Relationships Between Modes of Convergence

Core Idea

The convergence modes form a hierarchy: almost sure convergence implies convergence in probability, which implies convergence in distribution. L^p convergence implies convergence in L^q for p > q by Hölder's inequality. Convergence in probability and almost sure convergence are generally incomparable. Understanding these relationships helps select the appropriate convergence mode for applications.

How It's Best Learned

Draw the hierarchy diagram showing implications. Work examples showing non-implications (e.g., convergence in distribution does not imply convergence in probability). Construct explicit counterexamples.

Common Misconceptions

• Thinking all modes of convergence are equivalent. - Believing convergence in distribution implies convergence in probability. - Forgetting that almost sure and in-probability convergence are not directly comparable.

Explainer

You have studied four distinct ways a sequence of random variables Xₙ can converge to a limit X: almost surely (a.s.), in probability, in distribution, and in Lᵖ. Each definition makes a different type of claim about how Xₙ approaches X. Understanding these modes in isolation is necessary but not sufficient — the real analytical power comes from knowing which modes imply which others, and from developing the habit of asking "which type of convergence do I actually need for this theorem?"

The main hierarchy runs: a.s. ⟹ in probability ⟹ in distribution. Almost sure convergence says P(lim_{n→∞} Xₙ = X) = 1 — the convergence happens on a probability-1 set of sample paths. This is a strong pathwise statement, and it implies convergence in probability: if Xₙ → X on almost every path, then for any ε > 0, the probability that |Xₙ − X| > ε must vanish. Convergence in probability is weaker because it only asks that Xₙ is within ε of X *most* of the time, without requiring the exceptional excursions to disappear forever. Convergence in distribution is weaker still: it only requires that the CDFs Fₙ(t) → F(t) at continuity points of F — the limit X need not even live on the same probability space as the Xₙ.

None of these implications reverses. The canonical counterexample for "in probability ⟹ a.s." is the typewriter sequence: on [0,1] with Lebesgue measure, define Xₙ as the indicator of a sliding interval that covers every point infinitely often. Xₙ → 0 in probability (the interval's length shrinks to 0), but Xₙ(ω) fails to converge for any ω (every point is visited by infinitely many intervals). For "in distribution ⟹ in probability," let Xₙ = X for all n where X ~ N(0,1), and let Y be an independent N(0,1) copy. Then Xₙ → Y in distribution (both are N(0,1)) but P(|Xₙ − Y| > ε) = P(|X − Y| > ε) > 0 for all n — no convergence in probability.

The Lᵖ modes connect via two key facts: Lᵖ ⟹ in probability (by Markov's inequality: P(|Xₙ − X| > ε) ≤ E[|Xₙ − X|ᵖ]/εᵖ → 0), and Lᵖ ⟹ Lq for p > q (by Hölder's inequality). But Lᵖ convergence and a.s. convergence are independent of each other — neither implies the other in general. The complete diagram has a.s. and Lᵖ both pointing to in probability, which points to in distribution, with no arrows pointing backward. When a theorem requires one mode and you have another, checking this diagram immediately tells you whether your hypothesis is sufficient — or whether you need an additional condition like uniform integrability to bridge the gap.