Minimum Envy Graphical House Allocation

• The paper introduces ME-GHA, a framework integrating agent-specific valuations and social graph constraints to minimize aggregate envy.
• It generalizes NP-hard problems like minimum linear arrangement and demonstrates tractability under restrictions using dynamic programming and parameterized algorithms.
• Advanced strategies, including greedy swaps and decomposition techniques, offer practical methods for fair division in complex networked settings.

The Minimum Envy Graphical House Allocation (ME-GHA) problem is a foundational generalization of classical fair division and assignment models, integrating agent-specific utility functions with network-based (graphical) constraints on envy. In this framework, $n$ agents and $n$ houses are represented, with each agent's potential envy directed towards neighbors specified by an undirected social graph. The objective is to compute an allocation that minimizes the aggregate envy—quantified as the sum of excess utilities accrued along the edges of the social graph—encompassing both the combinatorial structure of agent interactions and heterogeneity in agents’ value systems. The ME-GHA framework explicitly generalizes the minimum linear arrangement problem and encompasses notable economic and algorithmic variants such as minimum envy in housing markets under local constraints and social networks.

1. Formal Problem Definition, Generalization, and Model

Given:

• A set $V = \{v_1, \ldots, v_n\}$ of $n$ agents,
• A set $H=\{h_1,\ldots,h_n\}$ of $n$ houses,
• An undirected social/interaction graph $G=(V,E)$,
• For each agent $v \in V$, a valuation function $v_v: H \rightarrow \mathbb{Q}_{\geq 0}$.

An allocation is a bijection $\pi: V \rightarrow H$. For each edge $(u, v) \in E$ and allocation $\pi$, the directed envy is

$$\mathrm{envy}_{u \rightarrow v}(\pi) = \max\{0, v_u(\pi(v)) - v_u(\pi(u))\}.$$

Since $G$ is undirected, both directions are counted on each edge.

The aggregate envy of an allocation $\pi$ is

$$\mathrm{Envy}(\pi) = \sum_{(u, v) \in E} \left[\mathrm{envy}_{u \rightarrow v}(\pi) + \mathrm{envy}_{v \rightarrow u}(\pi)\right].$$

ME-GHA Objective: Find $\pi^* = \arg\min_\pi \mathrm{Envy}(\pi)$ (Inamdar et al., 22 Jan 2026, Hosseini et al., 2023, Hosseini et al., 2023).

Special cases where all $v_v$ are identical and evenly spaced reduce ME-GHA to classical NP-hard problems, including the Minimum Linear Arrangement (MinLA) (Hosseini et al., 2023, Hosseini et al., 2023). The model naturally extends to arbitrary cardinal valuations and diverse social network topologies, thereby broadening applicability and computational complexity.

2. Computational Complexity, Hardness, and Parameterized Results

ME-GHA exhibits pronounced computational intractability, coupled with a fine-grained complexity landscape depending on both the structure of $G$ and the class of valuations.

Hardness Results

• General Hardness: NP-hard even with identical valuations (reduction from Optimal Linear Arrangement), and remains NP-hard on simple classes such as matchings, stars, cycles, and cliques (Hosseini et al., 2023, Inamdar et al., 22 Jan 2026).
• Even with Constraints: NP-hard on $G$ with maximum degree $1$, with binary (i.e., $\{0,1\}$) and identical valuations (Inamdar et al., 22 Jan 2026).
• Valuation Complexity: NP-hard for even two house-types (Theorem 2 in (Inamdar et al., 22 Jan 2026)).
• Fixed-Parameter Hardness: NP-hardness persists on graphs of small or quasi-polynomial vertex cover, split graphs, and complete bipartite graphs, even with bounded or unit preference sizes (Dey et al., 1 May 2025).

Parameterized and Moderate-Exponential Algorithms

ME-GHA admits tractability and exact exponential-time algorithms under several parameterizations:

Parameter/Restriction	Algorithmic Result	Reference
Max degree $\Delta(G) \le 1$, binary valuations	Polytime: total envy $\le 1$	(Inamdar et al., 22 Jan 2026)
Treewidth $tw$ ($\ell$ types)	FPT: $O^*(\ell^{tw\cdot (1 + n/\ell)^{2(\ell-1)}})$	(Inamdar et al., 22 Jan 2026)
Vertex-cover number $vc$	FPT: $O^*(\ell^{vc})$	(Inamdar et al., 22 Jan 2026)
Clique modulator $k$	FPT: $O^*(\ell^k)$	(Inamdar et al., 22 Jan 2026)
Trees	Exact: $O^*(3^n)$, $O^*(2^n)$ for int-valuations	(Inamdar et al., 22 Jan 2026)
Complete bipartite, $\ell=O(1)$	Polytime for $k = \min\{|L|,|R|\}$: $O^*((1+k/\ell)^\ell)$	(Inamdar et al., 22 Jan 2026)
Balanced-separable classes	$O^*(4^n 2^{O(f(n) \log n)})$	(Inamdar et al., 22 Jan 2026)

In general, ME-GHA is only tractable in polynomial or fixed-parameter time under strong restrictions on both $G$ and the set of valuations.

3. Algorithmic Paradigms

Central algorithmic paradigms for ME-GHA draw from graph decomposition, dynamic programming (DP), and combinatorial optimization.

Binary Valuations, Low Degree

• For $G$ a matching plus isolates and binary valuations, allocation proceeds by greedily assigning highest-value houses to isolated vertices, then resolving envy locally using a “three-house swap” to guarantee at most one unit of residual envy (Lemma 1) (Inamdar et al., 22 Jan 2026).
• This approach exploits edge-local operations and recycles unused allocations from isolated agents to quench envy on residual pairs.

Treewidth-Parameterized Dynamic Programming

• A DP is performed over a nice tree decomposition of $G$, with table entries indexed by partial assignments to the current bag and house-type counts. Envy is charged incrementally as variables are introduced or forgotten, enabling tractable computation for bounded treewidth and small $\ell$ (Inamdar et al., 22 Jan 2026).

Vertex-Cover and Clique-Modulator Reductions

• For bounded vertex cover, assignments for the cover are guessed; the residual assignment on the independent set is solved as a minimum-weight perfect matching, with edge weights encoding projected envy towards assigned cover vertices (Inamdar et al., 22 Jan 2026).

Exact DP on Trees and Component Decomposition

• On trees or disjoint unions, DP recurses on subtrees, tracking partial allocations and their envy, aggregated via subset convolution to efficiently cover all subset assignments. For disjoint union of components, global DP jets over all house partitions across components; time is $O^*(2^n \cdot T)$ if each component admits a $T$-time solution (Inamdar et al., 22 Jan 2026, Hosseini et al., 2023).

4. Key Structural Results and Theorems

Several foundational theorems characterize both the combinatorics of optimal solutions and the tractability boundaries.

• Three-House Envy-Resolution (Lemma 1): For any two adjacent agents and three houses under binary valuations, assignments can always be made so that mutual envy is zero. This is established by exhaustive case analysis of pairwise binary utilities (Inamdar et al., 22 Jan 2026).
• FPT Results: For constant types $\ell$, ME-GHA is FPT in (treewidth, vertex-cover, clique-modulator). Theorems 3–5 in (Inamdar et al., 22 Jan 2026) give explicit running times and DP constructions.
• Strong/Weak Separability: For identical valuations, the decomposability of $G$ into strongly separable components assures FPT algorithms parameterized by the number of components. This fails with non-identical $v_i$ (Hosseini et al., 2023).
• Approximation Hardness: There exists no polynomial-time algorithm achieving a $o(n)$-approximation even on trees, planar, or bounded-degree graphs unless P=NP; lower bounds match the best-known upper bounds for each structural class (Hosseini et al., 2023).

5. Connections, Generalizations, and Related Models

• Minimum Linear Arrangement (MinLA): Under identical, evenly-spaced valuations, ME-GHA reduces to MinLA, establishing tight relations to classical graph layout and bandwidth problems (Hosseini et al., 2023, Hosseini et al., 2023).
• Preference-Set Models: In the restricted model where each agent designates a set of acceptable houses and envy is defined as not receiving an acceptable house when a neighbor does, Min-Envy House Allocation is polynomial-time solvable for single-choice agents, but NP-hard for preference-lists of size two even on complete bipartite or small vertex-cover graphs (Dey et al., 1 May 2025).
• Cycle-Based Envy Metrics: The Min-Max-Average-Cycle-Weight objective considers the maximal average envy around cycles of the (potentially directed) envy graph, pertinent for models with pre-existing allocations and tenants; the clean-slate case admits a polynomial-time maximum-weight matching solution, but general pre-existing envy settings remain open in complexity (Elmalem et al., 27 Jul 2025).
• Alternative Fairness Norms: ME-GHA targets the $\ell_1$-sum of envies, in contrast to $\ell_\infty$ (maximum envy), which equates to graph bandwidth in the linear arrangement setting (Hosseini et al., 2023).

6. Open Questions and Directions

The landscape of ME-GHA is punctuated by several concrete open problems:

• Characterization of Tractable Graph Classes: The precise structural graph classes (beyond paths, cycles, forests) enabling polynomial-time or FPT algorithms for non-identical agent valuations remain to be fully identified (Hosseini et al., 2023, Inamdar et al., 22 Jan 2026).
• Algorithmic Frontiers for Cycle-Based Measures: The complexity of minimizing max-average envy over cycles with arbitrary pre-existing assignments is open, with no known hardness proof or approximation algorithms of guaranteed quality (Elmalem et al., 27 Jul 2025).
• Separability Beyond Identical Valuations: Strong/weak separability yields FPT algorithms in the identical valuations case, but breaks down immediately for heterogeneous agent values (Hosseini et al., 2023).
• Alternative Objectives and Norms: Extending ME-GHA algorithms and hardness to other social welfare or fairness objectives, such as maximizing the minimum agent utility, or minimizing alternative envy aggregations.
• Approximation Algorithms and Tight Bounds: Whether improved (sublinear-factor) approximations are possible in terms of treewidth, degree, or planarity, and for which subclasses of ME-GHA, remains an open field matching only recently obtained lower bounds (Hosseini et al., 2023).

7. Notation Summary and Core Formulas

• Agents: $V=\{v_1,\ldots,v_n\}$; Houses: $H=\{h_1,\ldots,h_n\}$.
• Social graph: $G=(V,E)$.
• Valuations: $v_v: H\to \mathbb{Q}_{\geq 0}$.
• Allocation: $\pi: V \to H$ bijective.
• Envy per edge: $$\mathrm{envy}_{u\to v}(\pi) = \max\{0, v_u(\pi(v)) - v_u(\pi(u))\}$$.
• Objective: $$\mathrm{Envy}(\pi) = \sum_{(u,v)\in E}\left(\mathrm{envy}_{u\to v}(\pi) + \mathrm{envy}_{v\to u}(\pi)\right)$$.
• Treewidth DP: $T[i, f, (t_1,\ldots, t_\ell)]$ stores min envy in subtree $i$ with bag-assignment $f:B_i\rightarrow[\ell]$ and used counts $t_j$ (Inamdar et al., 22 Jan 2026).
• Subset convolution: For $S\subseteq H$, $$h(S)=\min_{R\subseteq S}\left(f(R)+g(S\backslash R)\right)$$, computable in $O^*(2^n)$.

The Minimum Envy Graphical House Allocation problem thus provides a unified lens on fairness, combinatorial structure, and optimization under both value heterogeneity and social interaction constraints, synthesizing several classical and contemporary strands of algorithmic allocation research (Inamdar et al., 22 Jan 2026, Hosseini et al., 2023, Hosseini et al., 2023, Dey et al., 1 May 2025, Elmalem et al., 27 Jul 2025).