Mollified Christoffel-Darboux Kernels and Density Recovery on Varieties

This paper introduces mollified Christoffel-Darboux kernels on algebraic varieties to achieve a uniform boundedness property on the support and enable consistent, quantitatively controlled density recovery from moment data without prior knowledge of the equilibrium measure.

The Gist

Imagine you are a detective trying to solve a mystery, but you don't have a camera or a map. All you have is a bag of clues (data points) scattered on the floor, and you need to figure out two things:

1. Where the action is happening: Which part of the floor is actually covered by the clues (the "support")?
2. How crowded it is: In which spots are the clues piled up thickly, and where are they sparse? (This is the "density").

In the world of mathematics and data science, this is a classic problem called Density Recovery. The paper you provided introduces a new, smarter tool to solve this mystery, called the Mollified Christoffel-Darboux (MCD) Kernel.

Here is the breakdown of how this tool works, using simple analogies.

1. The Old Tool: The "Flashlight" with a Glitch

For a long time, mathematicians used a tool called the Christoffel-Darboux (CD) Kernel. Think of this as a magical flashlight that shines on your data.

• How it worked: If you shined it on a spot where data existed, the light would get brighter and brighter as you made the flashlight more powerful (increasing the "degree"). If you shined it on an empty spot, the light would explode into blinding, infinite brightness.
• The Problem: This "on/off" behavior was great for finding the edges of the data (the support), but it was terrible for measuring how crowded the data was.
  • Imagine trying to measure the crowd density in a stadium. The old flashlight would just scream "CROWD!" or "EMPTY!" but wouldn't tell you if it was a packed concert or a sparse gathering.
  • Worse, to get the crowd count right, you had to know a secret "equilibrium map" of the stadium beforehand. If you didn't have that map, your crowd count was wrong.

2. The New Tool: The "Soft Focus" Lens (Mollification)

The authors of this paper say, "Let's fix the flashlight." They introduce a Mollifier.

Think of a mollifier as a soft-focus lens or a smudge filter. Instead of checking a single, tiny point (which is noisy and rigid), the new tool looks at a small neighborhood around that point and averages everything out.

• The Analogy: Imagine trying to read a sign in a foggy room.
  • The old tool tried to read a single letter at a time. If the fog was thick, it couldn't tell if the letter was there or not.
  • The new tool (Mollified CD) blurs the image slightly. It looks at a small patch of the sign. If the patch is full of ink, it says "Text here." If it's blank, it says "No text." But because it's looking at a patch, it can also tell you how dark the ink is.

3. What This New Tool Achieves

The paper proves that by using this "soft focus" approach, they get two superpowers:

A. The "Goldilocks" Zone (Improved Dichotomy)

The old tool had a harsh "on/off" switch. The new tool is more nuanced:

• Inside the data: The signal stays stable and bounded. It doesn't explode. It tells you, "Yes, we are inside the data, and here is the density."
• Outside the data: The signal still grows exponentially fast, acting like a siren to warn you, "You are far away from the data!"
• Why it matters: This allows the tool to perfectly distinguish between "inside" and "outside" without getting confused by noise.

B. The "Crowd Counter" (Density Recovery)

This is the big win. Because the tool uses a "soft focus" (mollifier), it can now estimate how dense the data is without needing that secret "equilibrium map" mentioned earlier.

• The Result: If you have enough data points, this tool can reconstruct the shape of the crowd (the density function) with high precision.
• The Math Magic: The authors show that if you balance the "softness" of the lens (how big the neighborhood is) with the "power" of the flashlight (how complex the math is), you get the fastest possible speed of convergence. It's like tuning a radio: if you tune the frequency (the parameters) just right, the static disappears, and the music (the true density) comes through clearly.

4. Two Special Cases

The paper tests this on two different "stages":

1. Flat Ground (Euclidean Space): Imagine data scattered on a table. They use standard "local" smudges (like a Gaussian blur) to find the density.
2. The Ball (The Sphere): Imagine data scattered on the surface of a globe (like weather patterns on Earth). Here, they invent a special type of "algebraic smudge" that respects the curve of the sphere. They prove that on a sphere, this new method is even faster and more accurate than previous methods.

Summary: Why Should You Care?

In the real world, data is messy. We often have a bunch of points and want to know:

• "What is the shape of this data?" (Support Estimation)
• "Where are the hotspots?" (Density Estimation)

This paper provides a universal, robust, and mathematically proven recipe to answer these questions. It replaces a rigid, "all-or-nothing" tool with a flexible, "soft-focus" tool that gives you a clear, quantitative picture of your data, even if you don't know the underlying rules of the game beforehand.

In short: They took a blunt instrument, sharpened it with a "smoothing" technique, and showed that it can now map the hidden landscapes of data with incredible precision.

Technical Summary

Here is a detailed technical summary of the paper "Mollified Christoffel-Darboux Kernels and Density Recovery on Varieties".

1. Problem Statement

The paper addresses two fundamental challenges in approximation theory and data science regarding the Christoffel-Darboux (CD) kernel associated with a probability measure $\mu$ on a compact set $X \subset \mathbb{R}^n$:

1. Support Recovery Dichotomy: The classical CD polynomial exhibits a "dichotomy": it grows linearly on the support of $\mu$ and exponentially outside it. While useful for support estimation, this linear growth on the support prevents direct density recovery without normalization by the equilibrium measure of the domain (which is often unknown or difficult to compute).
2. Density Recovery: Recovering the probability density function $f$ (where $d\mu = f d\lambda$) from moment data using the classical CD kernel is problematic because the kernel converges to $f$ multiplied by the density of the equilibrium measure. Explicit knowledge of this equilibrium measure is required for direct recovery, limiting the method's applicability to highly symmetric domains (e.g., balls, boxes).

The authors aim to develop a regularized (mollified) version of the CD kernel that:

• Replaces the linear growth on the support with a uniform bound.
• Enables consistent density recovery without prior knowledge of the equilibrium measure.
• Provides quantitative convergence rates under Sobolev regularity assumptions.

2. Methodology

2.1 Mollified Christoffel-Darboux (MCD) Kernels

The core innovation is the replacement of the point-evaluation functional (used in the classical CD kernel definition) with a smoothing operator (mollifier).

• Varieties with Mollifiers: The authors define a real algebraic variety $Z$ endowed with a family of $L^2$-mollifiers $\{\phi_{z,\epsilon}\}$. These are probability densities that concentrate around a point $z$ as the resolution parameter $\epsilon \to 0$.
• Definition: For a measure $\mu$ and a polynomial subspace $V_d$, the MCD kernel $MCD_{d,\epsilon}^\mu(x, y)$ is defined via the inner product of Riesz representatives $\phi_x, \phi_y \in V_d$. These representatives satisfy:
  $$\int_A \phi_{z,\epsilon}(y) p(y) d\lambda(y) = \int_X \phi_z(y) p(y) d\mu(y) \quad \forall p \in V_d$$
  where $A$ is a Borel set (either the whole variety $Z$ for support detection or the support $X$ for density estimation).
• Computation: The kernel can be computed purely from the moment matrix of $\mu$ using linear algebra, avoiding the need to compute orthogonal polynomials explicitly.

2.2 Two Main Applications

The paper analyzes two specific instances of the MCD kernel:

1. Support Locator ($S_{MCD}$): Uses $A=Z$. It detects the support of $\mu$ by exploiting the dichotomy between bounded values inside the support and exponential growth outside.
2. Density Estimator ($D_{MCD}$): Uses $A=X$. It estimates the density $f$ by normalizing the diagonal of the kernel.

3. Key Contributions and Theoretical Results

3.1 Improved Dichotomy Property (Theorem 1)

The authors prove that for mollifiers with local support:

• Outside Support: If $z \notin X$, the MCD polynomial grows exponentially with the degree $d$.
• Inside Support: If $z$ is in the interior of $X$, the MCD polynomial is uniformly bounded with respect to $d$ (unlike the linear growth of the classical kernel).
• Significance: This property allows for robust support detection and eliminates the need for the equilibrium measure in the density recovery process.

3.2 Density Recovery Consistency (Theorem 2)

The paper establishes that the normalized MCD kernel converges pointwise to the reciprocal of the density:
$$\lim_{\epsilon \to 0} \lim_{d \to \infty} \frac{D_{MCD}^\mu(x, x)}{\|\phi_{x,\epsilon}\|_{L^2(\lambda)}^2} = \frac{1}{f(x)}$$
for $x$ in the interior of the support. This holds without requiring knowledge of the equilibrium measure.

3.3 Quantitative Convergence Rates

The authors derive explicit error bounds by balancing approximation error (due to mollification $\epsilon$) and projection error (due to finite degree $d$).

• Case A: Euclidean Space (Local Support Mollifiers)

  • Assumptions: $f \in C^2(X)$ and $f, 1/f \in H^s(X)$ (Sobolev regularity).
  • Result: By choosing $\epsilon \sim d^{-k/(k+1)}$, the convergence rate is:
    $$| \hat{g}_{d,\epsilon}(x) - g(x) | = O(d^{-2k/(k+1)})$$
    where $k$ is an integer related to the Sobolev smoothness. This improves upon previous estimates.

• Case B: The Unit Sphere (Algebraic Mollifiers)

  • Method: Uses zonal polynomials (constructed from Gegenbauer polynomials) as algebraic mollifiers.
  • Result: By optimizing the mollifier degree $k$ as a function of the polynomial degree $d$ (specifically $k \approx d^{4/3}$), the authors achieve:
    • $O(d^{-2/3})$ for $C^1$ densities.
    • $O(d^{-4/3})$ for $C^2$ densities.
  • Significance: This rate strictly improves upon the best-known rates for density estimation on spheres (previously $O(d^{-1})$ or similar).

4. Numerical Experiments

The paper validates the theoretical findings on the 2-sphere ($S^2$) using a mixture of von Mises-Fisher distributions.

• Setup: The authors construct an orthonormal basis of spherical harmonics and compute the MCD kernel using the moment matrix.
• Observation: The numerical experiments confirm the theoretical error decomposition. The $L^2$ error decay follows the predicted $O(d^{-4/3})$ rate for $C^2$ densities, demonstrating the practical efficacy of the algebraic mollifiers.

5. Significance and Impact

• Generalization: The framework generalizes Lasserre's modified Christoffel function to arbitrary algebraic varieties and general mollifier families.
• Removal of Equilibrium Measure Dependency: It solves a long-standing limitation in CD-based density estimation, making the method applicable to arbitrary compact domains where the equilibrium measure is unknown.
• Optimal Rates: The construction of algebraic mollifiers on spheres leads to convergence rates that surpass existing constructive approximation results.
• Computational Feasibility: The method relies on linear algebra (moment matrices), making it compatible with the Moment-SOS hierarchy and scalable to high-dimensional problems where explicit orthogonal polynomials are hard to compute.

In summary, this paper provides a rigorous, systematic, and computationally tractable framework for recovering probability densities and supports from moment data, offering improved theoretical guarantees and practical performance over classical methods.