Technical Summary: A Unified Framework for Rerandomization using Quadratic Forms

Problem Statement

In the design of randomized experiments, ensuring covariate balance between treatment and control groups is critical to reducing standard errors and increasing statistical power. While blocking is effective for categorical variables, it struggles with high-dimensional continuous covariates. Rerandomization addresses this by randomizing treatment assignments until a prespecified level of covariate balance is achieved.

Historically, most rerandomization methods rely on the Mahalanobis distance (Morgan and Rubin, 2012), which uses the inverse covariance matrix ($\Sigma^{-1}$) as a weighting matrix. However, in high-dimensional settings, Mahalanobis rerandomization applies an equal percentage of variance reduction to all covariates. As dimensionality ($d$) increases, the variance reduction term approaches 1, rendering the method ineffective. While extensions like Ridge Rerandomization (Branson and Shao, 2021) and PCA Rerandomization (Zhang et al., 2024) attempt to mitigate this by prioritizing specific eigenvectors, they require tuning parameters or arbitrary cutoffs. Furthermore, existing optimality results often assume knowledge of the relationship between covariates and outcomes (e.g., Liu et al., 2025), which is rarely available prior to an experiment.

Methodology

The authors propose a unified framework where rerandomization is defined by a general quadratic form balance metric:
$$Q_A(\sqrt{n}\hat{\tau}_X) = (\sqrt{n}\hat{\tau}_X)^T A (\sqrt{n}\hat{\tau}_X)$$
where $\hat{\tau}_X$ is the vector of covariate mean differences, and $A$ is a fixed, non-zero, positive semi-definite matrix ($A \in S^d_+$).

The paper derives general theoretical results for any such scheme under the potential outcomes framework, conditioning on fixed covariates $X$ and potential outcomes. The analysis relies on two conditions:

1. Condition 1: Finite-population variances and covariances of outcomes and covariates are well-defined and have limiting values.
2. Condition 2: The rerandomization criterion is symmetric in treatment assignment and places non-zero probability on some assignments.

The authors utilize the eigenstructure of the covariance matrix of covariate mean differences ($\Sigma$) and the matrix $\Sigma^{1/2}A\Sigma^{1/2}$ to characterize the asymptotic distribution of the rerandomized estimator.

Key Contributions

1. General Theoretical Results

The authors derive the asymptotic covariance of covariate mean differences under any quadratic form rerandomization (Theorem 1). They show that the covariance matrix is rotated and scaled by variance reduction factors $\nu_{j,\eta}$, which depend on the eigenvalues of $\Sigma^{1/2}A\Sigma^{1/2}$. This framework allows for the rapid re-derivation of properties for existing methods (Mahalanobis, Ridge, PCA) and the analysis of new schemes.

2. Optimality for Covariate Balance (Without Outcome Information)

The paper establishes optimal choices for $A$ based solely on covariate properties, without requiring knowledge of the outcome relationship ($\beta$):

• Minimizing Frobenius Norm: The Euclidean distance ($A = I_d$) minimizes the Frobenius norm of the covariance matrix of covariate mean differences (Theorem 2). This ensures all directions of variation have equal magnitude after rerandomization.
• Maximizing Total Variance Reduction: The Mahalanobis distance ($A = \Sigma^{-1}$) maximizes the total variance reduction across all principal components (Theorem 3).

3. Optimality for Estimating the Average Treatment Effect (ATE)

The authors link the choice of $A$ to the variance of the difference-in-means estimator ($\hat{\tau}$).

• Oracle Optimality: If the relationship between covariates and outcomes ($\beta$) were known, the optimal matrix would be $A^* = \beta\beta^T$ (Theorem 5). This weights covariate mean differences by their predictive power for the outcome.
• Minimax Optimality: Since $\beta$ is typically unknown, the authors prove that Euclidean Rerandomization ($A = I_d$) is minimax optimal (Theorem 6). In the sense that the precision of the estimator is never "too far" from the optimal choice, regardless of the true relationship between covariates and outcomes. This makes Euclidean Rerandomization a robust "safe" choice when outcome information is unavailable.

4. Principal Component Selection Rules

The paper extends the framework to selecting the top $k$ principal components. It derives a cost-benefit decision rule (Equation 8) for choosing $k$ when $\beta$ is unknown, which balances the variance reduction gained on top components against the loss of reduction on discarded components. This is termed the Weighted Eigenvalue rule.

Results

Simulation Study

The authors conducted simulations across varying dimensions ($d$) and eigenvalue distributions (sparse, average, uniform) with different relationships between principal components and outcomes:

• Sparse Eigenvalues: When eigenvalues are sparse and outcomes depend on bottom principal components, Mahalanobis rerandomization outperforms others. However, if outcomes depend on top components, Mahalanobis performs poorly.
• Robustness: Euclidean Rerandomization consistently provided the second-best variance reduction across all settings and was the best at constraining the covariance matrix. It demonstrated superior robustness compared to Mahalanobis, which suffered significantly in settings where its assumptions (equal weighting) were misaligned with the outcome structure.
• Squared Euclidean: Setting $A = \Sigma$ (Squared Euclidean) upweights top eigenvalues and performed well when top components were predictive, but lacked the robustness of standard Euclidean rerandomization.

Real Data Application

Using a randomized controlled trial of mentoring for German adolescents ($n=308, d=21$), the authors constructed synthetic non-linear and non-additive potential outcomes.

• Findings: The eigenvalues of the covariates were relatively uniform. In this setting, all methods performed similarly, but Euclidean Rerandomization yielded the lowest variance (0.565 relative to complete randomization), followed closely by Squared Euclidean (0.576).
• Implication: The results align with the theoretical claim that Euclidean rerandomization is minimax optimal and robust, particularly when the eigenstructure does not strongly favor a specific subset of components.

Significance and Claims

The paper claims to provide a general theoretical framework that unifies existing rerandomization methods under the umbrella of quadratic forms. Its primary significance lies in:

1. Practical Guidance: It offers clear, theoretically grounded guidance for practitioners on how to choose the matrix $A$ when outcome information is unavailable. Specifically, it advocates for Euclidean Rerandomization as a minimax optimal strategy that avoids the pitfalls of high-dimensional Mahalanobis distance while maintaining robustness.
2. Bridging Theory and Practice: It fills a gap in the literature by deriving optimality results that do not rely on prior knowledge of the outcome-covariate relationship, a common limitation in previous works (e.g., Liu et al., 2025).
3. Methodological Expansion: It demonstrates that the choice of $A$ dictates the eigenstructure of the resulting covariance matrix, allowing researchers to tailor rerandomization strategies (e.g., via Squared Euclidean or PCA rules) based on the specific eigenstructure of their data and prior beliefs about outcome relationships.

The authors remain modest, noting that their results apply specifically to rerandomization schemes expressible as quadratic forms and do not necessarily preclude the superiority of non-quadratic design strategies (e.g., kernel allocation). They suggest future work could extend these frameworks to multi-armed experiments and sequential designs.