📊 statistics

A Characterization of the Orthocomplement of the Tangent Space of Semiparametric Markov Models

This paper derives closed-form expressions for the orthogonal complement of the tangent space in general semiparametric Markov models, thereby enabling the characterization of all influence functions and facilitating efficient semi-parametric inference for models defined by undirected graphs, chain graphs, and acyclic directed mixed graphs where such characterizations were previously unavailable.

Original authors: Trung Phung, Ilya Shpitser

Published 2026-07-28

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Trung Phung, Ilya Shpitser

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but the clues you have are messy. Some clues point directly to the culprit, while others are red herrings or just background noise. In the world of statistics and data science, researchers often face a similar problem: they have a mountain of data and want to find a specific number (like the average effect of a medicine or the strength of a relationship between two variables). The challenge is that the data is governed by hidden rules of "independence"—meaning some things happen without being influenced by others. To get the most accurate answer possible, statisticians use a special toolkit called "semiparametric theory." Think of this toolkit as a way to build the perfect, most efficient detective kit that works no matter how messy the background noise gets. The key to building this kit is understanding the "shape" of the rules that govern the data. If you know the shape, you can filter out the noise and get a crystal-clear answer. But for a long time, while detectives knew the shape of the rules for some simple cases, they were completely lost when the rules got complicated and mixed together in weird ways.

This paper is about a team of researchers who finally figured out how to describe the shape of those complicated, mixed-up rules. They focused on a specific type of statistical model called a "Markov model," which is just a fancy way of saying a system defined by who is independent of whom. While scientists already knew how to handle these models when they were simple and straight-line (like a family tree), they hit a wall when the models got tangled, like a knot of strings or a web of connections. The researchers discovered a clever new way to untangle these knots. Instead of trying to solve the whole messy knot at once, they realized they could break it down into tiny, simple loops, solve each one, and then stitch the solutions together. They found a "closed-form" recipe—a clear, step-by-step instruction manual—for describing the part of the data that represents pure noise (which they call the "orthogonal complement of the tangent space"). This might sound like gibberish, but think of it as finding the exact formula for the static on a radio so you can tune it out perfectly.

The authors didn't just find a theoretical formula; they showed exactly how to use it to build better, more efficient estimators. In the paper, they demonstrate that by using their new recipe, you can take a rough guess at an answer and systematically improve it, step-by-step, until you get the most precise version possible. They tested this idea on several different types of tangled graphs (like squares and mixed networks) and showed that their method works for all of them. However, they are careful to note that while they have the recipe for the noise, they haven't yet found the magic key to instantly solve the entire puzzle in one go for every single case. They have given us the map to the noise, which is a huge leap forward, but the final step of finding the absolute perfect answer for every possible scenario remains an open challenge. Still, this work provides a powerful new compass for anyone navigating the complex, tangled forests of modern data.

Technical Summary: A Characterization of the Orthocomplement of the Tangent Space of Semiparametric Markov Models

Problem Statement

In semiparametric statistical inference, efficient estimation of finite-dimensional target parameters relies on the theory of influence functions (IFs). The class of all regular and asymptotically linear (RAL) estimators for a target parameter is characterized by the linear variety formed by a single influence function and the orthogonal complement of the model's tangent space ( $T^\perp$ ).

For Markov models defined by Directed Acyclic Graphs (DAGs), the tangent space and its orthogonal complement are well-understood and admit closed-form expressions. This is because DAG models possess a likelihood factorization into variationally independent components, allowing the tangent space to be expressed as a direct sum of orthogonal subspaces.

However, for general Markov models associated with other graphical structures—specifically Undirected Graphs (UGs), Chain Graphs (CGs), and Acyclic Directed Mixed Graphs (ADMGs)—the orthogonal complement of the tangent space has not been characterized in closed form. These models often lack a single likelihood factorization representing all constraints; instead, their constraints are defined implicitly (e.g., as solutions to systems of equations). Consequently, the tangent space for these models cannot be easily expressed in terms of orthogonal subspaces, impeding the derivation of efficient estimators and the full characterization of influence functions for target parameters in these settings.

Methodology

The authors propose a geometric approach to characterize the orthogonal complement of the tangent space for general semiparametric Markov models without requiring an explicit likelihood factorization.

Intersection of Single-Independence Models: The paper views a general Markov model defined by a list of $K$ marginal and conditional independence (CI) constraints as the intersection of $K$ simpler models, where each sub-model is defined by a single CI constraint (a "single independence DAG model").
Tangent Space Intersection: The tangent space ( $T$ ) of the general Markov model is established as the intersection of the tangent spaces ( $T_i$ ) of these individual single-independence models: $T = \bigcap_{i=1}^K T_i$ .
Orthogonal Complement Span: Leveraging a geometric result from Van Der Laan and Robins [2003], the authors demonstrate that the orthogonal complement of the intersection of closed subspaces is the direct sum of their individual orthogonal complements. Thus, $T^\perp = \bigoplus_{i=1}^K T_i^\perp$ .
Closed-Form Derivation: Since the orthogonal complement for a single independence constraint ( $X \perp\!\!\perp Y | Z$ ) is known in closed form (Lemma 1), the authors derive the orthogonal complement for the general model by summing the projection operators associated with each individual constraint.

Key Contributions and Results

1. Closed-Form Characterization of $T^\perp$

The primary contribution is Theorem 3, which provides an explicit expression for the orthogonal complement of the tangent space for any Markov model defined by $K$ CI constraints.
For a model defined by constraints $X_i \perp\!\!\perp Y_i | Z_i$ for $i=1, \dots, K$ , the orthogonal complement is:
$T^\perp = \left\{ \sum_{i=1}^K \Pi(h_i | T_i^\perp) : \forall h_1, \dots, h_K \in H \right\}$
where $\Pi(h | T_i^\perp)$ is the projection operator onto the orthogonal complement of the $i$ -th single-independence model, defined as:
$\Pi(h | T_i^\perp)(x_i, y_i, z_i) = E[h | x_i, y_i, z_i] - E[h | x_i, z_i] - E[h | y_i, z_i] + E[h | z_i]$
This result applies to graphical models associated with UGs, CGs, and ADMGs, as well as any Markov model defined purely by CI constraints, regardless of whether it is observationally equivalent to a DAG model.

2. Characterization of Influence Functions

The paper demonstrates that knowing $T^\perp$ allows for the complete characterization of the class of influence functions for any target parameter $\psi(p)$ . If $\phi$ is a known influence function, then any function of the form:
$\phi + \sum_{i=1}^K \Pi(h_i | T_i^\perp)$
is also a valid influence function for $\psi(p)$ . This is illustrated in Section 5.1 by deriving the class of influence functions for the conditional mean parameter across various graphical models (DAG, UG, CG, and ADMG) shown in Figure 1.

3. Iterative Efficiency Improvement

The authors propose an iterative method (Theorem 4) to improve the efficiency of an initial RAL estimator. Given an initial influence function $\phi_0$ , one can construct a sequence $\phi_m$ by projecting $\phi_{m-1}$ onto the tangent space of a specific constraint model (or equivalently, subtracting its projection onto $T_i^\perp$ ).
$\phi_m = \phi_{m-1} - \Pi(\phi_{m-1} | T_{i_m}^\perp)$
This sequence yields influence functions with non-increasing variances ( $E[\phi_m^2] \leq E[\phi_{m-1}^2]$ ). While the paper notes that finding the exact efficient influence function (the global minimum variance) requires projecting onto the full tangent space $T$ (which remains an open problem for general Markov models), this method provides a systematic way to approach efficiency.

Significance and Claims

The paper claims to resolve a long-standing gap in semiparametric theory for non-DAG Markov models. By bypassing the need for an explicit likelihood factorization or an explicit representation of the tangent space, the authors provide a tractable, closed-form characterization of the orthogonal complement of the tangent space.

The significance of this work lies in:

Enabling Inference: It allows researchers to derive the full class of influence functions for target parameters in complex graphical models (like ADMGs and CGs) where such derivations were previously impeded by the lack of a known tangent space structure.
Efficiency: It provides a concrete mechanism to construct more efficient estimators by iteratively reducing the variance of influence functions using the derived projection operators.
Generality: The approach is applicable to any Markov model defined by CI constraints, not just those that are observationally equivalent to DAGs.

The authors remain modest regarding the scope of their results. They explicitly state that while they have characterized $T^\perp$ , the derivation of the projection operator onto the tangent space $T$ itself (which is required to find the exact Efficient Influence Function for general Markov models) remains an open problem. Consequently, the exact form of the efficient influence function for a target parameter in a general Markov model is not yet provided, though the derived characterization of $T^\perp$ is presented as a critical step toward that goal.