{"id":"38094999-a858-4c02-853b-a9fac03dd836","arxiv_id":"2501.04871","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"RieszBoost uses gradient boosting with a data augmentation trick to estimate Riesz representers directly from the Riesz loss, matching or improving on indirect plug-in estimators in simulations.","lead":"RieszBoost is a new algorithm that trains gradient-boosted trees to directly estimate the Riesz representer, a weighting function used in efficient causal inference estimators. It replaces the usual two-step approach of deriving and plugging in analytical weights, and shows competitive coverage in simulations for four causal estimands.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Empirical Riesz loss is unbounded below for continuous covariates, so RieszBoost's output depends on unanalyzed early stopping; the claimed minimizer of the Riesz loss is not well-defined.","rationale":"The reader's verdict is CONDITIONAL on exactly this issue, and I agree it is the main soft spot. The central claim is that RieszBoost directly estimates the Riesz representer by minimizing the Riesz loss; for that claim to hold, the empirical Riesz loss must be a reasonable optimization target. It is not: with continuous covariates, the empirical loss is unbounded below, so there is no minimizer and no population-to-sample justification. The paper does not prove consistency of the boosting procedure or analyze early stopping; it only recommends tuning via the validation Riesz loss, which inherits the unboundedness. I checked the gradient derivations in Examples 1, 2, and Appendices B.1.2 and B.3.2; they are internally consistent up to absorbed constants, so the issue is not a sign error but the well-posedness of the objective. The simulation results are encouraging but cannot fully support the algorithmic claim without an account of why the unbounded directions are controlled. A concrete experiment varying M can settle whether the unboundedness is only a theoretical artifact or actually drives the estimates. Since the reader already flagged this and recommended CONDITIONAL, my stress-test does not change the verdict.","tokens_in":17446,"tokens_out":8435,"duration_ms":87598,"concrete_test":"Re-run the ATE simulation (Section 3.1) with the RieszBoost implementation, holding tree depth and learning rate fixed, and vary M ∈ {1, 10, 50, 100, 500, 2000}. On a holdout set, record the validation Riesz loss and the EEE ATE estimate at each M. If the validation loss decreases monotonically without leveling off while the ATE estimate drifts away from ψ_ATE = 29.502 (or coverage degrades), then the unbounded loss is not controlled by early stopping, and the paper's recommended tuning via validation Riesz loss is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing gap is in Section 2.1, where the population Riesz loss is replaced by the empirical Riesz loss and then minimized. For continuous covariates with no ties, this empirical loss is unbounded below. In the ATE example (Eq. 4), a counterfactual value such as α(1,X_j) for a control unit j appears in the loss only through the linear term −2α(1,X_j); there is no quadratic term at that point because no observed unit has (A=1,X=X_j). Taking α(1,X_j)=t and holding all other values fixed sends L_n to −∞. The same holds for the ASE and LASE. This is not an artifact of unrestricted function spaces: a depth-k tree can isolate a leaf containing only counterfactual rows (e.g., a leaf with A=1 and an X interval populated only by control-origin rows), so the boosting function class contains the unbounded direction. Consequently, the algorithm has no finite minimizer, and its output is governed by the unanalyzed choice of boosting iterations M and tree-induced smoothing. The recommended tuning criterion—validation Riesz loss—is also unbounded below, so it cannot select a stopping point. The simulations may still show good finite-sample behavior, but the central claim that RieszBoost estimates α0 by minimizing the Riesz loss lacks a well-defined objective.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RieszBoost, a gradient boosting algorithm that directly estimates the Riesz representer of a linear functional by minimizing the Riesz loss. The method constructs predictor and target matrices to evaluate the loss at counterfactual points, and the authors derive residual functions for the ATE, ATT, ASE, and LASE functionals. Extensive simulations compare RieszBoost with plug-in estimators based on propensity scores or conditional densities; RieszBoost reports lower RMSE/MAE for the Riesz representer in the binary treatment cases and comparable coverage for all targets.","tokens_in":17735,"tokens_out":6079,"duration_ms":58827,"significance":"If validated, RieszBoost would be a practical and user-friendly contribution to causal inference with tabular data, sidestepping analytical derivation of Riesz representers and density estimation. The simulation evidence is encouraging, and the paper explicitly includes derivations, cross-fitting, and hyperparameter-tuning details. However, the paper's central claim is that the algorithm estimates the Riesz representer by minimizing the empirical Riesz loss; because this loss is unbounded below for continuous covariates, the optimization objective is not well defined. This is a load-bearing gap that prevents acceptance in its current form, although it is likely addressable with regularization or a restricted function class.","major_comments":[{"comment":"The empirical Riesz loss is unbounded below for continuous covariates with no ties, so the claimed minimization of L_n is not a well-defined optimization problem. In the ATE loss (Eq. 4), for a control unit j, the counterfactual value α(1,X_j) appears only in the linear term −2α(1,X_j) with no quadratic term; setting α(1,X_j)=t → ∞ sends L_n → −∞ while holding all other α values fixed. The same phenomenon occurs for the ASE and LASE losses (Section 2.3, Example 2, and Appendix B.3.2). The boosting function class of shallow trees can contain such unbounded directions by isolating a leaf with only counterfactual rows. Therefore, the algorithm's output depends on unanalyzed early stopping and tree-induced smoothing rather than on minimization of the stated loss. The authors should either add explicit regularization (e.g., a penalty on α), restrict the function space (e.g., bounded leaf values), or provide an analysis of implicit regularization in gradient boosting. As written, the theoretical motivation for RieszBoost is incomplete.","section":"Section 2.1, Eq. (4)"},{"comment":"The recommended tuning criterion—the validation-set empirical Riesz loss—inherits the same unboundedness described above. Since the validation loss can be driven to −∞ by increasing model complexity or pushing counterfactual α values to extreme magnitudes, it cannot be used to select a stopping point or choose hyperparameters. The claim that 'tuning is relatively straightforward' and the suggested early-stopping procedure are not supported by a well-defined selection criterion. A bounded or regularized validation objective, or a theoretical analysis of early stopping in this setting, is needed.","section":"Section 2.3, Tuning paragraph"}],"minor_comments":[{"comment":"The text contains a duplicated word: 'formed by by first estimating µ0 and α0' should read 'formed by first estimating µ0 and α0.'","section":"Section 1"},{"comment":"Typos: 'esitmation' should be 'estimation', and 'assoicated' should be 'associated' in the sentence about avoiding conditional density estimation.","section":"Section 3.2.1"},{"comment":"The validation loss formula writes α(O_i)^2, but since α is a function of W, this should be α(W_i)^2 to match the Riesz loss definition.","section":"Section 2.3, Tuning formula"},{"comment":"The mapping p(j) from rows of the predictor matrix to rows of the original data is not explicitly defined; the notation is introduced informally and would be clearer with a formal definition.","section":"Section 2.3, predictor matrix construction"},{"comment":"No code repository or reproducibility details are provided, which would be helpful for a methods paper depending on a new algorithmic procedure.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The unboundedness of the empirical Riesz loss is a genuine technical gap, but it is fixable: the authors could add a penalty, constrain the function class, or reframe the method as an implicitly regularized gradient descent. The simulation results may still stand after such a revision, but the central theoretical claim must be corrected. I would encourage the editor to request a revision that addresses this issue, rather than a rejection, unless further analysis reveals that the empirical performance also collapses."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core contribution here is real: a gradient boosting algorithm that targets the Riesz representer directly, with a clean data-augmentation trick to handle the pseudo-data that appear in the Riesz loss. The gradient derivations for ATE, ATT, ASE, and LASE are internally consistent, the simulations are honest enough to show where the method wins and where it merely ties, and the exposition is clear. The pseudo-data construction is the sort of practical detail that makes an algorithmic contribution usable.\n\nThe problem is the objective. The empirical Riesz loss is not bounded below for continuous covariates. In the ATE case, for a control unit with unique X, the counterfactual value α(1,X) appears only through the linear term −2α(1,X); there is no squared term at that point. Sending that value to +∞ while holding everything else fixed drives L_n to −∞. The same happens for the other functionals. Since the relabeled rows in the predictor matrix can be isolated by trees, the function class actually contains these unbounded directions. So the claimed minimizer does not exist, and the algorithm's output is governed by unanalyzed early stopping and tree-induced smoothing. The recommended tuning criterion, validation Riesz loss, inherits the same unboundedness, so it cannot select a stopping point either. This doesn't invalidate the simulations—the method can still have good finite-sample behavior due to implicit regularization—but it means the paper's central claim, that RieszBoost estimates α0 by minimizing the Riesz loss, is not well-defined as stated.\n\nA few smaller issues: the abstract says the method performs “on par with or better than” indirect estimation, but for the continuous ASE/LASE the Riesz representer RMSE is actually worse; only the final coverage is similar. There is also no comparison to existing Riesz regression implementations like RieszNet or ForestRiesz, which would be the natural baseline.\n\nWho is this for? Applied causal inference researchers who want a user-friendly, GBM-based alternative to deriving Riesz representers by hand. The algorithm is likely to work in practice with proper regularization, but the theory needs repair before I would trust it in a methods paper. I would send it to a serious referee—the clarity and the usefulness justify the time—but I would expect the revision to tackle the unbounded loss head-on, add a penalty or a well-posed surrogate objective, and report what the early stopping is actually doing.","headline":"RieszBoost is a genuinely useful algorithmic idea for direct Riesz regression with boosted trees, but the empirical Riesz loss is unbounded below for continuous covariates, so the stated objective is not well-defined and the paper needs substantial revision before it can be taken as a rigorous method.","tokens_in":730,"tokens_out":892,"would_cite":false,"duration_ms":32137,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","62G08"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces RieszBoost, a gradient boosting algorithm that estimates the Riesz representer directly by minimizing the Riesz loss, matching or outperforming indirect plug-in estimators for ATE, ATT, ASE, and LASE.","keywords":["Riesz representer","Riesz regression","gradient boosting","causal inference","efficient estimating equations","balancing weights","average treatment effect","Riesz loss"],"falsifier":"Run RieszBoost on simulated data with continuous covariates and a known treatment assignment mechanism, without early stopping, and compare the fitted weights with the known inverse-probability weights as the number of boosting iterations grows; if the empirical Riesz loss keeps decreasing while the fitted weights move away from the truth, the central claim fails.","tokens_in":17242,"feed_emoji":"🌲","tokens_out":6635,"duration_ms":60034,"temperature":0.7,"pith_summary":"RieszBoost is a gradient boosting algorithm that learns the Riesz representer, the weighting function at the heart of doubly robust causal estimators, by minimizing the Riesz loss directly without deriving its analytic form such as inverse propensity scores. Because the empirical Riesz loss for causal functionals depends on counterfactual pseudo-data, the paper augments the predictor matrix with these pseudo-observations and defines a row-wise residual function so that standard tree boosting can be run almost unchanged. In simulations covering the average treatment effect, the average treatment effect among the treated, the average shift effect, and the local average shift effect, RieszBoost estimates are unbiased and produce confidence intervals with coverage at or above that of indirect plug-in estimators, with tighter intervals for the ATE. The contribution is a practical, user-friendly way to obtain the Riesz representer for tabular data, which can then be plugged into efficient estimating equations, TMLE, or double machine learning.","feed_headline":"Boosting learns causal weights directly, matching plug-in estimators","feed_subtitle":"RieszBoost minimizes the Riesz loss on augmented data, giving unbiased ATE, ATT, ASE, and LASE estimates with good coverage.","key_machinery":"The load-bearing identity is the Riesz representation of a continuous linear functional, $\\Psi(P_0)=E[m(O,\\mu_0)] = E[\\alpha_0(W)\\mu_0(W)]$, which turns the representer into the minimizer of the Riesz loss $E[-2m(O,\\alpha)+\\alpha(W)^2]$. The machinery is gradient boosting in function space, repeatedly fitting trees to the negative empirical gradient, plus a data-augmentation step that expands the predictor matrix with pseudo-data where the empirical gradient is nonzero; a target matrix and residual function encode the parts of the gradient that do not depend on the candidate function $\\alpha$. This lets standard boosting libraries minimize the Riesz loss with minimal code changes.","core_discovery":"The paper's central claim is that the Riesz representer $\\alpha_0$, which encodes the weights needed for doubly robust estimation of a linear functional $\\Psi(P_0)=E[m(O,\\mu_0)]$, can be estimated directly by gradient boosting on the Riesz loss $E[-2m(O,\\alpha)+\\alpha(W)^2]$. The minimizer of this loss is $\\alpha_0$ by the Riesz representation theorem, and the paper shows how to implement the minimization with tree boosting. The key implementation step is a data augmentation trick: build a predictor matrix from the observed $W$ plus pseudo-data, such as counterfactual treatments, where the empirical gradient is nonzero; build a target matrix recording the original and counterfactual exposure values; and define a residual function that combines these with the current prediction. This makes the Riesz loss compatible with standard gradient boosting code. Simulations show that the resulting RieszBoost estimates of $\\alpha_0$ for ATE, ATT, ASE, and LASE are comparable to or better than indirect plug-in estimates, and that downstream efficient estimating equations estimates of the causal parameters have good coverage.","pith_inferences":["An implicit consequence is that RieszBoost's practical success may depend on early stopping and tree smoothing, because with continuous covariates and no ties the empirical Riesz loss is unbounded below; formalizing this regularization could make the method more reliable.","The same predictor and target augmentation could be automated from the functional $m(O,\\mu)$, letting users apply RieszBoost to new estimands without hand-deriving residual functions.","The method may be especially useful for longitudinal modified treatment policies and continuous exposures, where the analytical representer requires density-ratio estimation; the simulations only cover one-dimensional confounders, so high-dimensional behavior remains an open test.","Combining RieszBoost with stochastic sub-sampling or other standard boosting extensions could reduce runtime and variance at large sample sizes."],"forward_implications":["For causal functionals whose Riesz representer has no tractable analytic form, such as shift effects with continuous treatments, RieszBoost provides an estimate directly from data.","Substituting RieszBoost's representer into an efficient estimating equations estimator gave unbiased ATE and ATT estimates with 95% confidence interval coverage at or above the indirect method in simulations.","Because tuning is done on the validation Riesz loss and the algorithm reuses standard boosting implementations, researchers can adopt it without new modeling machinery.","RieszBoost estimates can be used in cross-fitted efficient estimating equations, TMLE, or double machine learning procedures without modification."],"supporting_citations":[{"why":"Introduces the Riesz loss whose minimizer is the true Riesz representer; this is the objective RieszBoost optimizes.","marker":"[11]"},{"why":"Presents the existing direct Riesz regression algorithms that the new boosting approach extends.","marker":"[12]"},{"why":"Provides the gradient boosting framework in function space on which Algorithm 2 is built.","marker":"[18]"},{"why":"Defines the efficient estimating equations estimator used in the simulations to turn estimated Riesz representers into causal parameter estimates.","marker":"[2]"},{"why":"Gives the efficient influence function representation involving the Riesz representer, motivating why direct estimation of $\\alpha_0$ matters.","marker":"[8]"},{"why":"Supplies the cross-fitting rationale the paper recommends for using RieszBoost estimates in downstream inference.","marker":"[29]"}],"fun_headline_variants":["Direct Riesz representer learning via gradient boosting","Boosting replaces plug-in causal weight formulas","RieszBoost: Boost causal weights, skip the derivation","Tree-based Riesz regression: no explicit form needed","Gradient boosting for robust causal effect estimation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that fitting trees to the finite-sample Riesz loss recovers the true weighting function, even though for continuous covariates that loss is unbounded below and the result can depend on early stopping.","fun_headline_variants_meta":{"raw":{"variants":["Direct Riesz representer learning via gradient boosting","Boosting replaces plug-in causal weight formulas","RieszBoost: Boost causal weights, skip the derivation","Tree-based Riesz regression: no explicit form needed","Gradient boosting for robust causal effect estimation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1497,"prompt_tokens":982,"completion_tokens":515,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":443}},"tokens_in":598,"tokens_out":515,"duration_ms":4738,"temperature":1.0,"reasoning_tokens":443,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:24:34.561569+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run RieszBoost on simulated data with continuous covariates and a known treatment assignment mechanism, without early stopping, and compare the fitted weights with the known inverse-probability weights as the number of boosting iterations grows; if the empirical Riesz loss keeps decreasing while the fitted weights move away from the truth, the central claim fails.","supporting_citations":[{"cited_title":"Friedman","cited_arxiv_id":null,"evidence_quote":"Provides the gradient boosting framework in function space on which Algorithm 2 is built."},{"cited_title":"Robins, Andrea Rotnitzky, and Lue Ping Zhao","cited_arxiv_id":null,"evidence_quote":"Defines the efficient estimating equations estimator used in the simulations to turn estimated Riesz representers into causal parameter estimates."},{"cited_title":"Double/debiased machine learning for treatment and structural parameters","cited_arxiv_id":null,"evidence_quote":"Supplies the cross-fitting rationale the paper recommends for using RieszBoost estimates in downstream inference."}],"review_version":1}