📐algorithm-gradient-descent-quadratic

The groupant beside algorithms/src/gradient_descent_quadratic/mod.rs in the Mova repository: what the code computes, what it never does, how to go back from its outputs, its invariants, cost and checks.

Facts

Components

Obernants

The typed relations around the forward model: the behaviour complement (what the code never does), the reverse reconstruction (every input compatible with an output) and specialised views.

Model text

descriptant

gradient_descent_quadratic mod.rs

Original standard-library numerical implementation and finite tests. Prospective predictions were saved before code in ../prospective.md.

For q(x)=half x'Dx-b'x, use g=Dx-b and x_next=x-g/L with L=max D. In exact arithmetic coordinate errors contract by 1-D_i/L in [0,1). Fresh Euclidean gradient norm is the stopping observation, not coordinate error.

19 diagonal spectra compared to analytic coordinate divisions b_i/D_i; fresh residual check, zero budget, already stationary input, ill-conditioned representational stagnation, shapes and invalid coefficients.

Literature: Burden and Faires, Numerical Analysis (bisection); Kiefer (1953), Sequential minimax search for a maximum (golden section); Nocedal and Wright, Numerical Optimization, second edition, chapter 3 (gradient descent); Hestenes and Stiefel (1952), Methods of conjugate gradients for solving linear systems (CG). No source code was copied.

2026-10-30 epoch: scoped safe bounds and explicit failure statuses were retained; finite observations do not establish universal convergence.

model

Mathematical combiner

Inputs: authoritative mod.rs, partial syntax projection, forward model, typed complement and reconstruction relations, and named finite observations. The combiner intersects accepted-domain constraints with update invariants, then partitions outputs by convergence, budget, stagnation or error. Cost is O(n*k) arithmetic, O(n) auxiliary memory through checked vector reservations; allocation failures are explicit. Tests are sibling sources, never implementation children. Source hashes bind this descriptive epoch; syntax is not semantic AST or proof.

Observation retained: 19 diagonal spectra compared to analytic coordinate divisions b_i/D_i; fresh residual check, zero budget, already stationary input, ill-conditioned representational stagnation, shapes and invalid coefficients.

Limits: callback determinism/continuity/unimodality and global SPD are assumptions; residual stopping alone says nothing universal about solution error. Conditioning, underflow, finite rounding and callback cost remain material. No speed measurements or universal accuracy certificate are claimed.

approximant

Forward approximant and invariant

Domain: Dimension 1..256, diagonal in [1e-100,1e100], finite linear and initial coordinates bounded by 1e100; positive finite gradient norm tolerance; budget <=100000.

For q(x)=half x'Dx-b'x, use g=Dx-b and x_next=x-g/L with L=max D. In exact arithmetic coordinate errors contract by 1-D_i/L in [0,1). Fresh Euclidean gradient norm is the stopping observation, not coordinate error.

Cost: O(n*k) arithmetic, O(n) auxiliary memory through checked vector reservations; allocation failures are explicit.

Invariant: accepted reports contain finite values; successful scalar intervals stay nested and vector residuals correspond to returned coordinates. Global mathematical assumptions remain caller obligations.

Numerical claims concern additive interval widths or absolute gradient/residual norms. A relative coordinate error has undefined behavior at zero without a separate convention and is not claimed. No supported 1/2/5/10% executable surrogate is attached: this exact API is already iterative, and no independently measured faster transformation with a useful certified metric is supplied.

models

Typed obernant family

behavior_complement

On the scoped input/report universe, the complement excludes fabricated convergence when neither declared stopping observation holds, invalid accepted shapes or nonfinite arithmetic results, and reports whose estimate leaves the retained bracket or whose vector residual is inconsistent. Domain: Dimension 1..256, diagonal in [1e-100,1e100], finite linear and initial coordinates bounded by 1e100; positive finite gradient norm tolerance; budget <=100000. This relation is broader than enumerated rejection tests. It does not claim the solver detects violated global callback/SPD assumptions.

reverse_reconstruction

At zero gradient, every positive diagonal D paired with b=Dx is a compatible antecedent for stationary x. A finite nonzero residual adds a tolerance ball around Dx-b; output x alone cannot recover D, b or initial trajectory.

diagonal_contraction

Exact-real factors 1-D_i/L are in [0,1); therefore q decreases unless stationary. Floating underflow and subtraction can break strict progress, reported as stagnation only when every coordinate is unchanged.

termination_partition

The bounded loop checks convergence first, then budget, then representational stagnation. Zero budget can still converge on the initial stopping observation. Iteration count is completed updates (including an attempted unchanged vector update). No unbounded search or implicit retry.

finite_arithmetic

Inputs reject NaN/infinity and restricted magnitude violations. Scalar callback results must remain finite; bounded endpoints keep differences and convex combinations finite. Vector norms use hypot; inner products and updates check finite arithmetic. Finite precision is not an exact-real proof; CG squared norm underflow is stagnation.

resource_observation

O(n*k) arithmetic, O(n) auxiliary memory through checked vector reservations; allocation failures are explicit.

tests

Checks

Run cargo test -p mova-algorithms gradient_descent_quadratic::tests and its release equivalent. 19 diagonal spectra compared to analytic coordinate divisions b_i/D_i; fresh residual check, zero budget, already stationary input, ill-conditioned representational stagnation, shapes and invalid coefficients.

Compiled negative control must hit diagonal coordinate oracle. Full algorithms suite runs twice per profile; normalized ordered test outcomes are compared. Strict all-target Clippy and focused rustfmt check complete validation. See validation.md for observed evidence.

prospective

Prospective numerical solver model

2026-10-30, before implementation. Original standard-library implementation, following the classical interval and quadratic optimization recurrences (Burden and Faires, Numerical Analysis; Nocedal and Wright, Numerical Optimization, second edition, chapters 3 and 5).

Scalar routines accept finite endpoints with magnitude at most 1e100 and positive finite absolute x tolerance, up to 100000 iterations. Bisection requires a continuous deterministic caller function and opposite endpoint signs; golden-section requires a deterministic strictly unimodal function (possibly with a unique boundary minimum). Function evaluations must be finite. Neither routine can validate those global caller assumptions. Nested intervals are the invariant. Width stopping is an additive coordinate observation, not relative error. Golden-section ties deterministically discard the right portion; bisection exact computed zero reports convergence but does not certify the exact-real function.

Vector routines accept 1..256 coordinates bounded by 1e100, positive finite tolerance, and up to 100000 iterations. Diagonal gradient descent requires coefficients in [1e-100,1e100], uses step 1/max diagonal, and minimizes half x'Dx - b'x. Dense conjugate gradients require exact symmetry and caller-assumed positive definiteness, with runtime positive-curvature checks. Every arithmetic vector and scalar is checked for finiteness; fresh residuals determine convergence. No residual-to-solution-error assertion without conditioning. Budget exhaustion and representational stagnation are distinct statuses. Checked reservations report allocation failure. Cost: scalar O(iterations) evaluations/O(1) storage; diagonal O(niterations)/O(n); dense O(n²iterations)/O(n) auxiliary storage. Rounding, underflow and ill-conditioning can cause stagnation; no universal accuracy or speed claim. No useful separately measured executable surrogate is claimed.

validation

Validation evidence

2026-10-30. External records under $MOVA_DATA/runs/octopus-wave19-solvers-2026-10-30/. Two debug and two release runs each passed 501 library tests (12 new focused tests); sorted named outcomes agree across four runs. Strict algorithms all-target Clippy and focused rustfmt passed.

Source SHA-256: 654341c8fb4b6fb1125090f2448a853909ab7b55248085b99e32ee96d06e6887. Negative production mutant failed diagonal coordinate oracle; restored bytes passed the focused suite.

• negative-gradient_descent_quadratic.log SHA-256 155d6eb36448f4ab36eb6b7d2924c5083e14e4d858d3277e976f9f065b35bc73
• restored-gradient_descent_quadratic.log SHA-256 de6de777c098e634fee1a7039f978a4798a261239f2fa86882312fdf8abf349c
• debug1.log SHA-256 61428a2e0c6dbc43cde2fa1cf9a4b021a6cebd75c04de1ecb3cf761fa95eccbf
• debug2.log SHA-256 73ffe6afebfbf497bd448639d77d254de8142e4c170e346e976eff376e95e496
• release1.log SHA-256 9a28e6d7f1af6cfbf056323b4ec824aca1899d9ed52b1635d30ed01ed8785afb
• release2.log SHA-256 17cdc616d33e1ca8add5062ebed8e5303eebab267338c981dbc2aecba57afd43
• clippy.log SHA-256 71ef179d48cc2d1e07e85d2e77009ed79e6e543be1e409886afdb92b50363422

Partial syntax projection generated independently twice and compared byte-for-byte. These are finite checks; allocation exhaustion and arbitrary caller functions/matrices are not exhaustively verified.

This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/algorithms/src/gradient_descent_quadratic/mod.rs.groupant/