Single-Peaked Preferences

• Single-peaked preferences are defined by each agent having a unique most-preferred option (peak), with preferences decreasing consistently on either side of the spectrum.
• Efficient algorithms for recognition and elicitation leverage the structured one-dimensional or generalized graph axes, simplifying complex voting and allocation tasks.
• Recent research establishes robust combinatorial, algorithmic, and axiomatic frameworks that transform NP-hard voting problems into tractable, fair, and strategy-proof procedures.

Single-peaked preferences model situations in which agents' preference orderings over alternatives are unimodal along a one-dimensional structure, such as a linear spectrum or axis. Formally, a strict or weak preference order is single-peaked with respect to an axis if each agent has a most-preferred alternative ("peak"), and preferences strictly fall off moving away from this peak on either side. This domain restriction, foundational in social choice theory, has deep implications for computational social choice, mechanism design, fair division, and learning theory. Recent literature establishes a comprehensive combinatorial, algorithmic, and axiomatic framework for single-peaked preferences, spanning from recognition algorithms and structural enumeration to the complexity of voting and allocation, stochastic learning, and fair social choice rules.

1. Formal Definitions and Characterizations

Let $C = \{c_1,\ldots,c_m\}$ be a set of alternatives, and fix a total order (axis) $A$, written $c_1 \prec_A \ldots \prec_A c_m$. A strict order $V$ is single-peaked (SP) with respect to $A$ if there exists a unique alternative $p$ (the "peak") such that for any $x <_A y \leq_A p$ or $p \leq_A y <_A x$, $x \succ_V y$. Equivalently, each voter's utility increases along $A$ up to the peak and then decreases. For weak orders, single-peakedness generalizes to allow plateaus or ties, notably in Black's model (unique peak), single-plateaued (contiguous top block), and existential (some extension is SP) forms (Fitzsimmons, 2014, Fitzsimmons et al., 2019).

Key characterizations include:

• Interval Property: For any $A$0, the set of a voter's top $A$1 alternatives forms a contiguous interval on $A$2 (Zhan, 2022).
• Forbidden Substructures: Single-peakedness can be characterized by the avoidance of certain forbidden patterns (e.g. "v-valleys", u-valleys) in the preference profile (Lackner et al., 2015, Fitzsimmons et al., 2019).
• Graph Generalizations: The axis can be generalized to trees, cycles, or arbitrary connected graphs, where single-peakedness is defined such that top-initial segments induce connected subgraphs (Escoffier et al., 2020, Peters et al., 2020).

The sign representation provides a bijection between single-peaked orders and sequences of left/right extensions from the peak, yielding $A$3 on $A$4 alternatives (Zhan, 2022).

2. Recognition, Elicitation, and Structural Results

Testing whether a profile is single-peaked is tractable in several cases:

• Linear Orders: Classic algorithms (e.g., Bartholdi-Trick, Escoffier-Lang-Öztürk) solve recognition in $A$5 time; for approval ballots, the task reduces to the consecutive-ones property, solvable via PQ-trees (0909.3257, Fitzsimmons, 2014).
• Weak and Partial Orders: Existence of an SP axis can be tested in polynomial time for weak orders and various existential/plateau models via reduction to consecutive-ones (Fitzsimmons, 2014). For general partial orders, the problem is NP-complete (Fitzsimmons et al., 2019).
• Tree/Cycle/Graph Axes: For trees, recognition is solvable efficiently via decomposition or LP-based flow algorithms (Escoffier et al., 2020, Peters et al., 2020), and for pseudotrees (at most one cycle), direct algorithms exist (Escoffier et al., 2020).
• Query Complexity: Eliciting ordinally SP preferences along a known axis requires $A$6 queries per agent; for cardinally SP preferences with known positions, $A$7 queries suffice. On trees, the number of leaves, path cover, or distance from a path tightly governs the complexity; for path-like structures, $A$8 queries suffice to find a Condorcet winner (Conitzer, 2014, Dey et al., 2016).

Table: Recognition/Elicitation Complexity in Key Settings

Domain Structure	Recognition	Elicitation (per agent)
Linear order, total	$A$9	$c_1 \prec_A \ldots \prec_A c_m$0
Linear order, cardinal	N/A	$c_1 \prec_A \ldots \prec_A c_m$1
Weak order	$c_1 \prec_A \ldots \prec_A c_m$2	$c_1 \prec_A \ldots \prec_A c_m$3
General partial order	NP-complete	N/A
Tree	$c_1 \prec_A \ldots \prec_A c_m$4	$c_1 \prec_A \ldots \prec_A c_m$5

3. Combinatorics, Likelihood, and Structural Enumeration

The number of SP votes on a fixed axis is $c_1 \prec_A \ldots \prec_A c_m$6 (Zhan, 2022). For a profile to be SP, its votes must be SP on some axis; the number of SP elections (profiles) for $c_1 \prec_A \ldots \prec_A c_m$7 voters and $c_1 \prec_A \ldots \prec_A c_m$8 alternatives is tightly bounded (Lackner et al., 2015):

• For random elections (Impartial Culture, IC), the probability that a profile is SP is exponentially small in $c_1 \prec_A \ldots \prec_A c_m$9 and $V$0: for fixed $V$1, $V$2. For $V$3, this probability is less than $V$4.
• Under the Pólya–Eggenberger urn model and the Mallows model (lower dispersion, high reinforcement), the likelihood of SP increases substantially, demonstrating the necessity of structured assumptions for SP profiles to arise in practice (Lackner et al., 2015).

4. Algorithmic & Voting-Theoretic Implications

Restricting to SP domains yields profound algorithmic simplifications:

• Social Choice Rules: Many winner determination problems that are NP-complete in general become polynomial for SP (Kemeny, Young, Chamberlin–Courant, Proportional Approval Voting, OWA rules) (Peters, 2016, 0909.3257).
• Dynamic Programming and LP: Integer programming formulations for many multi-winner voting rules become totally unimodular under SP, so their LP relaxations yield integral solutions in polynomial time, e.g., for PAV and Chamberlin–Courant rules (Peters, 2016).
• Manipulation and Control: Classic NP-hardness "shields" against strategic manipulation and control collapse under SP: e.g., for plurality and approval voting, various constructive/destructive control tasks become tractable via greedy/dynamic programs exploiting the SP structure (0909.3257).
• Multiwinner Bloc and Condorcet Consistency: For Bloc voting, under SP and for large enough committees ($V$5), winners form contiguous intervals on the axis and are Gehrlein–stable (strong Condorcet) (Calver et al., 17 Feb 2026). Monte Carlo studies confirm that "interval" and Condorcet consistency are robust for uniform SP cultures, but less so in spatial models.
• Mechanism Design: The Crawler rule provides a strategy-proof, Pareto-optimal, and obviously-dominant assignment mechanism for object allocation under SP (Tamura et al., 2019, Beynier et al., 2020). Such rules admit optimal communication complexity and polynomial-time implementations.

5. Axiomatic and Fairness Foundations

Single-peakedness enables sharp axiomatic characterizations:

• Min–Max and Median Rules: In SP domains, unanimous and strategy-proof deterministic rules must be min-max rules; imposing anonymity yields the median voter rule (Sreedurga et al., 2022).
• Random Social Choice: Convex combinations of min-max/median rules exhaust all unanimous and strategy-proof random social choice rules; group-wise anonymity and group-fairness conditions admit systematic characterization via probabilistic fixed-group-ballot rules and their extreme points (Sreedurga et al., 2022).
• Pareto Efficiency: For facility location and fair division, SP ensures the set of Pareto optimal allocations is the class of non-wasteful, peak-preserving, connected allocations (under a common slope assumption for cake cutting) (Bhardwaj et al., 2020, Alcalde-Unzu et al., 2023).
• Monotonic Rules in Endowment Economies: Monotonic (iterative) reallocation rules in SP economies satisfy strong forms of manipulation-resistance (withdrawal-proofness, merging-proofness), with SP being pivotal for the directionality of those manipulations (Bonifacio, 2022).

6. Extensions: Trees, Graphs, Bandits, and Open Directions

The one-dimensional axis paradigm generalizes:

• SP on Trees and Graphs: Single-peakedness on trees requires that for each voter and candidate, the top-initial segment induces a connected subtree. Recognition, winner determination, and preference elicitation admit efficient parameterized algorithms, with the number of leaves, path cover, or internal vertices sharply governing complexity (Escoffier et al., 2020, Peters et al., 2020, Dey et al., 2016).
• Stochastic Online Learning and Bandits: In multi-user online matching under budget constraints, the SP structure enables polynomial-time offline matching and online regret bounds of $V$6 (unknown order) and $V$7 (known order). PQ-tree-based order approximation and dynamic programming are key; intractability returns in the absence of SP (Keinan et al., 10 Oct 2025).
• Combinatorial Structures and Bruhat Order: The sign representation for SP orders provides a compact encoding, admits closed enumeration formulas, and reveals the induced poset structure corresponding to restrictions of Bruhat order (Zhan, 2022). Connections to rhombus tilings and other maximal Condorcet domains are established.
• Incomplete Preferences: Recognizing possible single-peakedness with partial or weak order inputs is polynomial-time in some models and NP-complete in others; practical algorithms exist for identifying nearly SP profiles in real-world, incomplete, or noisy data (Fitzsimmons et al., 2019).

7. Limitations and Open Problems

Several major questions remain:

• Robustness and Approximate SP: Real-world preference profiles often deviate slightly from perfect single-peakedness; quantifying and exploiting "near" SP structure under various distance measures is a current research frontier (Fitzsimmons et al., 2019, Roy et al., 3 Jul 2025).
• Multi-dimensional and Generalized Structures: Extending tractable algorithms and axiomatic characterizations to multidimensional (or tree/cycle-based) domains remains open for many rules (Escoffier et al., 2020, Peters et al., 2020, Keinan et al., 10 Oct 2025).
• Joint Fairness and Efficiency: Achieving both envy-freeness and Pareto efficiency with finite query or communication complexity remains unresolved, even for SP cake-cutting in more general domains (Bhardwaj et al., 2020).
• Complexity Barriers: Although SP removes many classical worst-case intractability results, certain scoring rules retain NP-hardness even in SP domains, and full dichotomies are known only in special cases (0909.3257).
• Learning and Robust Order Extraction: Efficiently learning the hidden SP order from noisy or adversarial data, with provably optimal regret or approximation guarantees, is a challenging direction, especially for dynamic and online settings (Keinan et al., 10 Oct 2025).

Single-peakedness remains a paradigmatic domain restriction that both simplifies and structures a wide array of problems in social choice, computational economics, and learning theory, underpinning contemporary research in algorithmic mechanism design, fair division, and preference learning.