📈 economics

Sharp Minimax Theory for Randomized Experiments

This paper establishes that for binary potential outcomes in finite population randomized experiments, the minimax-optimal strategy involves Bernoulli randomization paired with a nonlinear shrinkage estimator, achieving a second-order risk expansion of $n^{-1} - Cn^{-4/3}$ and demonstrating that standard procedures like complete randomization are suboptimal beyond the first order.

Original authors: Timothy Sudijono, Edgar Dobriban, Eric Tchetgen Tchetgen

Published 2026-08-17

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Timothy Sudijono, Edgar Dobriban, Eric Tchetgen Tchetgen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to figure out if a new magic potion actually works. You have a group of people, and you want to know: does the potion make them faster, smarter, or happier? To find out, you can't just give the potion to everyone; you have to split the group into two teams. One team gets the potion (the "treatment"), and the other gets a sugar pill (the "control"). This is the heart of a randomized experiment.

The tricky part is that you don't know who is naturally fast or naturally slow before you start. If you accidentally give the potion to all the naturally fast people, you might think the potion works when it's actually just their natural talent. To fix this, you use randomization—like flipping a coin for every person to decide which team they join. This ensures the teams are fair.

But here is the big question: once you flip the coins and collect the data, how do you crunch the numbers to get the best possible answer? Scientists have been using standard ways to flip the coins and standard ways to do the math for decades. But what if there is a smarter, more secret way to do both? What if the "standard" way is just "okay," but there is a "perfect" way that gets closer to the truth, especially when you don't have a huge crowd of people to test on? This paper dives into that exact mystery, asking: What is the absolute best possible strategy for designing an experiment and analyzing the results, no matter how tricky the situation gets?

The Great Experiment Heist

In the world of science, researchers often play a game of "Minimax." Imagine you are playing a game against a sneaky opponent who controls the hidden traits of your test subjects. This opponent wants to make your experiment look as bad as possible. Your goal is to choose a design (how you flip the coins) and an estimator (how you do the math) that keeps your error as small as possible, even in the worst-case scenario. This smallest possible "worst-case error" is called the minimax risk.

For a long time, scientists thought the standard way of doing things was pretty much unbeatable. The usual method involves Complete Randomization (like drawing names from a hat to ensure exactly half the people get the potion) and the Difference-in-Means estimator (simply subtracting the average result of the control group from the treatment group). It's the "textbook" approach, taught in every class.

But this paper, by Timothy Sudijono, Edgar Dobriban, and Eric Tchetgen Tchetgen, pulls back the curtain to show that the textbook approach is only partially right. It turns out that while the standard method is good enough for very large crowds, it starts to stumble when the crowd is small or medium-sized (like the 100 people often found in real-world medical trials).

The Secret Weapon: Shrinkage and the Airy Function

The authors discovered that the true "champion" of experiments isn't the standard method. Instead, the winning strategy is a two-part combo:

Bernoulli Randomization: Instead of forcing exactly half the people to get the potion, you flip a coin for each person individually. It's like giving everyone their own private coin flip.
Nonlinear Shrinkage: This is the real magic trick. When you calculate the result, you don't just take the raw difference. You "shrink" the result toward zero. Imagine you are guessing the weight of a mystery box. If your scale is a bit wobbly, and you guess 500 pounds, but you know the box is probably light, you might adjust your guess to 450 pounds to be safer. The paper shows that mathematically "shrinking" the estimate toward zero (specifically using a complex curve related to something called the Airy function) protects you from the sneaky opponent's worst tricks.

The authors proved that this new method, which they call (BRE, Opt), is the absolute best possible strategy. It achieves a level of precision that the standard methods cannot match.

The "Airy" Connection and the Second-Order Secret

Here is where the story gets delightfully weird. The authors didn't just find a better method; they calculated exactly how much better it is using a second-order expansion. In plain English, they looked at the error formula and found a tiny, hidden term that everyone else had ignored.

The standard methods have an error that shrinks like $1/n$ (where $n$ is the number of people). The new method also shrinks like $1/n$ , but it has a secret bonus: it subtracts a tiny bit more error, specifically a term involving $n^{-4/3}$ .

Why is this cool? Because this tiny extra bit of precision is connected to the Airy function, a special mathematical curve that usually describes how light bends or how electrons behave in quantum physics. Finding this curve in the world of medical experiments is like finding a recipe for a soufflé inside a physics textbook. It's a completely new discovery in this field. The authors show that the constant $C$ in their formula is related to the largest negative zero of the derivative of this Airy function, which equals approximately 1.617.

The Reality Check: What Works and What Doesn't

The paper is very clear about what it rules out. It proves that the standard "Difference-in-Means" method paired with "Complete Randomization" is not the best possible choice. It is only "minimax optimal up to the first order." This means that if you look at the big picture, it's fine, but if you zoom in on the details (especially with sample sizes around 100), it is significantly worse than the new method.

In fact, for a sample size of 100, the standard method has a maximum risk that is about 40% larger than the optimal method. That's a huge difference in the world of science! If you are running a trial with 100 people, sticking to the old textbook method leaves you with a much fuzzier answer than you need to have.

However, the authors are careful not to say the old methods are "useless." They show that the standard methods are admissible, meaning there is no other method that is strictly better in every single possible scenario. The new method is better in the worst-case scenarios, but the old methods might perform slightly better in some very specific, unlikely situations. It's a trade-off.

The Verdict

This paper doesn't just suggest a new way to do things; it mathematically proves that a specific combination of Bernoulli Randomization and a nonlinear shrinkage estimator is the theoretical gold standard for finite population experiments.

While the "perfect" method involves complex math that might be hard to calculate for massive datasets, the authors point out that for small-to-medium trials (like many medical studies with fewer than 100 participants), this method is computationally feasible and offers a massive improvement in accuracy.

So, the next time you hear about a scientific study, remember: the way the researchers flipped their coins and the way they did their math matters more than you think. There is a hidden, Airy-function-powered "super-method" waiting to be used, one that shrinks errors and outsmarts the worst-case scenarios, proving that even in the rigid world of statistics, there is always room for a little bit of mathematical magic.

Technical Summary: Sharp Minimax Theory for Randomized Experiments

Problem Statement
This paper addresses the fundamental problem of estimating the sample average treatment effect (SATE) in finite population randomized experiments under the potential outcomes framework. The authors seek to characterize the minimax risk when optimizing jointly over the space of all feasible experimental designs (treatment assignment mechanisms) and estimators. Unlike prior work that often fixes the estimator (e.g., difference-in-means) or restricts the design, this study considers the unrestricted joint selection of design and estimator to determine the theoretical lower bound on estimation error (mean squared error) averaged over the design randomness. The analysis focuses on binary potential outcomes, with extensions to bounded outcomes.

Methodology and Theoretical Framework
The core methodological contribution is a reduction of the complex joint optimization problem (over $2^{2n}$ potential outcome configurations and all possible designs) to a simplified statistical model with fewer parameters.

Reduction to a Two-Parameter Model:
The authors demonstrate that the minimax risk for the original problem is equivalent to the minimax risk $\rho^*_n$ of a reduced model. In this reduced model, the unknown parameters are non-negative integers $p, q, r$ such that $p+q+r=n$ , representing the counts of units with individual treatment effects (ITE) of $+1$ , $-1$, and $0$, respectively. The observed data is $X \sim p + \text{Bin}(r, 1/2)$ , and the goal is to estimate the parameter $(p-q)/n$ .
- This equivalence is established via a specific transformation of the observed data $Y_i$ and treatment indicators $A_i$ into variables $S_i = A_i Y_i(1) + (1-A_i)(1-Y_i(0))$ .
- Under a specific "least-favorable" prior on the potential outcomes, the transformed variables $S_i$ become independent of the design $A$ , allowing the problem to be mapped to the reduced binomial experiment.
Second-Order Asymptotic Analysis:
To derive sharp constants and higher-order terms, the authors employ a proof strategy inspired by Levit's work on bounded normal means. They reparametrize the reduced model using $\theta = p-q$ and $r$ , observing $X = \theta + \sum_{i=1}^r \epsilon_i$ where $\epsilon_i$ are Rademacher variables.
- They propose a candidate estimator of the form $\delta_{n,h}(X) = X/n - n^{-\alpha}h(X/n^\alpha)$ .
- By analyzing the risk expansion, they determine that a second-order improvement over the standard $1/n$ rate is only possible when $\alpha = 2/3$ .
- The optimization of the function $h$ leads to a variational problem involving the differential operator $-4 \frac{d^2}{dx^2} + |x|$ . The solution to this problem is governed by the Airy function, specifically its derivative's largest negative zero.
Admissibility and Dominance Analysis:
The paper evaluates standard procedures—Complete Randomized Experiments (CRE) and Bernoulli Randomized Experiments (BRE) paired with Difference-in-Means (DIM), Horvitz-Thompson (HT), and centered HT estimators—against the derived minimax lower bound. Admissibility is assessed by checking if any procedure dominates another in terms of risk across all potential outcome configurations.

Key Results

Sharp Minimax Risk Expansion:
The minimax risk $\rho^*_n$ admits the following second-order expansion:
$\rho^*_n = \frac{1}{n} - \frac{C_A}{n^{4/3}} + o(n^{-4/3})$
where $C_A = -4^{1/3} a'_1 \approx 1.617$ , and $a'_1$ is the largest negative zero of the derivative of the Airy function. This $n^{-4/3}$ rate and the connection to the Airy function are identified as novel contributions to the literature on minimax risk expansions.
Optimal Procedure:
The minimax risk is attained by a specific procedure: Bernoulli randomization with probability $1/2$ (BRE) combined with a nonlinear shrinkage estimator. The optimal estimator is the posterior mean of the SATE under a least-favorable prior, which corresponds to a nonlinear function applied to the transformed data $\sum S_i$ . This estimator shrinks the unbiased estimator towards zero.
Sub-optimality of Standard Procedures:
- Standard procedures such as (CRE, DIM) and (BRE, DIM) are shown to be minimax optimal only up to the first order ( $1/n$ ).
- For moderate sample sizes (e.g., $n=100$ ), the maximum risk of (CRE, DIM) is approximately 40% larger than the theoretical optimum $\rho^*_n$ .
- The (BRE, HT) estimator is shown to be inadmissible and strictly dominated by (CRE, DIM).
Admissibility:
- The optimal procedure (BRE, Opt) is admissible.
- The standard procedure (CRE, DIM) is also admissible and dominates both (BRE, DIM) and (BRE, HT).
- Neither (BRE, Opt) nor (CRE, DIM) dominates the other; they are incomparable in terms of admissibility, though (BRE, Opt) offers a lower worst-case risk.

Significance and Claims
The paper claims to provide a precise characterization of the minimax risk for finite population randomized experiments under no interference, filling a gap in the joint optimization of design and estimator.

Theoretical Benchmark: The derived expansion $\rho^*_n$ serves as a theoretical benchmark for robustness and efficiency, distinguishing between first-order and second-order optimality.
Novelty of the Second-Order Term: The appearance of the $n^{-4/3}$ term and its connection to the Airy function is highlighted as a nonstandard and new result in this literature, contrasting with the $n^{-3/2}$ rates found in survey sampling or the constant variance assumptions in bounded normal mean problems.
Practical Implications: While acknowledging that minimaxity is a conservative criterion, the authors suggest that for small to moderately sized trials (e.g., $n \le 100$ ), the nonlinear shrinkage estimator (BRE, Opt) offers substantial reductions in maximum risk compared to standard methods, particularly when the true treatment effect is small. However, they note that (CRE, DIM) remains a valid, admissible, and unbiased choice with well-developed inferential theory, justifying its continued use in practice.

The authors explicitly state that their results are intended as a theoretical benchmark and that the practical development of (BRE, Opt) for small trials requires further research, particularly regarding inference and the incorporation of covariates.

The Great Experiment Heist

The Secret Weapon: Shrinkage and the Airy Function

The "Airy" Connection and the Second-Order Secret

The Reality Check: What Works and What Doesn't

The Verdict

More like this