Enabling Sustainable Freight Forwarding Network via Collaborative Games

The paper contains no qualifying numbered complexity result: Algorithm 1 has no worst-case guarantee and relies on an NP-hard value oracle. The proposed clone-limit Shapley problem is a plausible high-multiplicity extension, but it cannot rescue bit (a) under the stated rule. Therefore the paper is red.

The objection that survived

Algorithm 1 enumerates \(2^{|N(i)|}\) neighbor subsets and invokes an NP-hard coalition-value computation, so it supplies no qualifying computational result to continuize.

What the mirror covers

The mirror covers the FFCG cost-allocation problem and the locality exploited by FS-LCG, but not a formal complexity theorem, the paper's heuristic FFCP treatment, or its numerical comparisons.

Open questions for a prover

The case FOR (proponent)

The strongest positive case is one anchor: Algorithm 1, “Fast Shapley for Locally Collaborative Game (FS-LCG)”, proposed and derived in this paper. It is a numbered algorithmic result, though not a theorem with a formal complexity guarantee. The paper contains no numbered Theorem, Lemma, Corollary, or Proposition asserting a complexity classification. Its statement that FFCP is NP-hard is unnumbered, so I do not use it as an anchor.

My lead mirror is Mass-FS-LCG for Continuous Freight-Forwarder Cost Sharing.

A type \( \theta \) is a complete forwarder profile: its finite multiset of request sizes and routes, procured services, box capacities, service costs, and all other data entering FFCP. The society is a rational distribution \( \mu\in\mathbb{Q}^{\Theta} \), where \( \mu_\theta \) is the fraction of forwarders of type \( \theta \). Two types are adjacent when their service-route sets overlap. Thus the type graph is exactly the paper’s collaboration graph after identifying genuinely identical forwarders.

For a coalition-mass vector \( \lambda\leq\mu \), define \( \widehat v(\lambda) \) as its minimum shipping cost. A convenient formulation uses configuration variables \(z_{s,\kappa}\), where \(z_{s,\kappa}\geq0\) is the mass of service-\(s\) boxes using feasible integer packing configuration \(\kappa\). The objective is \( \sum_{s,\kappa}c_s z_{s,\kappa} \), subject to serving exactly \(b_{\theta r}\lambda_\theta\) copies of every request item \(r\) of type \( \theta \), and using at most \( \sum_\theta a_{\theta s}\lambda_\theta \) boxes on service \(s\). This is the scalable configuration form of the paper’s FFCP, not an unrelated fractional-welfare problem: \(z\) represents the aggregate of many identical boxes.

The continuous Shapley output for type \( \theta \) is its per-forwarder cost-share density

\( \Phi_\theta(\mu)=\int_0^1 \partial_\theta^+\widehat v(u\mu)\,du \),

where \( \partial_\theta^+\widehat v(\lambda) \) is the right marginal cost of adding an infinitesimal mass of type \( \theta \). The requested solution is the exact vector \( (\Phi_\theta(\mu))_{\theta:\mu_\theta>0} \), satisfying \( \sum_\theta\mu_\theta\Phi_\theta(\mu)=\widehat v(\mu) \). Operationally, it is the expected marginal shipping cost of a representative infinitesimal forwarder arriving in a uniformly random position.

This is recognisably the paper’s Shapley problem. If \( \mu_\theta=a_\theta/q \), instantiate \(q\mu_\theta\) identical forwarder clones and run the paper’s discrete FFCG. Under the usual fixed-profile, scalable-supply regime, the configuration value is the high-multiplicity limit \( \widehat v(\lambda)=\lim_{q\to\infty}q^{-1}v_q(q\lambda) \), and \( \Phi_\theta(\mu) \) is the limiting Shapley value of a type-\(\theta\) clone. The factorial permutation counts in Algorithm 1 become an integral over predecessor mass; locality survives because a type’s marginal depends only on overlapping service types.

The regime is credible for a digital freight marketplace containing thousands of small forwarders using standardized carrier contracts, recurring port pairs, and recurring LCL request profiles. The number of named forwarders can be very large while the number of complete operational profiles is comparatively small. This is stronger than merely saying that prices are “similar”: identical prices, routes, capacities, and request data are included in the same type, exactly as high multiplicity requires.

I expect the structured version of Mass-FS-LCG to be Class A when each type has a bounded neighborhood and configuration pricing is polynomial-time solvable, or fixed-parameter tractable in the number of local request and service patterns. Column generation can handle the implicit box configurations, while Algorithm 1’s local permutation counting becomes a type-level mass calculation. The unrestricted version exposes a useful boundary: bin-packing configuration pricing may retain hardness in the request-size and service dimensions, so exact integer variants could be Class B even after population continuization. That is a meaningful boundary, not a failure of the mirror.

The weakest point is that the paper allocates costs to named forwarders, whereas an atomless population gives each individual zero influence. Replacing individual Shapley values by the density \( \Phi_\theta \) is therefore an extension, not a literal restatement. It is nevertheless controlled: identical clones remain symmetric, the exact FFCP resource semantics are retained through scalable configurations, the route-overlap graph is preserved, and rational clone instances recover the finite problem asymptotically. The paper’s authors should recognise this as the continuous, high-multiplicity form of their FFCG/FS-LCG contribution, even though the convergence theorem and pricing complexity would be new results.

The case AGAINST (opponent, writing after the proponent)

The positive case overstates what Algorithm 1 actually contributes. FS-LCG is a named algorithm, but not a worst-case efficiency theorem: it enumerates \(2^{|N(i)|}\) neighbour subsets and treats \(v\) as an oracle. In the paper’s freight-forwarder application, evaluating \(v\) itself requires an NP-hard packing problem, and the experiments are empirical. Thus the paper supplies no complexity result whose population axis can be cleanly continuized.

Even granting Algorithm 1 as an anchor, the proposed mirror does not preserve its central object. If \(q\) identical forwarders are cloned, every overlapping clone is still a neighbour of every other clone. A type with one neighbouring type therefore produces a clique of degree \(\Theta(q)\), not a two-node type graph. The induced subgraphs used by FS-LCG become count vectors ranging over \(0,\ldots,q_\theta\) for each neighbouring type. Collapsing these to a type graph is not an application of Algorithm 1; it is a new aggregate algorithm whose central subproblem is evaluating many packing instances.

The proposed \(\widehat v\) also changes the integer FFCP. If integer boxes and requests are retained, a coalition mass \(\lambda\) is meaningful only when it corresponds to integral clone counts; the “continuum” is then notation for a family of discrete high-multiplicity instances. If fractional configuration variables are allowed, the object becomes an aggregate configuration LP, not the paper’s Shapley game. Jointly scaling demand and capacity may make this a sensible asymptotic model, but it is a substantive re-modelling.

The claimed Shapley correspondence is especially under-justified. Convergence of normalized costs,
\[ q^{-1}v_q(q\lambda)\to\widehat v(\lambda), \]
does not by itself imply convergence of unit-sized Shapley marginals. Packing integrality produces \(O(1)\) rounding effects, precisely the scale at which an individual forwarder’s cost share lives. A convergence theorem would be new work.

Moreover, the displayed density
\[ \Phi_\theta(\mu)=\int_0^1\partial_\theta^+\widehat v(u\mu)\,du \]
need not even satisfy the asserted efficiency identity at kinks. For example, with \(\mu=(1/2,1/2)\) and \(\widehat v(a,b)=\max\{a,b\}\), both coordinate right derivatives along the diagonal are \(1\), giving
\[ \sum_\theta \mu_\theta\Phi_\theta(\mu)=1, \]
whereas \(\widehat v(\mu)=1/2\). Finite-clone Shapley values instead select a symmetric subgradient, approximately \(1/2\) for each type. Choosing that subgradient requires a new limit construction; it does not follow from FS-LCG.

A stronger repair would define the mirror directly as the limit of exact clone Shapley values, while scaling all resources and proving existence and computability of the limit. That could be a legitimate high-multiplicity cost-sharing project, but it would be a new asymptotic theory of configuration packing. Its main mathematics would no longer be the paper’s local permutation-counting algorithm.

So I would reject Mass-FS-LCG as a direct, programme-ready mirror of this paper. The negative case is not airtight in the universal sense: a marketplace with many genuinely repeated forwarder profiles is a credible high-multiplicity regime, and a new clone-limit cost-sharing problem might be worthwhile. But that possibility supports only a new extension, not the claim that Algorithm 1 already furnishes the promised continuous computational mirror.

The adversarial triple: the proponent anchors on up to three named results; the opponent sees that case and must defeat every anchor; the judge decides which case convinced it. These are the pipeline’s own outputs, generated by tools/triple_run.py — no human edited them. The paper’s own text is not reproduced here beyond the quoted statement above.