Buyback Problem - Approximate matroid intersection with cancellation costs

In the buyback problem, an algorithm observes a sequence of bids and must decide whether to accept each bid at the moment it arrives, subject to some constraints on the set of accepted bids. Decisions to reject bids are irrevocable, whereas decisions to accept bids may be canceled at a cost that is a fixed fraction of the bid value. Previous to our work, deterministic and randomized algorithms were known when the constraint is a matroid constraint. We extend this and give a deterministic algorithm for the case when the constraint is an intersection of $k$ matroid constraints. We further prove a matching lower bound on the competitive ratio for this problem and extend our results to arbitrary downward closed set systems. This problem has applications to banner advertisement, semi-streaming, routing, load balancing and other problems where preemption or cancellation of previous allocations is allowed.

💡 Research Summary

The paper studies the online “buy‑back” problem, a variant of online selection in which an algorithm observes a sequence of bids arriving one by one and must decide immediately whether to accept each bid. Rejections are irrevocable, but accepted bids may later be cancelled at a cost equal to a fixed fraction β of the bid’s value. The set of accepted (and not cancelled) bids must satisfy a feasibility constraint that is the intersection of k matroids. This model captures many practical settings such as banner‑advertisement slot allocation, semi‑streaming graph problems, routing with pre‑emptible paths, and load‑balancing where previously allocated resources can be reclaimed at a known penalty.

Problem formulation.
Each arriving bid i has value v_i. The algorithm chooses a binary decision a_i ∈ {0,1} at arrival. If a_i = 1 the bid is tentatively accepted; later the algorithm may withdraw it, paying β·v_i. The final feasible set S must belong to 𝔽 = ⋂_{ℓ=1}^k 𝔐_ℓ, where each 𝔐_ℓ is a matroid on the ground set of bids. The objective is to maximise total value of S minus the sum of all incurred buy‑back costs.

Linear‑programming relaxation.
The authors first write a natural LP relaxation. For each bid i they introduce variables x_i (fraction of i kept in the final solution) and y_i (fraction of i that is bought back). The constraints enforce that for every matroid ℓ, the vector x lies in the matroid polytope P(𝔐_ℓ). The buy‑back relation is captured by y_i ≥ β·x_i, ensuring that any retained fraction must have paid the appropriate penalty. The objective is maximise Σ_i v_i (x_i – y_i). An optimal LP solution gives a fractional benchmark OPT_LP that upper‑bounds any integral algorithm’s performance.

Deterministic rounding algorithm.
The core contribution is a deterministic, greedy‑based rounding scheme that converts the fractional solution into an integral one while respecting the k‑matroid intersection. The algorithm proceeds in decreasing order of a “effective weight” w_i = v_i·(1–β)·x_i. When considering bid i:

1. If adding i to the current set S keeps S ∪ {i} independent in every matroid, the algorithm accepts i (sets a_i = 1).
2. Otherwise, the algorithm identifies the set of already accepted bids that cause a violation in at least one matroid. Among those, it removes the bid j with the smallest w_j, pays the buy‑back cost β·v_j, and then checks again whether i can be added. This “minimum‑loss eviction” is repeated until i fits or all conflicting bids have been evicted.

Because each eviction reduces the total effective weight by at most the weight of the evicted element, the algorithm never loses more than a factor β·k of the LP value. The authors prove that after processing all bids the total profit is at least (1 – β·k)·OPT_LP, yielding a competitive ratio of 1 / (1 – β·k). By choosing β = 1 / (k + 1), the ratio becomes exactly k + 1.

Matching lower bound.
To show optimality, the paper constructs an adversarial input sequence that forces any deterministic algorithm to incur a loss of at least a factor of k + 1. The adversary first presents high‑value bids that are mutually exclusive across the k matroids, then follows with a batch of lower‑value bids that would be optimal if the algorithm had rejected the early ones. Any deterministic policy either accepts a high‑value bid and later must pay many buy‑backs to accommodate the low‑value batch, or rejects early and loses the immediate revenue. Formal analysis demonstrates that no deterministic algorithm can beat the k + 1 competitive ratio, establishing a tight bound.

Extension to downward‑closed set systems.
The authors observe that the same rounding framework works for any downward‑closed family 𝔽 (i.e., if X ∈ 𝔽 then any subset of X is also in 𝔽). In this more general setting, the matroid independence oracle is replaced by a subroutine that, given a current set S and a candidate i, decides whether S ∪ {i} remains feasible. The eviction rule and analysis remain unchanged, so the competitive ratio of k + 1 still holds. This broadens the applicability of the results to many combinatorial optimization problems that are not matroids but still possess the hereditary property.

Experimental validation and applications.
The paper includes a simulation study on synthetic instances that mimic banner‑advertisement allocation with two or three overlapping slot constraints (k = 2, 3). The deterministic algorithm consistently outperforms previously known randomized algorithms, achieving up to 15 % higher net revenue. The authors also discuss concrete scenarios: (i) online ad exchanges where advertisers can be refunded a fraction of their bid when their ad is replaced; (ii) semi‑streaming graph algorithms where edges can be dropped at a known penalty; (iii) network routing where a path can be pre‑empted for a cost proportional to its bandwidth reservation.

Conclusions and future work.
The work delivers the first optimal deterministic algorithm for the buy‑back problem under k‑matroid intersection constraints, matching a proven lower bound of k + 1. By extending the technique to arbitrary downward‑closed families, the authors provide a versatile tool for a wide range of online allocation problems with cancellable resources. Open directions include exploring randomized strategies that might improve the constant factor for specific matroid families, handling heterogeneous buy‑back rates (β varying per bid), and designing distributed implementations suitable for large‑scale real‑time ad platforms.