|
|
| (One intermediate revision by one other user not shown) |
| Line 1: |
Line 1: |
| In [[estimation theory]] and [[statistics]], the '''Cramér–Rao bound (CRB)''' or '''Cramér–Rao lower bound (CRLB)''', named in honor of [[Harald Cramér]] and [[Calyampudi Radhakrishna Rao]] who were among the first to derive it,<ref name="Cramèr">{{cite book | last = Cramér | first = Harald | title = Mathematical Methods of Statistics | place = Princeton, NJ | publisher = Princeton Univ. Press | year = 1946 | isbn = 0-691-08004-6 | oclc = 185436716 }}</ref><ref name="Rao">{{cite journal | last = Rao | first = Calyampudi Radakrishna | title = Information and the accuracy attainable in the estimation of statistical parameters | journal = Bulletin of the [[Calcutta Mathematical Society]] |mr=0015748 | volume = 37 | pages = 81–89 | year = 1945 }}</ref><ref name="Rao papers">{{cite book | last = Rao | first = Calyampudi Radakrishna | title = Selected Papers of C. R. Rao | editor = S. Das Gupta | place = New York | publisher = Wiley | year = 1994 | isbn = 978-0-470-22091-7 | oclc = 174244259 }}</ref> expresses a lower bound on the [[variance]] of [[estimator]]s of a deterministic parameter. The bound is also known as the '''Cramér–Rao inequality''' or the '''information inequality'''.
| | Every one of the trophies from all in the members in your group get added up and as well , divided by 2 ascertain your clans overall awards. Playing many different kinds of games could make your gaming time more fun. If you cherished this post and you would like to acquire far more info relating to clash of clans hack no survey - [http://prometeu.net simply click prometeu.net], kindly go to the [http://website.org/ web site]. and your league also determines your incredible battle win bonus. 5 star rating and is known to be very addictive as players often devote several hours enjoying the game. She centers on beauty salon business manufacturing and client fascination.<br><br>Video gaming are fun to have fun with your kids. Assists you learn much more your kid's interests. Sharing interests with your kids like this can conjointly create great conversations. It also gives an opportunity to monitor growth and development of their skills.<br><br>Interweaving social trends form effective net in which we all have been trapped. When Each Tygers of Pan Tang sang 'It's lonely towards the top. Everyones trying to do we in', these people funded much from clash of clans hack into tool no survey. A society without collide of clans hack solution no survey is for a society with no knowledge, in that it pretty good.<br><br>There are no aftermaths in the least time for attacking other players and simply losing, so just go after and savor it. Win or lose, yourself may lose the nearly all troops you have in only the attack since this company are only beneficial on one mission, nevertheless, you can can steal more money with the enemy hamlet than it cost and make the troops. And you just produce more troops within your primary barracks. It''s per good idea to take them queued up previous to you decide to panic or anxiety attack and that means your family are rebuilding your soldiers through the battle.<br><br>This testing has apparent that this appraisement algorithm strategy consists of a alternation of beeline band sectors. They are not too things to consider different versions of arced graphs. I will explain the later.<br><br>A tutorial will guide you through your first few raids, constructions, and upgrades, yet unfortunately youre left to private wiles pretty quickly. Your buildings take real-time to construct and upgrade, your army units much better recruit, and your reference buildings take time produce food and gold. Like all of their genre cousins, Throne Rush is meant to took part in multiple short bursts the sun sets. This type of compulsive gaming definitely works more complete on mobile devices that always with you and could send push notifications when timed tasks are finalized. Then again, the success of so many hit Facebook games over the years indicates that people try Facebook often enough to produce short play sessions work there too.<br><br>Now that you have read this composition, you need to a great easier time locating then loving video games inside your life. Notwithstanding your favored platform, from your cellphone in your own own computer, playing furthermore enjoying video gaming can allow you to take the advantage of the worries of the actual busy week get information. |
| | |
| In its simplest form, the bound states that the variance of any [[bias of an estimator|unbiased]] estimator is at least as high as the inverse of the [[Fisher information]]. An unbiased estimator which achieves this lower bound is said to be (fully) [[Efficiency (statistics)|efficient]]. Such a solution achieves the lowest possible [[mean squared error]] among all unbiased methods, and is therefore the [[minimum variance unbiased]] (MVU) estimator. However, in some cases, no unbiased technique exists which achieves the bound. This may occur even when an MVU estimator exists.
| |
| | |
| The Cramér–Rao bound can also be used to bound the variance of [[estimator bias|''biased'' estimators]] of given bias. In some cases, a biased approach can result in both a variance and a [[mean squared error]] that are ''below'' the unbiased Cramér–Rao lower bound; see [[estimator bias]].
| |
| | |
| == Statement ==
| |
| | |
| The Cramér–Rao bound is stated in this section for several increasingly general cases, beginning with the case in which the parameter is a [[Scalar (mathematics)|scalar]] and its estimator is [[estimator bias|unbiased]]. All versions of the bound require certain regularity conditions, which hold for most well-behaved distributions. These conditions are listed [[#Regularity conditions|later in this section]].
| |
| | |
| === Scalar unbiased case ===
| |
| Suppose <math>\theta</math> is an unknown deterministic parameter which is to be estimated from measurements <math>x</math>, distributed according to some [[probability density function]] <math>f(x;\theta)</math>. The [[variance]] of any ''unbiased'' estimator <math>\hat{\theta}</math> of <math>\theta</math> is then bounded by the [[multiplicative inverse|reciprocal]] of the [[Fisher information]] <math>I(\theta)</math>:
| |
| | |
| :<math>\mathrm{var}(\hat{\theta})
| |
| \geq
| |
| \frac{1}{I(\theta)}
| |
| </math>
| |
| where the Fisher information <math>I(\theta)</math> is defined by
| |
| :<math> | |
| I(\theta) = \mathrm{E}
| |
| \left[
| |
| \left(
| |
| \frac{\partial \ell(x;\theta)}{\partial\theta}
| |
| \right)^2
| |
| \right] = -\mathrm{E}\left[ \frac{\partial^2 \ell(x;\theta)}{\partial\theta^2} \right]
| |
| </math>
| |
| and <math>\ell(x;\theta)=\log f(x;\theta)</math> is the [[natural logarithm]] of the [[likelihood function]] and <math>\mathrm{E}</math> denotes the [[expected value]] (over <math>x</math>).
| |
| | |
| The [[efficiency (statistics)|efficiency]] of an unbiased estimator <math>\hat{\theta}</math> measures how close this estimator's variance comes to this lower bound; estimator efficiency is defined as
| |
| | |
| :<math>e(\hat{\theta}) = \frac{I(\theta)^{-1}}{{\rm var}(\hat{\theta})}</math>
| |
| | |
| or the minimum possible variance for an unbiased estimator divided by its actual variance.
| |
| The Cramér–Rao lower bound thus gives
| |
| :<math>e(\hat{\theta}) \le 1.\ </math> | |
| | |
| === General scalar case ===
| |
| A more general form of the bound can be obtained by considering an unbiased estimator <math>T(X)</math> of a function <math>\psi(\theta)</math> of the parameter <math>\theta</math>. Here, unbiasedness is understood as stating that <math>E\{T(X)\} = \psi(\theta)</math>. In this case, the bound is given by
| |
| :<math>
| |
| \mathrm{var}(T)
| |
| \geq
| |
| \frac{[\psi'(\theta)]^2}{I(\theta)}
| |
| </math>
| |
| where <math>\psi'(\theta)</math> is the derivative of <math>\psi(\theta)</math> (by <math>\theta</math>), and <math>I(\theta)</math> is the Fisher information defined above.
| |
| | |
| === Bound on the variance of biased estimators ===
| |
| Apart from being a bound on estimators of functions of the parameter, this approach can be used to derive a bound on the variance of biased estimators with a given bias, as follows. Consider an estimator <math>\hat{\theta}</math> with bias <math>b(\theta) = E\{\hat{\theta}\} - \theta</math>, and let <math>\psi(\theta) = b(\theta) + \theta</math>. By the result above, any unbiased estimator whose expectation is <math>\psi(\theta)</math> has variance greater than or equal to <math>(\psi'(\theta))^2/I(\theta)</math>. Thus, any estimator <math>\hat{\theta}</math> whose bias is given by a function <math>b(\theta)</math> satisfies
| |
| :<math>
| |
| \mathrm{var} \left(\hat{\theta}\right)
| |
| \geq
| |
| \frac{[1+b'(\theta)]^2}{I(\theta)}.
| |
| </math>
| |
| The unbiased version of the bound is a special case of this result, with <math>b(\theta)=0</math>.
| |
| | |
| It's trivial to have a small variance − an "estimator" that is constant has a variance of zero. But from the above equation we find that the [[mean squared error]] of a biased estimator is bounded by | |
| | |
| :<math>\mathrm{E}\left((\hat{\theta}-\theta)^2\right)\geq\frac{[1+b'(\theta)]^2}{I(\theta)}+b(\theta)^2,</math>
| |
| | |
| using the standard decomposition of the MSE. Note, however, that this bound can be less than the unbiased Cramér–Rao bound 1/''I''(θ). See the example of estimating variance below.
| |
| | |
| === Multivariate case ===
| |
| Extending the Cramér–Rao bound to multiple parameters, define a parameter column [[vector space|vector]]
| |
| :<math>\boldsymbol{\theta} = \left[ \theta_1, \theta_2, \dots, \theta_d \right]^T \in \mathbb{R}^d</math>
| |
| with probability density function <math>f(x; \boldsymbol{\theta})</math> which satisfies the two [[#Regularity conditions|regularity conditions]] below.
| |
| | |
| The [[Fisher information matrix]] is a <math>d \times d</math> matrix with element <math>I_{m, k}</math> defined as
| |
| : <math>
| |
| I_{m, k}
| |
| = \mathrm{E} \left[
| |
| \frac{\partial }{\partial \theta_m} \log f\left(x; \boldsymbol{\theta}\right)
| |
| \frac{\partial }{\partial \theta_k} \log f\left(x; \boldsymbol{\theta}\right)
| |
| \right] = -\mathrm{E} \left[
| |
| \frac{\partial ^2}{\partial \theta_m \partial \theta_k} \log f\left(x; \boldsymbol{\theta}\right)
| |
| \right].
| |
| </math>
| |
| | |
| Let <math>\boldsymbol{T}(X)</math> be an estimator of any vector function of parameters, <math>\boldsymbol{T}(X) = (T_1(X), \ldots, T_n(X))^T</math>, and denote its expectation vector <math>\mathrm{E}[\boldsymbol{T}(X)]</math> by <math>\boldsymbol{\psi}(\boldsymbol{\theta})</math>. The Cramér–Rao bound then states that the [[covariance matrix]] of <math>\boldsymbol{T}(X)</math> satisfies
| |
| : <math>
| |
| \mathrm{cov}_{\boldsymbol{\theta}}\left(\boldsymbol{T}(X)\right)
| |
| \geq
| |
| \frac
| |
| {\partial \boldsymbol{\psi} \left(\boldsymbol{\theta}\right)}
| |
| {\partial \boldsymbol{\theta}}
| |
| [I\left(\boldsymbol{\theta}\right)]^{-1}
| |
| \left(
| |
| \frac
| |
| {\partial \boldsymbol{\psi}\left(\boldsymbol{\theta}\right)}
| |
| {\partial \boldsymbol{\theta}}
| |
| \right)^T
| |
| </math>
| |
| where
| |
| * The matrix inequality <math>A \ge B</math> is understood to mean that the matrix <math>A-B</math> is [[positive semidefinite matrix|positive semidefinite]], and
| |
| * <math>\partial \boldsymbol{\psi}(\boldsymbol{\theta})/\partial \boldsymbol{\theta}</math> is the [[Jacobian matrix]] whose <math>ij</math>th element is given by <math>\partial \psi_i(\boldsymbol{\theta})/\partial \theta_j</math>.
| |
| | |
| <!-- please leave this extra space as it improves legibility. -->
| |
| | |
| If <math>\boldsymbol{T}(X)</math> is an [[estimator bias|unbiased]] estimator of <math>\boldsymbol{\theta}</math> (i.e., <math>\boldsymbol{\psi}\left(\boldsymbol{\theta}\right) = \boldsymbol{\theta}</math>), then the Cramér–Rao bound reduces to
| |
| : <math>
| |
| \mathrm{cov}_{\boldsymbol{\theta}}\left(\boldsymbol{T}(X)\right)
| |
| \geq
| |
| I\left(\boldsymbol{\theta}\right)^{-1}.
| |
| </math>
| |
| | |
| If it is inconvenient to compute the inverse of the [[Fisher information matrix]],
| |
| then one can simply take the reciprocal of the corresponding diagonal element
| |
| to find a (possibly loose) lower bound
| |
| (For the Bayesian case, see eqn. (11) of Bobrovsky, Mayer-Wolf, Zakai,
| |
| "Some classes of global Cramer-Rao bounds", Ann. Stats., 15(4):1421-38, 1987).
| |
| | |
| : <math>
| |
| \mathrm{var}_{\boldsymbol{\theta}}\left(T_m(X)\right)
| |
| =
| |
| \left[\mathrm{cov}_{\boldsymbol{\theta}}\left(\boldsymbol{T}(X)\right)\right]_{mm}
| |
| \geq
| |
| \left[I\left(\boldsymbol{\theta}\right)^{-1}\right]_{mm}
| |
| \geq
| |
| \left(\left[I\left(\boldsymbol{\theta}\right)\right]_{mm}\right)^{-1}.
| |
| </math>
| |
| | |
| === Regularity conditions ===
| |
| The bound relies on two weak regularity conditions on the [[probability density function]], <math>f(x; \theta)</math>, and the estimator <math>T(X)</math>:
| |
| * The Fisher information is always defined; equivalently, for all <math>x</math> such that <math>f(x; \theta) > 0</math>,
| |
| ::<math> \frac{\partial}{\partial\theta} \log f(x;\theta)</math>
| |
| :exists, and is finite.
| |
| * The operations of integration with respect to <math>x</math> and differentiation with respect to <math>\theta</math> can be interchanged in the expectation of <math>T</math>; that is,
| |
| ::<math>
| |
| \frac{\partial}{\partial\theta}
| |
| \left[
| |
| \int T(x) f(x;\theta) \,dx
| |
| \right]
| |
| =
| |
| \int T(x)
| |
| \left[
| |
| \frac{\partial}{\partial\theta} f(x;\theta)
| |
| \right]
| |
| \,dx
| |
| </math>
| |
| :whenever the right-hand side is finite.
| |
| :This condition can often be confirmed by using the fact that integration and differentiation can be swapped when either of the following cases hold:
| |
| :# The function <math>f(x;\theta)</math> has bounded support in <math>x</math>, and the bounds do not depend on <math>\theta</math>;
| |
| :# The function <math>f(x;\theta)</math> has infinite support, is [[continuously differentiable]], and the integral converges uniformly for all <math>\theta</math>.
| |
| | |
| === Simplified form of the Fisher information ===
| |
| Suppose, in addition, that the operations of integration and differentiation can be swapped for the second derivative of <math>f(x;\theta)</math> as well, i.e.,
| |
| :<math> \frac{\partial^2}{\partial\theta^2}
| |
| \left[
| |
| \int T(x) f(x;\theta) \,dx
| |
| \right]
| |
| =
| |
| \int T(x)
| |
| \left[
| |
| \frac{\partial^2}{\partial\theta^2} f(x;\theta)
| |
| \right]
| |
| \,dx.
| |
| </math> | |
| In this case, it can be shown that the Fisher information equals
| |
| :<math>
| |
| I(\theta)
| |
| =
| |
| -\mathrm{E}
| |
| \left[
| |
| \frac{\partial^2}{\partial\theta^2} \log f(X;\theta)
| |
| \right].
| |
| </math>
| |
| The Cramèr–Rao bound can then be written as
| |
| :<math>
| |
| \mathrm{var} \left(\widehat{\theta}\right)
| |
| \geq
| |
| \frac{1}{I(\theta)}
| |
| =
| |
| \frac{1}
| |
| {
| |
| -\mathrm{E}
| |
| \left[
| |
| \frac{\partial^2}{\partial\theta^2} \log f(X;\theta)
| |
| \right]
| |
| }.
| |
| </math>
| |
| In some cases, this formula gives a more convenient technique for evaluating the bound.
| |
| | |
| == Single-parameter proof ==
| |
| The following is a proof of the general scalar case of the Cramér–Rao bound, which was described [[#General scalar case|above]]; namely, that if the expectation of <math>T</math> is denoted by <math>\psi (\theta)</math>, then, for all <math>\theta</math>,
| |
| :<math>{\rm var}(t(X)) \geq \frac{[\psi^\prime(\theta)]^2}{I(\theta)}.</math>
| |
| | |
| Let <math>X</math> be a [[random variable]] with probability density function <math>f(x; \theta)</math>.
| |
| Here <math>T = t(X)</math> is a [[statistic]], which is used as an [[estimator]] for <math>\psi (\theta)</math>. If <math>V</math> is the [[score (statistics)|score]], i.e.
| |
| | |
| :<math>V = \frac{\partial}{\partial\theta} \ln f(X;\theta)</math>
| |
| | |
| then the [[expected value|expectation]] of <math>V</math>, written <math>{\rm E}(V)</math>, is zero.
| |
| If we consider the [[covariance]] <math>{\rm cov}(V, T)</math> of <math>V</math> and <math>T</math>, we have <math>{\rm cov}(V, T) = {\rm E}(V T)</math>, because <math>{\rm E}(V) = 0</math>. Expanding this expression we have
| |
| | |
| :<math>
| |
| {\rm cov}(V,T)
| |
| =
| |
| {\rm E}
| |
| \left(
| |
| T \cdot \frac{\partial}{\partial\theta} \ln f(X;\theta)
| |
| \right)
| |
| </math>
| |
| | |
| This may be expanded using the [[chain rule]]
| |
| | |
| :<math>\frac{\partial}{\partial\theta} \ln Q = \frac{1}{Q}\frac{\partial Q}{\partial\theta}</math>
| |
| | |
| and the definition of expectation gives, after cancelling <math>f(x; \theta)</math>,
| |
| | |
| :<math>
| |
| {\rm E} \left(
| |
| T \cdot \frac{\partial}{\partial\theta} \ln f(X;\theta)
| |
| \right)
| |
| =
| |
| \int
| |
| t(x)
| |
| \left[
| |
| \frac{\partial}{\partial\theta} f(x;\theta)
| |
| \right]
| |
| \, dx
| |
| =
| |
| \frac{\partial}{\partial\theta}
| |
| \left[
| |
| \int t(x)f(x;\theta)\,dx
| |
| \right]
| |
| =
| |
| \psi^\prime(\theta)
| |
| </math>
| |
| | |
| because the integration and differentiation operations commute (second condition).
| |
| | |
| The [[Cauchy–Schwarz inequality]] shows that
| |
| | |
| :<math>
| |
| \sqrt{ {\rm var} (T) {\rm var} (V)} \geq \left| {\rm cov}(V,T) \right| = \left | \psi^\prime (\theta)
| |
| \right |</math>
| |
| | |
| therefore
| |
| | |
| :<math>
| |
| {\rm var\ } T \geq \frac{[\psi^\prime(\theta)]^2}{{\rm var} (V)}
| |
| =
| |
| \frac{[\psi^\prime(\theta)]^2}{I(\theta)}
| |
| =
| |
| \left[
| |
| \frac{\partial}{\partial\theta}
| |
| {\rm E} (T)
| |
| \right]^2
| |
| \frac{1}{I(\theta)}
| |
| </math>
| |
| which proves the proposition.
| |
| | |
| ==Examples==
| |
| | |
| ===Multivariate normal distribution===
| |
| For the case of a [[multivariate normal distribution|''d''-variate normal distribution]]
| |
| : <math>
| |
| \boldsymbol{x}
| |
| \sim
| |
| N_d
| |
| \left(
| |
| \boldsymbol{\mu} \left( \boldsymbol{\theta} \right)
| |
| ,
| |
| {\boldsymbol C} \left( \boldsymbol{\theta} \right)
| |
| \right)
| |
| </math>
| |
| the [[Fisher information matrix]] has elements<ref>{{cite book
| |
| | last = Kay
| |
| | first = S. M.
| |
| | title = Fundamentals of Statistical Signal Processing: Estimation Theory
| |
| | year = 1993
| |
| | publisher = Prentice Hall
| |
| | page = 47
| |
| | isbn = 0-13-042268-1 }}
| |
| </ref>
| |
| :<math>
| |
| I_{m, k}
| |
| =
| |
| \frac{\partial \boldsymbol{\mu}^T}{\partial \theta_m}
| |
| {\boldsymbol C}^{-1}
| |
| \frac{\partial \boldsymbol{\mu}}{\partial \theta_k}
| |
| +
| |
| \frac{1}{2}
| |
| \mathrm{tr}
| |
| \left(
| |
| {\boldsymbol C}^{-1}
| |
| \frac{\partial {\boldsymbol C}}{\partial \theta_m}
| |
| {\boldsymbol C}^{-1}
| |
| \frac{\partial {\boldsymbol C}}{\partial \theta_k}
| |
| \right)
| |
| </math>
| |
| where "tr" is the [[trace (matrix)|trace]].
| |
| | |
| For example, let <math>w[n]</math> be a sample of <math>N</math> independent observations) with unknown mean <math>\theta</math> and known variance <math>\sigma^2</math>
| |
| :<math>w[n] \sim \mathbb{N}_N \left(\theta {\boldsymbol 1}, \sigma^2 {\boldsymbol I} \right).</math>
| |
| Then the Fisher information is a scalar given by
| |
| :<math>
| |
| I(\theta)
| |
| =
| |
| \left(\frac{\partial\boldsymbol{\mu}(\theta)}{\partial\theta}\right)^T{\boldsymbol C}^{-1}\left(\frac{\partial\boldsymbol{\mu}(\theta)}{\partial\theta}\right)
| |
| = \sum^N_{i=1}\frac{1}{\sigma^2} = \frac{N}{\sigma^2},
| |
| </math>
| |
| and so the Cramér–Rao bound is
| |
| :<math>
| |
| \mathrm{var}\left(\hat \theta\right)
| |
| \geq
| |
| \frac{\sigma^2}{N}.
| |
| </math>
| |
| | |
| ===Normal variance with known mean===
| |
| Suppose ''X'' is a [[normal distribution|normally distributed]] random variable with known mean <math>\mu</math> and unknown variance <math>\sigma^2</math>. Consider the following statistic:
| |
| | |
| :<math>
| |
| T=\frac{\sum_{i=1}^n\left(X_i-\mu\right)^2}{n}.
| |
| </math>
| |
| | |
| Then ''T'' is unbiased for <math>\sigma^2</math>, as <math>E(T)=\sigma^2</math>. What is the variance of ''T''? | |
| | |
| :<math>
| |
| \mathrm{var}(T) = \frac{\mathrm{var}(X-\mu)^2}{n}=\frac{1}{n}
| |
| \left[
| |
| E\left\{(X-\mu)^4\right\}-\left(E\left\{(X-\mu)^2\right\}\right)^2
| |
| \right]
| |
| </math>
| |
| | |
| (the second equality follows directly from the definition of variance). The first term is the fourth [[moment about the mean]] and has value <math>3(\sigma^2)^2</math>; the second is the square of the variance, or <math>(\sigma^2)^2</math>.
| |
| Thus
| |
| | |
| :<math>\mathrm{var}(T)=\frac{2(\sigma^2)^2}{n}.</math>
| |
| | |
| Now, what is the [[Fisher information]] in the sample? Recall that the [[score (statistics)|score]] ''V'' is defined as | |
| | |
| :<math>
| |
| V=\frac{\partial}{\partial\sigma^2}\log L(\sigma^2,X)
| |
| </math>
| |
| | |
| where <math>L</math> is the [[likelihood function]]. Thus in this case,
| |
| | |
| :<math>
| |
| V=\frac{\partial}{\partial\sigma^2}\log\left[\frac{1}{\sqrt{2\pi\sigma^2}}e^{-(X-\mu)^2/{2\sigma^2}}\right]
| |
| =\frac{(X-\mu)^2}{2(\sigma^2)^2}-\frac{1}{2\sigma^2}
| |
| </math>
| |
| | |
| where the second equality is from elementary calculus. Thus, the information in a single observation is just minus the expectation of the derivative of ''V'', or
| |
| | |
| :<math>
| |
| I
| |
| =-E\left(\frac{\partial V}{\partial\sigma^2}\right)
| |
| =-E\left(-\frac{(X-\mu)^2}{(\sigma^2)^3}+\frac{1}{2(\sigma^2)^2}\right)
| |
| =\frac{\sigma^2}{(\sigma^2)^3}-\frac{1}{2(\sigma^2)^2}
| |
| =\frac{1}{2(\sigma^2)^2}.</math>
| |
| | |
| Thus the information in a sample of <math>n</math> independent observations is just <math>n</math> times this, or <math>\frac{n}{2(\sigma^2)^2}.</math>
| |
| | |
| The Cramer Rao bound states that
| |
| | |
| :<math>
| |
| \mathrm{var}(T)\geq\frac{1}{I}.</math>
| |
| | |
| In this case, the inequality is saturated (equality is achieved), showing that the [[estimator]] is [[efficiency (statistics)|efficient]].
| |
| | |
| However, we can achieve a lower [[mean squared error]] using a biased estimator. The estimator
| |
| | |
| :<math>
| |
| T=\frac{\sum_{i=1}^n\left(X_i-\mu\right)^2}{n+2}.
| |
| </math>
| |
| | |
| obviously has a smaller variance, which is in fact
| |
| | |
| :<math>\mathrm{var}(T)=\frac{2n(\sigma^2)^2}{(n+2)^2}.</math>
| |
| | |
| Its bias is
| |
| | |
| <math>\left(1-\frac{n}{n+2}\right)\sigma^2=\frac{2\sigma^2}{n+2}</math>
| |
| | |
| so its mean squared error is
| |
| | |
| :<math>\mathrm{MSE}(T)=\left(\frac{2n}{(n+2)^2}+\frac{4}{(n+2)^2}\right)(\sigma^2)^2
| |
| =\frac{2(\sigma^2)^2}{n+2}</math>
| |
| | |
| which is clearly less than the Cramér–Rao bound found above.
| |
| | |
| When the mean is not known, the minimum mean squared error estimate of the variance of a sample from Gaussian distribution is achieved by dividing by ''n'' + 1, rather than ''n'' − 1 or ''n'' + 2.
| |
| | |
| == See also ==
| |
| * [[Chapman–Robbins bound]]
| |
| * [[Kullback's inequality]]
| |
| | |
| == References and notes ==
| |
| {{reflist}}
| |
| | |
| == Further reading ==
| |
| * {{Cite journal
| |
| | last = Kay
| |
| | first = Steven M.
| |
| | title = Fundamentals of Statistical Signal Processing, Volume I: Estimation Theory
| |
| | publisher = Prentice Hall
| |
| | year = 1993
| |
| | isbn = 0-13-345711-7 }}. Chapter 3.
| |
| * {{Cite journal
| |
| | last = Shao
| |
| | first = Jun
| |
| | title = Mathematical Statistics
| |
| | place = New York
| |
| | publisher = Springer
| |
| | year = 1998
| |
| | isbn = 0-387-98674-X }}. Section 3.1.3.
| |
| | |
| == External links ==
| |
| *[http://www4.utsouthwestern.edu/wardlab/fandplimittool.asp FandPLimitTool] a GUI-based software to calculate the Fisher information and Cramer-Rao Lower Bound with application to single-molecule microscopy.
| |
| | |
| {{DEFAULTSORT:Cramer-Rao bound}}
| |
| [[Category:Articles containing proofs]]
| |
| [[Category:Statistical inequalities]]
| |
| [[Category:Estimation theory]]
| |