Approximability of Sparse Integer Programs

The main focus of this paper is a pair of new approximation algorithms for certain integer programs. First, for covering integer programs {min cx: Ax >= b, 0 <= x <= d} where A has at most k nonzeroes per row, we give a k-approximation algorithm. (We assume A, b, c, d are nonnegative.) For any k >= 2 and eps>0, if P != NP this ratio cannot be improved to k-1-eps, and under the unique games conjecture this ratio cannot be improved to k-eps. One key idea is to replace individual constraints by others that have better rounding properties but the same nonnegative integral solutions; another critical ingredient is knapsack-cover inequalities. Second, for packing integer programs {max cx: Ax <= b, 0 <= x <= d} where A has at most k nonzeroes per column, we give a (2k^2+2)-approximation algorithm. Our approach builds on the iterated LP relaxation framework. In addition, we obtain improved approximations for the second problem when k=2, and for both problems when every A_{ij} is small compared to b_i. Finally, we demonstrate a 17/16-inapproximability for covering integer programs with at most two nonzeroes per column.

💡 Research Summary

The paper investigates the approximability of two fundamental families of sparse integer programs: covering integer programs (CIPs) whose constraint matrix is k‑row‑sparse (each row contains at most k non‑zero entries) and packing integer programs (PIPs) whose matrix is k‑column‑sparse (each column contains at most k non‑zero entries). Both problems are defined with non‑negative rational data and optional multiplicity (capacity) constraints 0 ≤ x ≤ d.

Main Results

1. k‑row‑sparse CIPs admit a k‑approximation.

   • The authors introduce the notion of Z⁺‑equivalent constraints: two constraints are equivalent if they admit exactly the same set of non‑negative integer solutions. By replacing each original row with a Z⁺‑equivalent row whose coefficients sum to at most k, they obtain a k‑roundable system. A constraint is ρ‑roundable if, for any non‑negative real vector x satisfying the constraint, the integer vector ⌊ρ x⌋ also satisfies it.
   • For a k‑roundable system, rounding the optimal LP solution x* by factor k yields a feasible integer solution of cost at most k·OPT, giving a k‑approximation. The transformation relies on a constructive lemma (Lemma 11) that shows a specific family of constraints of the form Σ_{i=1}^{k‑1} x_i + v x_k ≥ 1 is k‑roundable, and on Proposition 9 which proves that any row with at most k non‑zeros can be replaced by a Z⁺‑equivalent k‑roundable row.
   • When multiplicity bounds d are finite, the algorithm augments the LP with knapsack‑cover inequalities (as in Carr et al. 2005) and applies the same rounding argument, preserving the k‑approximation factor.

2. k‑column‑sparse PIPs admit a (2k² + 2)‑approximation.

   • The authors employ the iterated LP relaxation framework. Starting from the natural LP relaxation, they iteratively fix variables that are integral or that can be rounded without violating any constraint by more than a factor of k. This process may leave a fractional solution that violates some constraints, but the total violation can be bounded by O(k²).
   • A coloring (or greedy partition) argument then decomposes the violating solution into at most O(k²) feasible integer solutions. Selecting the best among them yields a solution whose objective value is at least 1/(2k² + 2) of the LP optimum, i.e., a (2k² + 2)‑approximation.
   • The paper also presents refined bounds for the special case k = 2 and for instances where every coefficient A_{ij} is sufficiently small relative to the corresponding right‑hand side b_i, achieving better constant factors in those regimes.

3. Hardness for 2‑column‑sparse CIPs.

   • By adapting the classic reduction for k‑Set‑Cover, the authors construct instances where each column has exactly two non‑zero entries and the matrix consists of block‑diagonal 1‑column‑sparse submatrices. They prove that achieving a (17/16 − ε)‑approximation for such CIPs is NP‑hard for any ε > 0. This result strengthens known APX‑hardness for set cover with frequency two and shows that even the seemingly simple 2‑CS case is non‑trivial.

Complexity Lower Bounds

• For k‑row‑sparse CIPs, the paper re‑establishes that a (k − 1 − ε)‑approximation is impossible unless P = NP, and a (k − ε)‑approximation is ruled out under the Unique Games Conjecture.
• For k‑column‑sparse PIPs, they cite Hazan, Safra, and Schwartz’s Ω(k / log k) inapproximability result, showing that any algorithm with a factor substantially better than linear in k would imply major complexity breakthroughs.

Technical Contributions and Significance

• The Z⁺‑equivalence transformation and the associated k‑roundability concept provide a clean, LP‑based pathway to achieve the optimal k‑approximation for covering problems with row sparsity. This approach is conceptually simpler than earlier primal‑dual or local‑ratio methods and yields integrality‑gap bounds that match known hardness.
• The iterated LP relaxation combined with a coloring decomposition offers a novel way to control constraint violations in packing problems with column sparsity, leading to the first O(k²)‑approximation that depends only on k and not on the problem size. Subsequent work (Bansal et al., 2009) later improved this to O(k), but the present technique remains valuable for its deterministic nature and its applicability to problems with multiplicity constraints.
• The hardness proof for 2‑CS CIPs tightens the approximation landscape for covering problems with very low column sparsity, demonstrating that even when each variable participates in only two constraints, the problem does not admit a PTAS.

Broader Impact
Sparse integer programs arise naturally in network design, resource allocation, and scheduling where each constraint or variable interacts with only a few others. By exploiting this sparsity, the paper delivers algorithms that are both theoretically optimal (matching known lower bounds) and practically implementable via standard LP solvers. The techniques introduced—constraint replacement preserving integer feasibility, roundability analysis, and iterative relaxation with controlled violation—are likely to influence future work on other structured integer programs, such as those with submodular objectives or dynamic constraints.

Overall, the paper makes a substantial contribution to the field of combinatorial optimization by clarifying the exact approximability frontier for row‑ and column‑sparse integer programs, providing both algorithmic breakthroughs and tight hardness results.