37.04.04 · probability / 04-conditional-expectation-martingales

Kakutani's Theorem on Product Martingales and Absolute Continuity of Product Measures

shipped3 tiersLean: none

Anchor (Master): Williams — Probability with Martingales §14.12-14.18; Durrett §4.3.3; Williams — Probability with Martingales §14.18 (Kakutani's theorem, statement and proof); Shiryaev — Probability (Springer, 2nd ed., 1996) Ch. VII §6

Intuition Beginner

Imagine multiplying together a long string of random factors, each one a fair multiplier: on average each factor equals $1$ , but any single factor might come out a bit above or a bit below $1$ . Your running product is like a fortune in a fair game played by multiplication instead of addition — given everything so far, your best forecast for tomorrow's product is exactly today's. The question is what happens after infinitely many factors: does the product settle to a sensible positive number, or does it collapse to zero?

The surprising answer is that there is no middle ground. Either the product converges to a genuinely positive limit and behaves like a faithful final settlement, or it slides all the way to zero, with no possibility of hovering at some random positive value in an unreliable way. Which of the two happens is decided by a single number you can compute in advance: take the average of the square root of each factor, and multiply all those averages together. If that grand product stays positive, the fortune survives; if it shrinks to zero, the fortune dies.

Why the square root? Because the square root gently penalises spread. A factor that is sometimes much bigger and sometimes much smaller than $1$ has an average of $1$ , but the average of its square root is strictly less than $1$ — the square root sees the risk that the additive average hides. Each risky factor chips a little off the square-root product, and if the chips add up to everything, the fortune is doomed.

The takeaway: a product of independent fair factors lives or dies by a clean all-or-nothing rule, and the verdict is read off a single square-root bookkeeping number. This same rule decides when two ways of assigning probabilities to an endless sequence of experiments are close enough to be comparable, or so far apart that each calls impossible what the other calls certain.

Visual Beginner

Picture three running products plotted across many multiplication steps. The horizontal axis is the number of factors multiplied so far; the vertical axis is the running product, starting at $1$ . One path keeps fluctuating around a comfortable positive band and settles to a positive number. A second path drifts steadily downward, hugging zero. The square-root bookkeeping number — the running product of the average square roots — is drawn underneath each: it stays positive for the surviving path and decays to zero for the dying one.

The picture makes the dichotomy visible: the running product either finds a positive home or falls to zero, and the companion square-root number predicts which, computed from the factors alone before any path is drawn.

Worked example Beginner

We test the rule on a concrete product. Let each factor $X_{k}$ take the value $1 + a$ or $1 - a$ , each with probability one half, for a fixed number $a$ with $0 < a < 1$ . The factors are independent. The running product is $M_{n} = X_{1} X_{2} \dots X_{n}$ .

Step 1. Check the fair-game property. The average of each factor is $\frac{1}{2} (1 + a) + \frac{1}{2} (1 - a) = 1$ . So the average of the product is $1 \times 1 \times \dots \times 1 = 1$ for every $n$ : a fair multiplicative game.

Step 2. Compute the square-root bookkeeping number for one factor. We need the average of $X_{k}$ : $$ \mathbb{E}\big[\sqrt{X_k}\big] = \tfrac12\sqrt{1 + a} + \tfrac12\sqrt{1 - a}. $$ Take $a = 0.6$ . Then $1.6 \approx 1.2649$ and $0.4 \approx 0.6325$ , so the average square root is $\frac{1}{2} (1.2649 + 0.6325) = \frac{1}{2} (1.8974) \approx 0.9487$ .

Step 3. Multiply over all factors. The square-root product after $n$ factors is $(0.9487)^{n}$ . Since $0.9487 < 1$ , raising it to higher and higher powers drives it to zero: $(0.9487)^{10} \approx 0.59$ , $(0.9487)^{100} \approx 0.0055$ , and it keeps shrinking.

Step 4. Read the verdict. The square-root product collapses to zero, so the rule predicts the running product $M_{n}$ dies: $M_{n} \to 0$ . Each factor is genuinely risky (it really does swing by $0.6$ ), each swing chips the square-root product below $1$ , and infinitely many chips erase it.

What this tells us: even though every running product has average exactly $1$ , the typical product collapses to zero. The average is propped up by rare paths where the product stays large; the square-root number sees past that illusion and reports the fate of the typical path. Only if the factors were close enough to $1$ that the square-root product stayed positive would the fortune survive.

Check your understanding Beginner

Exercise (easy, multiple choice).

A product of independent fair factors (each with average $1$ ) is governed by Kakutani's rule. The rule says the running product converges to a strictly positive limit precisely when:

A. Each factor is greater than $1$ B. The product of the averages of the square roots of the factors stays positive C. The number of factors is finite D. The average of the product equals $1$

Hint

The square root is the quantity that detects hidden risk; multiply the average square roots together.

Answer

B. The fortune survives exactly when the product of the average square roots stays positive. Option A is impossible for fair factors averaging $1$ (some must dip below $1$ ). Option C confuses the infinite limit with a finite truncation. Option D is automatic for every fair product and so cannot be the deciding condition — every such product has average $1$ at all times, yet many collapse to zero.

Formal definition Intermediate+

Fix a probability space $(Ω, F, P)$ . Martingales, the a.s. convergence theorem for $L^{1}$ -bounded martingales 37.04.02, the upcrossing inequality and optional-stopping accounting 37.04.01, and the uniform-integrability / $L^{1}$ -closure theory 37.04.03 are taken as established. Conditional expectation is built from the Radon-Nikodym theorem 02.07.08, and product measures from Fubini-Tonelli 02.07.07.

Definition (product martingale). Let $(X_{k})_{k \geq 1}$ be independent, non-negative, integrable random variables with $E [X_{k}] = 1$ for every $k$ . Set $F_{n} = σ (X_{1}, \dots, X_{n})$ and $F_{0} = {\emptyset, Ω}$ . The product martingale is $$ M_0 = 1, \qquad M_n = \prod_{k=1}^n X_k . $$ That $(M_{n})$ is a martingale follows from independence: $E [M_{n + 1} ∣ F_{n}] = M_{n} E [X_{n + 1} ∣ F_{n}] = M_{n} E [X_{n + 1}] = M_{n}$ , using that $X_{n + 1}$ is independent of $F_{n}$ . Each $M_{n} \geq 0$ and $E [M_{n}] = 1$ , so $(M_{n})$ is a non-negative $L^{1}$ -bounded martingale and by 37.04.02 converges a.s. to a limit $M_{\infty} \in [0, \infty)$ with $E [M_{\infty}] \leq 1$ (Fatou).

Definition (Hellinger coefficient). For each $k$ set $$ a_k = \mathbb{E}\big[\sqrt{X_k}\big] \in (0, 1]. $$ By Jensen's inequality applied to the strictly concave map $x \mapsto x$ , $a_{k} = E [X_{k}] \leq E [X_{k}] = 1$ , with equality iff $X_{k} = 1$ a.s. The Hellinger product is $\prod_{k \geq 1} a_{k}$ , a decreasing product in $(0, 1]$ that either converges to a strictly positive limit or to $0$ .

Definition (Hellinger integral of two laws). Let $p, q$ be probability densities with respect to a common $σ$ -finite measure $μ$ (for instance $μ = p + q$ ). The Hellinger integral (affinity) is $$ H(p, q) = \int \sqrt{p, q}; d\mu \in [0, 1], $$ independent of the dominating $μ$ . The associated Hellinger distance is $h (p, q) = (\frac{1}{2} \int (p - q)^{2} d μ)^{1/2}$ , so that $h^{2} = 1 - H$ . When $X_{k} = q_{k} / p_{k}$ is a likelihood ratio under $P$ with $q_{k}, p_{k}$ the $k$ -th coordinate densities, $a_{k} = E_{P} [X_{k}] = \int q_{k} / p_{k} p_{k} d μ = H (p_{k}, q_{k})$ .

Counterexamples to common slips Intermediate+

Mean one does not protect positivity of the limit. The product $M_{n} = \prod_{k \leq n} (1 + ξ_{k})$ with $ξ_{k} = \pm 1$ equally likely satisfies $E [M_{n}] = 1$ , yet $M_{n} \to 0$ a.s. once any factor hits $0$ . The constant mean is propped up by an event of vanishing probability.
Pointwise convergence $a_{k} \to 1$ is not enough. Even with each $a_{k}$ close to $1$ , the verdict is the convergence of the infinite product $\prod a_{k}$ , equivalently the convergence of $\sum_{k} (1 - a_{k})$ . If $\sum (1 - a_{k}) = \infty$ the product collapses; the rate, not the limit of the terms, decides survival.
$L^{1}$ -boundedness is automatic and says nothing. Every product martingale here is $L^{1}$ -bounded ( $E ∣ M_{n} ∣ = 1$ ), so the a.s. limit always exists. The genuine content is whether the convergence is in $L^{1}$ (limit faithful, strictly positive) or merely a.s. (limit zero); the dichotomy is exactly the UI/non-UI split of 37.04.03.
Absolute continuity is all-or-nothing for products, unlike for single measures. Two laws on one coordinate can be neither equivalent nor singular (e.g. one supported on a sub-interval of the other). For infinite products of independent coordinates the Kakutani dichotomy forces $Q ≪ P$ or $Q ⊥ P$ with nothing in between, because the relevant event ${d Q / d P > 0}$ is a tail event.

Key theorem with proof Intermediate+

Theorem (Kakutani's dichotomy for product martingales). Let $(X_{k})_{k \geq 1}$ be independent, non-negative, with $E [X_{k}] = 1$ , and let $M_{n} = \prod_{k \leq n} X_{k}$ with a.s. limit $M_{\infty}$ . Set $a_{k} = E [X_{k}] \in (0, 1]$ . Then exactly one of the following holds.

(i) If $\prod_{k \geq 1} a_{k} > 0$ , then $(M_{n})$ is uniformly integrable, $M_{n} \to M_{\infty}$ in $L^{1}$ , $E [M_{\infty}] = 1$ , and $M_{\infty} > 0$ a.s.

(ii) If $\prod_{k \geq 1} a_{k} = 0$ , then $M_{\infty} = 0$ a.s.

Proof. Define the auxiliary process $$ N_n = \prod_{k=1}^n \frac{\sqrt{X_k}}{a_k}. $$ Each factor $X_{k} / a_{k}$ is non-negative with mean $E [X_{k}] / a_{k} = 1$ , and the factors are independent, so by the same computation as for $M_{n}$ , $(N_{n})$ is a non-negative martingale with $E [N_{n}] = 1$ . In particular $(N_{n})$ is $L^{1}$ -bounded and converges a.s. to some $N_{\infty} \in [0, \infty)$ .

Case (i): $A := \prod_{k \geq 1} a_{k} > 0$ . We show $(N_{n})$ is $L^{2}$ -bounded, hence UI by 37.04.03. Compute, using independence and $E [(X_{k} / a_{k})^{2}] = E [X_{k}] / a_{k}^{2} = 1/ a_{k}^{2}$ , $$ \mathbb{E}[N_n^2] = \prod_{k=1}^n \mathbb{E}\Big[\frac{X_k}{a_k^2}\Big] = \prod_{k=1}^n \frac{1}{a_k^2} = \Big(\prod_{k=1}^n a_k\Big)^{-2} \le A^{-2} < \infty, $$ since $\prod_{k \leq n} a_{k} \geq A$ as the $a_{k} \leq 1$ make the partial products decrease to $A$ . Thus $sup_{n} E [N_{n}^{2}] \leq A^{- 2} < \infty$ : $(N_{n})$ is $L^{2}$ -bounded. By Doob's $L^{p}$ theory 37.04.03 $(N_{n})$ converges in $L^{2}$ (hence in $L^{1}$ ) to a limit $N_{\infty}$ with $E [N_{\infty}^{2}] \leq A^{- 2}$ . The pointwise identity $$ M_n = \prod_{k=1}^n X_k = \Big(\prod_{k=1}^n a_k\Big)^2 N_n^2 \le N_n^2 $$ holds because $X_{k} = a_{k}^{2} (X_{k} / a_{k})^{2}$ and $\prod_{k \leq n} a_{k} \leq 1$ . The $L^{2}$ convergence $N_{n} \to N_{\infty}$ gives $N_{n}^{2} \to N_{\infty}^{2}$ in $L^{1}$ , and an $L^{1}$ -convergent sequence is uniformly integrable by Vitali's theorem 37.04.03; hence ${N_{n}^{2}}$ is UI. Domination by a UI family preserves UI, so $0 \leq M_{n} \leq N_{n}^{2}$ makes ${M_{n}}$ uniformly integrable. Therefore $M_{n} \to M_{\infty}$ in $L^{1}$ by 37.04.03, and $E [M_{\infty}] = lim_{n} E [M_{n}] = 1$ .

It remains to show $M_{\infty} > 0$ a.s. Since $N_{n} \to N_{\infty}$ in $L^{2}$ and $E [N_{n}] = 1$ , we have $E [N_{\infty}] = 1 > 0$ , so $N_{\infty}$ is not a.s. zero. The event ${N_{\infty} = 0}$ is a tail event of the independent sequence $(X_{k} / a_{k})$ : indeed ${N_{\infty} = 0} = {\sum_{k} lo g^{-} (X_{k} / a_{k}) = \infty}$ up to the behaviour of finitely many positive factors, hence measurable with respect to $σ (X_{m}, X_{m + 1}, \dots)$ for every $m$ . By the Kolmogorov 0-1 law 37.04.03 it has probability $0$ or $1$ ; as $E [N_{\infty}] = 1$ rules out probability $1$ , we get $P (N_{\infty} = 0) = 0$ , i.e. $N_{\infty} > 0$ a.s. Finally $M_{n} = (\prod_{k \leq n} a_{k})^{2} N_{n}^{2} \to A^{2} N_{\infty}^{2}$ a.s. (the partial products converge to $A > 0$ ), and matching this a.s. limit with $M_{\infty}$ gives $M_{\infty} = A^{2} N_{\infty}^{2} > 0$ a.s.

Case (ii): $\prod_{k \geq 1} a_{k} = 0$ . Then $\prod_{k \leq n} a_{k} \to 0$ . The process $(N_{n})$ is still a non-negative martingale with $E [N_{n}] = 1$ , so it converges a.s. to a finite $N_{\infty} < \infty$ a.s. (a.s. finiteness of the limit of a non-negative $L^{1}$ -bounded martingale, 37.04.02). Therefore $$ M_n = \Big(\prod_{k=1}^n a_k\Big)^2 N_n^2 \xrightarrow[n\to\infty]{} 0 \cdot N_\infty^2 = 0 \quad \text{a.s.}, $$ because $\prod_{k \leq n} a_{k} \to 0$ while $N_{n} \to N_{\infty} < \infty$ a.s. Hence $M_{\infty} = 0$ a.s. (Here the convergence is not in $L^{1}$ , since $E [M_{n}] = 1 \neq = 0$ ; the martingale fails UI exactly because the Hellinger product collapses.) $□$

Bridge. Kakutani's dichotomy builds toward the absolute-continuity-versus-singularity classification of infinite product measures and appears again in the theory of statistical contiguity and the Neyman-Pearson likelihood-ratio test, where $M_{n}$ is the likelihood ratio of two hypotheses after $n$ observations. The foundational reason the square root governs everything is that $a_{k} = E [X_{k}]$ is the Hellinger affinity $H (p_{k}, q_{k})$ , and $M_{n} = (\prod a_{k}) N_{n}$ is, up to the deterministic Hellinger factor, an $L^{2}$ -bounded martingale — this is exactly the move that converts an $L^{1}$ question about $M_{n}$ into an $L^{2}$ question about $M_{n} / \prod a_{k}$ , where the Hilbert-space geometry of 37.04.03 applies. The dichotomy is dual to the all-or-nothing tail behaviour of independent sequences: ${M_{\infty} > 0}$ is a tail event, so the Kolmogorov 0-1 law forces a clean split, and putting these together, the central insight is that uniform integrability of the product martingale, positivity of its limit, and the survival of the Hellinger product are one condition viewed through the $L^{2}$ lift, the tail 0-1 law, and the closure theorem respectively.

Exercises Intermediate+

Exercise 3 (medium, symbolic).

Prove that the survival condition $\prod_{k \geq 1} a_{k} > 0$ is equivalent to $\sum_{k \geq 1} (1 - a_{k}) < \infty$ , given $0 < a_{k} \leq 1$ .

Hint

Take logarithms and compare $- lo g a_{k}$ with $1 - a_{k}$ as $a_{k} \to 1$ , using $- lo g (1 - x) \geq x$ and a two-sided bound for $x$ small.

Answer

Write $a_{k} = 1 - b_{k}$ with $b_{k} = 1 - a_{k} \in [0, 1)$ . The product $\prod a_{k}$ is positive iff $\sum_{k} (- lo g a_{k}) < \infty$ . For $b_{k} \in [0, 1)$ , $- lo g a_{k} = - lo g (1 - b_{k}) \geq b_{k}$ , so $\sum b_{k} = \infty \Rightarrow \sum (- lo g a_{k}) = \infty \Rightarrow \prod a_{k} = 0$ . Conversely if $\sum b_{k} < \infty$ then $b_{k} \to 0$ , and for $b_{k} \leq 1/2$ one has $- lo g (1 - b_{k}) \leq b_{k} + b_{k}^{2} \leq 2 b_{k}$ , so $\sum_{k \geq K} (- lo g a_{k}) \leq 2 \sum_{k \geq K} b_{k} < \infty$ (only finitely many $b_{k} > 1/2$ since $b_{k} \to 0$ , and each contributes a finite term), giving $\prod a_{k} > 0$ . Rubric: full credit for the logarithm reduction, the lower bound $- lo g (1 - x) \geq x$ , and the matching upper bound for small $x$ . The condition $\sum (1 - a_{k}) < \infty$ is the practical form of the dichotomy.

Exercise 4 (medium, symbolic).

Verify that $(N_{n})$ with $N_{n} = \prod_{k \leq n} X_{k} / a_{k}$ is a martingale, and compute $E [N_{n}^{2}]$ in terms of the $a_{k}$ .

Hint

Use independence for the martingale property; for the second moment use $E [(X_{k} / a_{k})^{2}] = E [X_{k}] / a_{k}^{2}$ .

Answer

Each factor $Y_{k} = X_{k} / a_{k} \geq 0$ has $E [Y_{k}] = a_{k} / a_{k} = 1$ and is independent of $F_{n} = σ (X_{1}, \dots, X_{n})$ for $k > n$ . Hence $E [N_{n + 1} ∣ F_{n}] = N_{n} E [Y_{n + 1} ∣ F_{n}] = N_{n} E [Y_{n + 1}] = N_{n}$ , so $(N_{n})$ is a martingale with $E [N_{n}] = 1$ . For the second moment, by independence $E [N_{n}^{2}] = \prod_{k = 1}^{n} E [Y_{k}^{2}] = \prod_{k = 1}^{n} E [X_{k}] / a_{k}^{2} = \prod_{k = 1}^{n} a_{k}^{- 2} = (\prod_{k \leq n} a_{k})^{- 2}$ . Rubric: full credit for the unit conditional mean, the independence factorisation, and the exact $(\prod a_{k})^{- 2}$ formula. This is the bound that makes $(N_{n})$ $L^{2}$ -bounded precisely when the Hellinger product is positive.

Exercise 5 (medium, numeric).

Let $p_{k} = Bernoulli (1/2)$ and $q_{k} = Bernoulli (θ_{k})$ on ${0, 1}$ . Compute the Hellinger integral $H (p_{k}, q_{k})$ and find the condition on $(θ_{k})$ for the infinite product law $Q = ⨂ q_{k}$ to be absolutely continuous with respect to $P = ⨂ p_{k}$ .

Hint

$H (p_{k}, q_{k}) = p_{k} (0) q_{k} (0) + p_{k} (1) q_{k} (1)$ ; then apply the dichotomy via $\sum (1 - H) < \infty$ .

Answer

With $p_{k} (1) = p_{k} (0) = 1/2$ and $q_{k} (1) = θ_{k}$ , $q_{k} (0) = 1 - θ_{k}$ , $$ H(p_k, q_k) = \sqrt{\tfrac12(1-\theta_k)} + \sqrt{\tfrac12 \theta_k} = \tfrac{1}{\sqrt2}\big(\sqrt{\theta_k} + \sqrt{1-\theta_k}\big). $$ By Kakutani, $Q ≪ P$ iff $\prod_{k} H (p_{k}, q_{k}) > 0$ iff $\sum_{k} (1 - H (p_{k}, q_{k})) < \infty$ . Writing $θ_{k} = \frac{1}{2} + ε_{k}$ , a Taylor expansion gives $1 - H (p_{k}, q_{k}) ≍ ε_{k}^{2}$ for small $ε_{k}$ , so the precise condition is $\sum_{k} ε_{k}^{2} < \infty$ , i.e. $\sum_{k} (θ_{k} - \frac{1}{2})^{2} < \infty$ . If this sum diverges, $Q ⊥ P$ . Rubric: full credit for the Hellinger integral, the dichotomy criterion, and the $\sum ε_{k}^{2}$ reduction. This is the discrete analogue of the Gaussian-shift contiguity threshold.

Exercise 6 (medium, symbolic).

Show that in case (i) of Kakutani's theorem the limit $M_{\infty}$ equals $A^{2} N_{\infty}^{2}$ where $A = \prod_{k} a_{k}$ , and deduce $E [M_{\infty}] = 1$ directly from $E [N_{\infty}^{2}]$ .

Hint

Combine the a.s. limits of $\prod_{k \leq n} a_{k}$ and $N_{n}$ ; for the mean, $L^{2}$ convergence of $N_{n}$ gives $E [N_{\infty}^{2}] = lim E [N_{n}^{2}]$ .

Answer

Pointwise $M_{n} = (\prod_{k \leq n} a_{k})^{2} N_{n}^{2}$ . The deterministic partial products converge to $A > 0$ and $N_{n} \to N_{\infty}$ a.s., so $M_{n} \to A^{2} N_{\infty}^{2}$ a.s.; since $M_{n} \to M_{\infty}$ a.s. as well, $M_{\infty} = A^{2} N_{\infty}^{2}$ . For the mean, $N_{n} \to N_{\infty}$ in $L^{2}$ implies $E [N_{n}^{2}] \to E [N_{\infty}^{2}]$ , and $E [N_{n}^{2}] = (\prod_{k \leq n} a_{k})^{- 2} \to A^{- 2}$ , so $E [N_{\infty}^{2}] = A^{- 2}$ . Therefore $E [M_{\infty}] = A^{2} E [N_{\infty}^{2}] = A^{2} \cdot A^{- 2} = 1$ , confirming the mean is preserved. Rubric: full credit for the a.s. limit identification, the $L^{2}$ convergence of the second moments, and the cancellation to $1$ . The factor $A^{2}$ exactly undoes the $A^{- 2}$ in the second moment of $N_{\infty}$ .

Exercise 7 (hard, symbolic).

Let $P = ⨂_{k} p_{k}$ and $Q = ⨂_{k} q_{k}$ be infinite product measures on $\prod_{k} E_{k}$ with $q_{k} ≪ p_{k}$ and likelihood ratios $f_{k} = d q_{k} / d p_{k}$ . Set $M_{n} = \prod_{k \leq n} f_{k} (ω_{k})$ . Prove that under $P$ , $(M_{n})$ is the product martingale of Kakutani's theorem and that $\prod_{k} \int f_{k} d p_{k} = \prod_{k} H (p_{k}, q_{k})$ . Conclude the absolute-continuity dichotomy.

Hint

Under $P$ the coordinates are independent with $E_{P} [f_{k}] = \int f_{k} d p_{k} = \int d q_{k} = 1$ . Identify $a_{k} = E_{P} [f_{k}]$ with the Hellinger integral, then apply Kakutani.

Answer

Under $P$ the coordinate projections $ω_{k}$ are independent with law $p_{k}$ , so the $f_{k} (ω_{k})$ are independent, non-negative, with $E_{P} [f_{k}] = \int f_{k} d p_{k} = \int d q_{k} = 1$ . Hence $(M_{n})$ is exactly a Kakutani product martingale under $P$ , with $a_{k} = E_{P} [f_{k}] = \int f_{k} d p_{k} = \int (d q_{k} / d p_{k}) d p_{k} = \int q_{k} p_{k} d μ_{k} = H (p_{k}, q_{k})$ , where $μ_{k}$ dominates both. Now $M_{n} = d Q_{n} / d P_{n}$ where $Q_{n}, P_{n}$ are the laws of the first $n$ coordinates. If $\prod_{k} a_{k} > 0$ : case (i) gives $M_{n} \to M_{\infty}$ in $L^{1} (P)$ with $E_{P} [M_{\infty}] = 1$ and $M_{\infty} > 0$ a.s.; the $L^{1}$ convergence of the densities $d Q_{n} / d P_{n}$ identifies $M_{\infty} = d Q / d P$ , so $Q ≪ P$ with the stated derivative. If $\prod_{k} a_{k} = 0$ : case (ii) gives $M_{\infty} = 0$ $P$ -a.s., so $d Q / d P = 0$ $P$ -a.s.; by symmetry (or by the tail-event argument) $Q$ concentrates on ${M_{\infty} = \infty}$ , a $P$ -null set, hence $Q ⊥ P$ . Rubric: full credit for the independence-under- $P$ setup, the Hellinger identification of $a_{k}$ , the density-limit identification of $M_{\infty}$ with $d Q / d P$ , and the singular conclusion. This is the full Kakutani equivalence/singularity theorem for product measures.

Exercise 8 (hard, symbolic).

Two infinite sequences of independent Gaussians: under $P$ each coordinate is $N (0, 1)$ , under $Q$ each is $N (μ_{k}, 1)$ . Compute $H (p_{k}, q_{k})$ and show $Q ≪ P$ iff $\sum_{k} μ_{k}^{2} < \infty$ , and $Q ⊥ P$ otherwise.

Hint

$\int p_{k} q_{k} d x$ for two unit-variance Gaussians with means $0$ and $μ_{k}$ is a Gaussian integral; complete the square.

Answer

With $p_{k} (x) = \frac{1}{2 π} e^{- x^{2} /2}$ and $q_{k} (x) = \frac{1}{2 π} e^{- (x - μ_{k})^{2} /2}$ , $$ \sqrt{p_k q_k} = \tfrac{1}{\sqrt{2\pi}}\exp\Big(-\tfrac14\big(x^2 + (x-\mu_k)^2\big)\Big) = \tfrac{1}{\sqrt{2\pi}}\exp\Big(-\tfrac12\big(x - \tfrac{\mu_k}{2}\big)^2 - \tfrac{\mu_k^2}{8}\Big). $$ Integrating the Gaussian in $x$ (which integrates to $1$ against the normalised density) gives $H (p_{k}, q_{k}) = e^{- μ_{k}^{2} /8}$ . Then $\prod_{k} H (p_{k}, q_{k}) = exp (- \frac{1}{8} \sum_{k} μ_{k}^{2})$ , which is positive iff $\sum_{k} μ_{k}^{2} < \infty$ and zero iff $\sum_{k} μ_{k}^{2} = \infty$ . By Kakutani (Exercise 7), $Q ≪ P$ iff $\sum_{k} μ_{k}^{2} < \infty$ , and $Q ⊥ P$ otherwise. Rubric: full credit for the completed-square Gaussian integral, the closed form $e^{- μ_{k}^{2} /8}$ , and the $\sum μ_{k}^{2}$ dichotomy. This is the Feldman-Hájek dichotomy in its simplest shift form: equivalence of the infinite-dimensional Gaussian shifts is governed by the Cameron-Martin square-summability condition.

Advanced results Master

The product martingale of independent unit-mean factors is the prototype on which the entire likelihood-ratio theory of statistics is built, and Kakutani's dichotomy is the statement that the asymptotics admit no intermediate regime. The $L^{2}$ lift $M_{n} = (\prod_{k \leq n} a_{k}) N_{n}$ is the structural device: it separates the deterministic Hellinger decay $\prod a_{k}$ from a genuine $L^{2}$ -bounded martingale $N_{n}$ , so that the survival question for the $L^{1}$ object $M_{n}$ reduces to the Hilbert-space boundedness of $N_{n}$ , decided exactly by $\prod a_{k} > 0$ . The same lift shows the dichotomy is sharp: when $\prod a_{k} = 0$ the martingale $N_{n}$ still converges a.s. to a finite limit, and multiplying by the vanishing deterministic factor forces $M_{\infty} = 0$ , while the constant mean $E [M_{n}] = 1$ certifies the failure of $L^{1}$ convergence — the canonical UI/non-UI split of 37.04.03.

The signature application is the Kakutani dichotomy for infinite product measures. Let $P = ⨂_{k} p_{k}$ and $Q = ⨂_{k} q_{k}$ be product probability measures on a countable product of measurable spaces, with $q_{k} ≪ p_{k}$ on each coordinate. Under $P$ the finite-coordinate likelihood ratios $M_{n} = \prod_{k \leq n} (d q_{k} / d p_{k})$ form a product martingale with $a_{k} = H (p_{k}, q_{k}) = \int q_{k} p_{k} d μ_{k}$ . Kakutani's theorem then yields a perfect alternative: either $\prod_{k} H (p_{k}, q_{k}) > 0$ , in which case $Q ≪ P$ with Radon-Nikodym derivative $d Q / d P = M_{\infty} = lim_{n} M_{n}$ (the limit existing in $L^{1} (P)$ and strictly positive $P$ -a.s.); or $\prod_{k} H (p_{k}, q_{k}) = 0$ , in which case $Q ⊥ P$ . Because $q_{k} ≪ p_{k}$ on each coordinate is symmetric to $p_{k} ≪ q_{k}$ when the laws are mutually absolutely continuous, in the equivalent-coordinates case the dichotomy reads $Q \sim P$ (mutual equivalence) versus $Q ⊥ P$ (mutual singularity), with no possibility of one-sided absolute continuity for infinite products of equivalent factors. This is the Kakutani 0-1 phenomenon: the event ${d Q / d P > 0}$ is a tail event of the independent coordinates under $P$ , so the Kolmogorov 0-1 law forces $P (d Q / d P > 0) \in {0, 1}$ — the derivative is either a.s. positive (equivalence) or a.s. zero (singularity), never positive on a proper intermediate set.

The Hellinger geometry makes the criterion computable. The Hellinger affinity $H (p, q) = \int pq d μ$ and the Hellinger distance $h$ with $h^{2} = 1 - H$ turn the product condition into the additive condition $\sum_{k} (1 - H (p_{k}, q_{k})) < \infty$ , equivalently $\sum_{k} h^{2} (p_{k}, q_{k}) < \infty$ . Square-summability of the per-coordinate Hellinger distances is exactly the threshold separating equivalence from singularity. For Gaussian shifts $N (0, 1)$ versus $N (μ_{k}, 1)$ one computes $H = e^{- μ_{k}^{2} /8}$ , recovering the Cameron-Martin / Feldman-Hájek condition $\sum_{k} μ_{k}^{2} < \infty$ : two product Gaussian measures on a sequence space are either equivalent or mutually singular, equivalence holding iff the mean shift lies in the Cameron-Martin space $ℓ^{2}$ . The infinite-dimensional rigidity — no intermediate absolute continuity, only equivalence or orthogonality — is the measure-theoretic face of the fact that infinite-dimensional Gaussian measures have no translation-invariant reference measure.

In statistics this is the asymptotics of likelihood-ratio testing and contiguity. The product martingale $M_{n} = d Q_{n} / d P_{n}$ is the likelihood ratio after $n$ independent observations; $M_{\infty} > 0$ a.s. (the contiguous, mutually-absolutely-continuous regime) is the case where no test can perfectly separate $P$ from $Q$ even with infinitely many observations, while $M_{\infty} = 0$ $P$ -a.s. (singularity) is the case of asymptotically perfect discrimination, where a consistent test of power one against size zero exists. Le Cam's theory of contiguity is the local refinement: when $q_{k} = q_{k, n}$ depends on $n$ so that the Hellinger product tends to a positive constant, $P_{n}$ and $Q_{n}$ are mutually contiguous, and the log-likelihood ratio is asymptotically Gaussian by the central limit theorem applied to $\sum_{k} (f_{k} - 1)$ , the linearisation of $lo g M_{n}$ that the $L^{2}$ lift makes natural.

Synthesis. The foundational reason these results cohere is that $M_{n}$ divided by its deterministic Hellinger factor is an $L^{2}$ -bounded martingale precisely when the Hellinger product survives — this is exactly the device that converts the $L^{1}$ closure question of 37.04.03 into a Hilbert-space boundedness question, and it generalises the additive strong-law machinery to the multiplicative setting. The dichotomy is dual to the tail-event 0-1 law: ${M_{\infty} > 0}$ is a tail event, so the Kolmogorov law forces the clean split, and putting these together, the central insight is that uniform integrability of the likelihood-ratio martingale, mutual absolute continuity of the infinite product measures, and square-summability of the Hellinger distances are one condition read through the closure theorem, the Radon-Nikodym derivative, and the additive Hellinger geometry respectively. This is exactly the multiplicative companion of the additive Kolmogorov three-series and zero-one circle of 37.04.03: there the survival of a sum of independent terms was decided by summable variances; here the survival of a product is decided by summable Hellinger deficits, and the bridge is that taking the square root linearises the product into the sum, so Kakutani's theorem is the strong law of large numbers wearing the Hellinger metric, with the Radon-Nikodym derivative $d Q / d P$ playing the role that the sample mean plays for sums — the faithful limit when the system closes, the identically-zero collapse when it does not.

Full proof set Master

The dichotomy and its product-measure corollary are proved in full above. The remaining Master claims are recorded here.

Proposition (Hellinger criterion as a series). For probability densities $p_{k}, q_{k}$ with $0 < H (p_{k}, q_{k}) \leq 1$ , the infinite product $\prod_{k} H (p_{k}, q_{k})$ is positive iff $\sum_{k} h^{2} (p_{k}, q_{k}) < \infty$ , where $h^{2} = 1 - H$ .

Proof. Set $H_{k} = H (p_{k}, q_{k}) = 1 - h_{k}^{2}$ with $h_{k}^{2} = h^{2} (p_{k}, q_{k}) \in [0, 1)$ . The product $\prod_{k} H_{k}$ is positive iff $\sum_{k} (- lo g H_{k}) < \infty$ . Since $H_{k} \leq 1$ , $- lo g H_{k} = - lo g (1 - h_{k}^{2}) \geq h_{k}^{2}$ , so $\sum h_{k}^{2} = \infty \Rightarrow \prod H_{k} = 0$ . Conversely, if $\sum h_{k}^{2} < \infty$ then $h_{k}^{2} \to 0$ , and for $h_{k}^{2} \leq 1/2$ one has $- lo g (1 - h_{k}^{2}) \leq 2 h_{k}^{2}$ ; only finitely many indices have $h_{k}^{2} > 1/2$ and each contributes a finite term, so $\sum_{k} (- lo g H_{k}) < \infty$ and $\prod_{k} H_{k} > 0$ . $□$

Proposition (Kakutani equivalence/singularity for product measures). Let $P = ⨂_{k} p_{k}$ , $Q = ⨂_{k} q_{k}$ on $\prod_{k} E_{k}$ with $p_{k} \sim q_{k}$ (mutual equivalence) on each coordinate. Then either $Q \sim P$ or $Q ⊥ P$ ; the first holds iff $\prod_{k} H (p_{k}, q_{k}) > 0$ , and then $d Q / d P = lim_{n} \prod_{k \leq n} (d q_{k} / d p_{k})$ $P$ -a.s.

Proof. Under $P$ , the likelihood ratios $f_{k} = d q_{k} / d p_{k}$ are independent with $E_{P} [f_{k}] = \int f_{k} d p_{k} = \int d q_{k} = 1$ and $E_{P} [f_{k}] = \int f_{k} d p_{k} = \int q_{k} p_{k} d μ_{k} = H (p_{k}, q_{k}) =: a_{k}$ . So $M_{n} = \prod_{k \leq n} f_{k}$ is a Kakutani product martingale under $P$ and $M_{n} = d Q_{n} / d P_{n}$ on $F_{n} = σ (ω_{1}, \dots, ω_{n})$ . If $\prod a_{k} > 0$ , case (i) gives $M_{n} \to M_{\infty}$ in $L^{1} (P)$ , $M_{\infty} > 0$ $P$ -a.s., $E_{P} [M_{\infty}] = 1$ . For $A \in F_{n}$ , $Q (A) = \int_{A} M_{n} d P$ , and since $F_{n} ↑ F$ and $M_{n} \to M_{\infty}$ in $L^{1} (P)$ , the measures $A \mapsto \int_{A} M_{\infty} d P$ and $Q$ agree on the generating field $⋃_{n} F_{n}$ , hence on $F$ ; thus $Q ≪ P$ with $d Q / d P = M_{\infty}$ . As $M_{\infty} > 0$ $P$ -a.s., the symmetric argument with $P, Q$ exchanged (using $a_{k}^{'} = H (q_{k}, p_{k}) = a_{k}$ ) gives $P ≪ Q$ , so $Q \sim P$ . If $\prod a_{k} = 0$ , case (ii) gives $M_{\infty} = 0$ $P$ -a.s.; then $\int_{A} M_{\infty} d P = 0$ for all $A$ , yet $Q (\prod E_{k}) = 1$ , so $Q$ cannot be $≪ P$ . By the symmetric computation under $Q$ , the reciprocal martingale $1/ M_{n} = \prod_{k \leq n} (d p_{k} / d q_{k})$ has the same Hellinger product $\prod a_{k} = 0$ , so it collapses $Q$ -a.s.; hence $M_{n} \to \infty$ $Q$ -a.s. The set $S = {lim_{n} M_{n} = 0}$ has $P (S) = 1$ and $Q (S) = 0$ , witnessing $Q ⊥ P$ . $□$

Proposition (Kakutani 0-1 phenomenon). In the setting above, the event $D = {M_{\infty} > 0}$ is a $P$ -tail event, so $P (D) \in {0, 1}$ ; correspondingly $Q ≪ P$ or $Q ⊥ P$ with no intermediate case.

Proof. The a.s. limit $M_{\infty} = lim_{n} \prod_{k \leq n} f_{k} (ω_{k})$ satisfies, for each $m$ , $M_{\infty} = (\prod_{k \leq m} f_{k} (ω_{k})) \cdot lim_{n} \prod_{m < k \leq n} f_{k} (ω_{k})$ , and each $f_{k} > 0$ $P$ -a.s. (as $p_{k} \sim q_{k}$ ). Hence ${M_{\infty} > 0} = {lim_{n} \prod_{m < k \leq n} f_{k} > 0}$ up to a $P$ -null set, which is measurable with respect to $σ (ω_{m + 1}, ω_{m + 2}, \dots)$ for every $m$ ; thus $D$ lies in the tail $σ$ -algebra $⋂_{m} σ (ω_{m + 1}, \dots)$ . By the Kolmogorov 0-1 law 37.04.03 under the independent product $P$ , $P (D) \in {0, 1}$ . If $P (D) = 1$ then $M_{\infty} > 0$ $P$ -a.s. and the previous proposition gives $Q \sim P$ ; if $P (D) = 0$ then $M_{\infty} = 0$ $P$ -a.s. and $Q ⊥ P$ . No proper-subset positivity of the derivative is possible. $□$

Connections Master

The discrete martingale foundations of 37.04.01 supply the framework: the product $M_{n} = \prod_{k \leq n} X_{k}$ is a martingale by the independence-plus-take-out computation of that unit, and the optional-stopping and tail- $σ$ -algebra apparatus there is what lets ${M_{\infty} > 0}$ be analysed as a stopping-time-free tail event.

The almost-sure convergence theorem of 37.04.02 is the load-bearing input that guarantees $M_{n} \to M_{\infty}$ exists a.s. for the non-negative $L^{1}$ -bounded product martingale before any dichotomy is invoked; the a.s. finiteness of the limit of $N_{n}$ in case (ii) is exactly that unit's non-negative-martingale convergence statement, and the upcrossing mechanism is what forbids the product from oscillating instead of settling.

The uniform-integrability and $L^{p}$ theory of 37.04.03 is where the dichotomy lives: case (i) is precisely the UI/ $L^{1}$ -closure regime, proved here by lifting to the $L^{2}$ -bounded martingale $N_{n}$ and invoking Doob's $L^{2}$ bound, and case (ii) is the non-UI a.s.-but-not- $L^{1}$ regime; the Kolmogorov 0-1 law used to pin ${M_{\infty} > 0}$ is the same zero-one circle assembled in that unit.

The Radon-Nikodym theorem of 02.07.08 gives meaning to the product-measure corollary: $d Q / d P = M_{\infty}$ identifies the martingale limit with the density of the infinite product law, and the per-coordinate likelihood ratios $d q_{k} / d p_{k}$ are the Radon-Nikodym derivatives constructed there; the whole equivalence-versus-singularity statement is a Radon-Nikodym alternative read at infinity.

The Fubini-Tonelli and product-measure construction of 02.07.07 is what makes $P = ⨂_{k} p_{k}$ and the Hellinger integral $\int p_{k} q_{k} d μ_{k}$ well-defined objects, and the independence-under- $P$ factorisation of the likelihood ratio is the product-measure structure of that unit.

Historical & philosophical context Master

Shizuo Kakutani proved the dichotomy for infinite product measures in On equivalence of infinite product measures (Ann. of Math. 49, 1948, 214) ^{[Kakutani 1948]}, establishing that two infinite product probability measures are either equivalent or mutually singular, with the alternative decided by the convergence of the product of Hellinger integrals. The Hellinger integral and the associated quadratic-form distance go back to Ernst Hellinger's 1909 study of quadratic forms in infinitely many variables (J. Reine Angew. Math. 136, 1909, 210) ^{[Hellinger 1909]}, originally a tool in the spectral theory of operators rather than probability; its reinterpretation as a measure of statistical affinity came later. The martingale proof — recognising the finite-coordinate likelihood ratios as a non-negative product martingale and reading the dichotomy off uniform integrability — is due to Joseph Doob and was given its now-standard textbook form by David Williams in Probability with Martingales (Cambridge, 1991), where the square-root lift to an $L^{2}$ -bounded auxiliary martingale is the organising computation.

The Gaussian instance was settled independently by Jacob Feldman and Jaroslav Hájek in 1958, giving the Feldman-Hájek dichotomy: two Gaussian measures on a function space are equivalent or orthogonal, equivalence for shifts holding iff the mean difference lies in the Cameron-Martin space, the $ℓ^{2}$ condition $\sum μ_{k}^{2} < \infty$ that the Hellinger computation reproduces. The statistical descendants are Lucien Le Cam's theory of contiguity and asymptotic equivalence of experiments, in which the product martingale becomes the likelihood-ratio process and the Hellinger product its limiting affinity. The mathematical content is that infinite-dimensional measure spaces are rigid in a way finite-dimensional ones are not: a single coordinate admits a full continuum of mutual relationships between two laws, but an infinite independent product collapses all of them onto the binary equivalence-or-singularity, because the deciding quantity is a tail functional and tails of independent sequences are deterministic.

Bibliography Master

@article{kakutani1948,
  author  = {Kakutani, Shizuo},
  title   = {On equivalence of infinite product measures},
  journal = {Annals of Mathematics},
  volume  = {49},
  number  = {1},
  pages   = {214--224},
  year    = {1948}
}

@article{hellinger1909,
  author  = {Hellinger, Ernst},
  title   = {Neue Begr\"undung der Theorie quadratischer Formen von unendlichvielen Ver\"anderlichen},
  journal = {Journal f\"ur die reine und angewandte Mathematik},
  volume  = {136},
  pages   = {210--271},
  year    = {1909}
}

@book{williams1991,
  author    = {Williams, David},
  title     = {Probability with Martingales},
  publisher = {Cambridge University Press},
  series    = {Cambridge Mathematical Textbooks},
  year      = {1991}
}

@book{durrett2019,
  author    = {Durrett, Rick},
  title     = {Probability: Theory and Examples},
  edition   = {5th},
  series    = {Cambridge Series in Statistical and Probabilistic Mathematics},
  publisher = {Cambridge University Press},
  year      = {2019}
}

@book{shiryaev1996,
  author    = {Shiryaev, Albert N.},
  title     = {Probability},
  edition   = {2nd},
  series    = {Graduate Texts in Mathematics},
  volume    = {95},
  publisher = {Springer},
  year      = {1996}
}

@article{feldman1958,
  author  = {Feldman, Jacob},
  title   = {Equivalence and perpendicularity of Gaussian processes},
  journal = {Pacific Journal of Mathematics},
  volume  = {8},
  pages   = {699--708},
  year    = {1958}
}

@book{lecam1986,
  author    = {Le Cam, Lucien},
  title     = {Asymptotic Methods in Statistical Decision Theory},
  series    = {Springer Series in Statistics},
  publisher = {Springer},
  year      = {1986}
}

Prerequisites

37.04.01
37.04.02
37.04.03
02.07.07
02.07.08

Tier anchors

beginner: Williams — Probability with Martingales Ch. 14.12-14.13 (informal); Durrett — Probability: Theory and Examples (5th ed.) §4.3 (Kakutani dichotomy, informal)
intermediate: Williams — Probability with Martingales (CUP, 1991) §14.12-14.17; Durrett — Probability: Theory and Examples (5th ed.) §4.3.3
master: Williams — Probability with Martingales §14.12-14.18; Durrett §4.3.3; Williams — Probability with Martingales §14.18 (Kakutani's theorem, statement and proof); Shiryaev — Probability (Springer, 2nd ed., 1996) Ch. VII §6

References

Williams — Probability with Martingales (Cambridge University Press, 1991) · §14.12-14.18 (product martingales, Kakutani's theorem, absolute continuity vs singularity of product measures, the likelihood-ratio martingale)
Durrett — Probability: Theory and Examples (Cambridge University Press, 5th ed., 2019) · §4.3.3 (Kakutani's dichotomy for product martingales and product measures)
Kakutani — On equivalence of infinite product measures (Ann. of Math. 49, 1948) · pp. 214-224; the absolute-continuity-vs-singularity dichotomy for infinite product measures via the Hellinger integral
Shiryaev — Probability (Springer, 2nd ed., 1996) · Ch. VII §6 (the Hellinger integral, contiguity, and the Kakutani dichotomy)
Hellinger — Neue Begründung der Theorie quadratischer Formen von unendlichvielen Veränderlichen (J. Reine Angew. Math. 136, 1909) · pp. 210-271; the Hellinger integral and Hellinger distance between measures

Estimated time

beginner: 19m
intermediate: 53m
master: 96m