EconBase
← Back to paper

Learning Causal Representations from General Environments: Identifiability and Intrinsic Ambiguity

Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.

61,274 characters · 11 sections · 29 citation commands

Rendered from LaTeX for readability, not typeset faithfully. Citation keys are highlighted; maths is left as source; figures, tables and equation environments are summarised rather than reproduced; unrecognised commands are greyed out so nothing is silently dropped. Email addresses are removed.

Learning Causal Representations from General Environments: Identifiability and Intrinsic Ambiguity

abstractWe study causal representation learning, the task of recovering high-level latent variables and their causal relationships in the form of a causal graph from low-level observed data (such as text and images), assuming access to observations generated from multiple environments. Prior results on the identifiability of causal representations typically assume access to single-node interventions which is rather unrealistic in practice, since the latent variables are unknown in the first place. In this work, we provide the first identifiability results based on data that stem from general environments. We show that for linear causal models, while the causal graph can be fully recovered, the latent variables are only identified up to the surrounded-node ambiguity (SNA) varici2023score. We provide a counterpart of our guarantee, showing that SNA is basically unavoidable in our setting. We also propose an algorithm, LiNGCReL which provably recovers the ground-truth model up to SNA, and we demonstrate its effectiveness via numerical experiments. Finally, we consider general non-parametric causal models and show that the same identification barrier holds when assuming access to groups of soft single-node interventions.

Introduction

Artificial intelligence (AI) has achieved tremendous success in various domains in the past decade bengio2013representation,silver2016mastering,bubeck2023sparks. However, current approaches are largely based on learning the statistical structures and relationships in the data that we observe. As a result, it is not surprising that these approaches often capture spurious statistical dependencies between different features, resulting in poor performance in the presence of test distribution shift ovadia2019can,koh2021wilds or adversarial attacks akhtar2018threat,wang2023adversarial.

In view of these pitfalls, a recent line of work has explored the problem of causal representation learning scholkopf2021toward, the task of learning the causal relationships between high-level latent variables underlying our low-level observations. Notably, it is widely believed in cognitive psychology that humans take a causal approach to distill information from the world and make decisions to achieve their goals shanks1988associative,dunbar2004causal,holyoak2011causal. As a result, there is reason to believe that learning causal representations has the potential to significantly improve the power of AI, especially on tasks where performance lags far behind human level geirhos2020shortcut.

Despite such promise, a crucial challenge in causal representation learning is the identifiability of the data generating process; in other words, given the data that we observe, can we uniquely identify the underlying causal model. It has been shown that given observational data (i.e., i.i.d. data generated from a single environment), the model is already non-identifiable in strictly simpler settings where the latent variables are known to be independent locatello2019challenging,locatello2020sober, or where there is no mixing function and one directly observes the latent variables silva2006learning. A natural question that arises is what types of data do we need to acquire to make identification possible.

One line of works assumes access to counterfactual data locatello2020weakly,von2021self,brehmer2022weakly, where some form of weak supervision is typically required. A common assumption here is that one observes data in pairs, where each pair of data is related via sharing part of the latent representation. However, such data is hard to acquire since it requires direct control on the latent representation.

Another line of works ahuja2023interventional,von2023nonparametric,buchholz2023learning,varici2023general instead considers an interventional setting, where the learner observes data generated from multiple different environments. This is arguably a much more realistic setup and reflects common practices in robotics lippe2023biscuit and genomics lopez2023learning,tejada2023causal applications. However, a vast majority of identifiability guarantees assume that each environment corresponds to single-node, hard interventions, which is defined as interventions that isolate a single latent variable from its causal parents. Again, this is quite a restrictive assumption because of two reasons. First, since the latent variables are unknown and need to be learned from data, it is unclear how to perform interventions that only affect one variable. Second, even if one can perform single-node interventions, it may not be feasible to artificially remove causal effects in the data generating processes. This issue is ubiquitous in real-world applications as pointed out in campbell2007interventionist,eberhardt2014direct,eronen2020causal. Motivated by these challenges, we instead consider the following two settings:

itemize• Learning from single-node, soft interventions, which only change the dependency of each latent variable on its direct causes, but does not remove their causal relationships. This setting is considered in seigal2022linear,zhang2022towards,varici2023score,buchholz2023learning, which, however, make parametric assumptions on either the causal model or the mixing function. The most related paper is varici2023score, which proves identifiability under an ambiguity induced by the so-called \enquote{surrounded set}. In this paper, we show that this type of ambiguity is intrinsic in the soft intervention setting. • Learning from fully general and diverse environments. This is a significantly more general and challenging setting, and to the best of our knowledge, no identifiability guarantees are known. khemakhem2020variational,lu2021invariant also consider a multi-environment setup without assuming single-node interventions, but they still assume that the distributions of latent variables all come from a certain parametric family with a fixed set of sufficient statistics.

We make the following contributions:

itemize• For linear causal models, with a linear mixing function, we prove identification results assuming access to data from general and diverse environments ((ref)). To the best of our knowledge, this is the first identification guarantee that makes no assumption on the relationship between the environments. Interestingly, while we show that the causal graph can be exactly recovered, the latent variables are only recovered up to a surrounded-node ambiguity (SNA) ((ref)). • We propose an algorithm, LiNGCReL, in (ref) that provably recovers the ground-truth model up to SNA ((ref)) in the setting of (ref). To demonstrate the effectiveness of LiNGCReL, we present extensive experimental results in (ref) using it to learn causal representations from randomly generated causal models. • Going beyond the linear setting, we study the limit of identification for non-parametric causal models and general mixing functions, assuming access to single-node soft interventions. We show that the model is identifiable up to SNA ((ref)), and then prove that SNA is actually the best achievable guarantee in this setting ((ref)), thereby highlighting a key difference between soft and hard interventions.

Preliminaries

We consider the standard setup of causal representation learning from multiple environments $E\in\mathfrak{E}$. Let $\mathcal{G}=(\mathcal{V},\mathcal{E})$ be the ground-truth causal graph which is directed and acylic (DAG), where $\mathcal{V}=[d]$ and $\mathcal{E}$ describes the causal relationship between different nodes. Each node corresponds to a latent variable ${\bm{z}}_i\in\mathbb{R}$.

For any node $i$, we let $\mathrm{pa}_{\mathcal{G}}(i)$, $\mathrm{ch}_{\mathcal{G}}(i)$, $\mathrm{ans}_{\mathcal{G}}(i)$ and $\mathrm{nd}_{\mathcal{G}}(i)$ to be the set of all parents, children, ancestors and non-descendants of $i$ in $\mathcal{G}$ respectively. We also define $\overline{\mathrm{pa}}_{\mathcal{G}}(i)=\mathrm{pa}_{\mathcal{G}}(i)\cup\{i\}$ and similarly for $\overline{\mathrm{ch}}_{\mathcal{G}}(i),\overline{\mathrm{ans}}_{\mathcal{G}}(i)$ and $\overline{\mathrm{nd}}_{\mathcal{G}}(i)$. Assuming that all probability distributions have continuous densities, the joint density of the latent variables ${\bm{z}}$ can then be written as

equation[equation omitted — 150 chars of source]

where $p_i^E$ is the (unknown) latent generating distribution from environment $E$ at node $i$. Here for a given vector ${\bm{v}}$, we write ${\bm{v}}_i={\bm{e}}_i^\top{\bm{v}}$, and let ${\bm{v}}_S = \left({\bm{v}}_i: i\in S\right)\in\mathbb{R}^{|S|}$.

The causal graph model with density given by (ref) necessarily enjoys the following property:

definition[Causal Markov Condition] For any node $i$, conditioned on ${\bm{z}}_{\mathrm{pa}_{\mathcal{G}}(i)}$, ${\bm{z}}_i$ is independent of ${\bm{z}}_{\mathrm{nd}_{\mathcal{G}}(i)}$. As a consequence, for any node $i,j\in[d]$ and $S\subseteq[d]$, if $S$ $d$-separates $i$ from $j$ (cf. (ref)), then ${\bm{z}}_i\protect\mathpalette{\protect\independenT}{\perp}{\bm{z}}_j\mid{\bm{z}}_S$.

The latent variables ${\bm{z}}$ are unknown to the learner. Instead, the learner has access to observations ${\bm{x}}\in\mathbb{R}^n$ ($n\geqslant d$) from all environments $E\in\mathfrak{E}$ that are related to the latent ${\bm{z}}$ via an injective mixing function ${\bm{g}}$:

equation[equation omitted — 73 chars of source]

The main assumption here that the mixing function is the same across all environments:

assumptionAll environments $E\in\mathfrak{E}$ share the same diffeomorphic mixing function ${\bm{g}}:\mathbb{R}^d\mapsto\mathbb{R}^n$.

In causal representation learning, the goal of the learner is to 1) recover the inverse of the mixing function ${\bm{h}} = {\bm{g}}^{-1}$ (often called the unmixing function) which allows recovering the latent variables given any observations, and, 2) recover the underlying causal graph $\mathcal{G}$. In the remaining part of this paper, we refer to $({\bm{h}},\mathcal{G})$ as the causal model to be learned. Obviously, there would be some ambiguities in learning $({\bm{h}},\mathcal{G})$. For example, choosing a different permutation of the nodes in the causal graph would lead to a different model, and so does element-wise transformations on each component ${\bm{h}}_i$ of ${\bm{h}}$.

commentTo be concrete, the following ambiguities are intrinsic in our problem formulation von2023nonparametric: \begin{definition} Let $\mathcal{H}:\mathbb{R}^n \supseteq \mathcal{X}\mapsto\mathbb{R}^d$ be the space of diffeomorphic mappings from observation to latent, and $\mathfrak{G}$ be the space of all DAGs with $d$ nodes, then for $h,\hat{h}\in\mathcal{H}$ and $\mathcal{G},\hat{\mathcal{G}}\in\mathfrak{G}$, we write $(h,\mathcal{G})\sim_{\mathrm{CRL}} (\hat{h},\hat{\mathcal{G}})$ if there exists a permutation $\pi$ on $[d]$ and some element wise diffeomorphism $\phi:\mathbb{R}^d\mapsto\mathbb{R}^d$ such that \begin{itemize} • For $\forall i,j\in[d]$, $i\in\mathrm{pa}_{\mathcal{G}}(j) \Leftrightarrow\pi(i)\in\mathrm{pa}_{\hat{\mathcal{G}}}(\pi(j))$, and • ${\bm{P}}_{\pi}\circ \hat{{\bm{h}}}= \phi\circ {\bm{h}}$, where ${\bm{P}}_{\pi}$ is a permutation matrix satisfying $({\bm{P}}_{\pi})_{ij}=1 \Leftrightarrow j=\pi(i)$. \end{itemize} \end{definition}

A line of recent works show that the ground-truth model can be identified up to these ambiguities in various settings, assuming access to single-node hard interventions seigal2022linear,von2023nonparametric,varici2023general. On the other hand, some weaker notions of identifiability have also been proposed and studied in the literature seigal2022linear,varici2023score,liang2023causal for single-node soft interventions. Here, we provide a generic definition of single-node soft interventions that we will rely on in this paper.

definitionWe say that a collection of environments $\hat{\mathfrak{E}}$ is a set of (soft) interventions on a subset of latent variables ${\bm{z}}_i, i\in S$ if $\forall i\in[d]$, $\forall E_1, E_2 \in \hat{\mathfrak{E}}, E_1 \neq E_2$, we have $p_i^{E_1}=p_i^{E_2} \Leftrightarrow i\notin S$ (the notation $p_i^E$ comes from (ref)). Equivalently, we write $\mathcal{I}_{{\bm{z}}}^{\hat{\mathfrak{E}}} = S$.

We note that soft interventions are very different from hard interventions, since they do not remove causal relationships between latent variables. The goal of this paper is to address the following question:

center[center omitted — 160 chars of source]

The effect domination set and a notion of identifiability

One may expect that identifiability with soft interventions is not much different from hard interventions, since soft interventions can approximate hard interventions with arbitrary accuracy. However, we will show that this is not the case. At a high level, hard intervention is more powerful than soft intervention because it is capable of isolating a latent variable from its direct cause while soft interventions is not, so soft interventions can sometimes fail to identify the true causal relationship from a mixture of causal effects.

To quantify what kind of ambiguities may arise, we can define the surrounding set for each node in a causal graph $\mathcal{G}$ as follows:

definition(varici2023score) For two nodes $i,j\in[d]$ in $\mathcal{G}$, we say that $j$ is effect-dominated by $i$, or $i\in\mathrm{sur}_{\mathcal{G}}(j)$ if $i\in\mathrm{pa}_{\mathcal{G}}(j)$, and $\mathrm{ch}_{\mathcal{G}}(j)\subseteq\mathrm{ch}_{\mathcal{G}}(i)$. Moreover, we define $\overline{\mathrm{sur}}_{\mathcal{G}}(j)=\mathrm{sur}_{\mathcal{G}}(j)\cup\{j\}$.

In other words, the effect of $j$ on its child set $\mathrm{ch}_{\mathcal{G}}(j)$ is dominated by the effect of $i$. Intuitively, if there exists some $i\in\mathrm{sur}_{\mathcal{G}}(j)$, then ambiguities may arise for the causal variable at node $j$, since any effect of $j$ on any of its child $k$ can also be interpreted as an effect of $i$. In (ref) we discuss a three-node example to further illustrate such ambiguities.

figure[figure omitted — 930 chars of source]

(ref) naturally induces the following relationship between causal models:

definitionUsing the notations in (ref), we write $({\bm{h}},\mathcal{G})\sim_{\mathrm{sur}}(\hat{{\bm{h}}},\hat{\mathcal{G}})$ if there exists a permutations $\pi$ on $[d]$, and a diffeomorphism $\psi:\mathbb{R}^d\mapsto\mathbb{R}^d$ where the $j$-th component of $\psi$, denoted by $\psi_j({\bm{z}})$, is a function of ${\bm{z}}_{\overline{\mathrm{sur}}_{\mathcal{G}}(j)}$ for $\forall j\in[d]$, such that the following holds: \begin{itemize} • For $\forall i,j\in[d]$, $i\in\mathrm{pa}_{\mathcal{G}}\left(j\right) \Leftrightarrow \pi(i)\in\mathrm{pa}_{\hat{\mathcal{G}}}\left(\pi(j)\right)$, • ${\bm{P}}_{\pi}\circ\hat{{\bm{h}}}=\psi\circ{\bm{h}}$, where ${\bm{P}}_{\pi}$ is a permutation matrix satisfying $({\bm{P}}_{\pi})_{ij}=1 \Leftrightarrow j=\pi(i)$. \end{itemize}

In other words, $\sim_{\mathrm{sur}}$ requires that the causal graph to be exactly the same up to some permutation of nodes, but allows each latent variable ${\bm{v}}_i$ to be entangled with ${\bm{z}}_{\mathrm{sur}_{\mathcal{G}}(i)}$. Although not obvious from definition, one can actually check that $\sim_{\mathrm{sur}}$ defines an equivalence relation (see (ref)). Moreover, we will show later that $\sim_{\mathrm{sur}}$ is in general the best that we can hope for in our problem setting.

Identifiability of linear causal models from general environments

In this section, we consider learning causal models from general environments. Specifically, we assume that the environments $E_k, k\in[K]$ share the same causal graph, but the dependencies between connected nodes (latent variables) are completely unknown, and, in contrast with existing literature on single-node interventions, we impose no similarity constraints on the environments. We begin our investigation of identifiability in this setting in the context of linear causal models with a linear mixing function.

Problem setup

Formally, we assume the following generative model in $K$ distinct environments $\mathfrak{E}=\left\{E_k: k\in[K]\right\}$ with data generating process

equation[equation omitted — 165 chars of source]

where the matrix ${\bm{A}}_k$ satisfies $({\bm{A}}_k)_{ij}\neq 0$ if and only if $j\to i$ in $\mathcal{G}$. We refer to $({\bm{A}}_k,\bm{\Omega}_k)$ as the weight matrices of latent variables ${\bm{z}}$ in the environment $E_k$. It is easy to see that (ref) in our general setup translates into the following assumption:

assumptionThe mixing matrix ${\bm{G}}\in\mathbb{R}^{n\times d}$ has full column rank. Equivalently, the unmixing matrix ${\bm{H}}={\bm{G}}^{\dagger}$ has full row rank.

Let ${\bm{B}}_k = \bm{\Omega}_k^{-\frac{1}{2}}({\bm{I}}-{\bm{A}}_k), k\in[K]$, then we have $\epsilon = {\bm{B}}_k{\bm{z}} = {\bm{B}}_k{\bm{H}}{\bm{x}}$. Since in the linear case, there is an easy to see one-to-one correspondence between the matrix ${\bm{H}}$ and the un-mixing function ${\bm{x}}\mapsto{\bm{H}}{\bm{x}}$, we abuse the notation and write $({\bm{H}},\mathcal{G})$ to represent the model instead of $({\bm{h}},\mathcal{G})$. Using ${\bm{h}}_i$ to denote the $i$-th row of ${\bm{H}}$, the following lemma translates (ref) the the linear setting:

lemmaAccording to (ref), $({\bm{H}},\mathcal{G})\sim_{\mathrm{sur}} (\hat{{\bm{H}}},\hat{\mathcal{G}})$ if and only if there exists a permutation $\pi$ on $[d]$, such that $i\in\mathrm{pa}_{\mathcal{G}}(j)\Leftrightarrow \pi(i)\in\mathrm{pa}_{\hat{\mathcal{G}}}(\pi(j))$, and for $\forall i\in[d]$, $\hat{{\bm{h}}}_i\in\mathrm{span}\left\langle{\bm{h}}_j: \pi(j)\in\overline{\mathrm{sur}}_{\mathcal{G}}(i)\right\rangle$.

We also need to make the following assumption on noise.

assumptionThe noise vector $\epsilon\in\mathbb{R}^d$ has independent components, at most one component is Gaussian distributed, and any two components have different distribution.

The non-gaussianity of the noise vectors is a typical assumption in causal discovery within linear models comon1994independent,silva2006learning and is always assumed in the LinGAM setting shimizu2006linear. The assumption that all components have a different distribution is not so standard, but is quite natural in real-world scenarios.

Identifiability guarantee

For each node $i\in[d]$ of $\mathcal{G}$, we use ${\bm{w}}_{k}(i)$ to be the weight vector of environment $E_k$ at node $i$, i.e., ${\bm{w}}_k(i) = \left(({\bm{A}}_k)_{ij}:j\in\mathrm{pa}_{\mathcal{G}}(i)\right) \in \mathbb{R}^{\left\vert\mathrm{pa}_{\mathcal{G}}(i)\right\vert}$. In other words, the structural equation for node $i$ in environment $k$ is of the form:

equation[equation omitted — 142 chars of source]

To obtain our identifiability result, the main assumption we need to make is the non-degeneracy of the weights at each node:

assumptionFor each node $i\in[d]$ of $\mathcal{G}$, we have $\mathrm{aff}\left({\bm{w}}_{k}(i):k\in[K]\right) = \mathbb{R}^{\left\vert\mathrm{pa}_{\mathcal{G}}(i)\right\vert}$ where $\mathrm{aff}(\cdot)$ denotes the affine hull. Equivalently, the weights ${\bm{w}}_{k}(i), k=1,2,\cdots,K$ do not lie in a $\left(\left\vert\mathrm{pa}_{\mathcal{G}}(i)\right\vert-1\right)$-dimensional hyperplane of $\mathbb{R}^{\left\vert\mathrm{pa}_{\mathcal{G}}(i)\right\vert}$.

This assumption is quite mild since it only requires the weight vectors to be in general positions, and it holds with probability $1$ if the weights at each node are sampled from continuous distributions. Moreover, as shown in (ref), it is equivalent to the following assumption.

assumption[Node-level non-degeneracy] We say that the matrices $\left\{{\bm{B}}_k\right\}_{k=1}^K$ are node-level non-degenerate if for all node $i \in [d]$, we have $\dim\mathrm{span}\left\langle({\bm{B}}_k)_i: k\in[K]\right\rangle = \left\vert\mathrm{pa}_{\mathcal{G}}(i)\right\vert+1$, where $({\bm{B}}_k)_i$ is the $i$-th row of ${\bm{B}}_k$.

In the following, we state our main result in this section, which shows that $K=d$ non-degenerate environments suffices for the model to be identifiable up to $\sim_{\mathrm{sur}}$.

theoremSuppose that $K\geqslant d$ and we have access to observations generated from the linear causal model $({\bm{H}},\mathcal{G})$ across multiple environments $\mathfrak{E}=\left\{E_k: k\in[K]\right\}$ with observation distributions $\{\mathbb{P}_{{\bm{x}}}^E\}_{E\in\mathfrak{E}}$, and the data generating processes are given by (ref). Let $(\hat{{\bm{H}}},\hat{\mathcal{G}})$ be any candidate solution with the hypothetical data generating process \begin{equation} \notag {\bm{v}} = \hat{{\bm{A}}}_k{\bm{v}}+\hat{\bm{\Omega}}_k^{\frac{1}{2}}\hat{\epsilon},\quad {\bm{x}} = \hat{{\bm{H}}}^{\dagger}{\bm{v}} \quad in the environment $E_k$ \end{equation} where $\hat{{\bm{H}}}$ has full row rank, such that \begin{enumerate}[label={(\roman*)}] • the observation distribution that this hypothetical model generates in $E_k$ is exactly $\mathbb{P}_{{\bm{x}}}^{E_k}$; • all environments share the same causal graph: $\forall k\in[K]$ and $i,j\in[d]$, $({\bm{A}}_k)_{ij}\neq 0 \Leftrightarrow j\in\mathrm{pa}_{\mathcal{G}}(i)$, $(\hat{{\bm{A}}}_k)_{ij}\neq 0 \Leftrightarrow j\in\mathrm{pa}_{\hat{\mathcal{G}}}(i)$ and $\bm{\Omega}_k, \hat{\bm{\Omega}}_k$ are diagonal matrices with positive entries; • $\{{\bm{B}}_k\}_{k=1}^K$ and $\left\{\hat{{\bm{B}}}_k = \hat{\bm{\Omega}}_k^{-\frac{1}{2}}({\bm{I}}-\hat{{\bm{A}}}_k)\right\}_{k=1}^K$ are non-degenerate in the sense of (ref); • the noise variables $\epsilon$ and $\hat{\epsilon}$ satisfy (ref). \end{enumerate} Then we must have $({\bm{H}},\mathcal{G})\sim_{\mathrm{sur}}(\hat{{\bm{H}}},\hat{\mathcal{G}})$.

The proof of (ref) is given in (ref). In the next section, we will introduce an algorithm, LiNGCReL, that provably recovers the ground-truth up to $\sim_{\mathrm{sur}}$.

To the best of our knowledge, this is the first identifiability guarantee in the literature for causal representation learning from general environments. Remarkably, while the fact that existing works seigal2022linear,zhang2023identifiability focus on single-node interventions seem to suggest that learning from diverse environments is hard, our result indicates that such diversity is actually helpful. Specifically, we show that in the worst case, $\Theta(d^2)$ interventions are required for identifying the ground-truth model under $\sim_{\mathrm{sur}}$:

commentIn previous works seigal2022linear,zhang2023identifiability, it is common to consider single-node soft interventions in the following sense: \begin{assumption} For $\forall 2\leqslant k\leqslant K$, there exists $i_k\in[d]$, such that the structural equation in environment $k$ satisfies (ref) satisfies ${\bm{w}}_k(i) = {\bm{w}}_1(i)$ and $\omega_{k,i,i}=\omega_{1,i,i}$ for $\forall i\neq i_k$. \end{assumption} Let $S_i = \left\{k: 2\leqslant k\leqslant K, i_k=i\right\}, i\in[d]$ and $s_i = \left\vertS_i\right\vert$. Suppose that $\mathcal{G}$ has $e = \sum_{i=1}^d \left\vert\mathrm{pa}_{\mathcal{G}}(i)\right\vert$ edges, then we can view the weight vectors $\left\{({\bm{w}}_k(i), \omega_{k,i,i}): k=1 \text{ or } i=i_k\right\}$ as elements of the Euclidean space $\mathbb{R}^{e+\sum_{k=2}^K \left\vert\mathrm{pa}_{\mathcal{G}}(i_k)\right\vert}\times \mathbb{R}_{+}^{d+K-1}$. Under (ref), the models can be fully determined by these weight vectors. The following result states that if we restrict ourselves to single-node interventions, then in the worst case, $\Theta(d^2)$ interventions are required. \begin{theorem} There exists a causal graph $\mathcal{G}$ with $\Theta(d^2)$ edges, such that for any unmixing matrix ${\bm{H}}\in\mathbb{R}^{d\times n}$ with full row rank, any independent noise variables $\epsilon$, and any $s_i >0, i\in[d]$ such that $s_i \leqslant \left\vert\mathrm{pa}_{\mathcal{G}}(i)\right\vert$ for some $i$, the following holds: except from a null set of the weight space $\mathbb{R}^{e+\sum_{k=2}^K \left\vert\mathrm{pa}_{\mathcal{G}}(i_k)\right\vert}\times \mathbb{R}_{+}^{d+K-1}$ (w.r.t the Lebesgue measure), there must exist a candidate solution $(\hat{{\bm{H}}},\hat{\mathcal{G}})$ and a hypothetical data generating process \begin{equation} \notag \forall k\in[K],\quad {\bm{v}} = \hat{{\bm{A}}}_k {\bm{v}} + \hat{\bm{\Omega}}_k^{\frac{1}{2}}\epsilon,\quad {\bm{x}} = \hat{{\bm{H}}}^{\dagger}{\bm{v}} \end{equation} such that \begin{enumerate}[label={(\roman*$'$)}] • the unmixing matrix $\hat{{\bm{H}}}\in\mathbb{R}^{d\times n}$ has full row rank; • $\forall k\in[K]$ and $i,j\in[d]$, $(\hat{{\bm{A}}}_k)_{ij}\neq 0 \Leftrightarrow j\in\mathrm{pa}_{\hat{\mathcal{G}}}(i)$ and $\hat{\bm{\Omega}}_k$ is a diagonal matrix with positive entries; • for $\forall 2\leqslant k\leqslant K$, the weight matrices $\hat{{\bm{A}}}_k, \hat{\bm{\Omega}}_k$ of environment $E_k$ are from a single-node soft intervention on $E_1$ on node $i_k$, in the sense of (ref), \end{enumerate} but $\mathcal{G}$ is non-isomorphic to $\hat{\mathcal{G}}$. \end{theorem}
theorem[informal version of (ref)] There exists a causal graph $\mathcal{G}$ with $\Theta(d^2)$ edges, such that for any unmixing matrix ${\bm{H}}\in\mathbb{R}^{d\times n}$ with full row rank, any independent noise variables $\epsilon$, and any $0 < s_i \leqslant \left\vert\mathrm{pa}_{\mathcal{G}}(i)\right\vert, i\in[d]$, the ground-truth model $({\bm{H}},\mathcal{G})$ is non-identifiable up to $\sim_{\mathrm{sur}}$ with $s_i$ soft interventions for node $i$, unless the (ground-truth and intervened) weights of the causal model lie in a null set (w.r.t the Lebesgue measure).

A formal version and the proof of (ref) can be found in (ref). On the other hand, by having $d$ single-node interventions per node, (ref) can be satisfied as long as the weights are in general positions, so in this case we have $({\bm{H}},\mathcal{G})\sim_{\mathrm{sur}}(\hat{{\bm{H}}},\hat{\mathcal{G}})$ by (ref). Therefore, (ref) together imply that $\Theta(d^2)$ single-node interventions are necessary and sufficient for identification up to $\sim_{\mathrm{sur}}$.

Given that (ref) only guarantees identification up to $\sim_{\mathrm{sur}}$ that is strictly weaker than full identification, one might naturally ask whether (ref) can be further improved. Our last theorem in this section indicates that $\sim_{\mathrm{sur}}$ is indeed the best one can hope for in our setting, even assuming access to single node, soft intervention.

theorem[Counterpart to (ref), informal version of (ref)] For any linear causal model $({\bm{H}},\mathcal{G})$ and any set of environments $\mathfrak{E}=\left\{E_k: k\in[K]\right\}$ such that all conditions in (ref) are satisfied, there must exists a candidate solution $(\hat{{\bm{H}}},\mathcal{G})$ and a hypothetical data generating process that satisfy the same set of conditions, but \begin{equation} \notag \frac{\partial {\bm{v}}_i}{\partial {\bm{z}}_j} \neq 0,\quad \forall j\in\overline{\mathrm{sur}}_{\mathcal{G}}(i). \end{equation} Moreover, if we additionally assume that the environments are groups of single-node soft interventions, then we can guarantee the existence of $(\hat{{\bm{H}}},\mathcal{G})$ and weight matrices which, besides the properties listed above, are also groups of single-node soft interventions.

LinGCReL: Algorithm for linear non-Gaussian causal representation learning

In this section, we describe Linear Non-Gaussian Causal Representation Learning (LiNGCReL), an algorithm that provably recovers the underlying causal graph and latent variables up to $\sim_{\mathrm{sur}}$ in the infinite-sample limit. At this point, it is instructive to recall the celebrated LiNGAM algorithm shimizu2006linear for linear causal graph discovery. Different from their setting, we only observe some unknown linear mixture of the latent variables. Hence, running linear ICA as in LiNGAM only gives us ${\bm{M}}_k={\bm{B}}_k{\bm{H}}$ rather than the weight matrix ${\bm{B}}_k$ itself.

The key idea in our approach is an effect cancellation scheme that allows us to determine the \enquote{remaining degree of freedom} (RDF) of any node (a.k.a. latent variable) given any subset of its ancestors. This scheme allows us to not only find a topological order of the nodes, but also figure out direct causes by tracking the changes of the RDF. In the following, we present the main steps of LiNGCReL in more details.

Suppose that we are given samples of observations ${\bm{X}}^{(k)}=\left\{{\bm{x}}_i^{(k)}\right\}_{i=1}^N, k\in[K]$ where ${\bm{x}}_i^{(k)}$ is the $i$-th sample from the $k$-th environment.

Step 1. Recover the matrices ${\bm{M}}_k={\bm{B}}_k{\bm{H}}$ Since $\epsilon = {\bm{B}}_k{\bm{z}} = {\bm{B}}_k{\bm{H}}{\bm{x}}$ in the $k$-th environment, so we can use any identification algorithm for linear ICA to recover the matrix ${\bm{M}}_k$. Then we properly rearrange the rows of ${\bm{M}}_k$ so that all ${\bm{M}}_k{\bm{x}}, k=1,2,\cdots,K$ correspond to the same permutation of noise variables. This step is quite standard and details can be found in (ref).

Step 2. Causal representation learning based on ${\bm{M}}_k$ Now we have obtained ${\bm{M}}_k = {\bm{B}}_k{\bm{H}}$, but the unmixing matrix ${\bm{H}}$ is still unknown. We propose (ref) to learn ${\bm{H}}$ and the causal graph $\mathcal{G}$. The main part of (ref) contains a loop that maintains a node set $S$ which, we will show later, is ancestral, i.e., $i\in S \Rightarrow \mathrm{ans}_{\mathcal{G}}(i)\subseteq S$. In each round the algorithm finds a new node $i\notin S$ such that $\mathrm{ans}_{\mathcal{G}}(i)\subseteq S$, and a subroutine Identify-Parents ((ref)) is used to find all parents of $i$. After that, we append $i$ into $S$ and continue until $S$ contains all nodes in $\mathcal{G}$. Finally, the rows of the mixing matrix ${\bm{H}}$ is obtained by intersections of properly-chosen row spaces of ${\bm{M}}_k$.

Both (ref) and (ref) include a crucial step, which we call it orthogonal projection, as described in (ref). At a high level, it helps determine the minimal RDF for ${\bm{z}}_i$ after fixing the latent variables ${\bm{z}}_S$, and this exactly corresponds to the number of parents of ${\bm{z}}_i$ that are not in ${\bm{z}}_S$. We provide a simple example in (ref) to illustrate why this approach works.

The following result states that (ref) can recover the ground-truth causal model up to $\sim_{\mathrm{sur}}$:

theoremSuppose that ${\bm{M}}_k, k\in[K]$ are perfectly identified in Step 1. Let $(\hat{{\bm{H}}},\hat{\mathcal{G}})$ be the solution returned by (ref), then we must have $({\bm{H}},\mathcal{G})\sim_{\mathrm{sur}}(\hat{{\bm{H}}},\hat{\mathcal{G}})$.

The full proof of (ref) is given in (ref). It crucially relies on the following two propositions that reveal how (ref) and the subroutine (ref) work.

algorithm[algorithm omitted — 691 chars of source]
propositionThe following two propositions hold for (ref): \begin{itemize} • $\mathrm{ans}_{\mathcal{G}}(i)\subseteq S \Leftrightarrow$ the if condition in line 8 of (ref) is fulfilled; • the set $S$ maintained in (ref) is always an ancestral set, in the sense that $j\in S \Rightarrow \mathrm{ans}_{\mathcal{G}}(j)\subseteq S$. \end{itemize}
propositionGiven any ordered ancestral set $S$ that contains $\mathrm{pa}_{\mathcal{G}}(i)$ for some $i\notin S$, (ref) returns a set $P_i \subseteq S$ that is exactly $\mathrm{pa}_{\mathcal{G}}(i)$.
algorithm[algorithm omitted — 842 chars of source]
algorithm[algorithm omitted — 1,437 chars of source]

Experiments

In this section, we present our experimental setup and results for LiNGCReL. Note that LiNGCReL as described in the previous section only works in the population regime. When the number of samples is limited, two main challenges in implementing LiNGCReL are to accurately compute the dimension of a subspace (line 6 of (ref) and line 8 of (ref)), and to find a vector in the intersection of multiple subspaces (line 20, (ref)). Due to space limit, the implementation details are described in (ref).

Experimental setup. We generate the independent noise variables from generalized Gaussian distributions $p_{\beta}(x)\propto \exp\left(-\left\vertx\right\vert^{\beta}\right)$ with parameters $\beta_k = 0.2 k^2, k=1,2,\cdots,d$, multiplied by normalization constants to make their variances equal to $1$. The ground-truth causal graph is generated by first fixing a total order of the vertices, say $1,2,\cdots,d$, then add directed edges $i\to j (i<j)$ according to i.i.d. Bernoulli($p$) distributions, where $p\in(0,1)$. The non-zero entries of matrices ${\bm{B}}_k$ and ${\bm{H}}$ are all generated independently from Gaussian distributions. For simplicity, we focus on the case $n=d$ since recovery of the latent graphs only requires information from $d$ components of ${\bm{x}}$.

figure*[figure* omitted — 3,127 chars of source]

Metrics of estimation error. Since causal representation learning seeks to learn both the causal graphs and the latent variables, for each output of our algorithm we first check if it exactly recovers the ground-truth causal graph. Then, recall that the latent variables and the observations are related by ${\bm{z}} = {\bm{H}}{\bm{x}}$, given any output unmixing matrix $\hat{{\bm{H}}}$ from (ref), we define the relative estimation error $\Delta_i$ for ${\bm{z}}_i$ as the solution of the following optimization problem:

equation[equation omitted — 494 chars of source]

where signed permutation is allowed here since the noise distribution in our experiments is symmetric and the order of latent variables ${\bm{z}}_i, i=1,2,\cdots,d$ does not matter. We refer to the errors $\Delta_i$ defined in (ref) as the SNA error. The SNA error measures how much of the row $\hat{\hat{{\bm{h}}}}_i$ that we learn is contained in the span of the ground-truth rows ${\bm{h}}_j, j\in\overline{\mathrm{sur}_{\mathcal{G}}}(i)$. Indeed, recall that given any observation ${\bm{x}}$, the ground-truth latent variable is ${\bm{z}} = {\bm{H}}{\bm{x}}$ while our algorithm outputs $\hat{{\bm{v}}}_i = \hat{\hat{{\bm{h}}}}_i^\top{\bm{x}}$, so the SNA error essentially captures whether the recovered latent variable is close to some linear mixture of latent variables in the effect-dominating set of $i$. When the SNA error is zero for some node $i$, we know that the recovered latent variable at node $i$ is exactly a linear mixture of the ground-truth latent variables in $\overline{\mathrm{sur}_{\mathcal{G}}}(i)$, according to (ref).

We also define the true error for estimating each latent variable. Formally, let $\hat{\hat{{\bm{H}}}}$ be the unmixing matrix that corresponds to the solution of (ref), then we define the true estimation error $\tilde{\Delta}_i$ of ${\bm{z}}_i$ as

equation[equation omitted — 158 chars of source]

Results. We randomly sample $100$ causal models with size $d=5$, $30$ causal models with size $d=8$ ad $30$ causal models of size $d=10$. In light of (ref), for each $d\in\{5,8,10\}$, we sample data from $K=d$ randomly chosen environments; for $d=5$ we also consider $K=20$ to study how different choices of $K$ can affect the result. We run LiNGCReL for each model with different sample sizes, compute the SNA error and true error of the obtained solution from (ref) respectively for each latent variable, and check whether the ground-truth causal graph is exactly recovered.

(ref) shows how the average SNA error (over all latent variables) and the accuracy of graph recovery changes when sample size grows. We can see LiNGCReL successfully recovers about $80\%$ of all models within each category, and the median of the average SNA error is smaller than $1\%$. Moreover, by comparing (ref) with (ref), one can observe that if we fix the total number of samples but choose a larger $K$ (i.e., fewer samples per environment), LiNGCReL can still achieve the same level of performance compared with the choice $K=d$. Intuitively, this is because $K \gg d$ vectors sampled from an $r (r\leqslant d)$ dimensional subspace are unlikely to approximately lie in an ($r-1$)-dimensional subspace, so that the calculation of line 6 of (ref) and line 8 of (ref) can be more accurate. We leave a better and quantitative understanding of the trade-off between $d$ and $K$ to future work.

SNA error v.s. true error. To understand the implication of our theory, we dive deeper by looking into the learning outcome of LiNGCReL on a specific model, of which the causal graph is shown in (ref).

In (ref), we list the surrounding set of each node and the corresponding SNA error and true error. We can see that if $\mathrm{sur}_{\mathcal{G}}(i)=\emptyset$, the two errors equal and both are small, but if $\mathrm{sur}_{\mathcal{G}}(i)\neq\emptyset$, the true error is much larger than the SNA error. This indicates that LiNGCReL indeed learns the ground-truth model up to $\sim_{\mathrm{sur}}$, as (ref) predicts.

commentResults. We randomly sample $10$ causal graphs of size $d=5$ with $p=0.5$, and the models are generated as described above. These causal graphs are plotted in (ref) in (ref). For each model, we sample $N_{\max}=2\times 10^4$ data. The results for graph recovery and for both the SNA and the true estimation errors are given in (ref). Note that our algorithm only guarantees the SNA error to be small; indeed, we will show in (ref) that this is the best one can hope for. Nonetheless, we still include the true errors for comparison. We can see that with $N_{\max}$ samples, $9$ out of $10$ graphs are exactly recovered, while for all graphs the latent variables are estimated with low SNA errors. Meanwhile, by comparing with (ref), one can find that if $\mathrm{sur}_{\mathcal{G}}(i)=\emptyset$, then the SNA error is the same as the true error. Otherwise, the true error can be very large. This empirical observation is consistent with our theory. \begin{table*}[t] \caption{Results for $d=5$ and $p=0.5$. From each causal model we sample $N_{\max}=2\times 10^4$ data. We report the results for recovering the causal graphs and estimation errors of latent variables. The SNA errors are defined as $\Delta_i, i\in[d]$ where $\bm{\Delta}$ is the solution to (ref), and the true errors are defined in (ref).} \resizebox{0.8\textwidth}{!}{ \begin{tabular}{cccccccccccc} \toprule & & \multicolumn{2}{c}{$z_1$} & \multicolumn{2}{c}{$z_2$} & \multicolumn{2}{c}{$z_3$} & \multicolumn{2}{c}{$z_4$} & \multicolumn{2}{c}{$z_5$} \\ \cmidrule(l){3-12} \multirow{-2}{*}{Graph No.} & \multirow{-2}{*}{Recovery} & SNA err. & True err. & SNA err. & True err. & SNA err. & True err. & SNA err. & True err. & SNA err. & True err. \\ \midrule 1 & \Checkmark & 2.1e-2 & 2.1e-2 & 4.4e-3 & 4.4e-3 & 6.5e-3 & \texttt{2.5e-1} & \texttt{9.0e-2} & \texttt{7.5e-1} & \texttt{4.8e-2} & \texttt{9.9e-1} \\ 2 & \Checkmark & \texttt{7.3e-3} & \texttt{7.3e-3} & \texttt{1.4e-2} & \texttt{1.4e-2} & \texttt{1.1e-2} & \texttt{4.6e-2} & \texttt{3.4e-2} & \texttt{9.6e-1} & \texttt{2.4e-3} & \texttt{9.8e-1} \\ 3 & \Checkmark & \texttt{3.8e-3} & \texttt{3.8e-3} & \texttt{8.3e-3} & \texttt{8.3e-3} & \texttt{2.2e-3} & \texttt{2.2e-3} & \texttt{2.4e-2} & \texttt{4.4e-1} & \texttt{6.7e-3} & \texttt{6.2e-1} \\ 4 & \Checkmark & \texttt{3.2e-3} & \texttt{3.2e-3} & \texttt{1.7e-3} & \texttt{4.6e-1} & \texttt{1.1e-2} & \texttt{6.6e-1} & \texttt{2.2e-2} & \texttt{2.2e-2} & \texttt{2.8e-2} & \texttt{9.3e-1} \\ 5 & \XSolidBrush & \texttt{4.6e-3} & \texttt{4.6e-3} & \texttt{1.6e-2} & \texttt{1.6e-2} & \texttt{9.8e-2} & \texttt{7.3e-1} & \texttt{6.9e-3} & \texttt{9.0e-1} & \texttt{9.9e-13} & \texttt{9.7e-1} \\ 6 & \Checkmark & \texttt{4.8e-3} & \texttt{4.8e-3} & \texttt{8.2e-3} & \texttt{8.2e-3} & \texttt{4.2e-3} & \texttt{1.4e-1} & \texttt{2.4e-3} & \texttt{2.1e-1} & \texttt{1.2e-2} & \texttt{1.2e-2} \\ 7 & \Checkmark & \texttt{5.4e-3} & \texttt{5.4e-3} & \texttt{5.6e-3} & \texttt{5.6e-3} & \texttt{1.8e-3} & \texttt{1.9e-1} & \texttt{1.6e-3} & \texttt{9.9e-1} & \texttt{6.5e-3} & \texttt{6.5e-3} \\ 8 & \Checkmark & \texttt{2.8e-3} & \texttt{2.8e-3} & \texttt{4.4e-3} & \texttt{4.4e-3} & \texttt{3.4e-2} & \texttt{3.4e-2} & \texttt{2.8e-3} & \texttt{8.5e-1} & \texttt{1.5e-2} & \texttt{9.7e-1} \\ 9 & \Checkmark & \texttt{6.5e-3} & \texttt{6.5e-3} & \texttt{5.8e-3} & \texttt{5.8e-3} & \texttt{1.1e-3} & \texttt{5.3e-1} & \texttt{3.2e-2} & \texttt{8.3e-1} & \texttt{4.5e-3} & \texttt{9.7e-1} \\ 10 & \Checkmark & \texttt{6.9e-3} & \texttt{6.9e-3} & \texttt{1.7e-2} & \texttt{1.7e-2} & \texttt{1.0e-2} & \texttt{6.9e-1} & \texttt{1.2e-2} & \texttt{1.0} & \texttt{4.0e-4} & \texttt{9.6e-1} \\ \bottomrule \end{tabular} } \end{table*} The fifth graph fails to be recovered by our algorithm since there is an underestimation of the matrix rank in (ref), line 6, which is due to the existence an extremely small singular value. On the other hand, choosing a smaller tolerance level may lead to an overestimation of the matrix rank, especially when the number of samples is limited. In other words, there exists a dichotomy between identifying the \enquote{signal} and ignoring the \enquote{noise}. This tradeoff is further discussed in (ref), which also includes additional experiment results on larger graphs.

Identification limit of general causal models with soft interventions

While (ref) guarantees identifiability with general environments, it only applies to linear causal models. In this section, we show that if we have access to single-node soft interventions, then we can identify general non-parametric causal models up to $\sim_{\mathrm{sur}}$. To obtain our identifiability result, we also require that the environments are non-degenerate in the following sense:

definition[Non-degeneracy set of interventions] Let $\hat{p}_k\left({\bm{z}}_i\mid{\bm{z}}_{\mathrm{pa}_{\mathcal{G}}(i)}\right), k\in[K_i]$ be conditional probability densities at node $i$, then $\left\{\hat{p}_k\right\}_{k=1}^{K_i}$ is said to be non-degenerate on node $i$ at point $\hat{{\bm{z}}}\in\mathbb{R}^d$ if all these conditional densities are well-defined and positive at $\hat{{\bm{z}}}$, and the matrix \begin{equation} \notag \left[\frac{\partial\left(\hat{p}_1/\hat{p}_k\right)}{\partial {\bm{z}}_j}\right]_{2\leqslant k\leqslant K_i, j\in\bar{\mathrm{pa}}_{\mathcal{G}}(i)}\Biggm\vert_{{\bm{z}}=\hat{{\bm{z}}}} \in \mathbb{R}^{(K_i-1)\times \left(\left\vert\mathrm{pa}_{\mathcal{G}}(i)\right\vert+1\right)} \end{equation} has full row rank. Moreover, we say that $\left\{\hat{p}_k\right\}_{k=1}^{K_i}$ is non-degenerate in a point set ${\bm{O}}$ if for all $\hat{{\bm{z}}}\in{\bm{O}}$, it is non-degenrate at $\hat{{\bm{z}}}$.

The following lemma shows how (ref) is related to (ref) in the linear setting:

lemmaSuppose that $\hat{p}_k({\bm{z}})=\prod_{i=1}^d \hat{p}_k\left({\bm{z}}_i\mid{\bm{z}}_{\mathrm{pa}_{\mathcal{G}}(i)}\right), k\in[K]$ be probability distributions of latent variables ${\bm{z}}$ generated from the linear causal models (ref), such that for $\forall i\in[d]$, $\hat{p}_k\left({\bm{z}}_i\mid{\bm{z}}_{\mathrm{pa}_{\mathcal{G}}(i)}\right), k\in[K]$ are non-degenerate on node $i$ in the sense of (ref). Then the corresponding matrices ${\bm{B}}_k, k\in[K]$ satisfy (ref).

Now we are ready to state our main result in this section:

theoremSuppose that we have access to observations generated from multiple environments $\{P_{{\bm{X}}}^E\}_{E\in\mathfrak{E}}$. Let $\left(\hat{{\bm{h}}},\hat{\mathcal{G}}\right)$ be any candidate solution with data generated according to (ref) with latent variables ${\bm{v}} = \hat{{\bm{h}}}({\bm{x}})$ and joint distribution $q_E$ with factors $q_{i}^E$. Assuming that \begin{enumerate}[label={(\roman*)}] • the joint densities $\left\{p_E({\bm{z}})\right\}_{E\in\mathfrak{E}}$ are continuous differentiable on $\mathbb{R}^d$ with common support ${\bm{O}}_{{\bm{z}}}$, and $\left\{q_E({\bm{v}})\right\}_{E\in\mathfrak{E}}$ are continuous differentiable on $\mathbb{R}^d$ with common support ${\bm{O}}_{{\bm{v}}}$; • we have access to multiple single-node soft interventions on each node with unknown targets: there exists a partition $\mathfrak{E} = \cup_{i=1}^d\mathfrak{E}_i$ such that $\mathcal{I}_{{\bm{z}}}^{\mathfrak{E}_i} = \{\pi(i)\}, \mathcal{I}_{{\bm{v}}}^{\mathfrak{E}_i} = \{\pi'(i) \},\forall i\in[d]$ for some unknown permutations $\pi$ and $\pi'$ on $[d]$; • the intervention distributions on each node are non-degenerate in the sense of (ref): there exists ${\bm{N}}_{{\bm{z}}}\subseteq{\bm{O}}_{{\bm{z}}}$ and ${\bm{N}}_{{\bm{v}}}\subseteq{\bm{O}}_{{\bm{v}}}$ satisfying ${\bm{N}}_{{\bm{z}}}^{\mathrm{o}} = {\bm{N}}_{{\bm{v}}}^{\mathrm{o}} = \emptyset$ where $S^{\mathrm{o}}$ denotes the interior of a set $S$, such that for all $i\in[d]$, $\left\{p_{i}^E(\cdot): E\in\mathfrak{E}_{\pi^{-1}(i)}\right\}$ (resp. $\left\{q_{i}^E(\cdot): E\in\mathfrak{E}_{\pi'^{-1}(i)}\right\}$) is non-degenerate on node $i$ in ${\bm{O}}_{{\bm{z}}}\setminus{\bm{N}}_{{\bm{z}}}$ (resp. ${\bm{O}}_{{\bm{v}}}\setminus{\bm{N}}_{{\bm{v}}}$). \end{enumerate} Then we must have $({\bm{h}},\mathcal{G})\sim_{\mathrm{sur}}(\hat{{\bm{h}}},\hat{\mathcal{G}})$.

Previous works on the identifiability of non-parametric causal models typically require that all the joint distributions are supported on the whole space $\mathbb{R}^d$ von2023nonparametric,liang2023causal,varici2023general. In contrast, we only assume that the densities have common and unknown support across all interventions.

(ref) can be regarded as a soft-intervention version of von2023nonparametric, which assumes access to hard interventions and only need two paired interventions per node. While they are able to show full identifiability, we show in the following that identifiability up to $\sim_{\mathrm{sur}}$ is the best we can hope for with soft interventions.

theorem[Counterpart to (ref), informal version of (ref)] For any causal model $({\bm{h}},\mathcal{G})$ and any set of environments $\mathfrak{E}=\left\{E_k: k\in[K]\right\}$ such that all conditions in (ref) are satisfied, there must exists a candidate solution $(\hat{{\bm{h}}},\mathcal{G})$ and a hypothetical data generating process that satisfy the same set of conditions, but \begin{equation} \notag \frac{\partial {\bm{v}}_i}{\partial {\bm{z}}_j} \neq 0,\quad \forall j\in\overline{\mathrm{sur}}_{\mathcal{G}}(i). \end{equation} Finally, the ambiguity still exists if we additionally assume standard axioms such as causal minimality ((ref)) and faithfulness ((ref)) on the causal model.

Conclusions

This paper studies the limit of learning identifiable causal representations using data from multiple environments. When hard interventions are not available, we provide theory and algorithm for identification up to SNA, and also show that SNA is an intrinsic ambiguity in our setting.

It is interesting to further investigate the setting where we do not assume that the causal model is linear. Moreover, it is important to understand the concrete form of available interventions in real-world applications. For instance, it is suggested that for single-cell genomics, the intervention is sometimes a "mixture" of hard and soft interventions, and sometimes can even reverse the direction of an edge tejada2023causal. Modelling such more complicated interventions appears to be crucial to reveal the underlying causal mechanisms in real-world problems.

Broader Impact

This paper presents work whose goal is to advance the field of Machine Learning and in particular the sub-field Causal Representation Learning. There are many potential societal consequences of our work, especially as it pertains to building more reliable machine learning models, none which we feel must be specifically highlighted here.

\printbibliography