Skip to content

Latest commit

 

History

History
198 lines (163 loc) · 11.2 KB

reference-realization.md

File metadata and controls

198 lines (163 loc) · 11.2 KB

Native commitment, contraction and eligibility interfaces

The reference modules implement the original-optimum commitment clients of the main proof, the complete preprocessing of safe machine contraction, and the local certificates of eligibility padding. Inputs retain original physical machine IDs, original mandatory job bases and positive common type sizes. There is no assignment-cost or setup budget in this model. All counts and capacities remain binary numbers.

1. Full-marginal absorption and all threshold closures

Absorption validates a common original ground of complete job-rank tables. It repeatedly selects the minimum current total marginal workload, breaking ties by original machine ID. At that step it records every original type's rank increment on the selected machine. Greedy polymatroid feasibility makes these records a restricted base on every prefix. Submodularity makes each remaining machine's marginal nonincreasing as the prefix grows.

The stored bottleneck is the maximum marginal seen so far, rather than the most recent marginal. Thus decreases in individual marginal loads cannot make the recorded threshold decrease. At any threshold, the longest prefix whose bottleneck is at most that threshold is precisely the full-marginal closure: the next greedy minimum is too large, while every earlier step was legal. Comparing two arbitrary legal absorption orders by prefix inclusion proves that their final closed sets coincide. The canonical proof establishes both this order independence and monotonicity in the threshold.

for_budget locates the first prefix whose actual residual distinguished rank is at most the requested binary budget, then extends through the entire bottleneck plateau. It does not enumerate the numerical rank budget. This returns the smallest threshold among these closures, including threshold zero and zero-cost initial machines. It does not optimize arbitrary partial preassignments to machines.

2. Actual contracted bases, dominated recovery and lifting

On a closure, the reference retains the original remaining IDs and explicitly constructs each contracted rank:

Formula

dominated accepts a supplied original integer schedule and computes an integer contracted base inside its remaining-coordinate box. For a contracted rank and that box, it evaluates the usual box-restricted rank on successive prefixes and returns its greedy differences. The canonical contraction lemma proves that the maximum boxed total is the entire contracted rank:

Formula

Hence the constructed base is coordinatewise no larger than the original remaining assignment. Nonnegative sizes imply that its load at each remaining machine cannot increase. This is an actual producer for the original-optimum comparison, rather than merely an existence assertion about the optimum.

lift validates every supplied residual type base and combines it with the recorded committed rows. The canonical submodularity argument proves that any restricted base and any contracted base combine to an original base. The implementation additionally checks all original inequalities and totals. Its output therefore has makespan bounded by the maximum of the closure's recorded bottleneck and the supplied residual makespan.

3. Position commitments keep resource ranks intact

position_commitment is a distinct operation. It greedily selects an independent set of the required size in an ordinary player matroid, ordered by the supplied nonnegative singleton exposure. For each exposure threshold, matroid greedy reaches that threshold set's rank. Therefore the largest selected exposure is minimal among independent commitments of that size. It contracts only the player matroid and removes its resulting loops while retaining original IDs.

Every original player basis spans this contraction, so its remaining members contain a residual basis. This transfers the target promise. It does not contract the recipient workload resource or allocate private quota to the committed machines. The distinction from full-marginal machine removal in Sections 1–2 is essential.

4. Original-optimum many-size commitment client

CommitmentBackend fixes its bottleneck position set before target search. For a hard target it reuses the native full-machine workload construction and exact reconstruction. Its workload quantum is engineered from the divisor numerator so that both the full quota and its divided weak quota are integers.

The reference streams residual original player subsets and tests completion to the original largest-job basis together with all original dual inequalities. A full target supplies a full-quota basis of that contraction and hence a weak-quota candidate. Exhaustion is therefore a sound original rejection. This explicit subset enumeration has a separate cost; it does not implement the fast normalized-prefix unit search.

The dual lower vector puts quota only on that residual basis and zero on all position commitments. Zero extension preserves feasibility in the original full-machine dual. The common workload routine completes the vector, complements it, decomposes the binary-scaled workload, and restores every original smaller job base by joint upper-prefix rounding. It keeps the selected largest basis fixed and decomposes it into its original equal-largest-size types.

At each committed machine the final integer counts obey the original singleton ranks. Its load is bounded directly by the largest size plus the commitment exposure; there is no additional small-job rounding term there. The other machines retain the main proof's quota and additive-rounding estimate. This proves the original-optimum bound of the main workload commitment theorem.

5. Exact integer two-size client and target optimization

The separate two-size mode forms the small count dual with the exact floor capacity and floor-difference quota from the main proof. It does not assume that the largest size is an integer multiple of the smaller size. The residual weak quota is the exact finite-rank floor formula. Commitments again get zero dual lower quota.

In this mode all small scaling multipliers are one. The decomposition already returns integer original type bases. Although reconstruction calls the common upper-prefix routine, all these prefix values are integral. Its full-machine small count caps are tight because their sum equals the original mandatory small total. The final small count on every machine is consequently unchanged; all smaller types have the same size, so no additive rounding loss is incurred. This gives the main proof's exact two-size commitment guarantee.

The high-target branch uses the original all-job fractional LP and joint upper rounding. Boolean-rank loss only rejects the hard regime, where a full target would assign at most one largest unit per machine. One-size systems use exact capped bases; empty and zero-rank systems keep their original zero totals.

Both clients keep actual outputs in an original-load-grid search. Only genuine original failures raise the lower endpoint; there is no private-cap rejection in this adapter. Adjacent endpoints place the saved upper target at most the original optimum. The optional default fallback is an actual original all-job LP/additive-rounding schedule; returning the better output preserves its separate additive-largest-job bound. The external specialized two-size polynomial approximation is not implemented by these modules.

6. Local eligibility and exact scalar padding certificates

eligibility reads original singleton ranks. It retains the minimum and maximum eligible large sizes and the small singleton workload at each original machine. To find the smallest possible large collision, it retains at most two copies of each eligible type's size, independent of its numerical capacity. The minimum pair can occur in full bases because each original type's selected one- or two-unit vector is independent and extends to a base. Conversely every actual collision obeys these singleton capacities. Its returned threshold is therefore exact; None denotes an impossible collision.

The same routine constructs the Boolean recipient matroid by unit-box restriction of the sum of original large ranks. It does not identify the original total with its capped rank when the cap loses rank.

padding_scale implements the optional exact one-dimensional optimization in the eligibility manuscript. It scans intervals between maximum eligible sizes. Within each interval the active set is constant and the feasibility inequality is linear. At an endpoint, machines with maximum size equal to that endpoint are removed from the active set, exactly as its strict definition requires. The first feasible interval candidate is the minimum attainable overhead. For coefficient one the direct width solution applies. The returned positive scale realizes the computed overhead in the original piecewise expression.

This optimizes the scalar padding certificate, not the joint choice of a commitment set or an original schedule. The full heterogeneous-large padding allocator is not implemented here. Its proof remains separate, including the collision branch and the caveat about unconditional approximation coefficients.

7. Representation costs and independent evidence

After rank validation, the full absorption order uses O(q**2*m) original rank lookups and exact operations, and stores O(q*m) count records. Complete table validation and construction of contracted or Boolean rank tables have explicit subset costs, polynomial in the supplied complete-table size. Boxed greedy recovery similarly enumerates subsets of the remaining original ground. No loop depends on a count or rank budget's numerical magnitude. The scalar padding scan uses polynomially many comparisons and rational operations in the number of eligibility pairs. Actual integer, rational and original-ID encodings are charged throughout.

The workload clients reuse the streamed active-system LP solver with its explicit combinatorial factor. The residual selector enumerates combinations of original machines. These costs differ from the theorem's paid polynomial LP/intersection primitives and rank-parameter local search. Only the current candidate and active system are stored; capacities and workload multipliers are not expanded, and full solution or vertex families are not retained.

Three independent checkers compare arbitrary legal absorption orders, original job base schedules, original optimum loads, all feasible player commitments, and piecewise-affine scalar envelopes. They include a target feasible under position commitment but rejected by the stricter private-cap method, noninteger size ratios, decreasing marginal loads, entire bottleneck plateaus, exact active set endpoints and 100-digit capacities. These finite checks support the explicit proofs; they are not proof-assistant verification or external peer review.

Attribution: icekylinx and GPT-6 Astra Pro.