Khinchin inequalities for uniforms on spheres with a deficit

This paper refines moment comparison inequalities for sums of random vectors uniformly distributed on Euclidean spheres by establishing sharp constants and introducing an optimal deficit term, particularly in high dimensions.

The Gist

Imagine you are a chef trying to bake the perfect cake. You have a specific recipe (a mathematical formula) that tells you how "fluffy" or "dense" your cake will be based on the ingredients you mix together.

In the world of mathematics, specifically probability theory, there is a famous rule called the Khinchin inequality. It's like a master recipe that predicts how big a "sum" will get when you mix together random ingredients.

For a long time, mathematicians knew the exact size of the cake if the ingredients were simple, like flipping a coin (Heads or Tails). But what happens if your ingredients are more complex, like rolling a die in many dimensions at once? This paper tackles that exact problem.

Here is the breakdown of what Jacek Jakimiuk, Colin Tang, and Tomasz Tkocz discovered, explained through simple analogies.

1. The Ingredients: Random Spins on a Sphere

Imagine you are in a room with a giant, invisible ball (a sphere). You have a bunch of arrows ($\xi_1, \xi_2, \dots$) pointing in random directions on the surface of this ball.

• The Old Way: If you add these arrows together, you get a new arrow. The paper asks: "How long is this new arrow?"
• The Comparison: The authors compare this random arrow to a "Gaussian" arrow (a bell-curve distribution), which is the mathematical standard for "perfectly random" noise.

2. The Problem: The "Perfect" Recipe vs. Reality

The authors already knew a "perfect" recipe (an inequality) that said:

"The length of your random arrow sum will never exceed the length of a specific Gaussian arrow sum."

This is like saying, "No matter how you mix your random ingredients, your cake will never be bigger than this specific reference cake."

But here's the catch: That old rule was a bit too generous. It told you the maximum possible size, but it didn't tell you how much smaller your cake actually was if your ingredients weren't perfectly balanced. It was like a speed limit sign that just said "Don't go over 100 mph," without telling you how much slower you were actually going.

3. The New Discovery: The "Deficit" (The Missing Piece)

The main breakthrough of this paper is adding a "deficit term."

Think of it like a budget gap.

• The Old Rule: "Your spending will never exceed $100."
• The New Rule: "Your spending will never exceed $100, minus a penalty based on how unevenly you spent your money."

In the math world, the "penalty" depends on how "spiky" your coefficients (the weights you give to each random arrow) are.

• If you give every arrow the exact same weight (a perfect balance), the penalty is zero. You hit the maximum size.
• If you give one arrow a huge weight and the others tiny weights (an unbalanced mix), the penalty is huge. Your resulting arrow will be significantly smaller than the maximum possible size.

The authors calculated the exact size of this penalty (the "deficit") for high-dimensional spheres. They found a formula that tells you exactly how much "smaller" your result is compared to the theoretical maximum, depending on how unbalanced your ingredients are.

4. Why This Matters: Stability

The paper is about stability.
Imagine you are building a tower of blocks.

• Stability: If you wiggle the blocks slightly, does the tower fall?
• The Insight: The authors show that if your tower (your sum of random vectors) is not the maximum possible height, it's not just "close" to the max; it is guaranteed to be significantly lower if your blocks are uneven.

This is crucial because in the real world (physics, engineering, data science), things are rarely perfectly balanced. Knowing exactly how much the imbalance hurts the result allows scientists to make much more precise predictions.

5. The "High-Dimensional" Twist

The paper also notes something fascinating about dimensions.

• In 1 dimension (just a line), the rules are one way.
• In 2 dimensions (a flat circle), it's different.
• In high dimensions (like 100 dimensions), the rules change again.

The authors found that as the number of dimensions gets huge, their new "deficit" formula becomes incredibly efficient. It's like discovering that in a massive crowd, the effect of one person shouting is much more predictable than in a small room. Their formula works best when the "room" (the dimension) is very large.

Summary Analogy

Imagine you are trying to guess the total weight of a bag of marbles.

• The Old Rule: "The bag will weigh no more than 10 lbs." (True, but vague).
• The New Rule: "The bag will weigh no more than 10 lbs, minus 5 lbs for every time you put a giant marble in the bag instead of a small one."

This paper provides the exact math for that "minus 5 lbs" part. It turns a vague safety net into a precise measuring tape, showing exactly how much "slack" exists in the system when things aren't perfectly uniform.

In short: They took a known mathematical limit and added a "fine print" clause that tells you exactly how much you lose when your random ingredients aren't perfectly mixed. This makes the prediction much sharper and more useful for real-world applications.

Technical Summary

Here is a detailed technical summary of the paper "Khinchin inequalities for uniforms on spheres with a deficit" by Jacek Jakimiuk, Colin Tang, and Tomasz Tkocz.

1. Problem Statement

The paper addresses the problem of stability in moment comparison inequalities for sums of independent random vectors uniformly distributed on Euclidean spheres ($S^{d-1}$).

• Context: Sharp Khinchin-type inequalities compare the $L_p$ moments of a sum of independent random vectors $\sum a_j \xi_j$ to a Gaussian reference. Specifically, for $p \ge 2$, it is known that:
  $$\mathbb{E}\left| \sum_{j=1}^n a_j \xi_j \right|^p \le \mathbb{E}|Z|^p$$
  where $\xi_j$ are uniform on $S^{d-1}$, $Z$ is a Gaussian vector with covariance $\frac{1}{d}I_d$ (normalized so $\mathbb{E}|Z|^2 = \mathbb{E}|\xi|^2 = 1$), and $\sum a_j^2 = 1$. The constant $\mathbb{E}|Z|^p$ is sharp.
• The Gap: While the sharp inequality is established, the stability of this inequality—quantifying how much the moment drops when the coefficients $a_j$ deviate from the "worst-case" scenario (which is the uniform distribution $a_j = 1/\sqrt{n}$)—had not been fully resolved for spherical vectors.
• Goal: The authors aim to sharpen these inequalities by adding a deficit term (a negative term on the right-hand side) that depends on the deviation of the coefficients $a_j$ from uniformity. This extends previous stability results known only for Rademacher sums (the $d=1$ case) to higher dimensions ($d \ge 2$).

2. Methodology

The authors employ a combination of Lindeberg's swapping argument, convexity analysis, and quantitative estimates of derivatives.

A. Lindeberg's Swapping Argument

The core proof strategy involves iteratively replacing each spherical random vector $\xi_j$ with a Gaussian vector $Z_j$.

• Let $S_0 = \sum a_j \xi_j$ and $S_n = \sum a_j Z_j$.
• The total difference $\mathbb{E}|S_n|^p - \mathbb{E}|S_0|^p$ is decomposed into a telescoping sum of local differences:
  $$\mathbb{E}|S_n|^p - \mathbb{E}|S_0|^p = \sum_{k=1}^n \left( \mathbb{E}|S_k|^p - \mathbb{E}|S_{k-1}|^p \right)$$
• Each term in the sum represents the deficit $D_p(a_k, v_k) = \mathbb{E}|a_k Z_k + v_k|^p - \mathbb{E}|a_k \xi_k + v_k|^p$, where $v_k$ is the sum of the remaining terms.

B. Quantitative Convexity (The Deficit Lemma)

To obtain a non-trivial deficit, the authors analyze the function $h_{v}(t) = \mathbb{E}|v + \sqrt{t}\xi|^p$.

• They establish that the non-negativity of the deficit ($D_p \ge 0$) stems from the convexity of $h_v(t)$ on $(0, \infty)$.
• Key Innovation: Instead of just proving convexity, they derive an exact expression for the second derivative $h_v''(t)$ using the divergence theorem and polar coordinates (Lemma 3).
• They use Taylor expansion with Lagrange remainder to bound the deficit $D_p$ from below by the second derivative. This leads to Lemma 4, which provides a quantitative lower bound for the deficit involving terms like $a^4$ and powers of $|v|$.

C. Handling Different Regimes

The proof splits into cases based on the moment order $p$:

1. **Case $2 \le p \le 4$:** The function$x \mapsto x^{(p-4)/2}$ is convex. The authors use Jensen's inequality on the integral representation of the second derivative to bound the deficit.
2. Case $p > 4$: The function is no longer convex in the same way. The authors utilize monotonicity properties of the moments and Khinchin-type inequalities (Lemma 5) to bound the lower-order moments of the partial sums from below.

D. Schur-Concavity and Majorization

For Theorem 2 (stability relative to the uniform vector), the authors use a local exchange argument. They show that the moment functional is Schur-concave. By iteratively replacing the largest and smallest coefficients with values closer to the mean ($1/\sqrt{n}$), they demonstrate that the deficit accumulates in a telescoping manner related to the$\ell_4$ norm of the coefficients.

3. Key Contributions and Results

Theorem 1: Deficit Relative to Gaussian

The paper establishes a sharp inequality with an explicit deficit term depending on the $\ell_4$ norm of the coefficients:
$$\mathbb{E}\left| \sum_{j=1}^n a_j \xi_j \right|^p \le \mathbb{E}|Z|^p - c_{p,d} \sum_{j=1}^n a_j^4$$

• Constant $c_{p,d}$: The authors provide an explicit, optimal constant in high dimensions:
  $$c_{p,d} = \frac{(p+d-2)(p+d-4)}{24d^2(d+2)} \times \begin{cases} 3p(p-2) & 2 \le p \le 4 \\ 1 & p > 4 \end{cases}$$
• Optimality: The constant scales as $\Theta_p(1/d)$ as $d \to \infty$, which is shown to be the best possible rate.

Theorem 2: Stability Relative to Uniform Coefficients

The paper provides a stability result comparing any coefficient vector $a$ (with $\sum a_j^2=1$) to the uniform vector $(1/\sqrt{n}, \dots, 1/\sqrt{n})$:
$$\mathbb{E}\left| \sum_{j=1}^n a_j \xi_j \right|^p \le \mathbb{E}\left| \sum_{j=1}^n \frac{1}{\sqrt{n}}\xi_j \right|^p - \tilde{c}_{p,d} \sum_{j=1}^n \left( \frac{1}{n} - a_j^2 \right)^2$$

• The deficit term is proportional to the squared deviation of the coefficients from uniformity (related to the difference in $\ell_4$ norms).
• The constant $\tilde{c}_{p,d}$ is explicitly calculated, involving the Khinchin constant for $p > 4$.

4. Significance and Impact

1. Extension of Stability Theory: Prior to this work, stability results with explicit deficit terms were largely confined to the Rademacher case ($d=1$). This paper successfully generalizes these results to Euclidean spheres of arbitrary dimension $d \ge 2$.
2. Paradigm Shift: The results support the paradigm that methods effective for Rademacher variables often extend to broader parameter ranges in higher dimensions. Notably, the authors prove the result for all $p \ge 2$, whereas previous stability results for Rademacher sums were often restricted to $p \ge 3$ or $p \ge 4$.
3. Optimal Constants: The paper provides explicit, dimension-dependent constants that are optimal in the high-dimensional limit. This is crucial for applications in high-dimensional probability and geometry where precise bounds are required.
4. Methodological Advancement: The derivation of the second derivative of the moment functional and the subsequent quantitative convexity bounds offer a robust toolkit for analyzing moment inequalities for rotationally invariant distributions beyond just spheres.
5. Applications: These inequalities are relevant to the study of geometric functional analysis, specifically regarding the volume of sections of $\ell_p$-balls, polydiscs, and simplices, as well as concentration of measure phenomena in high-dimensional spaces.

In summary, the paper closes a significant gap in the theory of Khinchin inequalities by providing the first sharp, stable moment comparison inequalities for spherical random vectors with explicit, optimal deficit terms.