On strong law of large numbers for weakly stationary $\varphi$-mixing set-valued random variable sequences

This paper extends the concept of $\varphi$-mixing to set-valued random sequences in Banach spaces and establishes several strong laws of large numbers for such sequences, demonstrating the naturalness and sharpness of the underlying hypotheses through illustrative examples.

The Gist

Here is an explanation of the paper, translated from complex mathematical jargon into a story about crowds, clouds, and weather patterns.

The Big Picture: Predicting the Shape of the Future

Imagine you are trying to predict the weather. In the old days, you might have looked at a single thermometer (a single number) to guess tomorrow's temperature. This is what traditional statistics does: it deals with single numbers.

But in the real world, things are rarely just "numbers." Sometimes, you are dealing with a cloud. A cloud has a shape, a size, and a fuzzy edge. It isn't just "50 degrees"; it's a "cloud of uncertainty" that might be 50 degrees here and 52 degrees there.

This paper is about how to predict the average shape of these clouds over time, even when the clouds are "lazy" (they don't change their average shape) but also "chatty" (they influence each other).

The Cast of Characters

1. The Clouds (Set-Valued Random Variables):
   Instead of a single number, imagine every day you get a whole shape (like a circle, a line, or a blob). This is a "set-valued random variable."

   • Analogy: Instead of asking "What is the temperature?", you ask "What is the range of possible temperatures?" and you get a whole interval like [20°C, 25°C].

2. The Chatty Neighbors (φ-mixing):
   In a perfect world, today's weather has nothing to do with yesterday's. But in reality, if it's raining today, it's likely to rain tomorrow. The clouds "talk" to each other.

   • The Paper's Rule: The author calls this φ-mixing. It means the clouds talk, but they get tired. The further apart two days are, the less they influence each other. Eventually, the "chatter" dies down to zero.

3. The Lazy Clouds (Weak Stationarity):
   This means that while the clouds might wiggle and change shape day-to-day, their average shape stays the same forever.

   • Analogy: Imagine a cloud that sometimes looks like a fat circle and sometimes like a thin oval, but if you averaged all its shapes over a year, it always comes out to be a perfect circle.

The Main Question: The Law of Large Numbers for Clouds

You know the Law of Large Numbers from school: If you flip a coin enough times, the average of heads and tails will settle down to 50/50.

This paper asks: If I have a sequence of "clouds" that are chatty (mixing) but have a stable average shape (stationary), will the average of all these clouds eventually settle down to that stable shape?

The answer is YES, but with some specific rules.

The Three Rules for the Clouds to Settle Down

The author proves that for the clouds to settle into their average shape, three things must happen:

1. The Chatter Must Fade Fast Enough:
   The "mixing" (how much the clouds influence each other) has to drop to zero quickly. If the clouds are too chatty (influencing each other for too long), the average will never settle.

   • Metaphor: Imagine a room of people shouting. If they stop shouting after a few seconds, you can hear the average volume. If they keep shouting at each other for hours, you can't find a pattern.

2. The "Edges" Must Be Tame:
   The paper uses something called a "support function" (a fancy way of measuring the edge of the cloud from every angle). The paper requires that the "wiggles" of these edges don't get too wild too often.

   • Metaphor: If your cloud suddenly explodes into a giant spike once in a blue moon, it ruins the average. The spikes need to be small enough that they don't break the math.

3. The "Halo" Effect (For Weird Shapes):
   Sometimes the clouds are weird shapes (like a long, thin needle). The paper shows that even if the needle is long, as long as the "halo" (the fuzzy edge) shrinks fast enough, the average will still work.

The "Needle and Halo" Example (The Best Part)

The paper includes a cool example to show why these rules matter.

• The Setup: Imagine a long, thin needle pointing East. Every day, a tiny "halo" (a small speck) appears near the tip of the needle.
• The Magic: Even though the needle is infinite, the paper proves that if the halo gets smaller fast enough (shrinking like $1/n$), the average of all these needles will eventually look exactly like the original needle.
• The Warning: If the halo doesn't shrink fast enough, the average will get "stuck" in a weird shape and never settle down. This proves the rules in the paper are necessary.

Why Does This Matter?

You might ask, "Who cares about averaging clouds?"

Actually, this is huge for:

• Finance: Predicting the "range" of stock prices, not just the average price.
• Robotics: If a robot arm is shaky, it doesn't move in a straight line; it moves in a "cloud" of possible paths. Engineers need to know where the average path will be to avoid crashing.
• Data Mining: When you have messy, uncertain data (like "the customer is somewhere between 20 and 30 years old"), this math helps find the true trend.

The Conclusion

The author, Luc T. Tuyen, has successfully built a bridge. He took the old, boring math of averaging single numbers and upgraded it to handle complex, fuzzy, chatty shapes.

He proved that as long as the "chatter" between the shapes dies down fast enough and the shapes don't get too crazy, the average will always find its way home to the true, stable shape.

In short: Even if your data is messy, fuzzy, and talks to itself, if you wait long enough and the noise isn't too loud, the truth will eventually reveal itself.

Technical Summary

Here is a detailed technical summary of the paper "On strong law of large numbers for weakly stationary φ-mixing set-valued random variable sequences" by Luc T. Tuyen.

1. Problem Statement

The paper addresses a gap in probability theory regarding the Strong Law of Large Numbers (SLLN) for set-valued random variables (random sets) under conditions of dependence.

• Context: While the SLLN is well-established for independent and identically distributed (i.i.d.) real-valued and set-valued variables, and for dependent single-valued variables under mixing conditions (specifically $\phi$-mixing), the extension to dependent set-valued sequences remains limited.
• Specific Challenges:
  1. Degeneracy: If a set-valued random variable has an expectation of a single point $\{0\}$, it may degenerate into a single-valued variable, losing its set-valued nature.
  2. Selection Preservation: Selections (single-valued functions chosen from the set) of non-identically distributed random sets may fail to preserve the dependence structures required for standard SLLN proofs.
  3. Geometry: Set-valued sequences can be highly dispersed even if they "oscillate" around a mean set, complicating convergence definitions (Hausdorff vs. Kuratowski-Mosco).
• Goal: To define $\phi$-mixing and weak stationarity for set-valued sequences and prove SLLN results under these conditions without forcing the sets to degenerate into single points.

2. Methodology and Definitions

The author introduces a rigorous framework extending classical concepts to the space of closed subsets of a separable Banach space $X$.

Key Definitions

• $\phi$-Mixing for Set-Valued Variables:
  The dependence coefficient $\phi(n)$ is defined based on the $\sigma$-algebras generated by the set-valued variables $X_k$.
  $$\phi(n) = \sup_{k \ge 1} \sup_{A \in \mathcal{F}_1^k, B \in \mathcal{F}_{k+n}^\infty} |P(B|A) - P(B)|$$
  A sequence is $\phi$-mixing if $\phi(n) \to 0$ as $n \to \infty$. Crucially, the $\sigma$-algebra generated by a set-valued variable is defined via its support functions or measurable selections.
• Weak Stationarity:
  A sequence $\{X_n\}$ is weakly stationary if the Aumann expectation $E[X_n] = A$ is constant for all $n$, where $A$ is a non-empty closed subset of $X$. This is weaker than strict stationarity (which requires identical joint distributions).
• Support Functions:
  The proofs heavily rely on the support function $s(x^*, X) = \sup_{x \in X} \langle x^*, x \rangle$. A key property utilized is that if $\{X_n\}$ is $\phi$-mixing, then the scalar sequence of support functions $\{s(x^*, X_n)\}$ is also $\phi$-mixing for any $x^* \in X^*$.

Convergence Modes

The paper analyzes convergence in two distinct topologies:

1. Hausdorff Metric ($H$): Standard metric for compact sets.
2. Kuratowski-Mosco (K-M) Convergence: A stronger form of convergence involving both weak upper limits and strong lower limits, applicable to non-compact and unbounded sets.

3. Key Contributions and Results

The paper establishes three main theorems, each addressing different classes of set-valued variables and convergence types.

Theorem 3.1: SLLN for Compact Convex Sets (Hausdorff Convergence)

• Assumptions: $\{X_n\}$ is a weakly stationary, $\phi$-mixing sequence of compact convex sets in $L^2$.
  • Condition (i): $\sum \phi^{1/2}(n) < \infty$.
  • Condition (ii): $\sum \frac{E|s(x^*, X_n) - s(x^*, A)|^2}{n^2} < \infty$ for all $x^* \in S^*$.
• Result: The sample mean converges to the expectation set $A$ in the Hausdorff metric almost surely:
  $$\lim_{n \to \infty} H\left( \frac{1}{n}\sum_{k=1}^n X_k, A \right) = 0 \quad \text{a.s.}$$
• Method: The proof reduces the set-valued problem to a scalar problem using support functions. Since the support functions form a $\phi$-mixing scalar sequence, the classical SLLN (Theorem 1.1) is applied, and the result is lifted back to the set-valued domain using the properties of the Hausdorff metric on support functions.

Theorem 3.2: SLLN for Compact (Non-Convex) Sets

• Assumptions: $\{X_n\}$ is a weakly stationary, $\phi$-mixing sequence of compact (not necessarily convex) sets.
• Result: The sample mean of the sets converges to the closed convex hull of the expectation ($coA$) in the Hausdorff metric:
  $$\lim_{n \to \infty} H\left( \frac{1}{n}\sum_{k=1}^n X_k, coA \right) = 0 \quad \text{a.s.}$$
• Method: Utilizes the fact that the convex hull of the sum of sets is the sum of the convex hulls, reducing the problem to Theorem 3.1.

Theorem 3.3: SLLN for General Closed Sets (Kuratowski-Mosco Convergence)

• Assumptions: $\{X_n\}$ is a weakly stationary, $\phi$-mixing sequence of general closed sets (potentially unbounded).
  • Condition (i): $\sum \phi^{1/2}(n) < \infty$.
  • Condition (ii): Existence of square-integrable selections $x_n$ for every $a \in A$ such that $\sum \frac{E\|x_n - a\|^2}{n^2} < \infty$.
  • Condition (iii): Pointwise summability of the variance of support functions for $x^*$ where $s(x^*, A) < \infty$.
• Result: The sequence of sample means $S_n = \frac{1}{n} \text{cl} \sum X_i$ converges to $D = coA$ in the Kuratowski-Mosco sense almost surely.
• Method:
  • Lower Limit ($s\text{-lim inf}$): Constructed by carefully selecting specific selections $x_n$ that approximate points in $A$ and applying the scalar SLLN to these selections.
  • Upper Limit ($w\text{-lim sup}$): Proved by contradiction using the Hahn-Banach separation theorem and the convergence of support functions.

4. Illustrative Examples and Sharpness

The paper provides critical examples to demonstrate that the hypotheses are natural and necessary:

• Example 3.1 & 3.2: Demonstrate that a sequence can be weakly stationary and $\phi$-mixing without being strictly stationary or degenerating to single points.
• Example 3.5 ("Needle + shrinking halo"): Shows a case where the sets are unbounded (a ray plus a shrinking noise), satisfying all conditions, and converging to the ray.
• Example 3.6 (Violation of Condition iii): This is a crucial counter-example. It constructs a sequence where conditions (i) and (ii) hold, but condition (iii) (summability of support function variance) fails. In this case, the K-M limit fails to converge to the expected set. This proves that condition (iii) is genuinely necessary for unbounded sets, preventing the limit from "exploding" into a larger sector.

5. Significance and Conclusion

• Theoretical Advancement: This work successfully bridges the gap between classical SLLN for dependent scalar variables and the complex geometry of set-valued random variables. It generalizes previous results that were limited to i.i.d. cases or strictly stationary sequences.
• Practical Relevance: The results apply to fields involving uncertainty quantification, such as finance, logistics, and data mining, where data is often represented as sets (e.g., confidence intervals, feasible regions) and exhibits temporal dependence.
• Robustness: The paper clarifies that weak stationarity is sufficient for these laws, relaxing the need for strict stationarity.
• Future Directions: The author suggests extending these results to stronger mixing concepts (like $\rho$-mixing) and exploring summability methods for convergence.

In summary, the paper provides a rigorous mathematical foundation for the convergence of dependent set-valued data, establishing that under specific mixing and summability conditions, the empirical average of random sets converges to their theoretical expectation, even in complex, non-convex, and unbounded scenarios.