Cross-Fold Moments Overview

• Cross-fold moments are defined as structural and statistical quantities that capture inter-fold dependencies, enabling bias correction and decorrelation in experimental designs.
• In Gaussian process regression, they unify cross-validation and likelihood-based inference by recovering maximum-likelihood estimators while significantly reducing computational cost.
• In curved-crease origami mechanics, cross-fold moments quantify force transmission between creases, dictating folding pathways and lowering energy barriers through synchronous activation.

Cross-fold moments comprise a family of structural, statistical, and mechanical quantities that emerge when systems are organized into multiple distinct folds or partitions—either for cross-validation in machine learning, for bias correction in causal inference, or as physical moments transmitting force between folds in curved-crease origami. Despite disparate applications, all notions of cross-fold moments share the core property of capturing dependencies across partitioned data or structures, enabling decorrelation, feedback, or coupling phenomena that are inaccessible to naive fold-wise or aggregate treatments.

1. Cross-Fold Moments in Causal Inference

Cross-fold moments underpin a suite of estimation procedures for statistical surrogacy and long-term causal effect identification in regimes dominated by many weak randomized experiments. Consider $K$ cells (randomized experiments), within each of which $n$ units are independently assigned to one of $L$ folds, $V_i\sim\text{Unif}\{1,\ldots,L\}$. Observables include the surrogate $S_i\in\mathbb{R}^d$ and the long-term outcome $Y_i\in\mathbb{R}$.

Defining fold averages: $$\bar S_{a,v} = \frac{1}{n/L}\sum_{i:A_i=a,\,V_i=v} S_i, \qquad \bar S_{a,-v} = \frac{1}{n(1-1/L)}\sum_{i:A_i=a,\,V_i\neq v} S_i$$

$$\bar Y_{a,v} = \frac{1}{n/L}\sum_{i:A_i=a,\,V_i=v} Y_i,$$

the cross-fold moment condition for a linear bridge $h(S)=S\beta$ is

$$\sum_{a,v}\bar S_{a,-v}^\top \bigl(\bar Y_{a,v}-\bar S_{a,v}\beta\bigr) = 0,$$

with closed-form solution

$n$0

This estimator addresses the non-vanishing bias endemic to two-stage least squares (2SLS) in the many weak experiment regime by exploiting the independence between first-stage estimation (in the complement fold) and second-stage evaluation (in the evaluation fold) (Bibaut et al., 2023). In the nonparametric case, minimization of the “cross-fold risk”

$n$1

characterizes unbiased bridges.

2. Cross-Fold Residuals and Covariances in Gaussian Process Cross-Validation

Within Gaussian process regression, cross-fold moments arise as joint residuals and their covariances computed after partitioning observed data into $n$2 folds. Letting $n$3, $n$4 (kernel plus noise), and $n$5, the residual vector for fold $n$6 is

$n$7

and stacking all $n$8 folds produces $n$9 where $L$0 is block-diagonal with diagonal blocks $L$1. The cross-fold residual covariance is then

$L$2

This operator encodes the entire covariance structure among all cross-fold residuals. Crucially, in the noiseless case with scale parameter $L$3, accounting for $L$4 in the CV pseudo-likelihood precisely recovers the maximum-likelihood estimator of $L$5, unifying cross-validation-based and likelihood-based inference (Ginsbourger et al., 2021).

3. Cross-Fold Moments in Curved-Crease Origami Mechanics

In the mechanics of multi-crease origami, cross-fold moments refer to mechanical moments transmitted between adjacent or overlapping curved creases within a shell. Letting $L$6 denote the dihedral angle of crease $L$7, the total elastic bending energy becomes

$L$8

where $L$9 is the symmetric coupling matrix and $V_i\sim\text{Unif}\{1,\ldots,L\}$0 are spontaneous curvature contributions. The off-diagonal entries $V_i\sim\text{Unif}\{1,\ldots,L\}$1 quantify the cross-talk (mechanical cross-fold moments) coupling the deformation of crease $V_i\sim\text{Unif}\{1,\ldots,L\}$2 to the fold angle at crease $V_i\sim\text{Unif}\{1,\ldots,L\}$3: $V_i\sim\text{Unif}\{1,\ldots,L\}$4 signaling that crease $V_i\sim\text{Unif}\{1,\ldots,L\}$5 “feels” a bending moment in proportion to $V_i\sim\text{Unif}\{1,\ldots,L\}$6. In multi-crease structures, these cross-fold moments fundamentally determine the folding pathway and energetic landscape, with synchronous activation producing a lower energetic barrier than sequential folding (DeSimone et al., 2024).

4. Bias, Decorrelation, and Consistency via Cross-Fold Structures

Cross-fold procedures guarantee bias correction and unbiasedness in settings where canonical estimators fail. In many weak experiment settings, 2SLS suffers from error-in-variables attenuation bias that does not vanish even with infinite cells if cell sizes are finite: $V_i\sim\text{Unif}\{1,\ldots,L\}$7 By using complement-fold averages to instrument, cross-fold moments decorrelate first-stage and second-stage estimation noise, yielding estimators whose expectation matches the target parameter. In the nonparametric regime, minimization of cross-fold risk $V_i\sim\text{Unif}\{1,\ldots,L\}$8 under weak entropy and bounded range conditions ensures consistency: $V_i\sim\text{Unif}\{1,\ldots,L\}$9 as the number of experiments $S_i\in\mathbb{R}^d$0 (Bibaut et al., 2023).

5. Computational Algorithms and Efficiency Considerations

The cross-fold methodology enables dramatically accelerated algorithms. In Gaussian process cross-validation, naive computation of all $S_i\in\mathbb{R}^d$1-fold residuals requires $S_i\in\mathbb{R}^d$2 operations, whereas the fast cross-fold moment algorithm involves a single $S_i\in\mathbb{R}^d$3 Cholesky factorization and direct computation of $S_i\in\mathbb{R}^d$4 with $S_i\in\mathbb{R}^d$5 cost if folds are moderately sized (Ginsbourger et al., 2021). Empirically, speed-ups can approach orders of magnitude over naive implementations, with accuracy at machine precision levels. In cross-fold causal estimation, fold-averaged statistics can be precomputed or logged, further streamlining practical workflows.

6. Applications and Illustrative Examples

In causal inference, cross-fold moments enable the construction of estimators for long-term effects using weak, proxy, or imperfect surrogates in industrial-scale experimentation. For example, given $S_i\in\mathbb{R}^d$6 experiments with $S_i\in\mathbb{R}^d$7 units each split into $S_i\in\mathbb{R}^d$8 folds, the cross-fold estimator recovers target effects while naive regression is substantially biased toward zero (Bibaut et al., 2023).

In spatial statistics, cross-fold covariances provide diagnostic and model selection advantages—grouping clustered observations within the same fold prevents the overoptimism of leave-one-out approaches and allows for fold schemes matched to data heterogeneity (Ginsbourger et al., 2021).

Within origami mechanics, cross-fold moments dictate fold path cooperativity, with positive coupling terms $S_i\in\mathbb{R}^d$9 lowering energy barriers when folds actuate synchronously. This has implications for the design of morphable structures and the exploitation of distributed actuation.

7. Fundamental Principles and Interdisciplinary Significance

Cross-fold moments generalize the classical notion of moment or residuals to partitioned domains and intertwine structure, decorrelation, and coupling. They provide the mathematical and computational tools necessary for bias correction in statistical estimation, precise assessment of covariances in machine learning diagnostics, and quantification of mechanical couplings in engineered structures. Their unifying feature is the exploitation of foldwise dependencies—whether statistical, algorithmic, or physical—to achieve bias reduction, efficiency, control, or robustness across domains as diverse as causal inference, spatial statistics, and origami-based systems (Ginsbourger et al., 2021, Bibaut et al., 2023, DeSimone et al., 2024).