Estimators1 PDF

Bias, Variance, and MSE of Estimators
Guy Lebanon
September 4, 2010
We assume that we have iid (independent identically distributed) samples X (1) , . . . , X (n) that follow some
unknown distribution. The task of statistics is to estimate properties of the unknown distribution. In this
note we focus one estimating a parameter of the distribution such as the mean or variance. In some cases
the parameter completely characterizes the distribution and estimating it provides a probability estimate.
In this note, we assume that the parameter is a real vector θ ∈ Rd . To estimate it, we use an es-
timator which is a function of our observations θ̂(x(1) , . . . , x(n) ). We follow standard practice and omit
(in notation only) the dependency of the estimator on the samples, i.e. we write θ̂. However, note that
θ̂ = θ̂(X (1) , . . . , X (n) ) is a random variable since it is a function of n random variables.
A desirable property of an estimator is that it is correct on average. That is, if there are repeated
samplings of n samples X (1) , . . . , X (n) , the estimator θ̂(X (1) , . . . , X (n) ) will have, on average, the correct
value. Such estimators are called unbiased.
Definition 1. The bias of θ̂ is1 Bias(θ̂) = E(θ̂) − θ. If it is 0, the estimator θ̂ is said to be unbiased.
There is, however, more important performance characterizations for an estimator than just being unbi-
ased. The mean squared error is perhaps the most important of them. It captures the error that the estimator
makes. However, since the estimator is a RV, we need to average over its distribution thus capturing the
average performance if there are many repeated samplings of X (1) , . . . , X (n) .
Pd
Definition 2. The mean squared error (MSE) of an estimator is E(kθ̂ − θk2 ) = E( j=1 (θ̂j − θj )2 ).
Theorem 1.
E(kθ̂ − θk2 ) = trace(Var(θ̂)) + kBias(θ̂)k2 .
Pd
Note that Var(θ̂) is the covariance matrix of θ̂ and so its trace is j=1 Var(θ̂j ).
Pd
Proof. Since the MSE equals j=1 E((θ̂j − θj )2 ) it is sufficient to prove for a scalar θ, E((θ̂ − θ)2 ) =
Var(θ̂) + Bias2 (θ̂):
E((θ̂ − θ)2 ) = E(((θ̂ − E(θ̂)) + (E(θ̂) − θ)2 ) = E{(θ̂ − E(θ̂))2 + (E(θ̂) − θ)2 + (θ̂ − E(θ̂))(E(θ̂) − θ)}
= Var(θ̂) + Bias2 (θ̂) + E((θ̂ − E(θ̂))(E(θ̂) − θ)) = Var(θ̂) + Bias2 (θ̂) + E(θ̂E(θ̂) − (E(θ̂))2 − θθ̂ + E(θ̂)θ)
= Var(θ̂) + Bias2 (θ̂) + (E(θ̂))2 − (E(θ̂))2 − θE(θ̂) + θE(θ̂) = Var(θ̂) + Bias2 (θ̂).
Since the MSE decomposes into a sum of the bias and variance of the estimator, both quantities are
important and need to be as small as possible to achieve good estimation performance. It is common to
trade-off some increase in bias for a larger decrease in the variance and vice-verse.
1 Note here and in the sequel all expectations are with respect to X (1) , . . . , X (n) .
1
1
X (i) which estimates the vector E(X) and
P
Two important special cases are the mean θ̂ = X̄ = n
n
1 X (i)
θ̂ = S 2 , where Sj2 = (X − X̄j )2 j = 1, . . . , d
n − 1 i=1 j
which estimates the diagonal of the covariance matrix Var(X). We show below that both are unbiased and
therefore their MSE is simply their variance.
Theorem 2. X̄ is an unbiased estimator of E(X) and S 2 is an unbiased estimator of the diagonal of the
covariance matrix Var(X).
Proof. !
n
X n
X
−1 (i)
E(X̄) = E n X = E(X (i) )/n = nE(X (i) )/n.
i=1 i=1
To prove that S is unbiased we show that it is unbiased in the one dimensional case i.e., X, S 2 are scalars
2
(if this holds, we can apply this result to each component separately to get unbiasedness of the vector S 2 ).
We first need the following result (recall that below X is a scalar)
n
X X X X X
(X (i) − X̄)2 = (X (i) )2 − 2X̄ X (i) + nX̄ 2 = (X (i) )2 − 2nX̄ X̄ + nX̄ 2 = (X (i) )2 − nX̄ 2
i=1
and therefore
n
! n
! n
X X X
E (X (i) − X̄)2 =E (X (i) )2 − nX̄ 2 = E((X (i) )2 ) − nE(X̄ 2 ) = nE((X (1) )2 ) − nE(X̄ 2 ).
i=1 i=1 i=1
Substituting the expectations E(X ) = Var(X)+(E(X))2 = Var(X)+(EX)2 and E(X̄ 2 ) = Var(X̄)+(E(X̄))2 =

2
Var(X)
n + (EX)2 in the above equation we have
n
!
X
(i) 2 Var(X)
E (X − X̄) = n(Var(X) + (EX)2 ) − n + (EX)2 = (n − 1)Var(X).
i=1
n
Pn (i)
Returning now to the multivariate case, this implies E(Sj2 ) = E( i=1 (Xj − X̄j )2 /(n − 1)) = Var(Xj ) for
all j and therefore ES 2 = diag(Var(X)).
Pd Pd
Since the bias is zero, the MSE of X̄ as an estimator of EX is j=1 Var(X̄j ) = j=1 nn2 Var(Xj ). Thus,
if d is fixed and n → ∞ the MSE(X̄) → 0. For S 2 , the MSE is trace(Var(S 2 )) which may be computed with
some tedious algebra (it also decreases to 0 as n → ∞).
Another performance measure for estimators θ̂ is the probability that the estimator’s error is outside
some acceptable error range P (kθ̂ − θk ≥ ). However, to evaluate the above quantity, we need (i) the pdf
fθ̂ which depends on the pdf of X (which is typically unknown) and (ii) the true value θ (also typically
unknown). If θ̂ is unbiased we may obtain the following bound
 
Xd d
X d
X p
P (kθ̂ − θk2 ≥ ) = P  |θ̂j − θj |2 ≥  ≤ P (|θ̂j − θj |2 ≥ /d) = P (|θ̂j − θj | ≥ /d)
j=1 j=1 j=1
d
d X d
≤ Var(θ̂j ) = trace(Var(θ̂))
j=1

where we used Boole’s and Chebyshev’s inequalities. This again shows (but in a different way than the bias
variance decomposition of the MSE) that the quality of unbiased estimators is determined by trace(Var(θ̂)).

Estimators1 PDF

Uploaded by

Copyright:

Available Formats

Estimators1 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Estimators1 PDF

Uploaded by

Copyright:

Available Formats

Bias, Variance, and MSE of Estimators

Substituting the expectations E(X ) = Var(X)+(E(X))2 = Var(X)+(EX)2 and E(X̄ 2 ) = Var(X̄)+(E(X̄))2 =

You might also like