we-solve (beta)

Econ-150 · economics · Economics

Low-dimensional causal functionals of text and image treatments

LLM status Open unverified

Statement

CI-06 |  Econometric theory and causal inference

Research agenda formalization

Definitions and assumptions

Let the treatment space $\mathcal W$ and covariate space $\mathcal X$ be standard Borel spaces, and let $h:\mathcal W\to\R^k$ be a measurable representation. A representation value $z$ does not itself specify an intervention: declare a measurable kernel $K_z(dw\mid x)$ whose pushforward under $h$ is a specified law $\Lambda_z$. Exact fiber interventions have $\Lambda_z$ a point mass; neighborhood interventions can have continuous $\Lambda_z$. For jointly measurable bounded potential outcomes $Y(w)$, define \[ \theta(z)=\int\E[Y(w)\mid X=x]K_z(dw\mid x)P_X(dx). \] Assume consistency, no between-unit interference, and either genuine randomized assignment or the explicitly imposed conditional exchangeability of treatment and potential outcomes. For identification from observational treatment distributions, require $K_z(\cdot\mid x)\ll P(W\in\cdot\mid X=x)$ almost surely, with density ratio bounded by a specified constant.

Formal question

For specified candidate representations $\mathcal H$ and intervention kernels $\mathcal K$, characterize when low-dimensional targets $\theta(z)$ are well defined, identified, and invariant across scientifically equivalent treatment versions. If invariance fails, determine the sharp range of $\theta(z)$ over admissible lifts $K_z$. For mean functions factoring as $m(w,x)=r(h(w),x)$ with $r$ in a declared Hölder ball, determine optimal rates and simultaneous inference. Use a finite positivity-covered domain for exact fibers, or a compact continuous domain with supported neighborhood interventions. Include the statistical cost of learning $h$ on independent or appropriately separated data.

Scope and status

This makes precise the attached complex-treatment question. Scientific interpretability must be supplied through the admissible representations and interventions; it is not a consequence of low dimension. If a fiber has zero observational probability, deterministic fiber interventions can violate positivity. Representation learning cannot remove unmeasured confounding or establish version invariance by itself. These are agenda refinements, not precise conjectures quoted from the source.

References for the source of the problem

Carlos Cinelli, Avi Feller, Guido Imbens, Edward Kennedy, Sara Magliacane, and Jose Zubizarreta, Challenges in Statistics: A Dozen Challenges in Causality and Causal Inference, arXiv:2508.17099v1, 23 August 2025, Sections 3.1.2, paragraph Complex experimental designs, and 3.1.3. https://arxiv.org/html/2508.17099v1.

Common conventions: Source mathematical conventions

Well-defined objects. All probability spaces and measurable variables are specified or inherited from a local statistical model. Expectations used as finite targets require the stated integrability. A decision rule or estimator is measurable with respect to its available information; randomized rules may use an auxiliary independent random variable. Norms without a subscript are Euclidean in finite-dimensional spaces. An optimizer is requested only when attainment is assured; otherwise the question concerns the optimal value and $\varepsilon$-optimal admissible rules for every $\varepsilon>0$.

Identification. A structural model $m\in\mathcal M$ implies an observable law $P_m^{\rm obs}$ and target $\theta(m)$. The target is point identified on $\mathcal M$ if \[ P_{m_1}^{\rm obs}=P_{m_2}^{\rm obs} \quad\Longrightarrow\quad\theta(m_1)=\theta(m_2) \qquad(m_1,m_2\in\mathcal M). \] When this fails, the sharp identified set is \[ \Theta(P;\mathcal M)=\{\theta(m):m\in\mathcal M, P_m^{\rm obs}=P\}. \] Specifying an intervention or estimand does not establish identification. Positivity, exclusion, causal sufficiency, measurement restrictions, and transport assumptions are substantive local inputs, never automatic consequences of notation.

Minimax risk and nontrivial inference. For a statistical experiment with data law $P^{(n)}$, target $\theta(P)$, and nonnegative loss $\ell$, define \[ \mathcal R_n(\mathcal P,\ell) =\inf_{\widehat\theta_n}\sup_{P\in\mathcal P} \E_{P^{(n)}}\ell(\widehat\theta_n,\theta(P)). \] The infimum is over measurable estimators. A sharp rate is a sequence $r_n$ for which $0<c\leq C<\infty$, independent of $n$, satisfies $c r_n\leq\mathcal R_n\leq C r_n$ for all sufficiently large $n$. Squared-error parametric risk is of order $n^{-1}$; the corresponding root-mean-squared error is of order $n^{-1/2}$. These different conventions are distinguished locally. A uniform asymptotic confidence region $C_n$ must satisfy \[ \liminf_{n\to\infty}\inf_{P\in\mathcal P} \Pr_{P^{(n)}}\{\theta(P)\in C_n\}\geq 1-\alpha, \qquad 0<\alpha<1, \] together with an informative diameter or expected-width requirement. Coverage alone permits the vacuous whole parameter space. Testing questions similarly impose both size control and power against a declared separated alternative class.

Binary causal baseline. When an entry explicitly invokes this baseline, one observes independent copies of $O=(X,W,Y)$, with $W\in\{0,1\}$ and potential outcomes $Y(0),Y(1)$. Assume consistency $Y=Y(W)$, no interference, conditional exchangeability $(Y(0),Y(1))\perp W\mid X$, and overlap $\epsilon\leq e(X)\leq1-\epsilon$ almost surely for a fixed $0<\epsilon<1/2$, where \[ e(x)=\Pr(W=1\mid X=x),\qquad m_w(x)=\E[Y\mid W=w,X=x],\qquad \tau(x)=m_1(x)-m_0(x). \] These assumptions identify conditional mean effects almost everywhere under the covariate law. Pointwise or uniform effects require appropriate regularity and a specified support. Continuous treatment, longitudinal data, adaptive experiments, and network interference require their own assumptions in place of this baseline. Bounded outcomes, densities, and smoothness are additional assumptions only where stated.

Function classes. On $[0,1]^d$, $\mathcal H_d^s(L)$ is the isotropic Hölder ball of radius $L>0$: put $k=\lceil s\rceil-1$ and $\gamma=s-k\in(0,1]$, bound derivatives through order $k$, and require the order-$k$ derivatives to be $\gamma$-Hölder with constant at most $L$. Thus integer orders use a Lipschitz final derivative convention. The class $\operatorname{BV}_d(L)$ consists of functions $f\in L^1((0,1)^d)$ whose distributional derivative is a finite vector measure and for which $\|f\|_{L^1}+|Df|((0,1)^d)\leq L$. An $L^\infty$ bound or a particular pointwise version is imposed separately when needed. Sparse and neural-network classes require a declared dictionary or architecture and coefficient bounds.

An anisotropic Hölder class with declared block smoothness indices means coordinatewise Hölder regularity in each block, uniformly in the other blocks, with all corresponding derivative and Hölder bounds at most the stated radius. Derivative orders and the integer-order convention in each block are those just given. Mixed-derivative bounds are additional restrictions only when explicitly stated.

Decision and computational questions. Policy regret compares admissible feasible policies under the same law and constraints. In computational entries, finite data and thresholds have rational encodings; running time is measured in their total bit length. Real-valued equilibrium search uses the explicitly declared algebraic or FIXP encoding. An approximation ratio for maximization is written $\operatorname{OPT}/\operatorname{ALG}$ and for minimization as $\operatorname{ALG}/\operatorname{OPT}$, with zero optima and infeasible instances handled locally. Historical approximation bounds are not automatically current bounds.

Statement licence: CC-BY-4.0 · Source file

Contributions 1

Econ-150-A1 · Exploration · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Econ-150

Sharp version bounds for finite representation fibers

Status: Unresolved. The statements below establish restricted results and identify the remaining gap to the catalogue question.

Abstract

On a finite treatment and covariate domain, bounded-density interventions on one representation fiber form a capacitated simplex. We solve its lower and upper causal-mean problems exactly and characterize when the resulting range collapses. A fixed-density-ratio estimator gives a finite-sample risk bound, while a zero-probability fiber example shows why representation alone cannot create identification.

Short target and finite model

The catalogue asks when a low-dimensional representation of a complex treatment supports identified, version-invariant causal targets. Fix finite sets $\mathcal X,\mathcal W$, a representation $h$, and one representation value $z$. Let $F=h^{-1}(z)$. For each covariate value of positive mass $\pi_x$, the observational treatment probabilities $p_{w|x}>0$ are known. Consistency and conditional exchangeability identify the bounded conditional causal mean $m_{wx}={\mathbb{E}}[Y(w)\mid X=x]$, with $|Y(w)|\le M$.

An exact-fiber intervention chooses probabilities $q_{w|x}$ supported on $F$, satisfying $q_{w|x}\le R p_{w|x}$, for a specified finite density bound $R\ge1$. Its target is $\theta(q)=\sum_x\pi_x\sum_{w\in F}q_{w|x}m_{wx}$. Different choices of $q$ are different treatment versions consistent with the same $z$. Identification for a specified $q$ and equality across all such $q$ are separate questions.

Feasibility and exact sharp bounds

At a given $x$, write $c_w=R p_{w|x}$ and $C_x=\sum_{w\in F}c_w$.
Lemma (The intervention polytope): A feasible exact-fiber intervention at $x$ exists if and only if $C_x\ge1$. When $C_x=1$, it is unique and equals $q_w=c_w$. When $C_x>1$ and $F$ has at least two elements, its affine hull is the full probability hyperplane on $F$.
Proof: Summing the capacity constraints proves necessity. If $C_x\ge1$, $q_w=c_w/C_x$ is feasible, proving sufficiency. Equality forces every capacity to bind. For $C_x>1$, this construction has $0<q_w<c_w$ for every $w$, so a sufficiently small perturbation in any direction of coordinate sum zero remains feasible. This proves the affine-hull assertion. ∎
Order the treatments in $F$ so that $m_{1x}\le\cdots\le m_{kx}$, with a fixed tie rule. Let $j$ be the first index with $\sum_{r=1}^j c_r\ge1$. Define \[ L_x=\sum_{r<j}c_r m_{rx} +\left(1-\sum_{r<j}c_r\right)m_{jx}. \] Define $U_x$ by the same formula after reversing the order of means.
Theorem (Sharp version range): If $C_x\ge1$ for every relevant $x$, the range of causal means over all admissible lifts is exactly \[ \left[\sum_x\pi_x L_x,\ \sum_x\pi_x U_x\right]. \] Both endpoints are attained. The construction takes sorting time on each fiber. It applies to known population means; estimating them is a separate statistical problem.
Proof: The lower-endpoint allocation fills each smallest-mean treatment to capacity until unit mass has been assigned. If any feasible vector puts positive mass on a larger-mean treatment while a smaller-mean one has spare capacity, shifting the minimum of the two available masses weakly lowers the objective. Equivalently, writing $Q_s=\sum_{r\le s}q_r$, summation by parts gives \[ \sum_{r=1}^k q_r m_{rx} =m_{kx}-\sum_{s=1}^{k-1}(m_{s+1,x}-m_{sx})Q_s. \] The greedy vector simultaneously maximizes each $Q_s$ to $\min\{1,\sum_{r\le s}c_r\}$, hence minimizes the mean. Reversing order proves the upper formula. Choices at different $x$ have no coupling constraint, so their endpoints can be attained simultaneously. The product of the feasible simplices is convex; mixing the lower and upper kernels with a common scalar attains every intermediate target value. No value outside the interval is possible by the pointwise bounds. ∎
Corollary (Precisely when version dependence disappears): For each $x$ with $C_x>1$, the conditional target is the same for every admissible lift if and only if $m_{wx}$ is constant over $w\in F$. For $C_x=1$ it is invariant regardless of those means. The aggregate range is a singleton if and only if the corresponding condition holds at every $x$ of positive mass.
Proof: Constancy of means gives invariance directly. Conversely, at an interior feasible point from the lemma one can move a sufficiently small mass from any coordinate to any other. Invariance of the mean forces the two coefficients to agree. If capacities sum to one there is only one lift. Finally the aggregate width is the sum of the nonnegative widths $\pi_x(U_x-L_x)$, which is zero exactly when each is zero. ∎
The capacity-saturated exception matters: equality of a restricted version range need not imply that the representation retains all causal information. It may merely mean that the intervention class allows only one version.

Approximate invariance and estimation

Proposition: If $|m_{wx}-r(h(w),x)|\le\delta$ for every relevant $(w,x)$, any two admissible lifts on the same fiber have target means differing by at most $2\delta$.
Proof: Under either lift, the conditional mean differs from $r(z,x)$ by at most $\delta$. Apply the triangle inequality and average with weights $\pi_x$. ∎
For a specified kernel $q$, let $\omega(w,x)=q_{w|x}/p_{w|x}$, with $q=0$ off the fiber. On $N$ independent observations define $\widehat\theta=N^{-1}\sum_{i=1}^N\omega(W_i,X_i)Y_i$.
Theorem (Known-kernel risk): If $p$ and $q$ are supplied and $0\le\omega\le R$, then the estimator is unbiased and ${\mathbb{E}}(\widehat\theta-\theta(q))^2\le M^2R/N$. If the representation and kernel are learned on an independent training sample, the same statement holds conditionally on that sample for its realized target, provided its kernel satisfies the same bound almost surely.
Proof: Conditioning on $X$ and summing over $W$ proves the expectation identity. Since $\omega^2\le R\omega$ and ${\mathbb{E}}\omega=1$, ${\mathbb{E}}[\omega^2Y^2]\le M^2R$. Independence gives the variance bound. Conditional on independent training, the same calculation applies to the now fixed functions and their own target. It does not bound the difference between a learned representation's target and a prespecified scientific target. ∎

Zero-probability fibers and the remaining gap

Proposition (A positivity obstruction): There are two bounded, exchangeable treatment models with the same observational law but different means under an exact deterministic fiber intervention.
Proof: Let $W$ be uniform on $[0,1]$, let $h(w)=w$, and fix $z\in(0,1)$. In one model set $Y(w)=0$ for every $w$; in the other set $Y(w)={\mathbf{1}}\{w=z\}$. Potential outcomes are deterministic and jointly measurable, so exchangeability holds. Observed $Y(W)$ is zero almost surely under either model. Intervention $W=z$ has mean zero in the first and one in the second. Its point mass is not absolutely continuous with respect to the observational uniform law, exactly violating the maintained intervention restriction. ∎
The finite-fiber range is therefore a genuine identified-set result under explicit positivity, not a general consequence of low-dimensional embeddings. Optimal rates for learned continuous representations, neighborhood interventions, and simultaneous sharp-bound inference remain untreated. The finite formula can serve as a benchmark for those extensions.

References

  • [CF] C. Cinelli, A. Feller, G. Imbens, E. Kennedy, S. Magliacane, and J. Zubizarreta (2025), Challenges in Statistics: A Dozen Challenges in Causality and Causal Inference, arXiv:2508.17099v1, Sections 3.1.2--3.1.3. https://arxiv.org/html/2508.17099v1. Source of the complex-treatment agenda; the finite capacity-bound calculation specifies a restricted intervention experiment.

Completion Estimate

40%

Go to contribution →