A counter-example linked to Gaussian convex hulls

This paper demonstrates that if the assumption of weak convergence for a sequence of independent centered Gaussian random elements in a separable Banach space is relaxed, their normalized closed convex hulls can converge almost surely to an arbitrary convex compact set, rather than the concentration ellipsoid of a limiting distribution.

The Gist

Imagine you are throwing darts at a giant, invisible target board in a high-dimensional room. This isn't a normal game of darts, though. The "darts" are random numbers generated by a specific type of chaos called a Gaussian distribution (think of the classic bell curve, but in many directions at once).

Here is the story of Youri Davydov's paper, broken down into simple concepts and everyday analogies.

1. The Setup: The Dart Game

Imagine you have a machine that throws darts ($X_1, X_2, X_3...$) randomly.

• The Rule: The darts are "centered," meaning they aim for the exact middle of the room (zero).
• The Shape: Over time, if you look at where the darts land, they tend to cluster in a specific shape. In the old days of math, if the darts were all thrown by the same machine (identical distribution), they would always form a perfect ellipse (like a flattened circle or a rugby ball).

Mathematicians had a formula to predict this shape. If you take the first $n$ darts, draw the smallest rubber band that can wrap around all of them (this is called the convex hull), and then stretch or shrink that shape by a specific amount, it would eventually settle into that perfect ellipse.

2. The Twist: Breaking the Rules

For a long time, mathematicians thought: "If the darts are independent and centered, they will always eventually form that nice, smooth ellipse."

But Youri Davydov says: "Not necessarily."

He asks a simple question: What if we stop assuming every dart is thrown by the exact same machine? What if we change the machine slightly every few throws?

His answer is surprising: You can make the darts form ANY shape you want.

3. The Analogy: The Sculptor and the Clay

Think of the "limit shape" (the final shape the darts form) as a sculpture.

• The Old View: If you throw random clay balls, they will always settle into a smooth, round mound (the ellipse).
• Davydov's View: If you carefully control which clay balls you throw and when you throw them, you can force the pile to settle into a cube, a star, a pyramid, or any weird, jagged shape you can imagine.

In the paper, Davydov proves that if you have a specific shape in mind (let's say, a perfect cube or a diamond), you can design a sequence of random throws that, over a very long time, will naturally fill out that exact shape.

4. How Does He Do It? (The Secret Recipe)

Davydov's method is like a master chef following a recipe to bake a specific cake shape.

1. The Target: He starts with the shape he wants (let's call it $V$).
2. The Ingredients: He picks a bunch of points that are scattered all over the surface of that shape.
3. The Timing: He divides time into different "zones" (like different shifts in a factory).
   • In Shift A, he throws darts that are good at filling out the "left side" of the shape.
   • In Shift B, he throws darts for the "top side."
   • In Shift C, he throws darts for the "bottom right."
4. The Result: Because he throws the "left side" darts often enough, and the "top side" darts often enough, the rubber band (the convex hull) eventually stretches to touch every corner of his target shape.

He proves that even though every single dart is random, the pattern of the randomness is so carefully tuned that the final result is not a blob, but a precise geometric figure.

5. Why Does This Matter?

This might sound like a game, but it's actually a fundamental discovery about how randomness works in complex systems.

• The Lesson: Randomness doesn't always lead to "smoothness" or "averages." If you have enough control over the conditions of the randomness (even if the individual events are still random), you can engineer almost any outcome.
• The Counter-Example: Before this paper, people thought the "ellipse" was the only possible ending for these types of random problems. Davydov showed that the ellipse is just one special case. If you relax the rules slightly, the universe of possible shapes explodes.

Summary in One Sentence

Youri Davydov shows that by carefully timing your random throws, you can force a chaotic cloud of points to settle into any shape you can imagine, proving that randomness is far more flexible and controllable than we previously thought.

Technical Summary

1. Problem Statement and Context

The paper addresses the asymptotic behavior of the normalized closed convex hulls of a sequence of independent centered Gaussian random elements in a separable Banach space $B$.

Let $\{X_n\}$ be a sequence of independent centered Gaussian random elements in $B$, and let $W_n = \text{conv}\{X_1, \dots, X_n\}$ be their closed convex hull. The study focuses on the convergence of the normalized sequence:
$$\frac{1}{b(n)} W_n \quad \text{as} \quad n \to \infty,$$
where the normalizing factor is $b(n) = \sqrt{2 \ln n}$.

Background:

• Classical Result (Goodman, 1988): If $\{X_n\}$ are i.i.d. with a common distribution $P$, the normalized hulls converge almost surely (in the Hausdorff metric $d_H$) to the concentration ellipsoid $E$ of $P$.
• Extensions: This result was extended to Skorokhod spaces, stationary weakly dependent Gaussian fields, and cases where the sequence converges weakly ($X_n \Rightarrow X$).
• The Gap: Previous work showed that if the weak convergence assumption is dropped, the limit set might exist but differ from an ellipsoid (e.g., converging to a polytope). However, it was unknown whether the limit set could be arbitrary.

Objective:
Davydov aims to prove that if the assumption of weak convergence of the initial sequence is relaxed, the limit set of the normalized convex hulls can be any arbitrary convex compact set (specifically, any centrally symmetric one), not just an ellipsoid or a polytope.

2. Methodology

The author constructs a specific counter-example by carefully designing a sequence of independent Gaussian vectors $\{X_k\}$ that do not share a common distribution but are constructed to "fill out" a target shape.

Key Construction Steps:

1. Target Set: Let $V \subset B$ be an arbitrary convex, compact, centrally symmetric set.
2. Partitioning: The set of positive integers $\mathbb{N}$ is partitioned into disjoint subsets $\{T_k\}_{k=1}^\infty$ such that each $T_k$ has an asymptotic density $p_k > 0$ and $\sum p_k = 1$.
3. Dense Subset: A countable dense subset $\{s_k\}$ of the unit sphere in $B$ is selected.
4. Scaling Factors: For each $s_k$, a scaling factor $a_k$ is defined as the maximum radius such that $a_k s_k \in V$ (i.e., $a_k = \sup \{t > 0 \mid t s_k \in V\}$).
5. Random Variables:
   • Let $\{\xi_k\}$ be a sequence of independent standard Gaussian random variables.
   • Define the random vectors $X_k = a_k \xi_k s_k$.
   • Note: The distribution of $X_k$ is concentrated on the line spanned by $s_k$ with variance $a_k^2$.

Proof Strategy:
The proof relies on the theory of support functions. For a convex set $K$, the support function is $M_K(x^*) = \sup_{x \in K} \langle x, x^* \rangle$. Convergence in the Hausdorff metric is equivalent to the uniform convergence of support functions on the dual unit sphere.

The proof proceeds in two main stages:

1. Relative Compactness: Show that the sequence of normalized hulls $\{\frac{1}{b(n)}W_n\}$ is relatively compact almost surely. This involves proving that for any $\epsilon > 0$, the hulls are eventually contained within an $\epsilon$-neighborhood of a compact set $K$ (chosen as $V$). This is achieved using the Borel-Cantelli lemma and exponential tail bounds for Gaussian variables.
2. Convergence of Support Functions:
   • Lower Bound: Show that $\liminf M_n(x^*) \geq M_V(x^*)$. This is done by utilizing the fact that the points $\{a_k s_k\}$ are dense on the boundary of $V$ and that the maximum of Gaussian variables over the partition sets $T_k$ grows at the rate $b(n)$.
   • Upper Bound: Show that $\limsup M_n(x^*) \leq M_V(x^*)$. This follows because every $X_k$ lies within the cone defined by $V$, and the maximum of the Gaussian noise terms normalized by $b(n)$ converges to 1.

3. Key Contributions

• Arbitrariness of the Limit Set: The paper establishes that the limit set of normalized Gaussian convex hulls is not restricted to ellipsoids or specific polytopes. It can be any convex compact centrally symmetric set.
• Relaxation of Assumptions: It demonstrates that the weak convergence assumption ($X_n \Rightarrow X$) found in previous literature is crucial for the limit to be an ellipsoid. Without it, the "shape" of the limit is entirely determined by the construction of the variances and directions of the sequence.
• Uniform Boundedness: The constructed sequence satisfies $\sup_n \mathbb{E}|X_n|^2 < \infty$, ensuring the result holds even under uniform boundedness of second moments.

4. Main Result

Theorem 1:
Let $V \subset B$ be a convex compact centrally symmetric set. There exists a sequence of independent Gaussian vectors $\{X_k\}$ such that:

1. $\sup_n \mathbb{E}|X_n|^2 < \infty$.
2. With probability 1,
   $$\frac{1}{b(n)} W_n \xrightarrow{d_H} V \quad \text{as} \quad n \to \infty.$$

5. Significance

• Theoretical Insight: This result fundamentally alters the understanding of the asymptotic geometry of Gaussian samples in infinite-dimensional spaces. It shows that the "concentration ellipsoid" phenomenon is a specific artifact of stationarity or weak convergence, not a universal property of Gaussian processes.
• Counter-Example Utility: It provides a powerful counter-example for researchers studying limit theorems for convex hulls, demonstrating that without strict distributional assumptions, the limit set can be engineered to be any desired shape.
• Methodological Application: The technique of partitioning the index set and assigning specific variances/directions to "sample" the boundary of a target set is a novel and robust method for constructing Gaussian processes with specific asymptotic geometric properties.

In summary, Davydov proves that the shape of the limiting convex hull of independent Gaussian vectors is not intrinsic to the Gaussian nature of the variables but is entirely controllable via the sequence's covariance structure, provided the variables do not converge weakly to a single distribution.