arXiv paper2026-09-03· machine learning — causal discovery· diagnostic study (theory + synthetic experiments)

Guide, Not Bind: Why Defeasible Priors Fail in Augmented Lagrangian Causal Discovery

Q: What is the 'guide, not bind' failure in causal discovery?

It is the idea that a possibly-wrong expert rule (a forbidden edge between variables) can be enforced as a growing penalty during training while a data-adaptive check automatically relaxes it if the data support the edge. This paper proves two independent reasons the idea fails in differentiable causal discovery: geometric penalty ramping suppresses the true edge before any check can measure its cost, and (under correlation-matching objectives) the fitting loss cannot distinguish an edge from its reverse at all.

Q: Can correlation matching identify causal direction under equal-variance noise?

No. The paper proves that under correlation matching, the cost of fitting a true edge and its reverse is exactly 2r² — identical in both directions — because normalizing to correlation discards the marginal-variance information that equal-variance identifiability (Peters & Bühlmann, 2014) relies on. Covariance matching keeps that information and provably separates the two directions.

Q: Which fitting objective do the authors recommend after the study?

A Notears/Dagma-style least-squares objective. Across the four objectives tested (correlation, covariance, least-squares, likelihood), least-squares breaks the directional tie, gives the best whole-graph SHD, and avoids the likelihood objective's worst-in-class suppression of wrongly-forbidden edges (96–100% of trials).

Q: Was the failing mechanism (DADU) taken from a real system?

No. The authors introduce DADU themselves as the natural instantiation of the defeasible-prior idea and state explicitly that it is not drawn from any published implementation; correlation matching, the testbed objective, is likewise their own choice rather than a deployed system's.

Sairam Sundararaman, Sara Girdhar, Manit Narasimha Murthy, Samrudh N, Bhaskarjyoti Das

View original paper ↗arXiv:2609.03442

cs.LG

Quick verdict

Paper's own defeasible ALM rule shown to fail via suppression and an identifiability tie; synthetic-only study with honest limits, modest practical scope.

Summary

This paper examines a design pattern in differentiable causal discovery: encode an expert's forbidden-edge prior as an Augmented Lagrangian penalty, trusting a data-adaptive mechanism to relax it when data contradict it. The authors introduce DADU, a natural relaxation rule of their own construction, and identify two independent reasons the design fails. First, geometric penalty ramping drives a wrongly-forbidden true edge to zero before any counterfactual check can measure its cost; DADU violates all three necessary conditions of Proposition 1, and across 3,072 synthetic runs a single wrong prior suppresses the true edge in 87–97% of trials under correlation matching, independent of the threshold setting. Second, correlation matching ties an edge and its reverse to exactly the same cost, 2r² (Lemma 1), because normalization discards the variance information that equal-variance identifiability requires; covariance matching separates the directions by at least w0⁴ (Lemma 2). The authors build the fix their diagnosis prescribes — AdaptiveRelax with covariance matching — and in 768 paired trials it recovers the forbidden edge significantly more often than DADU (sign tests p=1.5×10⁻⁸ and 1.2×10⁻⁶), yet most weight is still absorbed by the unconstrained reverse edge. Re-running the sweep with Notears/Dagma-style least-squares and golem-EV-style likelihood objectives shows the tie is correlation-specific, but suppression persists and worsens under likelihood (96–100%). Main limitations: synthetic linear-Gaussian data only, one wrong prior per graph, a self-constructed baseline mechanism, and an untuned loss weight confounding the covariance whole-graph SHD comparison.

Review complete

Assessment at a glance

machine learning — causal discovery

risk flags identified

● high 0

● medium 2

● low 3

Reproducibility readinesshigh

Code & data availablenot stated

Review confidencehigh

AI text signal0/20 passages flagged

Deep research viableCode-reproducible

5 strengths · 5 weaknesses · 5 author questions

AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.

Levels reflect verification complexity — not a quality score. Evolve never rates your paper.

Key claims under review

Sequential penalty-ramping ALM suppresses a wrongly-forbidden true edge before any data-adaptive counterfactual check can measure its cost; DADU violates all three necessary conditions of Proposition 1, with 87–97% suppression under correlation matching across 3,072 runs (Table 1).
Correlation matching ties a true edge and its reverse to exactly 2r² because normalizing discards the variance signal equal-variance identifiability requires (Lemma 1), while covariance matching separates the directions by at least w0⁴ (Lemma 2).
The combined AdaptiveRelax + covariance fix recovers the wrongly-forbidden edge significantly more often than DADU (768 paired trials, p<10⁻⁵) but still in the minority, with the unconstrained reverse edge as the dominant attractor under every objective tested.
The identifiability tie is specific to correlation matching — it disappears under least-squares and likelihood objectives (Table 3) — while suppression is not, and the golem-EV-style likelihood is the worst objective, suppressing in 96–100% of trials (Table 4).
Covariance matching's whole-graph SHD cost leads the authors to revise their recommendation toward a Notears/Dagma-style least-squares objective as the best of the four tested.

Strengths

Tight theory-experiment closure
Lemma 1's exact 2r² tie is verified empirically to floating-point precision (Table 2: directional gap −6.8×10⁻¹⁸ to −1.5×10⁻¹² under correlation), and the δ-sweep in Section 4.1 shows final outcomes change <10⁻³ across five thresholds, confirming threshold-independence of the suppression trap.
Adversarial self-testing (Section 7.3)
The full 6,144-run sweep is re-run under population-form least-squares (Notears/Dagma) and golem-EV-style likelihood objectives, with the least-squares identity verified numerically to machine precision — and the inconvenient result (likelihood suppresses worst, 96–100%) is reported rather than buried.
Exceptional transparency
Matched-pair experimental design
Worked verification of the graphical claim

Weaknesses

Self-constructed baseline mechanism
Testbed objective also bespoke
Proposition 1's necessity is local and uncheckable a priori
Confounded covariance SHD comparison
Synthetic-only, single-prior scope

Full Review Report

$1.99 · one-time

Unlock the complete preliminary analysis Evolve produced for this paper.

Report preview

Scope & Claims

The paper targets a design pattern it names *guide, not bind*: a practitioner's forbidden-edge prior in differentiable causal discovery, enforced through an Augmented Lagrangian (ALM) with geometrically ramped penalty, plus a data-adaptive rule meant to relax the constraint when data contradict it. Crucially, the authors state upfront that both the relaxation rule they analyze (DADU, Eq. 8) and the testbed objective (correlation matching, Eq. 5) are their own constructions, not components of any published system; this honesty is exemplary, but it bounds practical significance. Four claims are made: (1) sequential penalty ramping suppresses a wrongly-forbidden true edge before any counterfactual check can measure its cost (Proposition 1 gives three necessary conditions; Corollary 1 shows DADU violates all three); (2) correlation matching ties an edge and its reverse at exactly 2r² (Lemma 1) because normalization discards the variance signal equal-variance identifiability requires, while covariance matching separates directions by ≥ w0⁴ (Lemma 2); (3) the prescribed fix (AdaptiveRelax + covariance matching) improves recovery significantly but leaves the reverse edge dominant; (4) the tie is correlation-specific while suppression is not — a golem-EV-style likelihood suppresses worst of four objectives (96–100%, Table 4).

Methodology Assessment

The theory-experiment chain is unusually tight for this area. Lemma 1 is a clean two-node closed-form computation, verified empirically to floating-point precision (Table 2: directional gap −6.8×10⁻¹⁸ to −1.5×10⁻¹² under correlation, versus covariance gaps growing 2.64→11.71 with d). Proposition 2's extension to covered edge reversals is backed by a worked three-node example in Appendix E (a=0.5, b=0.4, c=0.6 → a′≈0.2229, b′≈0.6002), demonstrating the reparameterization is not the identity on other edges. The sweep is well-designed: 6,144 runs over d∈{4,8,16,32}, two

The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.

All 5 strengths & 5 weaknesses, analyzed in detail
5 risk flags with severity, section & reviewer questions
Methods & statistical evidence assessment
5 questions a peer reviewer would ask
Full six-dimension evaluation report & recommendations

One-time purchase · this paper only

One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.

Frequently asked questions

What is the 'guide, not bind' failure in causal discovery?

It is the idea that a possibly-wrong expert rule (a forbidden edge between variables) can be enforced as a growing penalty during training while a data-adaptive check automatically relaxes it if the data support the edge. This paper proves two independent reasons the idea fails in differentiable causal discovery: geometric penalty ramping suppresses the true edge before any check can measure its cost, and (under correlation-matching objectives) the fitting loss cannot distinguish an edge from its reverse at all.

Can correlation matching identify causal direction under equal-variance noise?

No. The paper proves that under correlation matching, the cost of fitting a true edge and its reverse is exactly 2r² — identical in both directions — because normalizing to correlation discards the marginal-variance information that equal-variance identifiability (Peters & Bühlmann, 2014) relies on. Covariance matching keeps that information and provably separates the two directions.

Which fitting objective do the authors recommend after the study?

A Notears/Dagma-style least-squares objective. Across the four objectives tested (correlation, covariance, least-squares, likelihood), least-squares breaks the directional tie, gives the best whole-graph SHD, and avoids the likelihood objective's worst-in-class suppression of wrongly-forbidden edges (96–100% of trials).

Was the failing mechanism (DADU) taken from a real system?

No. The authors introduce DADU themselves as the natural instantiation of the defeasible-prior idea and state explicitly that it is not drawn from any published implementation; correlation matching, the testbed objective, is likewise their own choice rather than a deployed system's.

About this review

This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.

The provided text is a sampled excerpt with explicit '[...sampling gap...]' markers: mid-paper portions (parts of Section 5's derivations, the start of Section 7.4's paired-counts sentence, the tail of Table 4's likelihood row, and portions of Related Work/Section 10) were skipped by sampling. All numbers cited in this review appear in the visible text.
No code or data availability statement appears in the sampled text; one may exist in an unsampled portion, so codeAvailable=false reflects the provided excerpt only.

Go deeper on this paper

Wrote a paper? Stress-test it before reviewers do.

Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.

Scan your paper