---
title: Focused Weighted-Average Least Squares (FWALS)
url: https://www.emergentmind.com/topics/focused-weighted-average-least-squares-fwals
type: topic
---

# Focused Weighted-Average Least Squares (FWALS)

Focused Weighted-Average Least Squares (FWALS) is a focused model-averaging estimator for linear regression with model uncertainty over auxiliary regressors. It is designed for settings in which the inferential target is not the full parameter vector, nor generic prediction risk, but a specific smooth scalar function of the coefficients on regressors that are always included. Its defining feature is a semi-orthogonalization of the auxiliary regressors that reduces the weighting problem from \(2^{k_2}\) candidate submodels to at most \(k_2\) regressor-wise weights, yielding a tractable but sub-optimal alternative to full focused model averaging [2603.03008].

## 1. Target of inference and model-uncertainty setting

FWALS is formulated for the linear regression
\[
y_i=x_i^\top\beta+\epsilon_i=x_{1i}^\top\beta_1+x_{2i}^\top\beta_2+\epsilon_i,\qquad i=1,\dots,N,
\]
or in matrix form
\[
y=X\beta+\epsilon=X_1\beta_1+X_2\beta_2+\epsilon.
\]
Here \(X_1\) contains the primary or core regressors, which are always included, whereas \(X_2\) contains auxiliary regressors whose inclusion is uncertain [2603.03008].

The focused parameter is a smooth scalar function
\[
\mu(\beta_1),
\]
so the objective is explicitly target-specific. In the full focused model-averaging formulation, one considers all
\[
M=2^{k_2}
\]
submodels generated by including or excluding each of the \(k_2\) auxiliary regressors. For submodel \(m\), with selection matrix \(\Pi_m\), the least-squares estimator is denoted \(\hat\beta_m\), and the focused estimate is
\[
\hat\mu_m=\mu(\hat\beta_{1m}).
\]
Traditional focused averaging then forms
\[
\hat\mu(w)=\sum_{m=1}^M w_m\hat\mu_m,
\qquad
\mathcal H=\{w\in[0,1]^M:\sum_{m=1}^M w_m=1\}.
\]
FWALS keeps the same focused target but replaces this exponentially large weighting problem by a lower-dimensional construction [2603.03008].

A recurring misconception is that FWALS is simply ordinary model averaging with a new label. It is instead a focused estimator in the strict sense that the weights are chosen for \(\mu(\beta_1)\), not for overall fit, prediction error, or the full coefficient vector. A second misconception is that FWALS averages focused submodel estimates in the same order as standard focused model averaging. It does not; its aggregation acts on an orthogonalized coefficient representation before the focused function is applied [2603.03008].

## 2. Semi-orthogonalization and estimator construction

FWALS builds on the WALS orthogonalization idea of Magnus et al. and De Luca et al. by transforming the auxiliary regressors into a semi-orthogonalized representation
\[
X_2^*=X_2\hat\Lambda \hat P^{-1/2},
\]
where
\[
\hat\Lambda=\operatorname{DIAG}\!\left(\operatorname{DIAG}\!\left(\frac{X_2^\top M_1X_2}{N}\right)\right)^{-1/2},
\qquad
\hat P=\hat\Lambda \frac{X_2^\top M_1X_2}{N}\hat\Lambda,
\]
and
\[
M_1=I_N-X_1(X_1^\top X_1)^{-1}X_1^\top.
\]
This residualizes \(X_2\) with respect to \(X_1\), rescales it, and rotates it so that the transformed auxiliary directions are orthogonal in the residualized space [2603.03008].

In this representation, the submodel estimator for \(\beta_1\) can be written as
\[
\hat\beta_{1m}
=
\hat\beta_{1,\text{narrow}}
-
\hat\Xi \hat\Lambda \hat P^{-1/2}(\Pi_m^\top \Pi_m)\hat\beta_2,
\]
with
\[
\hat\Xi=(X_1^\top X_1)^{-1}X_1^\top X_2,
\qquad
\hat\beta_{1,\text{narrow}}=(X_1^\top X_1)^{-1}X_1^\top y,
\qquad
\hat\beta_2=\frac{X_2^{*\top}M_1y}{N}.
\]

A WALS-type average over submodels yields
\[
\hat\beta_{1,\mathrm{WALS}}
=
\hat\beta_{1,\text{narrow}}
-
\hat\Xi \hat\Lambda \hat P^{-1/2}\tilde W\hat\beta_2,
\]
where
\[
\tilde W=\sum_{m=1}^M w_m(\Pi_m^\top \Pi_m).
\]
Because of the semi-orthogonalized structure, \(\tilde W\) is diagonal:
\[
\tilde W=\operatorname{DIAG}(\tilde w_1,\dots,\tilde w_{k_2}).
\]
The crucial reduction is that the original \(2^{k_2}\)-dimensional model-weight vector enters only through the \(k_2\) partial-sum weights \(\tilde w_j\). These weights lie in
\[
\tilde{\mathcal H}=\{\tilde w\in[0,1]^{k_2}\},
\]
and, unlike the original model weights, they do not need to sum to one [2603.03008].

The FWALS estimator is then
\[
\hat\mu(\tilde w)=\mu(\hat\beta_{1,\mathrm{WALS}}).
\]
This ordering is essential: FWALS first aggregates the transformed coefficient contributions and only then applies the focused function \(\mu\) [2603.03008].

## 3. Asymptotic theory under local misspecification

FWALS is analyzed under a local-to-zero framework for the auxiliary coefficients,
\[
\beta_2=\frac{\delta}{\sqrt N},
\]
which makes the inclusion bias and estimation variance comparable asymptotically [2603.03008].

Let
\[
\frac{X^\top X}{N}\xrightarrow{p}Q=
\begin{bmatrix}
Q_{11}&Q_{12}\\
Q_{21}&Q_{22}
\end{bmatrix},
\qquad
\frac{X^\top\epsilon}{\sqrt N}\xrightarrow{d}R\sim N(0,\Omega),
\]
and define
\[
\Xi=Q_{11}^{-1}Q_{12},
\qquad
\Lambda=\operatorname{DIAG}\!\left(\operatorname{DIAG}(Q_{22}-Q_{21}Q_{11}^{-1}Q_{12})\right)^{-1/2},
\qquad
C=\Lambda P^{-1/2}.
\]
Then, for nonstochastic \(\tilde W\),
\[
\sqrt N(\hat\beta_{1,\mathrm{WALS}}-\beta_1)
\xrightarrow{d}
\Xi C(I-\tilde W)C^{-1}\delta+\Psi R,
\]
where
\[
\Psi=
\left[
Q_{11}^{-1}+\Xi C\tilde W C^\top \Xi^\top
\quad
-\Xi C\tilde W C^\top
\right].
\]
The first term is the asymptotic bias from partial omission of auxiliary effects; the second term is the asymptotic variance component [2603.03008].

For a smooth focused function, with gradient
\[
D_{\beta_1}=\frac{\partial\mu}{\partial\beta_1},
\]
the delta method gives
\[
\sqrt N\bigl(\mu(\hat\beta_{1,\mathrm{WALS}})-\mu(\beta_1)\bigr)
\xrightarrow{d}
D_{\beta_1}^\top \Xi C(I-\tilde W)C^{-1}\delta
+
D_{\beta_1}^\top \Psi R.
\]
Hence the asymptotic bias of FWALS depends jointly on the local signal \(\delta\), the correlation structure \(\Xi\), the orthogonalization geometry \(C\), and the sensitivity of the focused target through \(D_{\beta_1}\) [2603.03008].

This suggests a sharp distinction between FWALS and global shrinkage methods. Two weighting rules that look similar in coefficient space can behave differently once the target function changes, because \(D_{\beta_1}\) changes the relevant bias-variance tradeoff.

## 4. Plug-in AMSE criterion and data-driven weights

The infeasible optimal FWALS weights minimize the asymptotic mean squared error of the focused estimator over
\[
\tilde w\in[0,1]^{k_2}.
\]
The resulting AMSE is a quadratic form in \(\tilde w\) involving the bias term \(\delta\delta^\top\), the covariance matrix \(\Omega\), the focused gradient \(D_{\beta_1}\), and the transformed design objects \(C\), \(\Xi\), and \(B\) [2603.03008].

Operationally, the paper proposes plug-in AMSE minimization. The unknown local-bias component is estimated through
\[
\hat\delta=\sqrt N\,\hat C\hat\beta_2,
\]
but because \(\hat\delta\hat\delta^\top\) is not unbiased for \(\delta\delta^\top\), it is corrected as
\[
\hat\delta\hat\delta^\top-\hat C\hat B\hat\Omega \hat B^\top \hat C^\top.
\]
With this correction, the data-driven weights are chosen by
\[
\hat w
=
\arg\min_{\tilde w\in[0,1]^{k_2}}
\widehat{\mathrm{AMSE}}\bigl(\mu(\hat\beta_{1,\mathrm{WALS}})\bigr),
\]
and the final estimator is
\[
\hat\mu(\hat w)
=
\mu\!\left(
\hat\beta_{1,\text{narrow}}
-
\hat\Xi\hat\Lambda\hat P^{-1/2}\hat W\hat\beta_2
\right),
\qquad
\hat W=\operatorname{DIAG}(\hat w).
\]
The plug-in criterion is shown to be asymptotically unbiased for the infeasible AMSE, and the estimated weights converge in distribution to the AMSE-optimal limit [2603.03008].

A practical implication is that FWALS is not a fixed-weight procedure. Its weights are data-driven and target-driven, and the focus enters through the gradient of \(\mu\), not merely through model selection over \(X_2\).

## 5. Computational reduction, sub-optimality, and empirical behavior

The defining computational advantage of FWALS is that it replaces an optimization over
\[
2^{k_2}
\]
submodels by an optimization over at most \(k_2\) regressor-wise weights. This reduction is exact at the level of the transformed WALS representation but not at the level of full focused model averaging, because different model-weight vectors can produce the same diagonal partial-sum matrix \(\tilde W\). FWALS is therefore explicitly described as a tractable sub-optimal procedure rather than a full solution to the original focused model-averaging problem [2603.03008].

The paper also explains why conventional focused averaging does not admit the same collapse. Even after orthogonalization, the quantities \(\mu(\hat\beta_{1m})\) remain correlated through the common narrow-model component, so the full \(2^{k_2}\)-dimensional weight structure is not eliminated [2603.03008].

Simulation evidence is reported for a baseline linear-regression design with \(k_1=3\), \(k_2\in\{2,4,7\}\), pairwise regressor correlation \(\tau\in\{0.3,0.5,0.7\}\), \(N\in\{100,200\}\), varying \(R^2\), and focused parameter
\[
\mu(\beta_1)=\beta_{11}+\beta_{12}+\beta_{13}.
\]
Across 1000 Monte Carlo replications, FWALS, FIC, and mMSE are often similar; FWALS generally dominates SAIC and SBIC except sometimes when \(R^2\) is very small and \(k_2\) is large; prior-based WALS can be competitive when signals are weak but tends to perform worse when \(R^2\) is high because it is not designed for the focused parameter; and the differences between FWALS and FIC are usually small [2603.03008].

A second design uses impulse response functions as the focus:
\[
\mu_h(\beta_1)=\frac{\partial y_{t+h}}{\partial u_t}=e_1^\top A_\beta^h e_1.
\]
For horizons \(h=1,3,5,7\), FWALS and FIC remain very close, whereas mMSE becomes unstable across horizons, especially when the focused function is nonlinear. This is the basis for the claim that FWALS delivers stable performance when the focused function is designed for impulse response function [2603.03008].

The computational savings are substantial. For \(N=100\), reported timings are as follows.

| \(k_2\) | FWALS | FIC | mMSE |
|---|---:|---:|---:|
| 8 | 0.012 sec | 0.851 sec | 0.130 sec |
| 9 | 0.015 sec | 10.330 sec | 0.992 sec |
| 10 | 0.016 sec | 137.098 sec | 9.804 sec |
| 11 | 0.017 sec | 1407.540 sec | 97.055 sec |

These timings illustrate the central design goal of FWALS: focused inference with substantially lower computational burden than full focused model averaging [2603.03008].

## 6. Position within the broader weighted least-squares literature

FWALS belongs to a broader family of weighted least-squares methods, but its notion of focus is specific. In optimal weighted least-squares approximation, the weights and sampling law are coupled through Christoffel-type leverage scores so that stability and near-best approximation are obtained with sample size scaling linearly in model dimension up to a logarithmic factor [1608.00512]. On irregular domains, a related theory constructs weighted empirical least-squares estimators from surrogate discrete orthonormal bases while preserving stability and quasi-optimality [1907.12304]. These methods are weighted and often leverage-focused, but their target is approximation in \(L^2\), not a smooth focused function of a core parameter vector.

Another related strand concerns estimation under model or state-space pathologies. In the subcritical Heston process, a weighted least-squares estimator with weight
\[
w_t=\frac{1}{X_t+c}
\]
is introduced to regularize the singular region near \(X_t\approx 0\); the weighting is “focused” on the boundary problem rather than on a user-chosen target parameter [1509.09167]. In sufficient dimension reduction, Slice Weighted Average Regression constructs
\[
R=\sum_{h=1}^H w_h\,b_h b_h^\top
\]
from slice-wise least-squares slopes, with influence-based weights used to improve robustness [2209.04616]. These are close methodological relatives in the sense that they combine least-squares objects through structured weights, but they do not formulate the focused target as \(\mu(\beta_1)\).

A distinct but conceptually adjacent result studies weighted least squares for estimating a target linear projection of a misspecified regression function under a chosen weighting measure. There, weighted least squares with importance weights is required because ordinary least squares under a non-target design is inconsistent for the desired projection [2112.14178]. This suggests a broader interpretation: “focus” in weighted least squares can refer either to a chosen inferential target, as in FWALS, or to a chosen target measure, as in projection-based design problems.

Within this landscape, FWALS is characterized by three simultaneous commitments: a focused scalar target, model uncertainty restricted to auxiliary regressors, and computational reduction through semi-orthogonalization. What it does not claim is full optimality over the original \(2^{k_2}\) model-weight simplex. Its contribution is instead to provide a practical and robust alternative whose weights are directly tailored to the focused parameter and whose computational cost remains manageable even when the number of auxiliary regressors is moderately large [2603.03008].

Source: https://www.emergentmind.com/topics/focused-weighted-average-least-squares-fwals