acceptodds
Under review as a conference paper at ICLR 2027

IV-CRL: INSTRUMENTAL VARIABLE APPROACH FOR CAUSAL REPRESENTATION LEARNING

Abstract

We study causal representation via linear latent DAGs in which exogenous instruments are observed jointly with a high-dimensional vector , instrument targets are unknown, and observed variables may cross-load without pure measurements. Instruments identify the latent total-effect subspace from proportional IV total effects on . Because the latent total effects include descendant propagation, we recover the latent DAG and sparse loading matrix by reverse-topological residual peeling. We give necessary-and-sufficient conditions for population recovery and an example where is nonidentifying but is identifying. IV-CRL-Peel implements this argument using IV total-effect clustering and trimmed-regression peeling. It improves graph recovery over observational and IV-based baselines in simulations; with population IV total effects, it recovers every simulated graph. A masked OneK1K illustration exactly recovers the cis-eQTL groups and empirical latent DAG, with weaker score and loading recovery.

Primary area:
causal reasoning
Keywords:
causal representation learning, instrumental variables, latent causal graphs, sparse factor models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_TEDAOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2(+1)
Soundness
2(+1)
Read review
Paper summary

The paper studies linear latent causal models with observed instruments W and measurements X with unknown instrument targets, and sparse loadings that can include cross-loadings. A reverse-topological peeling procedure separates these directions into a latent DAG and loading matrix. The paper gives necessary-and-sufficient conditions for this population algorithm and a finite-sample clustering and trimmed-regression implementation. Empirical experiments include a simulation dataset with additional scaling and assumption stresses, dimension-selection experiments, and a three-gene masked OneK1K downstream genetic application.

Critical strengths

Reporting weak-IV failures and incomplete recovery provides a useful characterization of the method’s limitations.

Critical weaknesses
  1. The IV assumptions are difficult to justify in real-world data. Identification requires single-target instruments, coverage of every latent variable, and exclusion of pathways bypassing the modeled latent system. How can these assumptions be justified?
  2. The comparative evaluation is limited to basic representation-plus-GES pipelines. No existing interventional CRL method is evaluated.
  3. The empirical results are not comprehensive enough. The baselines (PCA, sparse-varimax and IV-SVD + GES) estimate a rotated basis and therefore fail by construction. The comparison is not convincing to show the advantage of the method.
Questions for authors
  1. The method requires a complete collection of valid, single-target instruments for unobserved latent variables, but provides limited guidance on how such instruments can be obtained or justified in practice.
  2. In real world datasets OneK1K genetic application, how to justify the model assumptions hold ?
  3. Do prop. 1 and theorems 2-3 allow an arbitrary Cov(ε)\mathrm{Cov}(\varepsilon), i.e., hidden confounding among the latent variables?
  4. How were s=4, q=2 and tau=0.06 chosen for OneK1K, and how stable are the conclusions to these choices? Sensitivity analysis is needed 5.The current comparisons largely rely on factor extraction followed by GES. The authors should compare against strong published latent-variable discovery methods, such as RLCD, LGES, FOFC/MIMBuild in a fair setting.
5 comments

Thank you for your valuable comments.

W1 and Q1: obtaining valid single-target instruments with full coverage

We agree that IV validity is difficult to establish from observed data alone and that obtaining valid single-target instruments covering every latent variable requires substantive justification. This is however not unique to this work but rather reflects a very broader inherent challenge in causal representation learning: many key identifying assumptions concern unobserved variables or mechanisms. Examples include independence of latent structural disturbances in non-Gaussian approaches [1], restrictions on which latent groups directly generate particular measurements in GIN [2], and interventions that change a single latent mechanism while preserving the measurement map in interventional CRL [3]. Even though some of these assumptions have testable implications that allow certain violations to be detected, satisfying those implications, however, does not establish that the underlying assumptions hold.

Despite these challenges, IV methods are widely used in econometrics, including studies of supply and demand and returns to schooling [4], and in genetics through Mendelian randomization [5]. Their use rests on scientific or institutional arguments for instrument validity, supported by empirical diagnostics and sensitivity analyses where available. The proposed IV-CRL brings this established source of causal information to representation learning, allowing Gaussian and correlated structural disturbances under its exogeneity and structural assumptions.

Our assumptions on instruments for latent variables are analogous to the relevance, exogeneity, and exclusion assumptions used in IV-based causal discovery with observed variables. This parallels non-Gaussian CRL methods [1], which extend the independent, non-Gaussian disturbance assumptions of causal discovery with observed variables [6] to latent systems. In IV-CRL, single targeting expresses exclusion relative to an instrument's designated latent target: its effects on other latents and measurements must pass through that target. Full coverage supplies relevance across the latent system and, under single targeting, ensures rank⁡(Γ)=d\operatorname{rank}(\Gamma)=d. A comparable coverage requirement appears in interventional CRL, where full identification may require an intervention on every latent variable [3].

We envisage applications to a focused set of latent traits or regulators for which external knowledge supports plausible instruments. For example, in genetically informed psychological measurement, WW contains candidate SNPs, ZZ contains latent traits such as anxiety and depression, and XX contains questionnaire measurements [7,8]. A valid single-target SNP directly affects one trait; its effects may propagate to other traits and measurements through that target.

In genetics, external association evidence, cis localization, conditional relevance screening, LD filtering, and ancestry adjustment provide practical starting points. Collecting multiple plausible candidates per target can improve the prospect of coverage; our model also already permits irrelevant candidates.

Q2: justification in OneK1K

As discussed in our response to W1/Q1, IV validity is generally difficult to establish from observed data alone. This reflects the broader challenge of assessing identifying assumptions about unobserved variables and mechanisms in CRL. The same limitation applies to OneK1K: the cis-associated SNPs provide scientifically motivated candidate instruments, while exogeneity and exclusion remain substantive assumptions. Our interpretation of the application is conditional on these assumptions.

Q3: arbitrary covariance among latent structural disturbances

Yes. For the IV response matrix Θ=Cov⁡(X,W)Cov⁡(W)−1\Theta=\operatorname{Cov}(X,W)\operatorname{Cov}(W)^{-1}, the moment assumptions give

Θ=MΓ\Theta=M\Gamma

under Cov⁡(ε,W)=Cov⁡(η,W)=0\operatorname{Cov}(\varepsilon,W)=\operatorname{Cov}(\eta,W)=0, regardless of within-ε\varepsilon or within-η\eta covariance. Both the original population-peeling guarantee under P1–P3 and the complementary sparsest-basis identification result under (S), defined in Q4 below, operate on this identified MM and do not need diagonal disturbance covariance.

Q4: choice and sensitivity of s, q, and tau

With three latent nodes, the largest possible outdegree is d−1=2d-1=2. Hence q=2q=2 imposes no additional degree restriction. The tolerance τ=.06\tau=.06 came from simulations as it works reasonably well across various settings. The choice of s=4s=4 was a difficult one: too small, the algorithm will exit (step 4) without completion; too big, it can absorb latent causal effects into the loading matrix and return an empty graph. One could in principle tune this parameter to achieve the desired sparsity of the latent graph. That said, requiring the user to specify s,q,τs,q,\tau is indeed a practical limitation of the original peeling algorithm. We now provide two alternatives that avoid these inputs, Orthogonal LP/refit and Orthogonal robust-order, that do not take s,q,τs,q,\tau as inputs.

  • Orthogonal LP/refit: estimate IV total-effect directions in whitened coordinates; propose sparse directions with linear programs; select one acyclic order; and refit/prune edges by robust regression that propagates first-stage uncertainty.
  • Orthogonal robust-order: use the same first stage and robust regression, but search over every candidate child set and order by dynamic programming. The implementation is intended for small latent dimension (d≤10d\le10).

For these two variants, the original P1-P3 are replaced by the identification condition on the loading matrix

Condition (S). For every vector vv with ∥v∥0≥2\lVert v\rVert_0 \ge 2,

∥Lv∥0>max⁡j∈supp⁡(v)∥ℓj∥0.\lVert Lv\rVert_0 > \max_{j\in\operatorname{supp}(v)} \lVert\ell_j\rVert_0.

where ℓj\ell_j is the jjth loading column. Thus mixing two or more loading columns must produce more nonzeros than every participating column. Under full-rank loadings and the IV assumptions, (S) makes the minimum-total-support basis unique up to scaling and permutation; acyclicity then identifies B,LB,L from MM up to latent scaling and relabeling. We retain this result alongside the original P1–P3 characterization of population peeling.

We reran the exact original main simulation grid and seeds, without changing its generating models: d=5,p=50d=5,p=50, n∈{250,500,1000}n\in\{250,500,1000\}, IV strength .3/.7.3/.7, measurement SNR ∞/5\infty/5, and one/two IVs per latent, with 100 replicates in each of 24 cells. The original peeler receives its original s=10,q=1s=10,q=1; the new variants receive neither value. All original datasets are retained: none was filtered to satisfy (S), and we have not certified (S) for every generated loading matrix.

MethodTPRFDRComplete fits
Original peeler, τ=.04\tau=.0450.46%13.99%1,123/2,400
Original peeler, τ=.06\tau=.0657.43%14.46%1,440/2,400
Original peeler, τ=.08\tau=.0857.96%14.95%1,685/2,400
Orthogonal LP/refit84.57%8.41%2,400/2,400
Orthogonal robust-order86.67%10.25%2,400/2,400

TPR is total true positives divided by all 5,274 true directed edges across the scheduled datasets. FDR is mean dataset-level FP/max⁡(TP+FP,1)FP/\max(TP+FP,1), with empty predictions contributing zero. Partial peeler outputs are retained and scored against the full true graph. Latent labels are aligned using IV-assignment overlap, without choosing labels to maximize graph agreement. Monte Carlo SEs for TPR/FDR, in percentage points, are .62/.44.62/.44 for LP/refit, .56/.45.56/.45 for robust-order, and .96/.59.96/.59 for the original .06.06 peeler.

In addition to being free of s,q,τs,q,\tau, both variants increase TPR and reduce FDR. For example, At n=1000n=1000, LP/refit reaches 97.14%/3.69%97.14\%/3.69\% TPR/FDR and robust-order 97.32%/5.33%97.32\%/5.33\%, compared with 78.47%/12.72%78.47\%/12.72\% for the original peeler.

W2: comparison with interventional CRL

We agree and have evaluated Squires et al. [3] and LSCALE-I [9] using their public numerical implementations on the same 2,400 original IV datasets, alongside the original peeling algorithm and both new IV-CRL variants. All methods receive the common supplied dimension d=5d=5; IVs define the environments for the interventional methods.

We first regress XX jointly on all instruments WW using the original nn observations. We normalize and cluster the estimated IV response columns into dd groups, treating opposite directions as equivalent and using no true target labels. Within each learned group, we select the IV with the largest estimated response-column norm.

Let m=⌊n/2⌋m=\lfloor n/2\rfloor. The baseline consists of the first mm observations in their original i.i.d. order. For each selected IV, we take the mm observations, from all nn rows, whose IV values are closest to its sample median. Their observed XX values form a central-half IV environment. For the independent Gaussian IVs used here, this restriction changes the selected IV's population variance relative to the baseline distribution. We center each environment separately and supply the resulting d+1d+1 environments to both methods. Each has mm observations; environments overlap and contain at most nn distinct subjects in total, so reused rows are not counted as independent additional observations.

Squires uses its public partial-order routine and estimated covariance matrices; LSCALE-I uses estimated Gaussian scores. The primary numerical thresholds and all sensitivity values are recorded before final aggregation.

We report direct edges and ancestral relations separately, since the targets differ:

Method/output at its primary thresholdEvaluation targetTPRFDR
Original peeler, τ=.06\tau=.06 (our method)Direct edges57.43%14.46%
Orthogonal LP/refit (our method)Direct edges84.57%8.41%
Orthogonal robust-order (our method)Direct edges86.67%10.25%
Squires, coefficient support, cutoff .1.1Direct edges7.38%60.97%
Squires, native partial-order outputAncestral relations46.88%86.96%
LSCALE-I, soft output, threshold .1.1Ancestral relations0.00%0.00%

All these runs returned outputs; the peeler completed all nodes in 1,440/2,400 datasets, and its partial outputs remain included. Native ancestral outputs are compared with the true transitive closure, not with the direct DAG. LSCALE-I's empty outputs have zero FDR and zero recovery. This result is threshold-dependent: its soft branch at .001.001 gives 85.39%85.39\% ancestral TPR with 75.89%75.89\% FDR, and at .01.01 gives 18.69%/55.27%18.69\%/55.27\%. Squires' direct cutoffs .05.05 and .2.2 give 19.64%/87.73%19.64\%/87.73\% and 0.93%/14.02%0.93\%/14.02\%, respectively. The separately labeled LSCALE-I hard refinement, applied to these soft IV environments, is misspecified and is not treated as a valid hard-intervention comparison.

Scope of this comparison. An additive IV changes a targeted disturbance distribution while preserving the latent parent coefficients. Such noise-only changes fall outside Squires' relevant genericity condition and, for nonzero parent effects, LSCALE-I's score-change condition [3,9]. In particular, a conditional log-density ratio ri(zi−bi⊤zpa⁡(i))r_i(z_i-b_i^\top z_{\operatorname{pa}(i)}) has proportional target and parent derivatives. Noisy measurement cells also depart from deterministic mixing, and Gaussian score estimation is approximate after IV stratification. Therefore, the proposed methods outperforming them is not surprising.

Reviewer_h6TvOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the identifiability of linear causal representations from observational data. The authors introduce instrumental variables (W), assuming that each latent variable is targeted by at least one instrument. Under additional structural and parameter assumptions, they establish identifiability of the latent dimension, latent DAG, structural coefficients, and loading matrix, up to permutation and scaling. The experiments provide empirical support for the proposed method.

Critical strengths
  1. The paper is clearly written and well organized.

  2. The use of instrumental variables to establish identifiability of latent causal representations is interesting and relevant to causal discovery.

  3. The experiments provide empirical support for the proposed method.

Critical weaknesses
  1. My main concern is the scope and practical plausibility of the instrumental-variable assumptions. The structure (W \to Z \to X) places the latent variables between the instruments and the observations, enabling their recovery through these relationships. Requiring each latent variable to be targeted by at least one instrument is a substantial assumption. It would be helpful to discuss concrete settings in which suitable instruments are available and this coverage requirement is reasonable.

  2. The role of the sparsity assumptions could be explained more clearly. In particular, how do sparsity in the loading matrix and sparsity in the latent DAG contribute to identifiability in the linear Gaussian setting? Separating their roles from those of the instruments and the parameter assumptions would help readers understand what resolves the usual ambiguities.

  3. I would appreciate a more precise characterization of the identification target. Does the method identify the latent representation and its distribution, or primarily the latent DAG and model parameters? This distinction would help clarify the contribution to causal representation learning. The connection to factor analysis also deserves further discussion: which ambiguities are resolved by the instruments, and which require the additional structural and parameter assumptions?

  4. Could the authors clarify whether latent variables are allowed to causally influence components of (W)? If such relationships are excluded, it would be useful to explain where this restriction enters the identification argument and what happens when it is violated.

  5. Equation (1) appears to rule out direct causal relationships among the observed variables. How essential is this restriction to the identifiability results? A discussion of whether the framework can accommodate such relationships would help clarify its scope.

Questions for authors

See Weaknesses.

2 comments

Thank you for your valuable comments.

W1: plausible instruments and full latent coverage

If some latent variables lack IVs, partial identification may still be possible; the precise characterization of such partial identifiability could be an interesting future work. Full coverage is not unique to our method though; it also appears in several interventional-CRL guarantees, albeit the required interventions and identification targets differ [1–3].

This coverage requirement also has a direct counterpart in IV-based causal discovery with observed variables: Li et al. [4, Assumption 1C] and Wang et al. [5, Assumption 1(B)] assume that every primary variable has at least one valid instrument. Our assumption applies the same requirement to the latent variables.

A larger candidate IV pool can increase the chance that a valid covering subset is present; our model already allows null IV columns, so not every candidate needs to affect a latent variable.

A potential application is genetically informed psychological measurement: WW contains candidate SNPs (germline genetic variants), ZZ contains latent psychological traits such as anxiety and depression, and XX contains questionnaire measurements of those traits [6,7]. A SNP satisfying our assumptions directly affects one latent trait, while its effects may propagate to other traits and measurements through the modeled causal pathways.

W2: distinct roles of IVs and the sparsity assumptions

Assumptions A1–A4 (full-rank loadings, nonsingular instrument covariance, IV exogeneity, full coverage, and single targeting) identify the latent dimension, IV target groups, and total-effect matrix M=L(I−B)−1M=L(I-B)^{-1} up to permutation and scaling. This step does not require sparsity.

To identify the latent DAG BB and loadings LL from MM, we additionally use loading sparsity in A5 together with P1–P3: P1 prevents premature fits, P2 identifies each node's child set, and P3 identifies its edge coefficients. The loadings then follow from L=M(I−B)L=M(I-B).

Between sparse loadings and sparse latent DAG, sparse loadings are the more central sparsity restriction for this identification argument. The outdegree bound qq primarily limits the computational search and can be relaxed to q=d−1q=d-1, removing the degree restriction, provided P1–P3 hold over the enlarged candidate family.

W3: structural parameters, latent distributions, and factor analysis

The identification target is (i) dd, (ii) the nonnull IV target partition, (iii) B,LB,L, and (iv) IV coefficients, up to permutation and nonzero scaling. With unrestricted measurement noise, the result does not generally identify P(Z)P(Z), P(Z∣X,W)P(Z\mid X,W), or each subject's latent realization. Instruments identify total-effect directions that ordinary factor analysis leaves rotationally ambiguous; additional restrictions identify their structural decomposition.

For example, take B=0,Γ=1,L=(1,0)⊤B=0,\Gamma=1,L=(1,0)^\top, Z=W+εZ=W+\varepsilon, and independent Gaussian variables with variances Var⁡(W)=1\operatorname{Var}(W)=1, Var⁡(ε)=a\operatorname{Var}(\varepsilon)=a, Var⁡(η1)=1−a\operatorname{Var}(\eta_1)=1-a, Var⁡(η2)=1\operatorname{Var}(\eta_2)=1, where 0<a<10<a<1. Every aa gives the same joint law of (X,W)(X,W) and the same fixed structural coefficients, but Var⁡(Z)=1+a\operatorname{Var}(Z)=1+a differs. This demonstrates the remaining ambiguity in ZZ. If measurement noise is absent, as considered in some existing CRL work, identified full-rank LL does allow the recovery of ZZ.

W4: causal effects from latent variables into instruments

Our method assumes WW to be exogenous (they are upstream instruments), which is encoded in condition A2 (moment exogeneity). The IV response matrix is Θ=Cov⁡(X,W)Cov⁡(W)−1\Theta=\operatorname{Cov}(X,W)\operatorname{Cov}(W)^{-1}. Dropping moment exogeneity changes the identity Θ=MΓ\Theta=M\Gamma to

\{M\operatorname{Cov}(\varepsilon,W)+\operatorname{Cov}(\eta,W)\} \operatorname{Cov}(W)^{-1}.$$ Latent-to-instrument effects generally induce the additional cross-moment term and can destroy the ray structure, which in turn destroys the IV total-effect direction identification in Proposition 1, which is prerequisite of the latent-graph and loading identification. ## W5: direct causal edges among measurements If $X=AX+LZ+\eta$, with $I-A$ invertible, then $$X=L_{\mathrm{eff}}Z+\eta_{\mathrm{eff}},\qquad L_{\mathrm{eff}}=(I-A)^{-1}L,\quad \eta_{\mathrm{eff}}=(I-A)^{-1}\eta.$$ The moment argument applies if its assumptions hold for this effective model. However, sparse direct loadings may become dense after multiplication by $(I-A)^{-1}$, and the effective loading matrix does not separately identify $A$ and $L$. Thus some systems with observed-to-observed edges can be represented, but recovering their additional measurement graph is outside the present guarantee.

References

  1. Squires et al. (2023). Linear Causal Disentanglement via Interventions.
  2. Zhang et al. (2023). Identifiability Guarantees for Causal Disentanglement from Soft Interventions.
  3. Varici et al. (2025). Score-based Causal Representation Learning: Linear and General Transformations.
  4. Li, Shen and Pan (2023). Inference for a Large Directed Acyclic Graph with Unspecified Interventions.
  5. Wang, Shen and Pan (2024). Causal Discovery with Generalized Linear Models through Peeling Algorithms.
  6. Grotzinger et al. (2019). Genomic structural equation modelling provides insights into the multivariate genetic architecture of complex traits.
  7. Thorp et al. (2021). Symptom-level modelling unravels the shared genetic architecture of anxiety and depression.
Reviewer_TmatOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper considers the problem of identifying the causal structure among unobserved and observed variables in a linear SCM. Unlike existing work, they assume that "instrumental variables" WW exist, where each unobserved variable is affected by at least one instrumental variable, but the correspondence is unknown. The authors propose a two-stage unique identification algorithm given samples of (X,W)(X, W). In the first step, they recover the influence matrix Θ\Theta from data, which represents the total causal effects from each instrumental variable to each observed variable. Subsequently, they show that the sink variable in the underlying DAG among unobserved variables can be recovered recursively given Θ\Theta, and provide necessary and sufficient conditions for unique identification of the underlying causal model. The authors evaluate the performance of the proposed algorithm using both synthetic and real data.

Critical strengths
  1. This work considers an interesting setting for uniquely recovering linear SCM with latent variables. Most existing methods assume that either there are no causal connections among latent variables, or the observed variables are pure children of the latent variables. This work relaxes both assumptions by introducing instrumental variables, although some proposed conditions may be overly restrictive.
  2. The overall presentation is clear and easy to follow. The theoretical results are well-explained with examples.
Critical weaknesses
  1. The proposed conditions may be overly restrictive. P2 and P3 in Theorem 2 seem to be non-trivial, which may insert strict assumptions on the underlying graph structure. (See Q1 below).
  2. Availability of matrix Θ\Theta, which represents the total causal effect from the exogenous noises of some latent variables to observed variables, could also reduce the problem complexity in recovering the underlying causal graph. It might be interesting to compare the derived conditions with existing conditions on causal discovery in linear SCM with non-Gaussian noises, where total causal effects among variables are heavily used to identify the underlying causal structure.
Questions for authors
  1. How likely are P2 and P3 in Theorem 2 satisfied? Do P2 and P3 require certain specific structures on the underlying model, or the coefficients?
  2. Are there any theoretical results on the extent of identifiability when P2 or P3 are not satisfied? Can the causal structure be partially recovered in this case?
  3. Can a continuous optimization algorithm be used in recovering the true causal structure under the theoretical guarantee?
3 comments

Thank you for your valuable comments.

W1 and Q1: restrictiveness and interpretation of P2 and P3

Let mjm_j and ℓj\ell_j denote the jjth columns of MM and LL, and let ss and qq denote the loading-support and latent-outdegree bounds. Write Cj=ch⁡(j)C_j=\operatorname{ch}(j), so that mj=ℓj+MCjbjm_j=\ell_j+M_{C_j}b_j. P2 distinguishes the child set; P3 distinguishes the coefficients conditional on that set. These restrict the joint latent and measurement model beyond acyclicity and sparsity.

P2 excludes child sets that can be reduced or replaced without violating loading sparsity. Among candidate sets available by the correct peeling round, every incorrect TT with ∣T∣≤∣Cj∣|T|\le |C_j| must satisfy

min⁡β∥mj−MTβ∥0>s.\min_\beta\|m_j-M_T\beta\|_0>s.

This rules out both omitting true children and substituting an equally small set of other available nodes. A concrete excluded support arrangement occurs when, for a proper subset U⊂CjU\subset C_j,

∣supp⁡(ℓj) ∪⋃k∈Cj∖Usupp⁡(mk)∣≤s.\left|\operatorname{supp}(\ell_j)\ \cup \bigcup_{k\in C_j\setminus U}\operatorname{supp}(m_k)\right|\le s.

Then subtracting only the children in UU leaves an ss-sparse residual: the omitted children's contributions can be absorbed into the loading vector. Thus those edges are not required by the permitted measurement sparsity.

To isolate P2, let an unrelated latent direction exactly replace the true child. Take d=3,p=4,s=2,q=1d=3,p=4,s=2,q=1, with only the edge 1→21\to2, whose coefficient is 2, and let

\ell_2=(0,1,1,0)^\top,\qquad \ell_3=(1,c,0,0)^\top,\qquad c\ne0.$$ Every loading column has exactly $s=2$ nonzeros, and $L$ has full column rank. The total-effect directions are $m_1=(1,2,2,1)^\top$, $m_2=\ell_2$, and $m_3=\ell_3$. At $c=2$, both the true singleton $\{2\}$ and the incorrect singleton $\{3\}$ are feasible: $$m_1-2m_2=(1,0,0,1)^\top, \qquad m_1-m_3=(0,0,2,1)^\top.$$ Both residuals are two-sparse, so P2 fails. Nevertheless, P1 holds because $m_1$ has four nonzeros and cannot be resolved before the sinks. P3 also holds: on the true child set, $$m_1-\beta m_2=(1,2-\beta,2-\beta,1)^\top$$ is two-sparse only when $\beta=2$. The ambiguity arises from exact alignment of the unrelated latent loading with the true child's propagated effect. Indeed, $$m_1-\gamma m_3=(1-\gamma,2-c\gamma,2,1)^\top$$ is two-sparse only if $\gamma=1$ and $c=2$. Perturbing $c$ away from 2 while preserving the graph and loading supports removes this alternative and restores P2. Hence P2 excludes interchangeable child representations from coefficient alignments or support redundancies, without prohibiting any latent DAG motif. **P3 excludes alternative edge coefficients that can be compensated by another sparse loading vector.** Replacing $b_j$ by $b_j+\delta$ gives residual $\ell_j-M_{C_j}\delta$. Hence P3 is exactly the requirement $$\|\ell_j-M_{C_j}\delta\|_0>s \qquad\text{for every }\delta\ne0.$$ Changing the coefficients must therefore introduce more nonzero loadings than can be removed through cancellation or accommodated by unused capacity under the bound $s$. In particular, full column rank of $M_{C_j}$ alone is insufficient: an alternative coefficient vector may change which loading entries are zero. For a concrete excluded parameter equality, consider the single edge $1\to2$ with coefficient 1, $p=6,s=4,q=1$, and $$\ell_1=(1,a,1,1,0,0)^\top,\qquad \ell_2=m_2=(1,1,0,0,1,1)^\top,\qquad a>0.$$ Here $m_1=(2,a+1,1,1,1,1)^\top$. P1 and P2 hold throughout this family. P3 holds for $a\ne1$, but fails at $a=1$: both coefficients 1 and 2 yield four-sparse residuals, $$m_1-m_2=(1,1,1,1,0,0)^\top, \qquad m_1-2m_2=(0,0,1,1,-1,-1)^\top.$$ Equality of the two shared-row loading ratios lets a wrong coefficient cancel two entries and introduce two others. Fixing the DAG and loading supports, this isolates a P3 coefficient restriction. P3 also excludes structural ambiguities: if a child's total-effect support lies entirely within its parent's loading support, changing that edge coefficient never introduces a loading outside the permitted support. Inspired by a comment from another reviewer, we now establish an alternative identification via a sparsest loading basis: **Condition (S).** For any $v$ with $\lVert v\rVert_0 \ge 2$, $$\lVert Lv\rVert_0 > \max_{j\in\operatorname{supp}(v)} \lVert\ell_j\rVert_0.$$ Any combination of two or more loading columns must be less sparse than every participating column. This rules out combinations whose cancellations or overlapping supports make them as sparse as an original loading direction. It imposes no prescribed $s,q$ or restriction to a particular DAG motif. Under full rank, (S) is equivalent to uniqueness, up to column scaling and permutation, of the minimum-total-support basis of $\operatorname{col}(L)$. Instruments identify the total-effect directions, and acyclicity then fixes their correspondence with this loading basis. Please see the general comment to all at the very top for a proof sketch.

Q2: what remains identified when P2 or P3 fails?

With P1–P3 dropped, we identify the observational-equivalence class of (B,L)(B,L) under A1–A5. Here A1–A4 require full-rank loadings, nonsingular instrument covariance, IV exogeneity, full coverage, and single targeting; A5 imposes the loading-support and latent-outdegree bounds s,qs,q. A1–A4 identify the total-effect directions MM and the nonnull IV target groups. Fixing the latent labels and scales, define equivalence between admissible structural representations by

\quad\Longleftrightarrow\quad L(I-B)^{-1}=\widetilde L(I-\widetilde B)^{-1}.$$ The equivalence class associated with the identified $M$ has the following characterization: $$\mathcal I_{s,q}(M) =\\{(A,M(I-A)):A\in\mathcal F_{s,q}(M)\\},$$ where $$\mathcal F_{s,q}(M)=\left\\{A\in\mathcal D_d: |\operatorname{ch}_A(j)|\le q,\quad \lVert[M(I-A)]_{\cdot j}\rVert_0\le s \quad\text{for every }j\right\\}.$$ and $\mathcal D_d$ denotes the set of acyclic coefficient matrices. Thus the class contains precisely the decompositions of $M$ that satisfy the original degree and loading-sparsity bounds. This is an observational-equivalence class for the full joint distribution $P(X,W)$. For any $A\in\mathcal F_{s,q}(M)$, retain $\Gamma$ and the entire joint law of $(W,\varepsilon,\eta)$, and set $Z^A=(I-A)^{-1}(\Gamma W+\varepsilon)$ and $L^A=M(I-A)$. Then $$L^AZ^A+\eta=M\Gamma W+M\varepsilon+\eta=X$$ samplewise, while A1–A5 remain satisfied. Conversely, any admissible representation compatible with the observed distribution must have the same normalized $M$ and satisfy these bounds. Consequently, $\mathcal I_{s,q}(M)$ is the sharp identified set: it includes all and only the admissible structural pairs compatible with $P(X,W)$. The class can contain different DAGs, including different skeletons, and need not be a Markov-equivalence class. Edges present in every member are compelled; a coefficient is identified when its value is constant across the class. If any of P1–P3 is retained, it will reduce the equivalence class size. Moreover, we recently realized that P3 is not needed for unique identification of the latent DAG. Under P1–P2, a modified population peeler identifies the graph while retaining the entire feasible coefficient set at each step. At each true reverse depth, P1 prevents premature resolution and P2 selects the unique child set. Later regressions use the fixed columns of $M$, so they do not depend on which feasible coefficient vector was retained. A feasible zero coefficient would permit dropping that child, contradicting P2. Induction therefore identifies the graph, while coefficients and loadings may remain set-valued. Under the alternative condition (S), stated in W1/Q1 above, the unique minimum of $$\min_{A\in\mathcal D_d}\ \sum_{j=1}^d\\|[M(I-A)]_{\cdot j}\\|_0$$ is the true $B$ in the fixed normalization of $M$, and $L=M(I-B)$. Thus point identification is possible under this alternative restriction even when the submitted peeling conditions fail. This identifies the model in the class satisfying (S).

W2: comparison with non-Gaussian linear causal discovery

Let us first clarify that Θ\Theta represents the total causal effects of the observed instruments WW (not exogenous noises of latent variables) on the measurements XX. From the structural equations,

\qquad M=L(I-B)^{-1},$$ and moment exogeneity gives $$\Theta=\operatorname{Cov}(X,W)\operatorname{Cov}(W)^{-1} =M\Gamma=L(I-B)^{-1}\Gamma.$$ Thus the columns of $\Theta$ correspond to observed IVs, rather than unobserved structural noise terms. Under single targeting, each nonzero column satisfies $\theta_k=\Gamma_{t(k),k}m_{t(k)}$; full IV coverage and full rank then identify the directions of $M$. Although $M$ also multiplies the latent disturbances $\varepsilon$, we identify its directions through IV cross-moments without requiring independent non-Gaussian disturbances. Gaussian and correlated disturbances are allowed under the stated exogeneity assumptions. Identifying $M$ still leaves its decomposition into $B$ and the unknown loading matrix $L$: every acyclic $A$ yields $L^A=M(I-A)$ unless the measurement restrictions exclude it. Adams, Hansen and Zhang [1] instead identify a mixing matrix from the noises attached to all observed and latent nodes to the observed variables, using independent non-Gaussian errors or their alternative heterogeneous-variance assumptions. To distinguish notation, write their mixing matrix as $H_X$ and their full structural matrix as $F$. Their decomposition satisfies $H_X(I-F)=I_X$, where $I_X$ is the row-selected identity; ours satisfies $M(I-B)=L$, with an unknown sparse residual $L$. This difference matters for the identifying assumptions. We compare the assumptions directly below. The correspondences describe their identifying roles, although there is no one-to-one equivalence between P1–P3 and their Conditions 1–3. | Assumption in Adams et al. [1] | Restriction | Closest role in IV-CRL | |---|---|---| | Condition 1: Bottleneck | Each node's children are the unique smallest set intercepting every directed path from those children to the observed variables. | P2 similarly rules out replacing the true child set by an equally small or smaller feasible set. | | Condition 2: Strong Non-Redundancy | For a latent $L_i$ and any distinct node $V_j$, $\operatorname{Ch}(L_i)\not\subseteq\operatorname{Ch}(V_j)\cup\{V_j\}$. | Related to the structural separation required by P1–P2: redundant latent roles must not permit premature fits or interchangeable child sets. | | Condition 3: Bottleneck Faithfulness | Each source-to-observed-target mixing submatrix has rank equal to the smallest directed bottleneck between those sets. | Related to the coefficient-alignment restrictions in P1–P3, which exclude spurious fits and coefficient ambiguity. | These assumptions act jointly: in their model, Conditions 1 and 3 ensure coefficient uniqueness on the true child set, while all three support child-set and layer identification [1, supplement, Section 1.4]. Our P3 additionally requires uniqueness while the unknown loading residual may change its support; full column rank of the child directions alone does not establish it. Consequently, their assumptions do not directly imply ours, and their generic bottleneck-faithfulness result does not establish generic P1–P3 for arbitrary loading supports. GIN [2] provides another comparison: it uses independent non-Gaussian errors, faithfulness and group-level measurement redundancy—at least $2r$ measurements pure with respect to each $r$-latent parent group whose members share the same observed-child set. Its independence restrictions and our IV-plus-sparse-residual restrictions address different model classes. ## Q3: continuous optimization with a guarantee In principle, yes: a NOTEARS-style continuous optimization formulation [3] could combine a sparsity-promoting objective with a continuous acyclicity constraint. However, a recovery guarantee would require showing that the sparsity surrogate still identifies the true causal structure and that the optimization procedure reaches the required solution despite nonconvexity. Exploring such formulations and their theoretical guarantees is an interesting direction for future work. ## References 1. Adams, Hansen and Zhang (2021). Identification of Partially Observed Linear Causal Models: Graphical Conditions for the Non-Gaussian and Heterogeneous Cases. 2. Xie et al. (2020). Generalized Independent Noise Condition for Estimating Latent Variable Causal Graphs. 3. Zheng, Aragam, Ravikumar and Xing (2018). DAGs with NO TEARS: Continuous Optimization for Structure Learning.

Comments on the paper

1 comment

We sincerely thank all the reviewers for their valuable comments. We provide point-by-point responses below. In addition, we have upload a revised paper, which includes, among others, two additional learning algorithms, their identifiability theory, and their empirical results. Here we provide the new identifiability condition (replacing the original P1-P3); see the revised paper for proof.

Condition (S). For any vv with ∥v∥0≥2\lVert v\rVert_0 \ge 2,

∥Lv∥0>max⁡j∈supp⁡(v)∥ℓj∥0.\lVert Lv\rVert_0 > \max_{j\in\operatorname{supp}(v)} \lVert\ell_j\rVert_0.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.