Thin Sets Are Not Equally Thin: Minimax Learning of Submanifold Integrals

This paper establishes a unified theory showing that the minimax optimal estimation rate for functionals identified by "thin sets" (submanifolds of dimension $m$ in a $d$-dimensional space) depends critically on the intrinsic dimensionality, specifically achieving a rate of $n^{-\frac{s}{2s+d-m}}$ for nonparametric functions with smoothness $s$, and provides valid inference procedures via sieve Riesz representation and Sobol points.

The Gist

Here is an explanation of the paper "Thin Sets Are Not Equally Thin" using simple language and creative analogies.

The Big Idea: Not All "Thin" Things Are Created Equal

Imagine you are a detective trying to solve a mystery. Usually, you have a huge map of a city (the data) and you are looking for a specific suspect. If the suspect is hiding in a big, open park, it's easy to find them. But what if the suspect is hiding on a single, invisible wire strung across the city? Or perhaps they are hiding on a flat sheet of paper floating in 3D space?

In economics, many important questions are like this. We want to know things that only exist on these "invisible wires" or "floating sheets." In math, these are called thin sets (submanifolds). They have zero volume in the big world, but they hold the key to the answer.

For a long time, economists thought: "If it's a thin set, it's hard to find. It's a needle in a haystack."

This paper says: "Wait a minute. Not all needles are the same."

Some needles are just a single point (very hard to find). Others are a long wire (easier). Others are a flat sheet (even easier). The authors show that the shape and dimension of this "thin set" changes exactly how fast we can learn the answer.

---

The Analogy: The "Shadow" Game

Imagine you are trying to guess the shape of a 3D object (like a sculpture) by looking at its shadow on a wall.

1. The Full Room (The Hard Way): If you try to guess the whole sculpture just by looking at random points in the 3D room, you need a lot of data. It's slow and messy.
2. The Shadow (The Thin Set): Now, imagine the "answer" you are looking for is actually painted on the shadow on the wall.
   • If the shadow is just a dot (0-dimensional), you still need a lot of data to pinpoint it exactly.
   • If the shadow is a line (1-dimensional), you have a path to follow. It's easier.
   • If the shadow is a surface (2-dimensional), you have a whole area to work with. It's much easier.

The Paper's Discovery: The authors proved that the speed at which you can learn the answer depends on the dimension of that shadow.

• If the "thin set" is a line, you learn at a certain speed.
• If it's a surface, you learn faster.
• The formula they found is like a "speed limit" for learning. It tells you the absolute fastest possible speed anyone could ever achieve, no matter how smart their computer algorithm is.

Why Does This Matter? (The "Policy" Example)

Let's use a real-world example: Job Training Programs.

Imagine you want to know: "What is the total benefit of a job training program for everyone who is 'on the fence' about taking it?"

• People who definitely want the job are already in.
• People who definitely don't want it won't join.
• The "magic" happens with the people who are indifferent. They are the ones where the decision is a toss-up.

Mathematically, these "indifferent" people form a thin set (a boundary line or surface) inside the big group of all people.

• Old View: "Oh, we are looking at a tiny group. We can't get a good answer. It's too hard."
• New View (This Paper): "Ah, but that group forms a surface. Because it's a surface and not just a dot, we can get a very precise answer, and we know exactly how much data we need to get there."

This allows economists to build better confidence intervals (like saying, "We are 95% sure the benefit is between $X and$Y") instead of just guessing.

The "Magic Tool": Sieve Estimators

How do they actually find these answers? They use a tool called Sieve Estimators.

Think of a Sieve like a kitchen strainer or a colander.

• You have a big pot of soup (your messy data).
• You want to find the specific ingredients (the pattern).
• You use a sieve with holes of a certain size.
  • If the holes are too big, you miss the small ingredients (too much error).
  • If the holes are too small, the soup gets stuck and you can't get anything out (too much noise).

The authors figured out the perfect hole size for different types of "thin sets."

• For a "dot" thin set, you need a very fine sieve.
• For a "surface" thin set, you can use a slightly coarser sieve and still get a perfect result.

They also invented a way to fix the "bias" (the slight error that happens when the sieve isn't perfect) by using a clever trick called Split-Sampling (dividing the data into two groups to check each other) or Leave-One-Out (checking the data by pretending one person isn't there).

The Bottom Line

1. Thin sets are everywhere: From the edge of a decision to the boundary of a policy, important economic answers often hide on these "thin" boundaries.
2. They aren't all equal: A line is easier to study than a point; a surface is easier than a line. The paper gives us the exact math to measure this difficulty.
3. We can do it: The authors didn't just say "it's possible." They built the specific tools (estimators) to do it and proved they are the fastest possible tools anyone could ever invent.
4. Real-world impact: This helps policymakers make better decisions with less data, knowing exactly how reliable their answers are.

In short: The paper teaches us that even when the answer is hidden on a "thin" slice of reality, if we understand the shape of that slice, we can find the answer faster and more accurately than we ever thought possible.

Technical Summary

Here is a detailed technical summary of the paper "Thin Sets Are Not Equally Thin: Minimax Learning of Submanifold Integrals" by Xiaohong Chen and Wayne Yuan Gao.

1. Problem Statement

The paper addresses the statistical challenge of estimating and making inference on functionals defined on "thin sets" (submanifolds with Lebesgue measure zero) within a higher-dimensional ambient space.

• Context: Many economic parameters (e.g., maximum score estimators, optimal treatment rules, marginal treatment effects) are identified by information concentrated on lower-dimensional manifolds (e.g., level sets, boundaries, or contours) where the covariate density is zero in the ambient space but positive on the manifold.
• The Gap: Existing literature often treats all "thin-set" identified parameters as irregular (non-$n^{-1/2}$ estimable) without distinguishing the impact of the manifold's intrinsic dimensionality. The authors argue that thin sets are not "equally thin"; the convergence rate depends critically on the intrinsic dimension $m$ of the submanifold relative to the ambient dimension $d$.
• Objective: To establish a unified theory for the minimax optimal rates of estimation and asymptotic inference for general integral functionals on $m$-dimensional submanifolds ($0 \le m < d$) for unknown functions$h_0$ (regression, density, or NPIV).

2. Methodology

The authors employ a combination of differential geometry, geometric measure theory, and sieve estimation techniques.

A. Mathematical Framework

• Submanifold Definition: The set of interest is $M = \{x \in X : g(x) = 0\}$, an $m$-dimensional submanifold in $\mathbb{R}^d$.
• Functionals: The paper studies:
  1. Linear Integrals: $L(h_0) = \int_M h_0(x) w(x) dH_m(x)$.
  2. Nonlinear Integrals: $\Gamma(h_0) = \int_M \phi(h_0(x), x) w(x) dH_m(x)$.
  3. Upper Contour Integrals: $V(h_0) = \int_{\{x: h_0(x) \ge 0\}} w(x) dx$, whose pathwise derivative is a submanifold integral over the level set $\{h_0(x)=0\}$.
• Hausdorff Measure: The integrals are defined with respect to the $m$-dimensional Hausdorff measure $H_m$, which generalizes Lebesgue measure to lower-dimensional sets.
• Decomposition: A key technical tool is the decomposition of the Hausdorff integral into a finite sum of standard Lebesgue integrals on $\mathbb{R}^m$ using partition of unity and local coordinate charts (implicit function theorem). This allows the application of standard nonparametric estimation theory to the "reduced" dimension.

B. Estimation Strategy

• Sieve Estimation: The unknown function $h_0$ is estimated using a sieve (e.g., B-splines or wavelets) in the full $d$-dimensional space.
• Plug-in vs. Debiased Estimators:
  • Plug-in: Directly substitutes the sieve estimator $\hat{h}$ into the functional.
  • Debiased (Split-Sample & Leave-One-Out): For nonlinear functionals, the authors construct estimators that subtract the leading bias terms (quadratic remainders) from the Taylor expansion. This is crucial for achieving optimal rates under weaker smoothness conditions.
• Inference: The paper utilizes Sieve Riesz Representers. Since the functionals are irregular (no $L^2$ Riesz representer exists), the authors construct finite-dimensional sieve Riesz representers $v^*_{K_n}$ which are well-defined and computable. These are used to derive asymptotic normality and construct confidence intervals.

3. Key Contributions

A. Minimax Lower Bounds (The "Not Equally Thin" Result)

The paper establishes that the minimax optimal rate of estimation is $n^{-\frac{s}{2s+d-m}}$, where:

• $s$: Smoothness of the unknown function $h_0$ (Hölder class).
• $d$: Dimension of the ambient space.
• $m$: Intrinsic dimension of the submanifold.
• Significance: This rate interpolates between pointwise estimation ($m=0 \implies n^{-s/(2s+d)}$) and full-dimensional integration ($m=d \implies n^{-1/2}$).
  • The integration over the $m$-dimensional manifold effectively "aggregates out" $m$ dimensions, reducing the effective curse of dimensionality from $d$ to $d-m$.
  • This result generalizes Stone's (1980) pointwise rate and Horowitz's (1993) smoothed maximum score rate ($m=d-1$).

B. Attainability (Optimality)

The authors prove that the lower bound is attainable:

• Linear Functionals: A simple plug-in sieve estimator achieves the optimal rate.
• Nonlinear Functionals:
  • Split-sample and Leave-one-out (LOO) debiased estimators achieve the optimal rate under mild smoothness ($s > m/2$).
  • Plug-in estimators achieve the optimal rate only under stronger smoothness ($s \ge m$).
• NPIV Models: The results are extended to Nonparametric Instrumental Variable (NPIV) settings, showing that the rate depends on the degree of ill-posedness and the codimension $d-m$.

C. Asymptotic Inference

• The paper establishes asymptotic normality for the proposed sieve estimators.
• It derives the growth rate of the norm of the sieve Riesz representer, showing it scales with the codimension ($K^{(d-m)/d}$), which is slower than the full-dimensional case.
• Valid confidence intervals are constructed using sieve student-t statistics and multiplier bootstrap methods (specifically a bootstrap-Lepski procedure for data-driven dimension selection).

4. Key Results

1. Convergence Rate: For a submanifold of dimension $m$ in $\mathbb{R}^d$, the optimal rate is $n^{-s/(2s+d-m)}$.
   • Example: For a level set ($m=d-1$), the rate is $n^{-s/(2s+1)}$, identical to a 1-dimensional nonparametric regression, regardless of the ambient dimension $d$.
2. Dimension Reduction Effect: Integration over a submanifold acts as a dimension reduction mechanism. The "effective" dimension of the estimation problem is $d-m$.
3. Debiasing Necessity: For nonlinear functionals (like quadratic integrals or upper contour sets), simple plug-in estimators suffer from bias that dominates the variance unless the function is very smooth ($s \ge m$). Debiased estimators (Split-sample/LOO) are required to achieve the minimax rate for moderate smoothness.
4. Inference Validity: The proposed sieve t-statistics are asymptotically standard normal, allowing for valid $95%$ confidence intervals even for irregular parameters.

5. Significance and Implications

• Unified Theory: Provides the first unified framework for a broad class of "thin-set" identified parameters, unifying disparate results in econometrics (e.g., maximum score, treatment effects, boundary discontinuities).
• Refined Understanding of Irregularity: Challenges the binary view of parameters being either "regular" ($n^{-1/2}$) or "irregular" (slower). It quantifies how irregular a parameter is based on the geometry of the identifying set.
• Practical Guidance:
  • Offers specific algorithms (sieve estimators with specific dimension choices) for practitioners dealing with level sets or boundaries.
  • Demonstrates that Sobol quasi-random sequences are superior to uniform random sampling for numerically computing these submanifold integrals in simulations.
• Policy Relevance: Directly applicable to estimating welfare functionals, optimal treatment rules, and marginal treatment effects where the decision boundary is unknown and must be estimated nonparametrically.

6. Empirical Validation

Monte Carlo simulations confirm the theoretical findings:

• The RMSE of the estimators shrinks at the predicted minimax rate.
• The realized coverage probabilities of the confidence intervals are close to the nominal 95% level.
• The bias-aware vs. undersmoothing comparison suggests that while undersmoothing (increasing sieve dimension) is robust for both linear and nonlinear functionals, bias-aware methods may be less efficient for nonlinear functionals with substantial bias.

In summary, this paper fundamentally advances the understanding of nonparametric estimation on lower-dimensional structures, proving that the "thinness" of a set is a continuous parameter ($m$) that dictates the statistical difficulty of the problem, and providing optimal, inferentially valid methods to tackle it.