Lexicographic Stacking in Preference Modeling

• Lexicographic stacking is a mathematical and algorithmic construct that models, composes, and analyzes hierarchically ordered preferences using lexicographic orders.
• It introduces a stacking operator that combines distinct lex models and ensures strong compositionality for efficient preference inference and consistency checks.
• The approach underpins practical applications in optimization, including precise convex hull descriptions in superincreasing knapsacks and integer programming.

Lexicographic stacking is a mathematical and algorithmic construct used to model, compose, and analyze systems of preferences or feasible sets where outcomes are ranked or defined via lexicographic orderings. It serves as the foundational mechanism for building composite preferences by stacking lexicographic models and underpins efficient algorithms for preference inference, optimization, and convex hull descriptions in combinatorial and optimization domains.

1. Lexicographic Models and Stacking Operator

Let $V = \{X_1, \dots, X_n\}$ denote a set of decision variables, each with a finite domain $D(X_i)$. The set of complete assignments, or outcomes, is $\mathcal{A} = \prod_{X \in V} D(X)$. A lexicographic model (lex-model) $\pi$ over $V$ is specified as an ordered sequence of distinct variables $Y_1, \dots, Y_k$ together with a total order $\geq_{Y_i}$ on $D(Y_i)$ for each:

$$\pi = ((Y_1, \geq_{Y_1}), \dots, (Y_k, \geq_{Y_k})).$$

This induces a total preorder $\leq_\pi$ over $D(X_i)$0, where for any $D(X_i)$1,

• $D(X_i)$2 if either $D(X_i)$3, or for some smallest $D(X_i)$4, $D(X_i)$5 for $D(X_i)$6 and $D(X_i)$7.

The strict part, denoted $D(X_i)$8, occurs when the first differing variable puts $D(X_i)$9 above $\mathcal{A} = \prod_{X \in V} D(X)$0 in the per-variable order.

The lexicographic stacking operator, denoted $\mathcal{A} = \prod_{X \in V} D(X)$1, is defined for two lex-models $\mathcal{A} = \prod_{X \in V} D(X)$2. If

$\mathcal{A} = \prod_{X \in V} D(X)$3

then

$\mathcal{A} = \prod_{X \in V} D(X)$4

is formed by appending each $\mathcal{A} = \prod_{X \in V} D(X)$5 to $\mathcal{A} = \prod_{X \in V} D(X)$6 where $\mathcal{A} = \prod_{X \in V} D(X)$7, preserving the original order within each model. The stacking operator is associative and strictly extends the variable support and ordering of $\mathcal{A} = \prod_{X \in V} D(X)$8 (Wilson et al., 2024).

2. Strong Compositionality and Preference Languages

A preference language $\mathcal{A} = \prod_{X \in V} D(X)$9 consists of preference statements $\pi$0. For lex-models, $\pi$1 denotes that $\pi$2 satisfies $\pi$3. A key property—strong compositionality—holds for $\pi$4 when, for any lex-models $\pi$5,

$\pi$6

where $\pi$7 means some extension of $\pi$8 satisfies $\pi$9. When strong compositionality holds, stacking a model that enforces a preference behind any compatible partial model preserves satisfaction of the preference.

This property is fundamental for the greedy algorithmic approaches for checking consistency of preferences and inferring composite lex-orders from local preference statements (Wilson et al., 2024).

3. Algorithms for Inference and Consistency

Given a preference base $V$0 (finite set of statements in a preference language such as $V$1) over variables $V$2, the lexicographic stacking mechanism enables polynomial-time greedy testing for consistency.

At each step, a partial lex-model $V$3 over $V$4 is extended by selecting an unassigned variable $V$5 and a total order $V$6 on $V$7. The choice of $V$8 is constrained by the set of all remaining unsatisfied preference statements. The local constraints are characterized by:

• $V$9: values required to be maximal by the current preferences,
• $Y_1, \dots, Y_k$0: values required to be minimal,
• $Y_1, \dots, Y_k$1: explicit pairwise constraints on $Y_1, \dots, Y_k$2,

which are deduced by examining the structure of $Y_1, \dots, Y_k$3 and the current model $Y_1, \dots, Y_k$4. Correctness and polynomial complexity are ensured by the strong compositionality property.

For domains of bounded size $Y_1, \dots, Y_k$5, the overall recursion is $Y_1, \dots, Y_k$6 local checks of $Y_1, \dots, Y_k$7 complexity, yielding polynomial time for fixed $Y_1, \dots, Y_k$8 (Wilson et al., 2024).

Example of Computational Workflow

Step	Variable Chosen	$Y_1, \dots, Y_k$9	$\geq_{Y_i}$0	Pairwise Constraints
1	airline	$\geq_{Y_i}$1	$\geq_{Y_i}$2	$\geq_{Y_i}$3
2	time	$\geq_{Y_i}$4	$\geq_{Y_i}$5	$\geq_{Y_i}$6
3	class	$\geq_{Y_i}$7	$\geq_{Y_i}$8	{biz > eco}

This workflow constructs a unique maximal lex-model satisfying all provided preference statements.

4. Optimality Notions Associated with Lexicographic Stacking

Given a consistent preference base $\geq_{Y_i}$9, several natural optimality selectors for a finite candidate set $D(Y_i)$0 are relevant. These are:

• Undominated: outcomes not strictly lex-inferior to any in $D(Y_i)$1
• Possibly-optimal: outcomes optimal for some compatible lex-model
• Necessarily-optimal: outcomes optimal for all compatible lex-models
• Possibly-strictly-optimal: outcomes uniquely optimal for some lex-model

For lexicographic stacking with strong compositionality, the undominated and necessarily-optimal sets coincide; so do the possibly-optimal and possibly-strictly-optimal sets. All can be captured by $D(Y_i)$2 consistency checks of the form: test whether $D(Y_i)$3 is consistent for each $D(Y_i)$4 (Wilson et al., 2024).

5. Lexicographic Stacking in Superincreasing Knapsack and Convex Hulls

Lexicographic orderings interact closely with the structure of superincreasing knapsack sets. Consider an integer knapsack $D(Y_i)$5 with strictly superincreasing weights $D(Y_i)$6, and upper bounds $D(Y_i)$7. The unique greedy solution $D(Y_i)$8 is computed in $D(Y_i)$9 time and maximizes $$\pi = ((Y_1, \geq_{Y_1}), \dots, (Y_k, \geq_{Y_k})).$$0 over $$\pi = ((Y_1, \geq_{Y_1}), \dots, (Y_k, \geq_{Y_k})).$$1.

A central result is that $$\pi = ((Y_1, \geq_{Y_1}), \dots, (Y_k, \geq_{Y_k})).$$2 in reverse lexicographic order, i.e., $$\pi = ((Y_1, \geq_{Y_1}), \dots, (Y_k, \geq_{Y_k})).$$3 is the set of integer vectors not lex-strictly above $$\pi = ((Y_1, \geq_{Y_1}), \dots, (Y_k, \geq_{Y_k})).$$4.

The convex hull of such $$\pi = ((Y_1, \geq_{Y_1}), \dots, (Y_k, \geq_{Y_k})).$$5 has only $$\pi = ((Y_1, \geq_{Y_1}), \dots, (Y_k, \geq_{Y_k})).$$6 nontrivial facets, each corresponding to a specific packing inequality parameterized via the position and value of $$\pi = ((Y_1, \geq_{Y_1}), \dots, (Y_k, \geq_{Y_k})).$$7. Explicitly, for $$\pi = ((Y_1, \geq_{Y_1}), \dots, (Y_k, \geq_{Y_k})).$$8 with $$\pi = ((Y_1, \geq_{Y_1}), \dots, (Y_k, \geq_{Y_k})).$$9, define coefficients $\leq_\pi$0 so that the facet-defining inequalities are:

$\leq_\pi$1

where $\leq_\pi$2 and $\leq_\pi$3 constructed inductively as described. This compact representation leverages the lexicographic structure and extends minimal cover inequalities known from binary knapsack polytopes (Gupte, 2015).

Moreover, for superincreasing knapsacks $\leq_\pi$4 and $\leq_\pi$5 on the same box, their intersection is precisely $\leq_\pi$6 for the respective greedy solutions $\leq_\pi$7, and the convex hull operator distributes: $\leq_\pi$8 (Gupte, 2015).

6. Illustrative Example and Computational Complexity

Suppose $\leq_\pi$9 with respective domains $D(X_i)$00, $D(X_i)$01, and $D(X_i)$02, and consider the preference base:

1. airline=KLM $D(X_i)$03 class=* (with context $D(X_i)$04),
2. time=day $D(X_i)$05 time=night,
3. $D(X_i)$06(class=eco $D(X_i)$07 class=biz).

Applying the greedy stacking algorithm constructs the lex-model

$D(X_i)$08

which satisfies $D(X_i)$09. For the resulting 8 possible assignments, the undominated outcome is the unique $D(X_i)$10 so Nec$D(X_i)$11 Pos$D(X_i)$12.

The total complexity of the greedy algorithm is $D(X_i)$13 for consistency, and $D(X_i)$14 for computing optimal alternatives (Wilson et al., 2024).

7. Connections, Extensions, and Practical Considerations

Lexicographic stacking generalizes and unifies efficient reasoning about preference languages, optimization sets, and polyhedral descriptions in domains governed by hierarchical or ordered attribute dependencies. In preference elicitation and inference, it allows polynomial-time consistency checks and direct computation of undominated alternatives.

In integer programming and combinatorial optimization, the lex-structured packing inequalities extend directly to multi-valued superincreasing knapsacks, generalizing classical 0–1 results. The distributive property—where convex hulls commute with intersection for stacked lex polytopes—facilitates separation and cutting-plane algorithms, especially when constraints arise from integer encoding or mixed lexicographic feasible sets.

A plausible implication is that lexicographic stacking can be systematically leveraged in any domain where attribute-wise orderings and their composition govern feasible or preferred sets, opening avenues for efficient algorithms and fine-grained polyhedral descriptions (Wilson et al., 2024, Gupte, 2015).