A review of Díaz and Rosenblum (2015) on constructing fluctuation submodels via exponential families for targeted maximum likelihood estimation.
\begin{equation} P(M = m \mid X, Y) = P(M = m \mid X) \end{equation}
Consider observing $n$ independent and identically distributed (i.i.d.) random vectors $O_1, \dots, O_n$ sampled from an unknown true distribution $P_0 \in \mathcal{M}$, where $\mathcal{M}$ represents a non-parametric or semiparametric statistical model. Typically, the observed data vector is decomposed into baseline covariates, exposure/treatment, and outcome:
$$O = (W, A, Y) \sim P_0$$
We are interested in estimating a finite-dimensional parameter $\theta_0 = \Psi(P_0) \in \mathbb{R}^k$, where $\Psi: \mathcal{M} \to \mathbb{R}^k$ is a pathwise differentiable target mapping (e.g., the average treatment effect $\mathbb{E}[Y(1) - Y(0)]$ or missing outcome mean $\mathbb{E}[Y]$).
Standard semiparametric efficiency theory dictates that any regular, pathwise differentiable parameter $\Psi$ possesses a unique canonical gradient (or efficient influence curve, EIC), denoted by $D^*(O; P) \in L_2^0(P)$.
| For any smooth parametric submodel ${P_\epsilon : \epsilon \in \mathbb{R}^k} \subset \mathcal{M}$ passing through $P$ at $\epsilon = 0$ with score $\mathbf{S}(O) = \left. \frac{\partial}{\partial \epsilon} \log dP_\epsilon(O) \right | _{\epsilon=0}$, pathwise differentiability implies: |
| <p>$$\left. \frac{d}{d\epsilon} \Psi(P_\epsilon) \right | _{\epsilon=0} = \mathbb{E}_P \left[ D^*(O; P) \mathbf{S}(O)^T \right]$$</p> |
An estimator $\hat{\theta}_n$ is asymptotically efficient at $P_0$ if it is asymptotically linear with influence function equal to $D^*(O; P_0)$:
$$\sqrt{n}(\hat{\theta}_n - \theta_0) = \frac{1}{\sqrt{n}} \sum_{i=1}^n D^*(O_i; P_0) + o_p(1) \quad \xrightarrow{d} \quad \mathcal{N}\left(0, \operatorname{Var}_{P_0}(D^*(O; P_0))\right)$$
To achieve efficiency via substitution, a TMLE updated distribution $\hat{P}^*$ must solve the empirical score equation:
$$\mathbb{E}_{P_n} \left[ D^*(O; \hat{P}^*) \right] = \frac{1}{n} \sum_{i=1}^n D^*(O_i; \hat{P}^*) = o_p(n^{-1/2})$$
Let $p^0(y \mid a, w)$ denote an initial conditional density estimate of $Y$ given $(A,W)$, typically obtained via flexible machine learning algorithms (e.g., Super Learner, Random Forests, or Neural Networks).
Instead of adding a standard clever covariate $H(A,W)$ to a link function, (missing reference) introduce a parametric fluctuation submodel ${p(\epsilon) : \epsilon \in \mathbb{R}^k}$ defined as a canonical exponential family tilted away from the initial density $p^0$:
$$p(\epsilon)(y \mid a, w) = p^0(y \mid a, w) \exp\left( \epsilon^T T(y, a, w) - A(\epsilon; a, w) \right)$$
where:
$$A(\epsilon; a, w) = \log \int \exp\left( \epsilon^T T(y, a, w) \right) p^0(y \mid a, w) \, dy$$
Taking the derivative of the log-density of $p(\epsilon)$ with respect to the fluctuation parameter $\epsilon$:
$$\nabla_\epsilon \log p(\epsilon)(y \mid a, w) = T(y, a, w) - \nabla_\epsilon A(\epsilon; a, w)$$
Using classic exponential family identities, the gradient of the log-partition function equals the conditional expectation of the sufficient statistic under $p(\epsilon)$:
$$\nabla_\epsilon A(\epsilon; a, w) = \mathbb{E}_{p(\epsilon)} \left[ T(Y, a, w) \mid A=a, W=w \right]$$
Evaluating the score at $\epsilon = 0$ yields:
$$\left. \nabla_\epsilon \log p(\epsilon)(y \mid a, w) \right|_{\epsilon=0} = T(y, a, w) - \mathbb{E}_{p^0} \left[ T(Y, a, w) \mid A=a, W=w \right]$$
To ensure that the fluctuation path spans the required component of the efficient influence curve $D^(O; P)$, the sufficient statistic $T(y, a, w)$ is explicitly constructed so that its centered expectation matches the score component of $D^$.
For example, when estimating a conditional mean parameter where the EIC takes the form $D^*(O; P) = H(A,W) \left( Y - \mathbb{E}_P[Y \mid A,W] \right)$, setting $T(y, a, w) = H(a, w) y$ gives:
$$\left. \nabla_\epsilon \log p(\epsilon)(y \mid a, w) \right|_{\epsilon=0} = H(a, w) \left( y - \mathbb{E}_{p^0}[Y \mid A=a, W=w] \right) = D^*(O; P^0)$$
The algorithm proceeds iteratively using maximum likelihood estimation over the fluctuation parameter $\epsilon$, guaranteeing computational efficiency through strict log-likelihood concavity.
+-------------------------------------------------------+
| 1. Initial ML Estimation |
| - Estimate p^0(Y|A,W) via Super Learner |
| - Estimate nuisance mechanism g^0(A|W) |
+-------------------------------------------------------+
|
v
+-------------------------------------------------------+
| 2. Exponential Family Submodel Construction |
| - Define canonical gradient D*(O; p^k, g^0) |
| - Set sufficient statistic T_k(y,a,w) |
| - Form p^k(\epsilon) \propto p^k exp(\epsilon^T T) |
+-------------------------------------------------------+
|
v
+-------------------------------------------------------+
| 3. Convex Maximum Likelihood Update |
| - Solve \hat{\epsilon}_k = argmax \sum log p^k(\epsilon)|
| - Update p^{k+1} = p^k(\hat{\epsilon}_k) |
+-------------------------------------------------------+
|
Is ||\hat{\epsilon}_k|| < tol?
/ \
No Yes
/ \
v v
[ Repeat Step 2 & 3 ] +-------------------------+
| 4. Plug-in Estimation |
| \hat{\theta} = \Psi |
+-------------------------+
$$\hat{\epsilon}_k = \arg\max_{\epsilon \in \mathbb{R}^k} \frac{1}{n} \sum_{i=1}^n \log p^k(\epsilon)(Y_i \mid A_i, W_i)$$
(missing reference) demonstrate the exponential family TMLE framework across three key problem domains:
| Problem Domain | Challenge in Standard TMLE | Exponential Family Solution |
|---|---|---|
| Missing Outcome Mean | Non-standard outcome bounds | Canonical exponential tilting directly adjusts the conditional density $p(y \mid w)$. |
| Median Regression | Non-differentiable indicator objective | Exponential family submodel creates a smooth fluctuation path spanning the non-smooth EIC. |
| Continuous Exposure | Infinite-dimensional nuisance functions | Multi-dimensional exponential family submodels handle kernel-smoothed density fluctuations seamlessly. |
The exponential family approach introduced by (missing reference) unifies targeted maximum likelihood estimation by replacing ad-hoc clever covariates with a principled, convex exponential tilting mechanism. This framework provides robust theoretical guarantees, numerical stability, and seamless scalability for complex causal parameters in modern data science and semiparametric inference.