Weighted Ordinal Cone: Theory & Applications

• The weighted ordinal ordering cone is a convex polyhedral cone that generalizes the classical isotonic ordering cone by incorporating explicit weights to quantify preference intensities.
• It facilitates statistical regression under monotonicity constraints and combinatorial optimization by modeling ordered categorical data and dominance relations.
• Efficient algorithms like the weighted Pool Adjacent Violators method achieve O(n) complexity, making practical large-scale computations feasible.

The weighted ordinal ordering cone is a convex polyhedral cone used to model problems involving ordinal data with explicit preference intensities between ordered categories. It generalizes the classical isotonic (ordinal) ordering cone by introducing explicit weights that quantify the tradeoff between consecutive ordinal categories. This construction supports both statistical regression tasks under monotonicity constraints and combinatorial optimization scenarios where ordinal, rather than cardinal, data is the basis for modeling constraints and dominance relations.

1. Formal Definition and Construction

In its classical form, the ordinal ordering cone in $\R^n$ consists of all vectors $x$ such that $x_1 \le x_2 \le \dots \le x_n$. This cone can be represented as

$K = \{ x \in \R^n : A x \le 0 \}$

where $A$ is the $(n-1) \times n$ matrix with entries $A_{i,j} = 1$ if $j=i$, $-1$ if $j=i+1$, and $0$ otherwise (Dimiccoli, 2015).

The weighted generalization introduces strictly ordered categories $$\mathcal{C} = \{\eta_1 \prec \eta_2 \prec \dots \prec \eta_K\}$$ and two sequences of nonnegative weights, $\omega = (\omega_1, \dots, \omega_{K-1})$ and $\gamma = (\gamma_1, \dots, \gamma_{K-1})$, with the restriction $\omega_i\,\gamma_i < 1$ for all $i$ (Klamroth et al., 6 Jan 2026). These weights encode the preference intensity: $\omega_i$ describes how many of category $\eta_i$ are needed to compensate one of category $\eta_{i+1}$; $\gamma_i$ quantifies, dually, how many of category $\eta_{i+1}$ are at most as good as one of category $\eta_i$.

For count vectors $y, y' \in \R^K$, the weighted dominance relation is defined as:

$y' \preceqq_{(\omega, \gamma)} y \quad \Longleftrightarrow \quad \nu^\top y' \le \nu^\top y \quad \forall \nu \in V_{(\omega, \gamma)}$

where $$V_{(\omega, \gamma)} = \{\nu \ge 0 : \omega_i \nu_i \le \nu_{i+1},\; \nu_i \ge \gamma_i \nu_{i+1}\;\forall i\}$$. The weighted ordinal ordering cone is then

$W_{(\omega, \gamma)} = \{ y' - y : y' \preceqq_{(\omega, \gamma)} y \} \subset \R^K$

which is a polyhedral cone (Klamroth et al., 6 Jan 2026).

2. Polyhedral and Algebraic Representations

Extreme Rays

The cone $W$ is generated by $2(K-1)$ extreme rays:

$$\begin{align*} u^i &= (0,\ldots,0, -\omega_i, 1, 0,\ldots,0)^\top \ g^i &= (0,\ldots,0, 1, -\gamma_i, 0,\ldots,0)^\top \ \end{align*}$$

where for $u^i$, $(-\omega_i, 1)$ occupy $(i, i+1)$; for $g^i$, $(1, -\gamma_i)$ occupy $(i, i+1)$. Thus,

$$W = \mathrm{vcone}\left(u^1, \dots, u^{K-1},\; g^1, \dots, g^{K-1}\right)$$

with $\mathrm{vcone}$ denoting the conic hull (Klamroth et al., 6 Jan 2026).

Facet-Defining Inequalities

Every facet corresponds to a unique choice $r^i \in \{u^i, g^i\}$ for $i=1,\dots,K-1$. The normal vector $n$ for each facet is given by the product rule:

$n_k = \prod_{i=1}^{K-1} d^i_k$

where

$$d^i_k = \begin{cases} \omega_i & \text{if } r^i = u^i,\, k>i \ 1 & \text{if } r^i = u^i,\, k\leq i \ \gamma_i & \text{if } r^i = g^i,\, k\leq i \ 1 & \text{if } r^i = g^i,\, k>i \ \end{cases}$$

The facet inequality is $n_1 y_1 + \cdots + n_K y_K \ge 0$. There are $2^{K-1}$ such facets (possibly fewer if some weights vanish) (Klamroth et al., 6 Jan 2026).

3. Optimization and Projection Algorithms

Weighted Least-Squares Projection

A central statistical application is the projection of data $y \in \R^n$ onto $K$, the weighted ordinal (isotonic) cone, typically formulated as:

$$x^* = \arg\min_x \frac{1}{2}(x - y)^\top W (x-y)\ \ \text{s.t.}\ Ax \le 0$$

with $W = \mathrm{diag}(w_1, \ldots, w_n)$ positive diagonal (Dimiccoli, 2015). The Karush–Kuhn–Tucker conditions fully characterize the solution.

Pool Adjacent Violators Algorithm (PAV)

The weighted PAV algorithm computes $x^*$ in $O(n)$ time and $O(n)$ space without matrix inversion. It iteratively enforces blockwise monotonicity, merging adjacent blocks when monotonicity is violated and recomputing weighted means. The algorithm can be further enhanced by multiscale binning and bound tightening for large-scale or nearly-flat-region data (Dimiccoli, 2015).

Pseudo-code Outline

The algorithm uses a stack of blocks, each labeled by its index range, total weight, weighted sum, and mean. In each sweep, adjacent blocks violating monotonicity are merged, their statistics updated, and the scan repeated until monotonicity is global. The final projection assigns the block mean to each element within its block.

4. Connections to Classical Dominance and Cones

The weighted ordinal ordering cone forms a bridge between various classical dominance concepts:

• Pareto dominance: For $\omega_i = \gamma_i = 0$, $W$ reduces to the nonnegative orthant $\{y: y_k \ge 0\;\forall k\}$, representing standard componentwise (Pareto) ordering.
• Weighted-sum dominance: For $\omega_i\,\gamma_i=1$ (e.g., $\gamma_i=1/\omega_i$), $W$ becomes the halfspace $\{y: \nu^\top y \ge 0\}$, induced by the weighted sum vector $$\nu = (1, \omega_1, \omega_1\omega_2, ..., \prod_{i=1}^{K-1}\omega_i)$$.
• Lexicographic dominance: As $\gamma_i \to +\infty$ and $\omega_i = 0$, ordering is enforced in the pure lexicographic sense, that is, $y' \le_{\text{lex}} y \iff y'$ is less than $y$ in the lexicographic order (Klamroth et al., 6 Jan 2026).

5. Linear Transformation and Multi-objective Optimization

Given the normal matrix $A$ defining the facets of $W$, the map $T\!: y \mapsto Ay$ embeds the weighted ordinal problem in a $p$-dimensional ($p \le 2^{K-1}$) space. In this transformed space, $T(y') \le_P T(y)$ if and only if $y' \preceqq_{(\omega, \gamma)} y$, where $P$ is the nonnegative orthant in $\R^p$. Thus, Pareto-minimizing $Ay$ over a feasible set is equivalent to finding $W$-minimal elements in the original space. This reduction enables the use of standard multi-objective optimization algorithms on problems initially posed in the ordinal framework (Klamroth et al., 6 Jan 2026).

6. Applications and Worked Example

Safest-path Problem

In combinatorial optimization, safest-path problems assign edges to ordered safety categories $\eta_1, \ldots, \eta_K$ (e.g., separate bike lane, shared lane, no lane). A path is summarized by its count vector $c(x)$. Standard ordinal dominance with no weighting ($\omega_i=1, \gamma_i=0$) often results in a very large non-dominated set, including unsafe or excessive paths. By adjusting $\omega$ and $\gamma$, one can prune infeasible or impractical alternatives. For example, increasing $\omega$ makes multiple “better” categories compensate for “worse” ones, reducing the set of efficient paths and eliminating short but unsafe or unnecessarily long detour paths (Klamroth et al., 6 Jan 2026).

Explicit Example (Small $n$)

For $n=5$, $y = [4, 1, 3, 2, 5]$ and $w = [1, 2, 1, 1, 1]$, the weighted PAV algorithm yields the isotonic projection $x^* = [2, 2, 2.5, 2.5, 5]$, which is blockwise constant and nondecreasing. The minimal weighted sum of squared deviations is achieved and the KKT conditions are satisfied (Dimiccoli, 2015).

7. Computational Complexity and Numerical Considerations

Weighted PAV admits $O(n)$ time and space complexity. Each block merge involves a single weighted average computation; no matrix factorization is required, rendering the algorithm robust to ill-conditioning. For extremely large $n$, a hierarchical or coarse-to-fine application of PAV, combining blockwise applications with refinement at block boundaries, maintains optimality while reducing computational constants. For nearly flat regions, “bound tightening” by near-merge on round-off proximity improves numerical stability (Dimiccoli, 2015). In the context of polyhedral cones for multi-objective optimization, the transformation to the image $Ay$ incurs exponential growth in $K$ in the worst case, but exploiting sparsity or category structure may mitigate this in practice (Klamroth et al., 6 Jan 2026).