Automatic Regressor Balancing (ARB) Techniques

• Automatic Regressor Balancing (ARB) is a set of methods that combine regression adjustment with explicit balancing procedures to reduce bias in various estimators.
• It encompasses diverse formulations such as augmented balancing weights, residual balancing for treatment effects, and fairness-aware post-processing.
• Its flexible framework targets covariate or regressor-induced moments, enabling robust bias correction in high-dimensional and nonlinear modeling settings.

Searching arXiv for the cited papers to ground the article. Automatic Regressor Balancing (ARB) denotes a family of balancing-based procedures that couple regression adjustment with explicitly constructed weights, balance equations, or regressor transformations. In the cited literature, the label is used for at least five related constructions: augmented balancing weights written as a single linear plug-in estimator (Bruns-Smith et al., 2023), approximate residual balancing for high-dimensional treatment-effect estimation (Athey et al., 2016), its extension to nonlinear generalized linear models with a second-order correction (Meza, 31 Oct 2025), model-agnostic post-processing for Wasserstein-based fairness metrics (Miroshnikov et al., 2021), and a generator-induced link construction in generalized Riesz regression implemented in genriesz (Kato, 19 Feb 2026). This suggests a common organizing idea—balancing regressors or regressor-induced moments so that plug-in estimators inherit reduced bias—while also indicating that the acronym does not refer to one universally standardized algorithm.

1. Terminological scope and principal variants

The acronym is not used uniformly across the cited literature. Some papers use ARB for causal or semiparametric estimators based on doubly robust or orthogonal-score constructions, whereas another uses it for fairness-oriented post-processing of a fixed predictor, and a further paper uses it as a design principle for generalized Riesz regression.

Usage	Balanced object	Characteristic form
Augmented balancing weights	Target-population feature averages	Single plug-in estimator with coefficients interpolating between a penalized base learner and OLS
Approximate residual balancing	Treated and control covariate moments	Regularized regression adjustment plus residual weighting
Nonlinear residual balancing	First- and second-order GLM moments	Convex weighting program with curvature correction
Fairness ARB	Predictor distributions driving Wasserstein bias	Monotone post-processing of selected regressors
Generalized Riesz regression ARB	Riesz moments in a user-chosen basis	Generator-induced link with exact sample moment matching when $\lambda=0$

A common conceptual thread is that balance is imposed directly rather than solely through explicit propensity-score modeling. A plausible implication is that ARB is better treated as a methodological motif—regression plus balancing—than as a single named estimator.

2. Augmented balancing weights as a linear regression representation

In the formulation developed for augmented balancing weights, also called automatic debiased machine learning (AutoDML), the target is a linear functional of an unknown regression $m(x,z)=E[Y\mid X=x,Z=z]$,

$\psi(m)=E[h(X,Z,m)]=E[\alpha(X,Z)m(X,Z)],$

where $\alpha(X,Z)$ is the Riesz representer. The doubly robust estimator is

$$\hat\psi_{DR}=\frac1n\sum_{i=1}^n\Bigl\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\bigl(Y_i-\hat m(X_i,Z_i)\bigr)\Bigr\},$$

and the ARB specialization chooses a penalized linear outcome learner together with balancing weights that directly target feature balance in a common feature map $\phi(X,Z)$ (Bruns-Smith et al., 2023).

The outcome learner is

$$\hat\beta_{\rm out}=\arg\min_\beta \frac1n\sum_i\bigl(Y_i-\Phi_i^\top\beta\bigr)^2+\lambda\|\beta\|^2 \quad(\text{or }\|\beta\|_1),$$

with $\Phi_i=\phi(X_i,Z_i)$. The balancing weights solve

$$\min_{w\in\mathbb R^n}\Bigl\|\tfrac1n w^\top\Phi-\bar\Phi_q\Bigr\|_*^2+\delta\|w\|_2^2,$$

and the resulting ARB estimator takes the plug-in-plus-residual-weighting form

$$\hat\psi_{ARB}=\bar\Phi_q^\top\hat\beta_{\rm out}+\frac1n\sum_i\hat w_i\bigl(Y_i-\Phi_i^\top\hat\beta_{\rm out}\bigr).$$

The key representation is that $m(x,z)=E[Y\mid X=x,Z=z]$0 is itself a single plug-in estimator: $m(x,z)=E[Y\mid X=x,Z=z]$1 with coordinatewise coefficients

$m(x,z)=E[Y\mid X=x,Z=z]$2

When $m(x,z)=E[Y\mid X=x,Z=z]$3 is diagonal, these $m(x,z)=E[Y\mid X=x,Z=z]$4. In this sense, ARB shrinks each coordinate of the base learner toward OLS, with the shrinkage determined by the balancing regularizer.

The paper derives explicit regularization paths. Under $m(x,z)=E[Y\mid X=x,Z=z]$5 weighting,

$m(x,z)=E[Y\mid X=x,Z=z]$6

so ARB follows the standard ridge path. If the base learner is unregularized OLS, then $m(x,z)=E[Y\mid X=x,Z=z]$7 for any $m(x,z)=E[Y\mid X=x,Z=z]$8. If $m(x,z)=E[Y\mid X=x,Z=z]$9, exact balancing implies $\psi(m)=E[h(X,Z,m)]=E[\alpha(X,Z)m(X,Z)],$0, and the estimator again collapses to OLS. Under $\psi(m)=E[h(X,Z,m)]=E[\alpha(X,Z)m(X,Z)],$1 weighting, the dual problem yields soft-thresholding of the feature shift, and if the base learner is a lasso with support $\psi(m)=E[h(X,Z,m)]=E[\alpha(X,Z)m(X,Z)],$2 while the nonzero weighting coordinates form $\psi(m)=E[h(X,Z,m)]=E[\alpha(X,Z)m(X,Z)],$3, then

$\psi(m)=E[h(X,Z,m)]=E[\alpha(X,Z)m(X,Z)],$4

recovering a double-selection property.

The same paper extends the equivalence to kernel ridge regression. In an RKHS with kernel matrix $\psi(m)=E[h(X,Z,m)]=E[\alpha(X,Z)m(X,Z)],$5, double-ridge ARB is algebraically

$\psi(m)=E[h(X,Z,m)]=E[\alpha(X,Z)m(X,Z)],$6

Under $\psi(m)=E[h(X,Z,m)]=E[\alpha(X,Z)m(X,Z)],$7, one finds $\psi(m)=E[h(X,Z,m)]=E[\alpha(X,Z)m(X,Z)],$8, the undersmoothing regime associated in the paper with $\psi(m)=E[h(X,Z,m)]=E[\alpha(X,Z)m(X,Z)],$9-consistent functional estimation. Practical guidance in that formulation is to cross-validate $\alpha(X,Z)$0 for prediction, often set $\alpha(X,Z)$1, and avoid selecting $\alpha(X,Z)$2 by imbalance alone because that can under-regularize.

3. Approximate residual balancing for high-dimensional linear treatment-effect estimation

Approximate residual balancing was introduced for de-biased inference of average treatment effects in high-dimensional linear models under unconfoundedness and overlap. The setup observes $\alpha(X,Z)$3, $\alpha(X,Z)$4, and

$\alpha(X,Z)$5

with overlap $\alpha(X,Z)$6 and linear outcome models

$\alpha(X,Z)$7

The target is the sample average treatment effect on the treated,

$\alpha(X,Z)$8

with the counterfactual control mean for treated units estimated by combining lasso adjustment and balancing weights (Athey et al., 2016).

The procedure has three parts. First, it computes approximately balancing control weights $\alpha(X,Z)$9 by solving

$$\hat\psi_{DR}=\frac1n\sum_{i=1}^n\Bigl\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\bigl(Y_i-\hat m(X_i,Z_i)\bigr)\Bigr\},$$0

subject to

$$\hat\psi_{DR}=\frac1n\sum_{i=1}^n\Bigl\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\bigl(Y_i-\hat m(X_i,Z_i)\bigr)\Bigr\},$$1

Second, it fits a control-group lasso

$$\hat\psi_{DR}=\frac1n\sum_{i=1}^n\Bigl\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\bigl(Y_i-\hat m(X_i,Z_i)\bigr)\Bigr\},$$2

Third, it forms

$$\hat\psi_{DR}=\frac1n\sum_{i=1}^n\Bigl\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\bigl(Y_i-\hat m(X_i,Z_i)\bigr)\Bigr\},$$3

The purpose of the residual-weighting term is to correct shrinkage bias left by penalized regression when $$\hat\psi_{DR}=\frac1n\sum_{i=1}^n\Bigl\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\bigl(Y_i-\hat m(X_i,Z_i)\bigr)\Bigr\},$$4. The paper emphasizes that, given linearity, it is not necessary to assume that the treatment propensities are estimable or that the average treatment effect is a sparse contrast of the outcome-model parameters; beyond standard assumptions for lasso consistency under $$\hat\psi_{DR}=\frac1n\sum_{i=1}^n\Bigl\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\bigl(Y_i-\hat m(X_i,Z_i)\bigr)\Bigr\},$$5-error, only overlap is additionally required.

Under linearity, unconfoundedness, overlap, sparsity of $$\hat\psi_{DR}=\frac1n\sum_{i=1}^n\Bigl\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\bigl(Y_i-\hat m(X_i,Z_i)\bigr)\Bigr\},$$6, sub-Gaussian design with restricted eigenvalue conditions, and homoskedastic noise or mild heteroskedastic robust variants, the estimator is $$\hat\psi_{DR}=\frac1n\sum_{i=1}^n\Bigl\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\bigl(Y_i-\hat m(X_i,Z_i)\bigr)\Bigr\},$$7-consistent and asymptotically normal: $$\hat\psi_{DR}=\frac1n\sum_{i=1}^n\Bigl\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\bigl(Y_i-\hat m(X_i,Z_i)\bigr)\Bigr\},$$8 The lasso error rate

$$\hat\psi_{DR}=\frac1n\sum_{i=1}^n\Bigl\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\bigl(Y_i-\hat m(X_i,Z_i)\bigr)\Bigr\},$$9

enters the bias bound through

$\phi(X,Z)$0

Variance estimation may use

$\phi(X,Z)$1

along with the treated-sample contribution.

The original paper already sketches a generalized-linear-model extension in which one balances $\phi(X,Z)$2 weighted by $\phi(X,Z)$3. That sketch becomes a full second-order theory in later work.

4. Nonlinear residual balancing in high-dimensional generalized linear models

The nonlinear extension answers an open problem posed by Athey et al. (2018): how to extend approximate residual balancing to high-dimensional settings in which the outcome follows a generalized linear model. The estimand is the average treatment effect on the treated,

$\phi(X,Z)$4

with

$\phi(X,Z)$5

Observed data are i.i.d. $\phi(X,Z)$6, unconfoundedness is assumed, and the link $\phi(X,Z)$7 is three-times continuously differentiable with controlled growth of $\phi(X,Z)$8 and $\phi(X,Z)$9. Sample-splitting or cross-fitting is used: one part estimates $$\hat\beta_{\rm out}=\arg\min_\beta \frac1n\sum_i\bigl(Y_i-\Phi_i^\top\beta\bigr)^2+\lambda\|\beta\|^2 \quad(\text{or }\|\beta\|_1),$$0 by a high-dimensional penalized GLM lasso, and another estimates balancing weights $$\hat\beta_{\rm out}=\arg\min_\beta \frac1n\sum_i\bigl(Y_i-\Phi_i^\top\beta\bigr)^2+\lambda\|\beta\|^2 \quad(\text{or }\|\beta\|_1),$$1 (Meza, 31 Oct 2025).

The estimator for the control-group counterfactual mean is

$$\hat\beta_{\rm out}=\arg\min_\beta \frac1n\sum_i\bigl(Y_i-\Phi_i^\top\beta\bigr)^2+\lambda\|\beta\|^2 \quad(\text{or }\|\beta\|_1),$$2

A Taylor expansion around $$\hat\beta_{\rm out}=\arg\min_\beta \frac1n\sum_i\bigl(Y_i-\Phi_i^\top\beta\bigr)^2+\lambda\|\beta\|^2 \quad(\text{or }\|\beta\|_1),$$3 yields an exact decomposition of the estimation error into a first-order imbalance term, a second-order Taylor remainder, and a variance term: $$\hat\beta_{\rm out}=\arg\min_\beta \frac1n\sum_i\bigl(Y_i-\Phi_i^\top\beta\bigr)^2+\lambda\|\beta\|^2 \quad(\text{or }\|\beta\|_1),$$4 Because the link is nonlinear, first-order balance alone does not eliminate the Taylor remainder. The paper therefore derives a second-order correction targeting the Hessian-weighted moment

$$\hat\beta_{\rm out}=\arg\min_\beta \frac1n\sum_i\bigl(Y_i-\Phi_i^\top\beta\bigr)^2+\lambda\|\beta\|^2 \quad(\text{or }\|\beta\|_1),$$5

Weights are chosen by a convex program that trades off variance against two imbalance penalties. In constrained form, one solves

$$\hat\beta_{\rm out}=\arg\min_\beta \frac1n\sum_i\bigl(Y_i-\Phi_i^\top\beta\bigr)^2+\lambda\|\beta\|^2 \quad(\text{or }\|\beta\|_1),$$6

subject to a first-order balance constraint $$\hat\beta_{\rm out}=\arg\min_\beta \frac1n\sum_i\bigl(Y_i-\Phi_i^\top\beta\bigr)^2+\lambda\|\beta\|^2 \quad(\text{or }\|\beta\|_1),$$7, a second-order balance constraint $$\hat\beta_{\rm out}=\arg\min_\beta \frac1n\sum_i\bigl(Y_i-\Phi_i^\top\beta\bigr)^2+\lambda\|\beta\|^2 \quad(\text{or }\|\beta\|_1),$$8, and

$$\hat\beta_{\rm out}=\arg\min_\beta \frac1n\sum_i\bigl(Y_i-\Phi_i^\top\beta\bigr)^2+\lambda\|\beta\|^2 \quad(\text{or }\|\beta\|_1),$$9

The first constraint controls

$\Phi_i=\phi(X_i,Z_i)$0

while the second controls

$\Phi_i=\phi(X_i,Z_i)$1

Under unconfoundedness, GLM smoothness, sub-Gaussian design, sparsity

$\Phi_i=\phi(X_i,Z_i)$2

dispersion constraints on $\Phi_i=\phi(X_i,Z_i)$3, and the stated balance rates, the dominant term is $\Phi_i=\phi(X_i,Z_i)$4, and the paper establishes

$\Phi_i=\phi(X_i,Z_i)$5

Since $\Phi_i=\phi(X_i,Z_i)$6 and $\Phi_i=\phi(X_i,Z_i)$7 is itself $\Phi_i=\phi(X_i,Z_i)$8-normal, the ATT estimator is overall $\Phi_i=\phi(X_i,Z_i)$9-consistent and asymptotically normal.

Practical details are explicit. The convex program is a quadratic or second-order cone problem with $$\min_{w\in\mathbb R^n}\Bigl\|\tfrac1n w^\top\Phi-\bar\Phi_q\Bigr\|_*^2+\delta\|w\|_2^2,$$0 constraints; CVX, Gurobi, OSQP, MOSEK, and ADMM are listed as viable solvers. The paper also describes a “beta-min” alternative that balances only the support of $$\min_{w\in\mathbb R^n}\Bigl\|\tfrac1n w^\top\Phi-\bar\Phi_q\Bigr\|_*^2+\delta\|w\|_2^2,$$1, thereby avoiding the full second-order constraint at the cost of a strong signal assumption. A plausible implication is that the nonlinear extension preserves the residual-balancing logic of the linear method while replacing ordinary covariate balance with curvature-aware balance.

5. Model-agnostic ARB for Wasserstein-based fairness metrics

In the fairness literature, ARB denotes a post-processing methodology rather than a semiparametric treatment-effect estimator. The setup takes original predictors $$\min_{w\in\mathbb R^n}\Bigl\|\tfrac1n w^\top\Phi-\bar\Phi_q\Bigr\|_*^2+\delta\|w\|_2^2,$$2, a binary protected attribute $$\min_{w\in\mathbb R^n}\Bigl\|\tfrac1n w^\top\Phi-\bar\Phi_q\Bigr\|_*^2+\delta\|w\|_2^2,$$3, a response $$\min_{w\in\mathbb R^n}\Bigl\|\tfrac1n w^\top\Phi-\bar\Phi_q\Bigr\|_*^2+\delta\|w\|_2^2,$$4, and a pre-trained regressor $$\min_{w\in\mathbb R^n}\Bigl\|\tfrac1n w^\top\Phi-\bar\Phi_q\Bigr\|_*^2+\delta\|w\|_2^2,$$5 trained without access to $$\min_{w\in\mathbb R^n}\Bigl\|\tfrac1n w^\top\Phi-\bar\Phi_q\Bigr\|_*^2+\delta\|w\|_2^2,$$6. Regressor-level bias is measured by the $$\min_{w\in\mathbb R^n}\Bigl\|\tfrac1n w^\top\Phi-\bar\Phi_q\Bigr\|_*^2+\delta\|w\|_2^2,$$7-Wasserstein distance

$$\min_{w\in\mathbb R^n}\Bigl\|\tfrac1n w^\top\Phi-\bar\Phi_q\Bigr\|_*^2+\delta\|w\|_2^2,$$8

with a more general $$\min_{w\in\mathbb R^n}\Bigl\|\tfrac1n w^\top\Phi-\bar\Phi_q\Bigr\|_*^2+\delta\|w\|_2^2,$$9 metric available by weighting regions of the feature space. The bias can be decomposed into positive and negative components via the quantile functions $$\hat\psi_{ARB}=\bar\Phi_q^\top\hat\beta_{\rm out}+\frac1n\sum_i\hat w_i\bigl(Y_i-\Phi_i^\top\hat\beta_{\rm out}\bigr).$$0 and a sign $$\hat\psi_{ARB}=\bar\Phi_q^\top\hat\beta_{\rm out}+\frac1n\sum_i\hat w_i\bigl(Y_i-\Phi_i^\top\hat\beta_{\rm out}\bigr).$$1 indicating whether larger scores are favorable (Miroshnikov et al., 2021).

ARB’s goal in this setting is to construct a post-processed regressor $$\hat\psi_{ARB}=\bar\Phi_q^\top\hat\beta_{\rm out}+\frac1n\sum_i\hat w_i\bigl(Y_i-\Phi_i^\top\hat\beta_{\rm out}\bigr).$$2 that lies in a small, continuous family of monotonic transformations of the original $$\hat\psi_{ARB}=\bar\Phi_q^\top\hat\beta_{\rm out}+\frac1n\sum_i\hat w_i\bigl(Y_i-\Phi_i^\top\hat\beta_{\rm out}\bigr).$$3, guarantees a pre-specified bias-performance trade-off, and does not depend on $$\hat\psi_{ARB}=\bar\Phi_q^\top\hat\beta_{\rm out}+\frac1n\sum_i\hat w_i\bigl(Y_i-\Phi_i^\top\hat\beta_{\rm out}\bigr).$$4 at prediction time. Rather than retraining, the method reshapes only the $$\hat\psi_{ARB}=\bar\Phi_q^\top\hat\beta_{\rm out}+\frac1n\sum_i\hat w_i\bigl(Y_i-\Phi_i^\top\hat\beta_{\rm out}\bigr).$$5 most bias-impactful predictors $$\hat\psi_{ARB}=\bar\Phi_q^\top\hat\beta_{\rm out}+\frac1n\sum_i\hat w_i\bigl(Y_i-\Phi_i^\top\hat\beta_{\rm out}\bigr).$$6 by a family of compressive maps

$$\hat\psi_{ARB}=\bar\Phi_q^\top\hat\beta_{\rm out}+\frac1n\sum_i\hat w_i\bigl(Y_i-\Phi_i^\top\hat\beta_{\rm out}\bigr).$$7

The simplest global map is

$$\hat\psi_{ARB}=\bar\Phi_q^\top\hat\beta_{\rm out}+\frac1n\sum_i\hat w_i\bigl(Y_i-\Phi_i^\top\hat\beta_{\rm out}\bigr).$$8

Asymmetric and local variants are also given. The perturbed predictor is passed through the existing regressor,

$$\hat\psi_{ARB}=\bar\Phi_q^\top\hat\beta_{\rm out}+\frac1n\sum_i\hat w_i\bigl(Y_i-\Phi_i^\top\hat\beta_{\rm out}\bigr).$$9

and then calibrated by an isotonic-regression map $m(x,z)=E[Y\mid X=x,Z=z]$00: $m(x,z)=E[Y\mid X=x,Z=z]$01

A dimensionality-reduction stage selects the bias drivers. For each predictor $m(x,z)=E[Y\mid X=x,Z=z]$02, the method computes positive and negative bias explanations $m(x,z)=E[Y\mid X=x,Z=z]$03 and $m(x,z)=E[Y\mid X=x,Z=z]$04 using a single-feature explainer such as PDP, marginal or conditional Shapley, or “individual bias explanations,” and forms sets

$m(x,z)=E[Y\mid X=x,Z=z]$05

Optimization over the bias-performance frontier then proceeds in the low-dimensional parameter space $m(x,z)=E[Y\mid X=x,Z=z]$06 using Gaussian-process Bayesian optimization with expected improvement. For a fairness penalty $m(x,z)=E[Y\mid X=x,Z=z]$07, the scalarized objective is

$m(x,z)=E[Y\mid X=x,Z=z]$08

The paper reports experiments on four synthetic data-generating models with $m(x,z)=E[Y\mid X=x,Z=z]$09 predictors and a logistic link. Asymmetric transforms attain strictly lower $m(x,z)=E[Y\mid X=x,Z=z]$10-bias for the same AUC than global compression in one comparison; local maps focusing on the region of distributional disparity yield a wider efficient frontier in another; and, relative to random search and hyperparameter tuning of a GBM, ARB finds frontier models with up to $m(x,z)=E[Y\mid X=x,Z=z]$11 lower $m(x,z)=E[Y\mid X=x,Z=z]$12-bias at equal AUC and up to $m(x,z)=E[Y\mid X=x,Z=z]$13 higher AUC at equal $m(x,z)=E[Y\mid X=x,Z=z]$14-bias, without retraining. Typical runtimes on $m(x,z)=E[Y\mid X=x,Z=z]$15 held-out samples with $m(x,z)=E[Y\mid X=x,Z=z]$16 to $m(x,z)=E[Y\mid X=x,Z=z]$17 parameters are about $m(x,z)=E[Y\mid X=x,Z=z]$18 seconds for preprocessing and roughly $m(x,z)=E[Y\mid X=x,Z=z]$19–$m(x,z)=E[Y\mid X=x,Z=z]$20 minutes total for $m(x,z)=E[Y\mid X=x,Z=z]$21 and $m(x,z)=E[Y\mid X=x,Z=z]$22.

This usage of ARB is therefore distinct from residual-balancing estimators for causal inference. It is model-agnostic, post-processing, and explicitly organized around an efficient frontier between a Wasserstein fairness metric and predictive performance.

6. ARB as a generator-induced balancing principle in generalized Riesz regression

In generalized Riesz regression, ARB is a design principle for estimating Riesz representers in automatic debiased machine learning. The target is a linear functional

$m(x,z)=E[Y\mid X=x,Z=z]$23

which admits a Riesz representer $m(x,z)=E[Y\mid X=x,Z=z]$24 satisfying

$m(x,z)=E[Y\mid X=x,Z=z]$25

for all test functions $m(x,z)=E[Y\mid X=x,Z=z]$26 in the model class. The associated Neyman-orthogonal score is

$m(x,z)=E[Y\mid X=x,Z=z]$27

The genriesz package implements this framework via empirical Bregman-divergence minimization, returning regression adjustment (RA), Riesz weighting (RW), augmented Riesz weighting (ARW), and TMLE-style estimators with cross-fitting, confidence intervals, and $m(x,z)=E[Y\mid X=x,Z=z]$28-values (Kato, 19 Feb 2026).

The generalized Riesz regression objective is built from a convex Bregman generator $m(x,z)=E[Y\mid X=x,Z=z]$29. In primal form,

$m(x,z)=E[Y\mid X=x,Z=z]$30

and in dual form one writes $m(x,z)=E[Y\mid X=x,Z=z]$31 and minimizes

$m(x,z)=E[Y\mid X=x,Z=z]$32

ARB enters through a generalized-linear-model-style parameterization

$m(x,z)=E[Y\mid X=x,Z=z]$33

together with the key link choice

$m(x,z)=E[Y\mid X=x,Z=z]$34

This generator-induced link makes the dual objective convex in $m(x,z)=E[Y\mid X=x,Z=z]$35 and, when $m(x,z)=E[Y\mid X=x,Z=z]$36, the KKT conditions yield exact empirical balancing equations

$m(x,z)=E[Y\mid X=x,Z=z]$37

Thus the fitted representer balances user-chosen basis moments against the target functional’s Riesz moments.

The package implements cross-fitted estimation. For each training fold, it solves for $m(x,z)=E[Y\mid X=x,Z=z]$38, forms

$m(x,z)=E[Y\mid X=x,Z=z]$39

fits outcome regressions if needed, predicts on the hold-out fold, and then computes RA, RW, ARW, or TMLE from the pooled out-of-fold predictions. Standard errors are based on the empirical variance of the orthogonal score. Under standard DML conditions, including cross-fitting and

$m(x,z)=E[Y\mid X=x,Z=z]$40

the ARW and TMLE estimators are $m(x,z)=E[Y\mid X=x,Z=z]$41-asymptotically normal.

The package supports several bases—polynomials, RKHS approximations, random-forest leaf indicators, frozen neural-network embeddings, and a nearest-neighbor catchment basis—and several generators, including squared distance, UKL, BKL, Basu’s power, and PU-divergence. For ATE estimation, the paper notes that ARB can recover entropy- or KL-balancing style weights that exactly balance any user-specified polynomial or tree-based function of covariates.

7. Shared structure, distinctions, and recurrent misconceptions

Across the causal and DML variants, ARB repeatedly appears as a combination of a regression fit with an explicit balancing correction. In approximate residual balancing and its nonlinear extension, the correction is literally a residual-weighted term added to a plug-in counterfactual estimate (Athey et al., 2016, Meza, 31 Oct 2025). In augmented balancing weights, the same structure can be rewritten as a single plug-in estimator with coefficients between a regularized base learner and OLS (Bruns-Smith et al., 2023). In generalized Riesz regression, ARB is encoded in KKT moment equations induced by a generator-specific link (Kato, 19 Feb 2026). In the fairness setting, by contrast, ARB is a post-processing family that reshapes selected regressors and calibrates the resulting scores, without retraining and without using the protected attribute at prediction time (Miroshnikov et al., 2021).

Several distinctions are therefore essential. First, “balancing” does not have a single mathematical meaning across these papers: it can mean balancing treated and control covariate moments, balancing $m(x,z)=E[Y\mid X=x,Z=z]$42- and $m(x,z)=E[Y\mid X=x,Z=z]$43-weighted moments in a GLM, matching Riesz moments in a basis, or reshaping predictor distributions to reduce a Wasserstein fairness metric. Second, “collapse to OLS” is specific to the augmented-balancing-weights linear representation when the base learner is OLS or when the balancing penalty tends to zero; it is not a generic property of all ARB formulations. Third, exact moment matching is formulation-dependent: in generalized Riesz regression it arises from the KKT conditions when $m(x,z)=E[Y\mid X=x,Z=z]$44, while in nonlinear residual balancing first-order balance is insufficient and must be supplemented by an explicit second-order correction.

Taken together, the cited literature portrays ARB as a broad balancing paradigm spanning semiparametric efficiency, high-dimensional causal inference, regularization bias correction, and fairness-aware post-processing. The family resemblance is substantive, but so are the differences in estimands, optimization problems, and guarantees.