Linear Weightings: Methods & Applications

• Linear weightings are schemes in which coefficients enter linearly to construct estimators, mixtures, or graded geometric structures, ensuring normalized and interpretable contributions.
• They are employed in convex combinations, arithmetic progressions, and optimized weighting estimators across fields like data compression, regression, portfolio construction, and causal inference.
• Linear weightings also underpin graded filtrations in differential geometry and vector bundles, uniting applied statistical methods with rigorous geometric and dynamical system frameworks.

Searching arXiv for relevant papers on “linear weighting(s)” and closely related formulations across statistics, learning, geometry, and dynamical systems. Linear weightings are schemes in which coefficients, scores, or filtration degrees enter linearly in the construction of an estimator, a mixture, a geometric structure, or a dynamical update. Across the literature, the phrase denotes several distinct but mathematically related ideas: convex combinations of predictors or model probabilities, position-dependent arithmetic progressions for author credit, weighted scalarizations over normalized features, weighted regression formulations for portfolio construction and causal inference, and graded or filtered linear structures on vector bundles and weighted manifolds. In each case, the defining feature is that the relevant object is controlled by a linear profile—either directly through formulas such as $\sum_i w_i x_i$, through arithmetic sequences of weights, or through filtrations whose local generators have prescribed weighted degrees (Abbas, 2010, Mattern, 2013, Kakushadze et al., 2016, Chattopadhyay et al., 2023, Hudson, 2023).

1. Linear weighting as convex combination and scalarization

A basic meaning of linear weighting is the formation of a convex combination. In data compression, the linear mixture combines model distributions by

$$P(x \mid x_{k-1}) = \sum_{i=1}^m w_i\, P_i(x \mid x_{k-1}), \qquad w_i \ge 0,\ \sum_{i=1}^m w_i = 1,$$

so the combined predictor is linear in the component probabilities (Mattern, 2013). The same paper interprets this as a switching source: at each step, one model is selected with probability given by the normalized weights, and the next symbol is generated from that model (Mattern, 2013). It further formulates the linear mixture as the solution of

$$P := \arg\min_{Q \in \mathcal{P}} \sum_{i=1}^m w_i\, D(P_i \| Q),$$

which yields the same convex combination (Mattern, 2013).

In multi-objective data analysis, linear weighting appears as scalarization on the simplex. The feature relevance vector

$$\gamma=(\gamma_1,\dots,\gamma_m),\qquad \gamma_j\ge 0,\qquad \sum_j \gamma_j=1$$

is learned from normalized data $\Phi\in[0,1]^{n\times m}$ and then used in the weighted sum

$r_i=\sum_{j=1}^m \gamma_j \Phi_{ij}$

to rank samples or identify Pareto-optimal points under a coordinatewise order-preserving normalization (Daniilidis et al., 9 Nov 2025). The paper emphasizes that the final output is still a linear scalarization method, even though the weights are generated endogenously by a replicator-type dynamic rather than chosen manually (Daniilidis et al., 9 Nov 2025).

Online regression with weighted averaging uses the same linear-combination principle at the level of models rather than observations. OLR-WA combines a previously learned model and an incremental model through

$$V\text{-Avg} = \frac{(w\text{-base} \cdot v\text{-base} + w\text{-inc} \cdot v\text{-inc})}{w\text{-base} + w\text{-inc}},$$

with user-defined weights controlling the influence of old and new data (Abu-Shaira et al., 2023). The method is explicitly described as model integration rather than data integration, and its flexibility comes from choosing whether the weighted average favors the base model, the incremental model, or a symmetric blend (Abu-Shaira et al., 2023).

These formulations share a common algebraic template: a weighted linear combination with nonnegative normalized coefficients. This suggests that, across disparate application domains, linear weighting is often used when interpretability, normalization, and direct control of relative influence are required.

2. Arithmetic and positional linear weightings

In bibliometrics, linear weighting takes the form of an arithmetic progression over author positions. The Arithmetic: Type-2 scheme assigns the $j$-th author weight

$$w_j = w_{j-1} - a, \qquad w_j = w_1 - (j-1)a, \qquad 1 \le j \le k,$$

with weights summing to one (Abbas, 2010). The paper derives

$$w_1 = \frac{1}{k} + \frac{a(k-1)}{2}, \qquad w_k = \frac{1}{k} - \frac{a(k-1)}{2},$$

and the closed form

$$P(x \mid x_{k-1}) = \sum_{i=1}^m w_i\, P_i(x \mid x_{k-1}), \qquad w_i \ge 0,\ \sum_{i=1}^m w_i = 1,$$0

Hence the weights form an arithmetic series, decreasing linearly in position when $$P(x \mid x_{k-1}) = \sum_{i=1}^m w_i\, P_i(x \mid x_{k-1}), \qquad w_i \ge 0,\ \sum_{i=1}^m w_i = 1,$$1 and increasing linearly when $$P(x \mid x_{k-1}) = \sum_{i=1}^m w_i\, P_i(x \mid x_{k-1}), \qquad w_i \ge 0,\ \sum_{i=1}^m w_i = 1,$$2 (Abbas, 2010).

The decrement parameter $$P(x \mid x_{k-1}) = \sum_{i=1}^m w_i\, P_i(x \mid x_{k-1}), \qquad w_i \ge 0,\ \sum_{i=1}^m w_i = 1,$$3 controls the slope of the author-credit profile. If $$P(x \mid x_{k-1}) = \sum_{i=1}^m w_i\, P_i(x \mid x_{k-1}), \qquad w_i \ge 0,\ \sum_{i=1}^m w_i = 1,$$4, then

$$P(x \mid x_{k-1}) = \sum_{i=1}^m w_i\, P_i(x \mid x_{k-1}), \qquad w_i \ge 0,\ \sum_{i=1}^m w_i = 1,$$5

for all authors, so equal weighting is recovered as a special case (Abbas, 2010). The paper also shows that the positional arithmetic scheme called Arithmetic: Type-1,

$$P(x \mid x_{k-1}) = \sum_{i=1}^m w_i\, P_i(x \mid x_{k-1}), \qquad w_i \ge 0,\ \sum_{i=1}^m w_i = 1,$$6

is recovered from Type-2 by choosing

$$P(x \mid x_{k-1}) = \sum_{i=1}^m w_i\, P_i(x \mid x_{k-1}), \qquad w_i \ge 0,\ \sum_{i=1}^m w_i = 1,$$7

so Type-1 is another special case of the more general linear family (Abbas, 2010).

The same paper gives interpolation and admissibility formulas that make the linear structure explicit: $$P(x \mid x_{k-1}) = \sum_{i=1}^m w_i\, P_i(x \mid x_{k-1}), \qquad w_i \ge 0,\ \sum_{i=1}^m w_i = 1,$$8 To keep the last author’s weight nonnegative, it imposes

$$P(x \mid x_{k-1}) = \sum_{i=1}^m w_i\, P_i(x \mid x_{k-1}), \qquad w_i \ge 0,\ \sum_{i=1}^m w_i = 1,$$9

and obtains

$$P := \arg\min_{Q \in \mathcal{P}} \sum_{i=1}^m w_i\, D(P_i \| Q),$$0

In the positivity-only case $$P := \arg\min_{Q \in \mathcal{P}} \sum_{i=1}^m w_i\, D(P_i \| Q),$$1, this becomes

$$P := \arg\min_{Q \in \mathcal{P}} \sum_{i=1}^m w_i\, D(P_i \| Q),$$2

The article’s comparative point is that equal weighting and Type-1 arithmetic weighting are embedded in Type-2, whereas geometric and harmonic schemes are positional but nonlinear (Abbas, 2010).

The same linear-versus-nonlinear contrast appears elsewhere. In linear RNN analysis, the recurrent model is shown to be equivalent in the wide limit to a weighted 1D convolutional network

$$P := \arg\min_{Q \in \mathcal{P}} \sum_{i=1}^m w_i\, D(P_i \| Q),$$3

where the lag weights $$P := \arg\min_{Q \in \mathcal{P}} \sum_{i=1}^m w_i\, D(P_i \| Q),$$4 determine how strongly each delay contributes (Emami et al., 2021). Here the “weighting” is not an arithmetic progression, but it is still a linear lag-indexed profile acting on a convolutional representation. Under $$P := \arg\min_{Q \in \mathcal{P}} \sum_{i=1}^m w_i\, D(P_i \| Q),$$5, the coefficients $$P := \arg\min_{Q \in \mathcal{P}} \sum_{i=1}^m w_i\, D(P_i \| Q),$$6 decay geometrically with $$P := \arg\min_{Q \in \mathcal{P}} \sum_{i=1}^m w_i\, D(P_i \| Q),$$7, producing an implicit bias toward short memory (Emami et al., 2021). A plausible implication is that the broad notion of linear weighting includes both explicitly linear profiles over position and linear operators whose coefficients encode positional preference.

3. Weight estimation through optimization and dynamics

Several papers treat linear weighting not as a fixed prescription but as the solution of an optimization problem or a deterministic dynamical system. In data compression, the linear mixture weights are estimated by minimizing cumulative code length: $$P := \arg\min_{Q \in \mathcal{P}} \sum_{i=1}^m w_i\, D(P_i \| Q),$$8 The paper proves that this objective is strictly convex, so there is a single global minimizer and no local minima other than the global one (Mattern, 2013). Because the problem is strictly convex, first-order methods are sufficient, and the paper gives the online gradient update

$$P := \arg\min_{Q \in \mathcal{P}} \sum_{i=1}^m w_i\, D(P_i \| Q),$$9

initialized at $$\gamma=(\gamma_1,\dots,\gamma_m),\qquad \gamma_j\ge 0,\qquad \sum_j \gamma_j=1$$0 (Mattern, 2013).

In feature weighting for data analysis, the update takes a multiplicative replicator form on the simplex: $$\gamma=(\gamma_1,\dots,\gamma_m),\qquad \gamma_j\ge 0,\qquad \sum_j \gamma_j=1$$1 The indices $$\gamma=(\gamma_1,\dots,\gamma_m),\qquad \gamma_j\ge 0,\qquad \sum_j \gamma_j=1$$2 are built from the normalized data matrix through column averages $$\gamma=(\gamma_1,\dots,\gamma_m),\qquad \gamma_j\ge 0,\qquad \sum_j \gamma_j=1$$3 and combine a dominance term and a balance term into

$$\gamma=(\gamma_1,\dots,\gamma_m),\qquad \gamma_j\ge 0,\qquad \sum_j \gamma_j=1$$4

For any interior initialization, the sequence converges globally to a unique interior equilibrium $$\gamma=(\gamma_1,\dots,\gamma_m),\qquad \gamma_j\ge 0,\qquad \sum_j \gamma_j=1$$5, with explicit formula

$$\gamma=(\gamma_1,\dots,\gamma_m),\qquad \gamma_j\ge 0,\qquad \sum_j \gamma_j=1$$6

The limiting weights are therefore inversely proportional to $$\gamma=(\gamma_1,\dots,\gamma_m),\qquad \gamma_j\ge 0,\qquad \sum_j \gamma_j=1$$7 (Daniilidis et al., 9 Nov 2025).

Optimization-based weight determination also appears in ensemble theory. A general framework formalizes the ensemble predictor as

$$\gamma=(\gamma_1,\dots,\gamma_m),\qquad \gamma_j\ge 0,\qquad \sum_j \gamma_j=1$$8

but enlarges the feasible set from uniform weights to structured admissible spaces satisfying normalization, monotone decay, and $$\gamma=(\gamma_1,\dots,\gamma_m),\qquad \gamma_j\ge 0,\qquad \sum_j \gamma_j=1$$9-boundedness (Fokoué, 25 Dec 2025). Optimal weights arise from constrained quadratic programs of the form

$\Phi\in[0,1]^{n\times m}$0

with $\Phi\in[0,1]^{n\times m}$1 the empirical covariance matrix of base learner predictions (Fokoué, 25 Dec 2025). The paper’s claim is that, for stable low-variance learners, the point of weighting is often not variance reduction but approximation geometry and spectral reshaping (Fokoué, 25 Dec 2025).

These results show that linear weighting schemes are often analytically tractable precisely because the weights enter linearly in the objective or update, allowing closed-form equilibria, strict convexity statements, or quadratic programming formulations.

4. Linear weightings in regression, portfolio construction, and causal inference

In large-scale portfolio construction, optimal linear alpha weights are derived from a regression reduction. The standard unconstrained Sharpe-optimal solution is

$\Phi\in[0,1]^{n\times m}$2

but direct inversion is infeasible when the number of alphas is large and the covariance matrix is singular (Kakushadze et al., 2016). Under a factor model and a large-$\Phi\in[0,1]^{n\times m}$3 no-clustering regime, the weights simplify to

$\Phi\in[0,1]^{n\times m}$4

where $\Phi\in[0,1]^{n\times m}$5 are residuals from a weighted cross-sectional regression of normalized expected returns over factor loadings (Kakushadze et al., 2016). The computational point is that the algorithm avoids $\Phi\in[0,1]^{n\times m}$6 and $\Phi\in[0,1]^{n\times m}$7 operations, avoids explicit PCA and matrix inversion, and scales effectively linearly in $\Phi\in[0,1]^{n\times m}$8 because the dominant costs are $\Phi\in[0,1]^{n\times m}$9 and $r_i=\sum_{j=1}^m \gamma_j \Phi_{ij}$0 with $r_i=\sum_{j=1}^m \gamma_j \Phi_{ij}$1 (Kakushadze et al., 2016).

In causal inference, ordinary least squares is recast as a weighting estimator. For a regression

$r_i=\sum_{j=1}^m \gamma_j \Phi_{ij}$2

the fitted treatment effect can be written as

$r_i=\sum_{j=1}^m \gamma_j \Phi_{ij}$3

where the implied linear model weights depend only on treatment and covariates, not on the outcome (Chattopadhyay et al., 2023). For URI and MRI, the paper gives closed forms that sum to one within treatment groups and exactly balance covariate means relative to their implied target populations (Chattopadhyay et al., 2023). It uses these weights to diagnose balance, representativeness, boundedness, extrapolation, and effective sample size, with software reporting SMD, TSMD, KS, TKS, and

$r_i=\sum_{j=1}^m \gamma_j \Phi_{ij}$4

A central concern is that weights can be negative, which indicates extrapolation and can yield non-bounded estimates (Chattopadhyay et al., 2023).

Random weighting in optimization provides a further regression-based perspective. Gradient descent with randomly weighted data points uses

$r_i=\sum_{j=1}^m \gamma_j \Phi_{ij}$5

and in expectation optimizes the weighted least-squares problem

$r_i=\sum_{j=1}^m \gamma_j \Phi_{ij}$6

The paper stresses that random weighting changes both optimization dynamics and the statistical target, and that weightings that accelerate convergence may still harm statistical performance (Clara et al., 11 Dec 2025).

Taken together, these works show that linear weighting in regression settings is not merely a matter of averaging. It determines target populations, computational tractability, and the trade-off between optimization speed, statistical stability, and extrapolation.

5. Linear weightings as geometric and graded structures

In differential geometry, “weighting” has a different but precise meaning: a filtration encoding weighted orders of vanishing along a submanifold. A weighting of order $r_i=\sum_{j=1}^m \gamma_j \Phi_{ij}$7 along a closed submanifold $r_i=\sum_{j=1}^m \gamma_j \Phi_{ij}$8 is a decreasing filtration

$r_i=\sum_{j=1}^m \gamma_j \Phi_{ij}$9

locally modeled by coordinates with assigned weights, so that products add weighted degree (Loizides et al., 2020). This induces a filtration on the normal bundle

$$V\text{-Avg} = \frac{(w\text{-base} \cdot v\text{-base} + w\text{-inc} \cdot v\text{-inc})}{w\text{-base} + w\text{-inc}},$$0

and leads to the weighted normal bundle

$$V\text{-Avg} = \frac{(w\text{-base} \cdot v\text{-base} + w\text{-inc} \cdot v\text{-inc})}{w\text{-base} + w\text{-inc}},$$1

which is canonically a graded bundle over $$V\text{-Avg} = \frac{(w\text{-base} \cdot v\text{-base} + w\text{-inc} \cdot v\text{-inc})}{w\text{-base} + w\text{-inc}},$$2 (Loizides et al., 2020).

The vector-bundle analogue is a linear weighting. For a vector bundle $$V\text{-Avg} = \frac{(w\text{-base} \cdot v\text{-base} + w\text{-inc} \cdot v\text{-inc})}{w\text{-base} + w\text{-inc}},$$3, a linear weighting is a $$V\text{-Avg} = \frac{(w\text{-base} \cdot v\text{-base} + w\text{-inc} \cdot v\text{-inc})}{w\text{-base} + w\text{-inc}},$$4-graded filtration of sections

$$V\text{-Avg} = \frac{(w\text{-base} \cdot v\text{-base} + w\text{-inc} \cdot v\text{-inc})}{w\text{-base} + w\text{-inc}},$$5

compatible with the base weighting, with local model

$$V\text{-Avg} = \frac{(w\text{-base} \cdot v\text{-base} + w\text{-inc} \cdot v\text{-inc})}{w\text{-base} + w\text{-inc}},$$6

where $$V\text{-Avg} = \frac{(w\text{-base} \cdot v\text{-base} + w\text{-inc} \cdot v\text{-inc})}{w\text{-base} + w\text{-inc}},$$7 is a weighted frame and $$V\text{-Avg} = \frac{(w\text{-base} \cdot v\text{-base} + w\text{-inc} \cdot v\text{-inc})}{w\text{-base} + w\text{-inc}},$$8 are vertical weights (Hudson, 2023). An equivalent polynomial characterization uses fiber coordinates $$V\text{-Avg} = \frac{(w\text{-base} \cdot v\text{-base} + w\text{-inc} \cdot v\text{-inc})}{w\text{-base} + w\text{-inc}},$$9 dual to the weighted frame and monomials

$j$0

satisfying

$j$1

This is presented as the correct coordinate-free characterization of a weighted vector bundle (Hudson, 2023, Hudson, 14 Aug 2025).

From this filtration one constructs the weighted normal bundle of the vector bundle,

$j$2

and the weighted deformation bundle,

$j$3

which interpolate between $j$4 and its weighted linearization (Hudson, 2023). A linear weighting also induces a filtration on differential operators: $j$5 has weighted order $j$6 if

$j$7

for all $j$8, and the associated graded operator gives the weighted linearization of $j$9 (Hudson, 2023).

This framework recovers concrete geometric constructions. The tangent bundle of a weighted pair carries a natural linear weighting, with vertical weights the negatives of the coordinate weights, and the cotangent bundle carries the dual weighting (Hudson, 14 Aug 2025). The Higson–Yi rescaled spinor bundle is identified as a weighted deformation bundle, and Ševera’s Clifford algebroid fits the same formalism (Hudson, 2023, Hudson, 14 Aug 2025).

A plausible implication is that the geometric literature uses “linear weighting” in a more structural sense than the applied literature: not as a vector of coefficients but as a filtration compatible with linear bundle operations and graded approximations.

6. Extensions, comparisons, and recurring themes

Across fields, linear weightings are repeatedly contrasted with nonlinear alternatives. In author-credit allocation, Arithmetic: Type-2 is compared with geometric and harmonic weighting, which are positional but nonlinear and do not offer a simple decrement parameter controlling the slope (Abbas, 2010). In data compression, the experiments on a binary implementation report that geometric weighting outperforms linear weighting and that both outperform 3-weighting (Mattern, 2013). In local polynomial regression near a boundary, the asymptotically optimal boundary kernel is shown to be equivalent to a local polynomial regression with a nonnegative linear weighting function such as

$$w_j = w_{j-1} - a, \qquad w_j = w_1 - (j-1)a, \qquad 1 \le j \le k,$$0

whereas common choices like the Bartlett–Priestley weight $$w_j = w_{j-1} - a, \qquad w_j = w_1 - (j-1)a, \qquad 1 \le j \le k,$$1 are not generally optimal in the boundary region (Sidorenko et al., 2018).

Other literatures use linear weighting to express implicit inductive bias. In linear RNNs, lag-dependent coefficients

$$w_j = w_{j-1} - a, \qquad w_j = w_1 - (j-1)a, \qquad 1 \le j \le k,$$2

define the equivalent convolutional NTK and decay geometrically in the stable regime $$w_j = w_{j-1} - a, \qquad w_j = w_1 - (j-1)a, \qquad 1 \le j \le k,$$3, giving a short-memory bias (Emami et al., 2021). In nonlinear feedback systems, exponential weighting operators

$$w_j = w_{j-1} - a, \qquad w_j = w_1 - (j-1)a, \qquad 1 \le j \le k,$$4

convert exponential convergence questions into ordinary $$w_j = w_{j-1} - a, \qquad w_j = w_1 - (j-1)a, \qquad 1 \le j \le k,$$5-stability of weighted signals (Su et al., 2022). In ergodic theory, linear sequences

$$w_j = w_{j-1} - a, \qquad w_j = w_1 - (j-1)a, \qquad 1 \le j \le k,$$6

serve as deterministic weights for multiple return-time and polynomial ergodic theorems when the operator has relatively weakly compact orbits; each such sequence decomposes into an almost periodic part and a Cesàro-null remainder (Eisner, 2013).

A recurring theme is that linear weightings tend to be analytically transparent. They admit normalization constraints, interpolation formulas, convex objectives, closed-form equilibria, graded models, or regression reductions. The corresponding limitation, stated explicitly in several papers, is that linear weighting need not be optimal in every regime: geometric mixtures can compress better than linear mixtures (Mattern, 2013), and weighting choices that speed optimization can worsen asymptotic statistical error (Clara et al., 11 Dec 2025). In causal inference, exact mean balance does not preclude negative weights or poor distributional balance (Chattopadhyay et al., 2023).

In this broader sense, linear weightings are best understood as a family of mathematically controlled weighting mechanisms. Their common core is linearity of combination, interpolation, or filtration; their differences lie in the object being weighted—authors, features, predictors, lags, observations, sections, or functions—and in the criteria governing admissibility, positivity, stability, and optimality.