“Jensen's inequality”的意思、由来-开放百科全书

In mathematics, Jensen's inequality, named after the Danish mathematician Johan Jensen, relates the value of a convex function of an integral to the integral of the convex function. It was proven by Jensen in 1906.^[1] Given its generality, the inequality appears in many forms depending on the context, some of which are presented below. In its simplest form the inequality states that the convex transformation of a mean is less than or equal to the mean applied after convex transformation; it is a simple corollary that the opposite is true of concave transformations.

Jensen's inequality generalizes the statement that the secant line of a convex function lies above the graph of the function, which is Jensen's inequality for two points: the secant line consists of weighted means of the convex function (for t ∈ [0,1]),

In the context of probability theory, it is generally stated in the following form: if X is a random variable and {{mvar|φ}} is a convex function, then

Statements

The classical form of Jensen's inequality involves several numbers and weights. The inequality can be stated quite generally using either the language of measure theory or (equivalently) probability. In the probabilistic setting, the inequality can be further generalized to its full strength.

Finite form

For a real convex function

, numbers

in its domain, and positive weights

, Jensen's inequality can be stated as:

For instance, the function {{math|log(x)}} is concave, so substituting

in the previous formula (4) establishes the (logarithm of the) familiar arithmetic mean-geometric mean inequality:

A common application has

as a function of another variable (or set of variables)

, that is,

. All of this carries directly over to the general continuous case: the weights {{math|a_i}} are replaced by a non-negative integrable function {{math| f (x)}}, such as a probability distribution, and the summations are replaced by integrals.

Measure-theoretic and probabilistic form

Let

be a probability space, such that

. If

is a real-valued function that is

-integrable, and if

is a convex function on the real line, then:

where

, and

is a non-negative Lebesgue-integrable function. In this case, the Lebesgue measure of

need not be unity. However, by integration by substitution, the interval can be rescaled so that it has measure unity. Then Jensen's inequality can be applied to get^[2]

The same result can be equivalently stated in a probability theory setting, by a simple change of notation. Let

be a probability space, X an integrable real-valued random variable and {{mvar|φ}} a convex function. Then:

In this probability setting, the measure {{mvar|μ}} is intended as a probability

, the integral with respect to {{mvar|μ}} as an expected value

, and the function

as a random variable X.

Notice that the equality holds if X is constant (degenerate random variable) or if {{mvar|φ}} is linear, and even if there is

(a Borel set, in fact) such that

General inequality in a probabilistic setting

More generally, let T be a real topological vector space, and X a T-valued integrable random variable. In this general setting, integrable means that there exists an element

in T, such that for any element z in the dual space of T:

, and

. Then, for any measurable convex function {{mvar|φ}} and any sub-σ-algebra

Here

stands for the expectation conditioned to the σ-algebra

. This general statement reduces to the previous ones when the topological vector space {{mvar|T}} is the real axis, and

is the trivial {{mvar|σ}}-algebra {{math|{∅, Ω}.}}^[3]

A sharpened and generalized form

Let X be a one-dimensional random variable with mean

and variance

. Let

be a twice differentiable function, and define the function

In particular, when

is convex, then

, and the standard form of Jensen's inequality immediately follows.

Proofs

Jensen's inequality can be proved in several ways, and three different proofs corresponding to the different statements above will be offered. Before embarking on these mathematical derivations, however, it is worth analyzing an intuitive graphical argument based on the probabilistic case where {{mvar|X}} is a real number (see figure). Assuming a hypothetical distribution of {{mvar|X}} values, one can immediately identify the position of

and its image

in the graph. Noticing that for convex mappings {{math|Y {{=}} φ(X)}} the corresponding distribution of {{mvar|Y}} values is increasingly "stretched out" for increasing values of {{mvar|X}}, it is easy to see that the distribution of {{mvar|Y}} is broader in the interval corresponding to {{math|X > X₀}} and narrower in {{math|X < X₀}} for any {{math|X₀}}; in particular, this is also true for

. Consequently, in this picture the expectation of {{mvar|Y}} will always shift upwards with respect to the position of

. A similar reasoning holds if the distribution of {{mvar|X}} covers a decreasing portion of the convex function, or both a decreasing and an increasing portion of it. This "proves" the inequality, i.e.

with equality when {{math|φ(X)}} is not strictly convex, e.g. when it is a straight line, or when {{mvar|X}} follows a degenerate distribution (i.e. is a constant).

Proof 1 (finite form)

If {{math|λ₁}} and {{math|λ₂}} are two arbitrary nonnegative real numbers such that {{math|λ₁ + λ₂ {{=}} 1}} then convexity of {{mvar|φ}} implies

This can be easily generalized: if {{math|λ₁, ..., λ_n}} are nonnegative real numbers such that {{math|λ₁ + ... + λ_n {{=}} 1}}, then

for any {{math|x₁, ..., x_n}}. This finite form of the Jensen's inequality can be proved by induction: by convexity hypotheses, the statement is true for n = 2. Suppose it is true also for some n, one needs to prove it for {{math|n + 1}}. At least one of the {{math|λ_i}} is strictly positive, say {{math|λ₁}}; therefore by convexity inequality:

one can apply the induction hypotheses to the last term in the previous formula to obtain the result, namely the finite form of the Jensen's inequality.

In order to obtain the general inequality from this finite form, one needs to use a density argument. The finite form can be rewritten as:

where μ_n is a measure given by an arbitrary convex combination of Dirac deltas:

Since convex functions are continuous, and since convex combinations of Dirac deltas are weakly dense in the set of probability measures (as could be easily verified), the general statement is obtained simply by a limiting procedure.

Proof 2 (measure-theoretic form)

Let g be a real-valued μ-integrable function on a probability space Ω, and let {{mvar|φ}} be a convex function on the real numbers. Since {{mvar|φ}} is convex, at each real number {{mvar|x}} we have a nonempty set of subderivatives, which may be thought of as lines touching the graph of {{mvar|φ}} at {{mvar|x}}, but which are at or below the graph of {{mvar|φ}} at all points (support lines of the graph).

because of the existence of subderivatives for convex functions, we may choose a and b such that

for all {{mvar|x}}. Since we have a probability measure, the integral is monotone with {{math|μ(Ω) {{=}} 1}} so that

Proof 3 (general inequality in a probabilistic setting)

Let X be an integrable random variable that takes values in a real topological vector space T. Since

is convex, for any

, the quantity

is decreasing as {{mvar|θ}} approaches 0⁺. In particular, the subdifferential of {{mvar|φ}} evaluated at {{mvar|x}} in the direction {{mvar|y}} is well-defined by

It is easily seen that the subdifferential is linear in {{mvar|y}} {{Citation needed|date=October 2013}} (that is false and the assertion requires Hahn-Banach theorem to be proved) and, since the infimum taken in the right-hand side of the previous formula is smaller than the value of the same term for {{math|θ {{=}} 1}}, one gets

In particular, for an arbitrary sub{mvar|σ}algebra

we can evaluate the last inequality when

to obtain

Now, if we take the expectation conditioned to

on both sides of the previous expression, we get the result since:

by the linearity of the subdifferential in the y variable, and the following well-known property of the conditional expectation:

Applications and special cases

Form involving a probability density function

Suppose {{math|Ω}} is a measurable subset of the real line and f(x) is a non-negative function such that

Then Jensen's inequality becomes the following statement about convex integrals:

If g is any real-valued measurable function and

is convex over the range of g, then

If g(x) = x, then this form of the inequality reduces to a commonly used special case:

Example: even moments of a random variable

In particular, if some even moment 2n of X is finite, X has a finite mean. An extension of this argument shows X has finite moments of every order

dividing n.

Alternative finite form

Let {{math|Ω {{=}} {x₁, ... x_n},}} and take {{mvar|μ}} to be the counting measure on {{math|Ω}}, then the general form reduces to a statement about sums:

Statistical physics

Jensen's inequality is of particular importance in statistical physics when the convex function is an exponential, giving:

where the expected values are with respect to some probability distribution in the random variable {{mvar|X}}.

The proof in this case is very simple (cf. Chandler, Sec. 5.5). The desired inequality follows directly, by writing

and then applying the inequality {{math|e^X ≥ 1 + X}} to the final exponential.

Information theory

If {{math|p(x)}} is the true probability density for {{mvar|X}}, and {{math|q(x)}} is another density, then applying Jensen's inequality for the random variable {{math|Y(X) {{=}} q(X)/p(X)}} and the convex function {{math|φ(y) {{=}} −log(y)}} gives

It shows that the average message length is minimised when codes are assigned on the basis of the true probabilities p rather than any other distribution q. The quantity that is non-negative is called the Kullback–Leibler divergence of q from p.

Since {{math|−log(x)}} is a strictly convex function for {{math|x > 0}}, it follows that equality holds when {{math|p(x)}} equals {{math|q(x)}} almost everywhere.

Rao–Blackwell theorem

If L is a convex function and

a sub-sigma-algebra, then, from the conditional version of Jensen's inequality, we get

So if δ(X) is some estimator of an unobserved parameter θ given a vector of observables X; and if T(X) is a sufficient statistic for θ; then an improved estimator, in the sense of having a smaller expected loss L, can be obtained by calculating

the expected value of δ with respect to θ, taken over all possible vectors of observations X compatible with the same value of T(X) as that observed.