probability p): I(p) = 1 / (p(1 − p)) per trial, so n / (p(1 − p)) for n trials (Ly et al., 2017).
- Normal with known σ (mean μ): I(μ) = n / σ² for n observations. Derived: log f = −(x − μ)² / (2σ²) + constant, the score is (x − μ)/σ², its variance is σ²/σ⁴ = 1/σ², and n observations add n/σ².
How to Find Fisher Information: Worked Example With Numbers
Suppose 30 of n = 100 participants respond to a treatment, so the estimate is p̂ = 30/100 = 0.30. The model is Bernoulli with parameter p.
- Log-likelihood: ℓ(p) = k log p + (n − k) log(1 − p), with k = 30.
- Information per trial: 1 / (p(1 − p)). At p = 0.30 this is 1 / (0.30 × 0.70) = 1 / 0.21 = 4.762.
- Information from all 100 trials: 100 / 0.21 = 476.19.
- Standard error of p̂: √(1 / 476.19) = √0.0021 = 0.0458.
- Approximate 95% confidence interval: 0.30 ± 1.96 × 0.0458 = 0.30 ± 0.0898, that is 0.210 to 0.390.
This is the Wald interval, and it is only an approximation (see confidence interval for a proportion for better-behaved alternatives when p is near 0 or 1 or n is small).
Using Fisher information to plan a sample size
Because I(p) = 1/(p(1 − p)) is smallest at p = 0.5 (value 4), that is the worst case. To get a standard error of 0.05 whatever the true p is, you need n × 4 ≥ 1 / 0.05² = 400, so n = 100. The same logic with a standard error of 0.1 gives n = 25, which is the worst-case example in Ly et al. (2017, Example 2.1).
What Does the Fisher Information Tell You? Interpretation
It tells you how precisely the data can pin down the parameter. High information means a narrow, sharply curved likelihood and a small standard error; low information means a flat likelihood and a wide interval.
| What you see | Reading |
|---|---|
| Large I(θ) | Parameter is estimated precisely; variance of the MLE, 1/(nI), is small |
| Small I(θ) | Likelihood is flat near θ; the data say little about it |
| I grows with n | Information from independent observations adds: n × I(θ) |
| Bernoulli I smallest at p = 0.5 | Proportions near 0.5 need the largest samples for a given standard error |
| Normal, I(μ) = n/σ² | More noise (larger σ) means less information about the mean |
The Cramér-Rao bound makes the link to efficiency explicit: the variance of any unbiased estimator is at least 1/I(θ). See also the related ideas in the likelihood function, the covariance matrix (the inverse of the Fisher information matrix is the asymptotic covariance of the estimates when there are several parameters) and Kullback-Leibler divergence.
Fisher Information vs Hessian: Observed vs Expected Information
The Hessian is the matrix of second derivatives of the log-likelihood. Fisher (expected) information is the expectation of the negative Hessian, taken over the model; observed information is the negative Hessian itself, evaluated at the actual data, usually at the maximum likelihood estimate.
| Quantity | Definition | Depends on observed data? |
|---|---|---|
| Hessian | ∂²ℓ/∂θ² (second derivative of the log-likelihood) | Yes |
| Observed information | −∂²ℓ/∂θ², evaluated at the realised data, typically at θ̂ | Yes |
| Expected (Fisher) information | E[−∂²ℓ/∂θ²] under the model | No, only on θ and n |
In the Bernoulli example, ℓ”(p) = −k/p² − (n − k)/(1 − p)². At p̂ = k/n this gives observed information n / (p̂(1 − p̂)) = 476.19, identical to the expected value there. For other models the two differ, and either can be used for standard errors. This is why software that optimises a likelihood numerically and returns a Hessian is reporting observed information.
Assumptions and Common Mistakes
- Regularity conditions. The score and curvature forms agree only under standard regularity conditions, broadly that the support does not depend on θ and that the log-density is smooth enough to differentiate.
- One observation vs n. I(p) = 1/(p(1 − p)) is per trial; the standard error needs n times that.
- Reparametrisation. Information depends on how the parameter is written. Information about p is not information about log-odds, so quote the parameter you mean.
- Plugging in estimates. The standard error uses I evaluated at p̂, not at the unknown true value; this is an approximation that improves with n.
- Wrong Fisher. Fisher information is unrelated to Fisher’s exact test, which compares proportions in a 2 × 2 table. See the FAQ below.
How to Report It
Journals rarely want the information itself; they want the estimate and its uncertainty. An APA-style sentence: The response rate was 30% (SE = 0.046, 95% CI [0.21, 0.39]), with standard errors derived from the inverse of the observed Fisher information. See how to report confidence intervals in a manuscript.
How to Calculate Fisher Information in Excel, R or Python
Excel
A1: 100 (n)
A2: 0.3 (p-hat)
A3: =A1/(A2*(1-A2)) -> 476.19 (information)
A4: =SQRT(1/A3) -> 0.0458 (standard error)
A5: =A2-1.96*A4 -> 0.210
A6: =A2+1.96*A4 -> 0.390
R
n <- 100; k <- 30
nll <- function(p) -(k*log(p) + (n-k)*log(1-p))
fit <- optim(0.5, nll, method = "Brent", lower = 0.01, upper = 0.99, hessian = TRUE)
fit$par # 0.30
fit$hessian # observed information, about 476.19
sqrt(1/fit$hessian) # standard error, about 0.0458
Python
import math
n, k = 100, 30
nll = lambda p: -(k*math.log(p) + (n-k)*math.log(1-p))
p, h = k/n, 1e-4
hess = (nll(p+h) - 2*nll(p) + nll(p-h)) / h**2 # finite-difference second derivative
print(hess, 1/math.sqrt(hess)) # 476.19, 0.0458
FAQ
What does the Fisher information tell you?
How much the data reveal about a parameter. Its inverse, 1/(nI(θ)), is the asymptotic variance of the maximum likelihood estimate, so it sets the standard error and the interval width.
How do you find Fisher information?
Write the log-likelihood, differentiate twice with respect to the parameter, take the negative expectation (or the variance of the first derivative), and multiply by n for n independent observations. For Bernoulli data the result is n / (p(1 − p)).
What is the difference between Fisher information and Hessian information?
The Hessian is the second-derivative matrix of the log-likelihood and depends on the data. Fisher information is the expected negative Hessian and depends only on the model and parameter. Observed information is the negative Hessian at the estimate.
What does a Fisher test tell you?
That is a different method. Fisher’s exact test (R. A. Fisher) tests association in a 2 × 2 contingency table. It is not the same thing as Fisher information, although both are named after the same statistician.
What is the Fisher information matrix and quantum Fisher information?
With several parameters, Fisher information becomes a matrix, and its inverse gives the asymptotic covariance of the estimates. Quantum Fisher information is the analogue used in quantum estimation and is outside the scope of this page.
Sources
- Ly, A., Marsman, M., Verhagen, J., Grasman, R. P. P. P., and Wagenmakers, E.-J. (2017). A tutorial on Fisher information. Journal of Mathematical Psychology, 80. arXiv:1705.01064, arxiv.org/abs/1705.01064. Read 2026-11-07.
- Fisher information, Wikipedia, en.wikipedia.org/wiki/Fisher_information. Read 2026-11-07, used to corroborate the score, curvature and Cramér-Rao statements.
- Cao, X., and Spall, J. C. Relative performance of expected and observed Fisher information in covariance estimation for maximum likelihood estimates. arXiv:1305.1056.







