gpt 6 astra pro attempt on CS-37

Polynomial pivot rule for the simplex method

Definitions and mathematical statement

In standard form a rational linear program has $Ax=b$, $x\ge0$, and a linear objective. A feasible basis selects an invertible basic submatrix and yields a feasible basic solution. A simplex pivot exchanges one basic and one nonbasic variable according to the rule while maintaining the algorithm's feasibility and objective conventions. For bounded nondegenerate inputs, seek a rule with \[ \sup_{\text{allowed LPs and feasible bases}}\mathbb E[\text{pivots to an optimum}] \le p(m,n) \] for one polynomial independent of coefficient bit lengths. For a deterministic rule omit expectation. The catalogue requests a pivot-count bound and does not separately impose an efficient per-pivot selection model.

Formulation and concept references: [S1].

Short English statement

Is there a simplex pivot rule that always reaches an optimum after polynomially many pivots, independent of coefficient magnitudes?

Research attempt

Current frontier and scope.

The catalogue asks for a polynomial pivot count; it does not require polynomial-time computation of the next pivot. That omission is mathematically important, not a defect to be silently repaired. [S1] The August 2026 revision of Disser--Loho--Maat--Mosis proves lower bounds for specified ranking-based classes of pivot rules, not for every rule allowed by this broad selection model. [E2] Black's September 2026 preprint constructs lattice polytopes in $[0,3]^{6d}$ with monotone diameter at least $2^d-1$. Its parameter is geometric dimension with a coordinate bound, not immediately the catalogue's pair $(m,n)$ for nondegenerate equality form. [E1] These recent statements are recorded with their hypotheses and are not treated as independently audited universal resolutions.

Attack map.

One route constructs a potential decreasing at every legal pivot and bounded by a polynomial in $m,n$; its missing property is a uniform range independent of numerical coefficients. A second searches all paths and chooses a shortest improving one, exploiting the absence of a per-pivot efficiency condition. That route reduces the problem to a geometric distance question, but does not automatically bound the distance. A counterexample route must force every legal path to be long in the correct input parameters; an exponentially long trajectory of one favorite rule is insufficient. The shortest-path route and its exact quantifiers are investigated here.

Attempt I: the directed feasible-basis graph.

Assume redundant equations have been removed so that $A$ has row rank $m$; this operation cannot increase either relevant input parameter. For a bounded nondegenerate feasible rational program in standard form, let $\mathcal B$ be its feasible bases. Every basis corresponds to a vertex, and nondegeneracy makes that correspondence one-to-one. Join two bases when a feasible single-variable exchange connects them. Orient an edge toward strictly larger objective value; optimal bases are terminal targets.

A nonoptimal vertex has an improving incident edge. Here is the usual geometric reason rather than an assumption of termination: if $v$ is nonoptimal and $w$ is better, the feasible direction $w-v$ lies in the polyhedral tangent cone at $v$. That cone is generated by the incident edge directions for a bounded polytope. A positive value of the objective on $w-v$ forces a positive value on at least one generator. Nondegeneracy realizes that edge as a simplex pivot. Every improving walk is finite because the objective strictly increases on a finite vertex set. Hence each feasible basis has a directed path to an optimum.

Let $\ell(B)$ be the length of a shortest such path. If the convention also permits zero-objective pivots, include those edges in the legal graph and use shortest distances there instead. Improving paths still ensure reachability. The equivalence proved next works with either specified convention; it does not exploit worsening pivots.

Pivot-count equivalence. In the catalogue's model, a deterministic or randomized polynomial pivot-count rule exists if and only if there is a polynomial $p(m,n)$ such that every allowed input and every feasible starting basis satisfy \begin{equation}\label{wsref-0-33-30-30-30-36-37-34-32-3a-64-69-73-74-61-6e-63-65} \ell(B)\leq p(m,n). \end{equation} Moreover, if any randomized rule has this expected bound, a uniform deterministic rule with no worse pivot-count bound exists, although its per-pivot running time may be exponential.

Proof. Every realized sequence of legal pivots reaching an optimum has length at least $\ell(B)$. Thus an expected-length bound implies (30006742:distance); finite expected termination also excludes nontermination with positive probability. Conversely, enumerate all $\binom nm$ candidate bases using exact rational arithmetic, test invertibility and feasibility, and construct the legal exchange graph. Compare the finitely many exact objective values to identify all optimal bases. Reverse breadth-first search from them computes $\ell$ on every feasible basis. Choose, with a deterministic lexicographic tie-break, an outgoing edge from $B$ to a basis with distance $\ell(B)-1$. Repeating reaches an optimum in exactly $\ell(B)$ pivots. The enumeration is finite and uniformly computable from the input. It need not be efficient, and no length-dependent advice or existence-only oracle is being used. $\square$

The construction is therefore a legitimate rule in the specified model. The unproved part is (30006742:distance), not the existence or computability of the shortest-path selector. Enumerating all bases without proving a short path would yield only the elementary bound \[ \ell(B)\leq|\mathcal B|-1\leq\binom nm-1. \] This is polynomial when $m$ or $n-m$ is fixed, but not for jointly varying $m,n$.

Attempt II: geometric consequences and a calibration family.

A full-dimensional simple rational $d$-polytope with $f$ facets admits a slack-coordinate representation in ${\mathbb{R}}^f$: \[ s_i=b_i-a_i\cdot x\geq0\quad(1\leq i\leq f),\qquad As=b, \] where the affine relations among the slacks have rank $f-d$. The affine slack map is injective for a bounded full-dimensional polytope: a direction annihilated by all facet normals would otherwise be a lineality direction. A simple vertex has exactly $d$ zero slacks and $f-d$ positive ones, giving a nondegenerate feasible basis in this representation. Edges and objective improvements are preserved. Thus a polynomial bound in (30006742:distance) would imply polynomial monotone distances to an optimum in terms of $f,d$ for such polytopes.

With a generic rational objective there is a unique optimum. Routing two arbitrary vertices to that same vertex would also bound their undirected graph distance by $2p(f-d,f)$. Hence the proposed pivot bound has a polynomial polytope-diameter consequence. The converse inference is not valid: a short undirected path may decrease the objective, and the route required by simplex is directed. No polynomial-diameter conjecture is assumed as an established theorem.

For an exact small-model check, consider \[ x_i+y_i=1,\quad x_i,y_i\geq0\quad(1\leq i\leq d),\qquad \max\sum_{i=1}^d w_i x_i,\quad w_i>0. \] Here $m=d,n=2d$, all $2^d$ vertices are nondegenerate, and an improving pivot changes one zero $x_i$ to one. The distance to the unique optimum is exactly the number of zero $x_i$, at most $d$. The number of bases is exponential, but the shortest-path distance is linear. This separates the cost of the brute-force selector from its pivot count and prevents an exponential state count from being mistaken for a geometric lower bound.

Attempt III: audit a plausible counterexample transfer.

Black's recent preprint concerns genuine shortest monotone distances, not just one inefficient pivot trajectory: its construction traps improving paths in an exponentially long shadow path. This is a relevant obstruction to any argument bounding monotone diameter by a polynomial in geometric dimension and a fixed lattice coordinate range. [E1], Theorem 1.1 and Section 2.

To turn it into a counterexample to the selected statement, one would still need to give an allowed nondegenerate equality-form family with $m,n$ polynomial in the parameter $d$, preserve the long directed distance, and verify that the bound beats every polynomial in those $m,n$. The displayed construction starts with exponentially many listed vertices and uses a Minkowski sum. Its geometric dimension alone supplies no verified polynomial bound on the number of facets in the resulting description, and a change to slack variables need not preserve the small lattice-coordinate range. Nor may nondegeneracy be imposed by an arbitrary perturbation without checking that new short improving paths have not appeared.

The preprint itself distinguishes equality-form lattice polytopes in its introduction. This notebook does not prove either that its family has polynomially many facets or that such a conversion is impossible. It records the missing conversion rather than claiming a counterexample to a stronger or differently parametrized statement. Likewise, the ranking restrictions in [E2] exclude the arbitrary global shortest-path selector above, so those lower bounds do not refute (30006742:distance).

Stress test.

Multiple optimal bases are handled by reverse search from all of them; a tie does not require an illegal objective decrease. The proof uses exact rational feasibility and objective comparisons, not a numerical tolerance that could change the directed graph. Randomness cannot reduce the expected number of traversed edges below shortest-path distance. On the other hand, the deterministic selector's arithmetic work is unbounded by a polynomial in $m,n$, so it would not yield a strongly polynomial-time linear-programming algorithm merely from a polynomial pivot bound.

The companion script enumerates the complementary-variable cube bases for small $d$, computes the improving graph and reverse distances, and checks the distance formula against every starting basis. It also checks a finite directed-path example where an extra undirected edge is unusable because it decreases the objective. Such graph examples are diagnostics only, not claimed realizations as arbitrary LP graphs.

Outcome.

Outcome: Conditional reduction.

For the exact pivot-count model, existence of any expected-polynomial rule is equivalent to a uniform polynomial shortest legal path bound, and that bound would already admit a deterministic uniform selector. The reduction isolates the geometric statement that must be proved or disproved and audits the parameter and nondegeneracy obstacles in a recent potential counterexample transfer. No polynomial bound on those distances, or superpolynomial lower bound in the required parameters, has been established. No priority claim is made.

Sources

• S1 Math Conjectures, OPT-002, catalog snapshot 31 July 2026. https://mathconjectures.com/conjectures/OPT-002. Catalogue formulation source.
• E1 Alexander E. Black. Monotone Diameters of Lattice Polytopes, arXiv:2609.08647v1, 8 September 2026. Theorem 1.1 and Section 2; the geometric-dimension parameter and equality-form distinction in the introduction. Recent preprint; no conversion to the catalogue's required input family is presumed. https://arxiv.org/html/2609.08647v1. Consulted 27 September 2026 for the indicated result or scope.
• E2 Yann Disser, Georg Loho, Matthew Maat, and Nils Mosis. Lower bounds for ranking-based pivot rules, arXiv:2512.16684v2, 10 August 2026. Abstract and introduction; restrictions on the information used by the pivot rule. https://arxiv.org/abs/2512.16684v2. Consulted 27 September 2026 for the indicated result or scope.