Guarantee
With D_k=3-2/k and \beta(D)=\max{3/2,2-1/D}, returns
For two sizes, the bound improves to \max{\alpha_k\mathrm{OPT},a+\lambda_k}, where \alpha_k=3/2 for k\le2 and 5/3 for k\ge3.
Objective and output
Commit a computable independent set of largest-job machines, solve the remaining rank-k problem, and recover every original job base.
Randomness
Deterministic greedy commitment selection, residual unit search, exact decomposition and additive rounding.
Time
For D=D_k, time is
For D=D_k+\eta with rational 0<\eta\le1, time is
Space
Polynomial in the complete binary input and paid oracle workspace; numerical resource multiplicities are not expanded into individual units.
Parameters
a is the largest positive size; u is the largest smaller size; h is the required largest-job total; 1\le k\le h. For W_p=\sum_{j:p_j<a}p_j f_j({p}), \lambda_k=\min{z:\mathrm{rank}_M({p:W_p\le z})\ge h-k}, where M is the rank-h Boolean-cap largest-job matroid. Set \lambda_h=0.
Model and representation
Integral job polymatroids with common type sizes, binary ranks, mandatory original bases, and no additional assignment-cost budget. The Boolean-cap rank-loss case is handled by a polynomial 3/2 branch.
Lower-bound scope
No lower bound is claimed for this algorithm.
Review status
Ready for review. All three canonical commitment manuscripts: original-optimum homogeneous-largest workload and exact two-size bottleneck commitments; full-marginal machine contraction with actual dominated residual bases, order-independent closures and the all-budget greedy profile; and local eligibility/collision padding including exact scalar optimization, its correct grid and approximation scope. The native modules realize both main original-optimum clients, safe absorption and lifting, position contractions and local structural certificates on original complete rank tables.
Dependencies and scope. The mathematical bounds pay for rank/SFM, integral decomposition/intersection, LP and the normalized-prefix unit search. Native clients instead stream residual original subsets and exact LP active systems with separately charged finite time and polynomial binary workspace. No capacity or artificial workload multiplier is expanded into individual jobs. All models use common positive type sizes and no assignment/setup budget. Commitment bottlenecks are explicit load terms, with no implicit bound by original optimum; private-cap and original-optimum comparisons remain distinct. Exact two-size recovery admits noninteger size ratios and incurs no smaller-job additive loss. The complete heterogeneous-large padding allocator and specialized public two-size approximation are not implemented. Scalar optimization is only for the stated padding certificate, and full-marginal closure optimality applies only to its absorption rule. No external peer review or proof-assistant verification is claimed.
The verification note identifies executable coverage. This status does not claim external peer review or proof-assistant verification.
Source
Manuscript and proof, with dependencies and verification sources.
The commitment term is a load bound, not an approximation ratio. The two-size rank-two result is a full 3/2 approximation precisely when a+\lambda_2\le3\mathrm{OPT}/2. The k=0, empty-job, and one-size cases have separate routines.
These companion theorems use the same common-type job-polymatroid model with mandatory original bases and no assignment or setup budget.
Full-marginal machine contraction
A polynomial greedy pass computes a bottleneck \beta_k and contracts every job polymatroid to residual large-job rank at most k. Every residual guarantee \alpha\mathrm{OPT}_F+e becomes \max{\beta_k,\alpha\mathrm{OPT}+e}, because \mathrm{OPT}_F\le\mathrm{OPT}. Machines may absorb multiple large units and original h may exceed q.
Homogeneous-large residual scheduling
Returns \min{\mathrm{OPT}+a,\max{\beta_k,(2-1/(3+\eta))\mathrm{OPT}+u}} in \mathrm{poly}(H)\min{2^{O(k)},(k+1)^{O((1+\eta^{-1})\log(k+1))}} time for rational 0<\eta\le1. With exactly two positive sizes, the residual term improves to 5\mathrm{OPT}/3 in 2^{O(k)}\mathrm{poly}(H) time.
Zero residual large rank
With exactly two positive sizes, \max{\beta_0,\mathrm{OPT}}; with arbitrary smaller sizes, \max{\beta_0,\mathrm{OPT}+u}. The original additive candidate \mathrm{OPT}+a may also be retained.
Eligibility padding and collision thresholds
Given an aggregate-independent commitment F_0 of size h-k, returns \min{\mathrm{OPT}+a_+,\max{\beta_D(F_0)\mathrm{OPT}+u,B(F_0)}}. Here \beta_D=\max{1+a_+/C_2,1+E_D/a_+}, E_D=((D-1)A+H_{\mathrm{width}})/D, and B(F_0)=\max_{p\in F_0}(W_p+a_p).
Eligibility-sensitive hard-target bound
If \mathrm{OPT}<C_2, the preceding construction gives \max{\mathrm{OPT}+E_D+u,B(F_0)}. The exact scalar refinement can replace E_D by its smaller optimized padding certificate C_\star.
For eligibility padding, a_+ is the largest attained type size, C_2 is
the minimum possible load of two large units on one machine, and
a_+/\infty=0. On the nonloops of the contracted large-recipient
matroid, A=\max_p a_p and H_{\mathrm{width}}=\max_p(a_p-\ell_p), where [\ell_p,a_p]
is the range of large sizes eligible at p. With D=D_k=3-2/k the
time is 2^{O(k)}\mathrm{poly}(H); with D=D_k+\eta, 0<\eta\le1, it has the same
rank-quasipolynomial alternative as the table above. Boolean-rank loss
is handled by the separate additive collision branch. Every bottleneck
is a load certificate; no promise that it is small compared with \mathrm{OPT}
is implicit.
The native commitment and contraction backends implement both original-optimum main clients, safe machine absorption and exact structural certificates. Their realization proof and original-input checks distinguish numerical-count execution from the fast theorem's paid primitives.
Attribution. icekylinx · GPT-6 Astra Pro