Posterior consistency under (possible) misspecification

We assume, without too much loss of generality, that our priors are discrete. When dealing with Hellinger separable density spaces, it is possible to discretize posterior distributions to study consistency (see this post about it).

Let \Pi be a prior on a countable space \mathcal{N} = \{f_1, f_2, f_3, \dots\} of probability density functions, with \Pi(f) > 0 for all f \in \mathcal{N}. Data X_1, X_2, X_3, \dots follows (independently) some unknown distribution P_0 with density f_0.

We denote by D_{KL}(f_0, f) = \int f_0 \log\frac{f_0}{f} the Kullback-Leibler divergence and we let D_{\frac{1}{2}}(f_0, f) = 1 - \int \sqrt{f_0 f} be half of the squared Hellinger distance.

The following theorem states that the posterior distribution of \Pi accumulates in Hellinger neighborhoods of f_0, assuming the prior is root-summable (i.e. \sum_{f \in \mathcal{N}} \Pi(f)^\alpha < \infty for some \alpha > 0) . In the well-specified case (i.e. \inf_{f \in \mathcal{N}} D_{KL}(f_0, f) = 0), the posterior accumulates in any neighborhood of f_0. In the misspecified case, small neighborhoods of f_0 could be empty, but the posterior distribution still accumulates in sufficiently large neighborhoods (how large exactly is a function of \alpha and \inf_{f \in \mathcal{N}} D_{KL}(f_0, f)).Read More »

The choice of prior in bayesian nonparametrics – part 2

See part 1. Most proofs are omitted; I’ll post them with the complete pdf later this week.

The structure of \mathcal{M}

Recall that \mathbb{M} is is a Polish space (ie. a complete and separable metric space). It is endowed with its borel \sigma-algebra \mathfrak{B} which is the smallest family of subsets of \mathbb{M} that contains its topology and that is closed under countable unions and intersections. All subsets of \mathbb{M} we consider in the following are supposed to be part of \mathfrak{B}. A probability measure on \mathbb{M} is a function \mu : \mathfrak{B} \rightarrow [0,1] such that for any countable partition A_1, A_2, A_3, \dots of \mathbb{M} we have that \sum_{i=1}^\infty \mu(A_i) = 1. The set \mathcal{M} consists of all such probability measures.

Note that since \mathbb{M} is complete and separable, every probability measure \mu \in \mathcal{M} is regular (and tight). It means that the measure of any A\subset \mathbb{M} can be well approximated from the measure of compact subsets of A as well as from the measure of open super-sets of A:

\mu(A) = \sup \left\{\mu(K) \,|\, K \subset A \text{ is compact}\right\}\\ = \inf \left\{\mu(U) \,|\, U \supset A \text{ is open}\right\}.

Metrics on \mathcal{M}

Let me review some facts. A natural metric used to compare the mass allocation of two measures \mu, \nu \in \mathbb{M} is the total variation distance defined by

\|\mu - \nu\|_{TV} = \sup_{A \subset \mathbb{M}}|\mu(A) - \nu(A)|.Read More »

Remark on the asymptotics of the likelihood ratio and the K.-L. divergence

The problem.

Let f, g, h be three densities and suppose that, x_i \sim h, i \in \mathbb{N}, independently. What happens to the likelihood ratio

\prod_{i=1}^n \frac{f(x_i)}{g(x_i)}

as n\rightarrow \infty?

Clearly, it depends. If h = g \not = f, then

\prod_{i=1}^n \frac{f(x_i)}{g(x_i)} \rightarrow 0

almost surely at an exponential rate. More generally, if h is closer to g than to f, in some sense, we’d expect that \prod_{i=1}^n \frac{f(x_i)}{g(x_i)} \rightarrow 0. Such a measure of “closeness” of “divergence” between probability distributions is given by the Kullback-Leibler divergence

D_{KL}(f, g) = \int f \log\frac{f}{g}.

It can be verified that D_{KL}(f,g) \geq 0 with equality if and only if f=g, and that

D_{KL}(h,g) < D_{KL}(h,f) \Longrightarrow \prod_{i=1}^n \frac{f(x_i)}{g(x_i)} \rightarrow 0 \qquad (1)

almost surely at an exponential rate. Thus the K.L.-divergence can be used to solve our problem.

Better measures of divergence?

There are other measures of divergence that can determine the asymptotic behavior of the likelihood ratio as in (1) (e.g. the discrete distance). However, in this note, I give conditions under which the K.-L. divergence is, up to topological equivalence, the “best” measure of divergence.Read More »

The choice of prior in bayesian nonparametrics – Introduction

In preparation for the 11th Bayesian nonparametrics conference, I’m writing (and rewriting) notes on the background of our research (i.e. some of the general theory of bayesian nonparametrics). There are some good books on the subject (such as Bayesian Nonparametrics (Ghosh and Ramamoorthi, 2003)), but I wanted a more introductory focus and to present Choi and Ramamoorthi’s very clear point of view on posterior consistency (Remarks on the consistency of posterior distributions, 2008).

1. Introduction

Let \mathbb{X} be a complete and separable metric space and let \mathcal{M} be the space of all probability measures on \mathbb{X}. Some unknown distribution P_0\in \mathcal{M} is generating observable data \mathcal{D}_n = (X_1, X_2, \dots, X_n) \in \mathbb{X}^n, where each X_i is independently drawn from P_0. The problem is to learn about P_0 using only \mathcal{D}_n and prior knowledge.

Example (Discovery probabilities).
A cryptographer observes words, following some distribution P_0, in an unknown countable language \mathcal{L}. What are the P_0-probabilities of the words observed thus far? What is the probability that the next word to be observed has never been observed before?

1.1 Learning and uncertainty

We need an employable definition of learning. As a first approximation, we can consider learning to be the reduction of uncertainty about what is P_0. This requires a quantification of how uncertain we are to begin with. Then, hopefully, as data is gathered out uncertainty decreases and we are able to pinpoint P_0.

This is the core of Bayesian learning, alghough our definition is not yet entirely satisfactory. There are some difficulties with this idea of quantifying uncertainty, at least when using information-theoric concepts. The solution we adopt here is the use of probabilities to quantify uncertain knowledge (bayesians would also talk of subjective probabilities quantifying rational belief). For example, you may know that a coin flip is likely to be fair, although it is not impossible the two sides of the coin are both the same. This is uncertain knowledge about the distribution of heads and tails in the coin flips, and you could assign probabilities to the different possibilities.

More formally, prior uncertain knowledge about what is P_0 is quantified by a probability measure \Pi on \mathcal{M}. For any A \subset \mathcal{M}, \Pi(A) is the the prior probability that “P_0 \in A“. Then, given data \mathcal{D}_n, prior probabilities are adjusted to posterior probabilities: \Pi becomes \Pi_n, the conditional distribution of \Pi given \mathcal{D}_n. The celebrated Bayes’ theorem provides a formula to calculate \Pi_n from \Pi and \mathcal{D}_n. Thus we have an operational definition of learning in our statistical framework.

Learning is rationally adjusting uncertain knowledge in the light of new information.

For explanations as to why probabilities are well suited to the representation of uncertain knowledge, I refer the reader to Pearl (Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference, 1988). We will also see that the operation of updating the prior to posterior posterior probabilities does work as intended.

1.2 The choice of prior

Specifying prior probabilities, that is quantifying prior uncertain knowledge, is not a simple task. It is especially difficult when uncertainty is over the non-negligeable part \mathcal{M} of an infinite dimensional vector space. Fortunately, “probability is not about numbers, it is about the structure of reasoning”, as Glenn Shafer puts it (cited in Pearl, 1988, p.15). The exact numbers given to the events “P_0 \in A” are not of foremost importance; what matters is how probabilities are more qualitatively put together, and how this relates to the learning process.

Properties of prior distributions, opening them to scrutiny, criticism and discussion, must be identified and related to what happens as more and more data is gathered.

Part 2.

Sudden jumps and almost surely slow convergence

Estimating a mean can be an arduous task. Observe, for instance, the “sudden jumps” in trajectories of the Monte Carlo estimate \overline X_n of \mathbb{E}X, where \mathbb{E}X^2 = \infty.

sudden jumpsProposition.
Let X_1, X_2, X_3, \dots be a sequence of i.i.d. and integrable random variables, and let \overline X_k = \frac{1}{k}\sum_{i=1}^k X_i. For all M > 0 we have that

P\left( \left| \overline X_{n} - \overline X_{n-1}\right| > Mn^{-1/2} \;\; \text{ i.o.} \right)

is 1 or 0 according to \text{Var}(X_n) =\infty or \text{Var}(X_n) < \infty.

Corollary.
If \text{Var}(X_k) = \infty, then \overline X_n converges towards \mathbb{E}X_k at a rate almost surely slower than \frac{1}{\sqrt{n}} :

\left|\overline X_n - \mathbb{E}X_n \right| \not = \mathcal{O}\left(n^{-1/2}\right) \quad \text{a.s.}

PDF proof.

 

Many thanks to Xi’an who wondered about the occurence of these jumps.

Comment on The Sample Size Required in Importance Sampling

I summarize and comment part of The Sample Size Required in Importance Sampling (Chatterjee and Diaconis, 2015). One innovative idea is to bound the mean estimation error in terms of the tail behavior of d\mu/d\lambda, where \mu and \lambda are the importance sampling target and proposal distributions, respectively.

The problem is to evaluate

I = I(f) = \int f d\mu,

where \mu is a probability measure on a space \mathbb{M} and where f: \mathbb{M} \rightarrow \mathbb{R} is measurable. The Monte-Carlo estimate of I is

\frac{1}{n}\sum_{i=1}^n f(x_i), \qquad x_i \sim \mu.

When it is too difficult to sample \mu, for instance, other estimates can be obtained. Suppose that \mu is absolutely continuous with respect to another probability measure \lambda, and that the density of \mu with respect to \lambda is given by \rho. Another unbiaised estimate of I is then

I_n(f) = \frac{1}{n}\sum_{i=1}^n f(y_i)\rho(y_i),\qquad y_i \sim \lambda.

This is the general framework of importance sampling, with the Monte-Carlo estimate recovered by taking \lambda = \mu. An important question is the following.

How large should n be for I_n(f) to be close to I(f)?

An answer is given, under certain conditions, by Chatterjee and Diaconis (2015). Their main result can be interpreted as follows. If X \sim \mu and if \log \rho(X) is concentrated around its expected value L=\text{E}[\log \rho(X)], then a sample size of approximately n = e^{L} is both necessary and sufficient for I_n to be close to I. The exact sample size needed depends on \|f\|_{L^2(\mu)} and on the tail behavior of \log\rho(X). I state below their theorem with a small modification.

Theorem 1. (Chatterjee and Diaconis, 2015)
As above, let X \sim \mu. For any a > 0 and n \in \mathbb{N},

\mathbb{E} |I_n(f) - I(f)| \le \|f\|_{L^2(\mu)}\left( \sqrt{a/n} + 2\sqrt{\mathbb{P} (\rho(X) > a)} \right).

Conversely, for any \delta \in (0,1) and b > 0,

\mathbb{P}(1 - I_n(1) \le \delta) \le \frac{n}{b} + \frac{\mathbb{P}(\rho(X) \le b)}{1-\delta}.

Remark 1.
Suppose \|f\|_{L^2(\mu)} \le 1 and that \log\rho(X) is concentrated around L = \mathbb{E} \log\rho(X), meaning that for some t > 0 we have that \mathbb{P}(\log \rho(X) > L+t/2) and \mathbb{P}(\log\rho(X) < L-t/2) are both less than an arbitrary \varepsilon > 0. Then, taking n \geq e^{L+t} we find

\mathbb{E} |I_n(f) - I| \le e^{-t/4} + 2\varepsilon.

However, if n \leq e^{L-t} , we obtain

\mathbb{P}\left(1 - I_n(1) \le \tfrac{1}{2}\right) \le e^{-t/2} + 2 \varepsilon.

meaning that there can be a  high probability that I(1) and I_n(1) are not close.

Remark 2.
Let \lambda = \mu, so that \rho = 1. In that case, \log\rho(X) only takes its expected value 0. The theorem yields

\mathbb{E} |I_n(f) - I(f)| \le \frac{\|f\|_{L^2(\mu)}}{\sqrt{n}}

and no useful bound on \mathbb{P}(1-I_n(1) \le \delta).

Comment.
For the theorem to yield a sharp cutoff, it is necessary that L = \mathbb{E} \log\rho(X) be relatively large and that \log\rho(X) be highly concentrated around L. The first condition is not aimed at in the practice of importance sampling. This difficulty contrasts with the broad claim that “a sample of size approximately e^{L} is necessary and sufficient for accurate estimation by importance sampling”. The result in conceptually interesting, but I’m not convinced that a sharp cutoff is common.

Example

I consider their example 1.4. Here \lambda is the exponential distribution of mean 1, \mu is the exponential distribution of mean 2,\rho(x) = e^{x/2}/2 and f(x) = x. Thus I(f) = 2. We have L = \mathbb{E}\log\rho(X) = 1-\log(2) \approxeq 0.3, meaning that the theorem yields no useful cutoff. Furthermore, {}\mathbb{P}(\rho(X) > a) = \tfrac{1}{2a} and \|f\|_{L^2(\mu)} = 2. Optimizing the bound given by the theorem yields

\mathbb{E}|I_n(f)-2| \le \frac{4\sqrt{2}}{(2n)^{1/4}}.

The figure below shows 100 trajectories of I_k(f). The shaded area bounds the expected error.

fig1.png

This next figure shows 100 trajectories for the Monte-Carlo estimate of 2 = \int x d\mu, taking \lambda = \mu and \rho = 1. Here the theorem yields

\mathbb{E}|I_n(f)-2| \le \frac{2}{\sqrt{n}}.

fig2.png

References.

Chatterjee, S. and Diaconis, P. The Sample Size Required in Importance Sampling. https://arxiv.org/abs/1511.01437v2

The asymptotic behaviour of posterior distributions – 1

I describe key ideas of the theory of posterior distribution asymptotics and work out small examples.

1. Introduction

Consider the problem of learning about the probability distribution that a stochastic mechanism is following, based on independent observations. You may quantify your uncertainty about what distribution the mechanism is following through a prior P defined on the space of all possible distributions, and then obtain the conditional distribution of P given the observations to correspondingly adjust your uncertainty. In the simplest case, the prior is concentrated on a space \mathcal{M} of distributions dominated by a common measure \lambda. This means that each probability measure in \mu \in \mathcal{M} is such that there exists f with \mu(E) = \int_E f \,d\lambda for all measurable E. We may thus identify \mathcal{M} with a space \mathbb{F} of probability densities. Given independent observations X = (X_1, \dots, X_n), the conditional distribution of P given X is then

P(A \mid X) = \frac{\int_A f(X) P(df)}{\int_{\mathbb{F}} f(X) P(df)}, \quad A \subset \mathbb{F},

where f(X) = \Pi_{i=1}^nf(X_i) and A is measurable. This conditional distribution is called the posterior distribution of P, and is understood as a Lebesgue integral relative to the measure P defined on \mathbb{F} \simeq \mathcal{M}.

The procedure of conditionning P on X is bayesian learning. It is expected that as more and more data is gathered, that the posterior distribution will converge, in a suitable way, to a point mass located at the true distribution that the data is following. This is the asymptotic behavior we study.

1.1 Notations

The elements of measure theory and the language of mathematical statistics I use are standard and very nicely reviewed in Halmos (1949). To keep notations light, I avoid making explicit references to measure spaces. All sets and maps considered are measurable.

2. The relative entropy of probability measures

Let \mu_1 and \mu_2 be two probability measures on a metric space \mathbb{M}, and let \lambda be a common \sigma-finite dominating measure such as \mu_1 + \mu_2. That is, \mu_i(E) = 0 whenever E \subset \mathbb{M} is of \lambda-measure 0, and by the Radon-Nikodym theorem this is equivalent to the fact that there exists unique densities f_1, f_2 with d\mu_i = f_i d\lambda. The relative entropy of \mu_1 and \mu_2, also known as the Kullback-Leibler divergence, is defined as

K(\mu_1, \mu_2) = \int \log \frac{f_1}{f_2} d\mu_1\,.

The following inequality will be of much use.

Lemma 1 (Kullback and Leibler (1951)). We have K(\mu_1, \mu_2) \geq 0, with equality if and only if \mu_1 = \mu_2.

Proof: We may assume that \mu_2 is dominated by \mu_1, as otherwise K(\mu_1, \mu_2) = \infty and the lemma holds. Now, let \phi (t) = t\log t, g = f_1/f_2 and write {}K(\mu_1, \mu_2) = \int \phi \circ g d\mu_2. Since \phi (t) = (t-1) +(t-1)^2 / h(t) for some h(t) > 0, we find

K(\mu_1, \mu_2) = \int (g-1) d\mu_2 + \int \frac{(g-1)^2}{h} d\mu_2 = \int \frac{(g-1)^2}{h\circ g} d\mu_2 \geq 0,

with equality if and only if f_1/f_2 = 1, \mu_1-almost everywhere. QED.

We may already apply the lemma to problems of statistical inference.

2.1 Discriminating between point hypotheses

Alice observes X = (X_1, \dots, X_n), X_i \sim^{ind}\mu_*, \mu_* \in \{\mu_1, \mu_2\}, and considers the hypotheses H_i = \{\mu_i\}, i=1,2, representing “\mu_* = \mu_i“. A prior distribution on H_1 \cup H_2 takes the form

P = \alpha \delta_1 + (1-\alpha)\delta_2, \quad 0 < \alpha < 1,

where \delta_i(A) is 1 when \mu_i \in A and is 0 otherwise. Given a sample X_1 \sim \mu_1, the weight of evidence for H_1 versus H_2 (Good, 1985), also known as “the information in X_1 for discrimination between H_1 and H_2” (Kullback, 1951), is defined as

W(X_1) = \log\frac{f_1(X_1)}{f_2(X_1)}.

This quantity is additive for independent sample: if X = (X_1, \dots, X_n), X_i \sim^{ind} \mu_*, then

W(X) = \sum_{i=1}^n W(X_i);

and the posterior log-odds are given by
\log\frac{P(H_1 \mid X)}{P(H_2 \mid X)} = W(X) + \log\frac{P(H_1)}{P(H_2)}.
Thus K(\mu_1, \mu_2) is the expected weight of evidence for H_1 against H_2 brought by an unique sample X_1 \sim \mu_1 and is strictly positive whenever \mu_1 \not = \mu_2. By the additivity of W, Alice should expect that as more and more data is gathered, the weight of evidence grows to infinity. In fact, this happens \mu_1-almost surely.

Proposition 2. Almost surely as the number of observations grows, W(X) \rightarrow \infty and P(H_1 | X)\rightarrow 1.

Proof: By the law of large numbers (the case K(\mu_1, \mu_2) = \infty is easily treated separatly), we have

\frac{1}{n}\sum_{i=1}^n \log\frac{f_1(X_i)}{f_2(X_i)}\rightarrow K(\mu_1, \mu_2) > 0, \quad \mu_1\text{-a.s.}.

Therefore

W(X) = n \left( \frac{1}{n}\sum_{i=1}^n \log\frac{f_1(X_i)}{f_2(X_i)} \right) \rightarrow \infty \quad \mu_1\text{-a.s.}

and P(H_1\mid X) \rightarrow 1 \mu_1-almost surely. QED.

2.2 Finite mixture prior under misspecification

Alice wants to learn about a fixed unknown distribution \mu_* through data X =(X_1, \dots, X_n), where X_i \sim^{ind.} \mu_*. She models what \mu_* may be as one of the distribution in \mathcal{M} = \{\mu_1, \dots, \mu_k\}, and quantifies her uncertainty through a prior P on \mathcal{M}. We may assume that K(\mu_*, \mu_i) < \infty for some i, as otherwise she will eventually observe data that is impossible under \mathcal{M} and adjust her model. (Indeed, K(\mu_*, \mu_i)=\infty implies that there exists a \mu_* non-negligible set E_i such that \mu_i(E_i) = 0 < \mu_*(E_i). Alice will \mu_*-almost surely observe X_m \in E_i and conclude that \mu_* \not = \mu_k.) If \mu_* \not\in \mathcal{M}, she may not realise that her model is wrong, but the following proposition ensures that the posterior distribution will concentrate on the \mu_i‘s that are closest to \mu_*.

Proposition 3. Let A = \{\mu_i \in \mathcal{M} \mid K(\mu_*, \mu_i) = \min_j K(\mu_*, \mu_j)\}. Almost surely as the number of observations grows, {} P(A | X) \rightarrow 1.

Proof: The prior takes the form {}P = \sum_{i=1}^{k} \alpha_i \delta_i, {}\alpha_i > 0, where {}\delta_i is a point mass at {}\mu_i. The model {}\mathcal{M} is dominated by a {}\sigma-finite measure {}\lambda, such as {}\sum \mu_i, so that the posterior distribution is

P(A\mid X) = \frac{\sum_{\mu_i \in A} f_i(X) \alpha_i }{\sum_{\mu_i \in \mathcal{M}} f_i(X) \alpha_i}, \quad d\mu_i = f_i d\lambda.

Because {}K(\mu_\star, \mu_i) < \infty for some {}i, {}\mu_\star is also absolutely continuous with respect to {}\lambda and we let {}d\mu_\star = f_\star d\lambda. Now, let {}\varepsilon = \min_j K(\mu_\star, \mu_j) and {}\delta = \min_j \{K(\mu_\star, \mu_j \mid \mu_j \in A^c)\} > \varepsilon. Write

\log \frac{P(A | X)}{P(A^c | X)} \geq \log \frac{f_A(X)}{f_{A^c}(X)} + \log \frac{\sum_{\mu_i \in A} \alpha_i}{ \sum_{\mu_i \in A^c} \alpha_i},

where {}f_A = \arg \min_{f_i : \mu_i \in A} f_i(X) and {}f_{A^c} = \arg \max_{f_i : \mu_i \in A^c} f_i(X) both depend on {}n. Using the law of large numbers, we find

\liminf_{n \rightarrow \infty} \frac{1}{n} \sum_{i=1}^n \log\frac{f_\star(X_i)}{f_{A^c}(X_i)} \geq \delta > \varepsilon = \lim_{n \rightarrow \infty} \frac{1}{n}\sum_{i=1}^n \log\frac{f_\star(X_i)}{f_A (X_i)}\quad \mu_\star\text{-a.s.},

and therefore

{}\log \frac{f_A(X)}{f_{A^c}(X)} = n\left( \frac{1}{n} \sum_{i=1}^n \log\frac{f_\star(X_i)}{f_{A^c}(X_i)} - \frac{1}{n}\sum_{i=1}^n \log\frac{f_\star(X_i)}{f_A (X_i)} \right) \rightarrow \infty\quad \mu_*\text{-a.s.}.

This implies {}P(A \mid X) \rightarrow 1, {}\mu_\star-almost surely. QED.

2.3 Properties of the relative entropy

The following properties justify the interpretation of the relative entropy K(\mu_1, \mu_2) as an expected weight of evidence.

Lemma 4. The relative entropy {}K(\mu_1, \mu_2) = \int f_1 \log\frac{f_1}{f_2} d\lambda, {}d\mu_i = f_id\lambda, does not depend on the choice of dominating measure {}\lambda.

Let {}T: \mathbb{M} \rightarrow \mathbb{M}' be a mapping onto a space {}\mathbb{M}'. If {}X \sim \mu, then {}T(X) \sim \mu T^{-1}, where {}\mu T^{-1} is the probability measure on {}\mathbb{M}' defined by {}\mu T^{-1}(F) = \mu (T^{-1}(F)). The following proposition, a particular case of (Kullback, 1951, theorem 4.1), states that transforming the data through {}T cannot increase the expected weight of evidence.

Proposition 5 (see Kullback and Leibler (1951)). For any two probability measures {}\mu_1, \mu_2 on {}\mathbb{M}, we have

{}K(\mu_1 T^{-1}, \mu_2 T^{-1}) \le K(\mu_1, \mu_2).

Proof:  Let {}\lambda be a dominating measure for {}\{\mu_1, \mu_2\}, {}d\mu_i = f_i d\lambda. Therefore, {}\mu_i T^{-1} is dominated by {}\lambda T^{-1}, and we may write {}d\mu_i T^{-1} = g_i d\lambda T^{-1}, {}i=1,2. The measures on the two spaces are related by the formula

{} \int_{T^{-1}(F)} h \circ T \, d\alpha = \int_{F} h \,d\alpha T^{-1},

where {}\alpha may be one of {}\mu_1, {}\mu_2 and {}\lambda and {}h is any measurable function (Halmos, 1949, lemma 3). Therefore,

{}K(\mu_1, \mu_2) - K(\mu_1 T^{-1}, \mu_2 T^{-1}) = \int_{\mathbb{M}} \log\frac{f_1}{f_2} d\mu_1 - \int_{\mathbb{M}'} \log \frac{g_1}{g_2} d\mu_1T^{-1} \\ = \int \log\frac{f_1}{f_2} d\mu_1 - \int \log\frac{g_1 \circ T}{g_2 \circ T} d\mu_1 = (\star).

By letting {}h = \frac{f_1 g_2 \circ T}{f_2 g_1 \circ T} we find

{} (\star) = \int h \log h \, \frac{g_1\circ T}{g_2 \circ T} d\mu_2 = \int h \log h \,d\gamma,

where {}\gamma is such that {}d\gamma = \frac{g_1\circ T}{g_2 \circ T} d\mu_2. It is a probability measure since {}\int \frac{g_1\circ T}{g_2 \circ T} d\mu_2 = 1.

By convexity of {}\phi : t \mapsto t \log t, we find

{} \int h \log h \,d\gamma \geq \phi\left( \int h \,d\gamma \right) = \phi(1) = 0

and this finishes the proof.

References.

Good, I. (1985). Weight of evidence: A brief survey. In A. S. D. L. J.M. Bernardo, M.H. DeGroot (Ed.), Bayesian Statistics 2, pp. 249–270. North-Holland B.V.: Elsevier Science Publishers.

Halmos, P. R. and L. J. Savage (1949, 06). Application of the radon-nikodym theorem to the theory of sufficient statistics. Ann. Math. Statist. 20 (2), 225–241.

Kullback, S. and Leibler, R. A. (1951). On information and sufficiency. Ann. Math. Statist. 22(1), 79-86