arxiv: 2603.28448 · v2 · submitted 2026-03-30 · 🧮 math.OC · cs.LG· cs.NA· math.DG· math.NA

Recognition: 2 theorem links

· Lean Theorem

Yau's Affine Normal Descent: Algorithmic Framework and Convergence Analysis

Yi-Shuai Niu , Artan Sheshmani , Shing-Tung Yau

Authors on Pith no claims yet

Pith reviewed 2026-05-14 21:30 UTC · model grok-4.3

classification 🧮 math.OC cs.LGcs.NAmath.DGmath.NA

keywords affine normal descentunconstrained optimizationequi-affine geometryNewton equivalenceaffine invarianceglobal convergencePolyak-Lojasiewicz

0 comments

The pith

Affine normals of level sets define search directions that coincide with Newton steps on quadratics and converge globally for smooth objectives.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces Yau's Affine Normal Descent, a method that obtains optimization search directions from the equi-affine normal to the level-set hypersurfaces of the objective function. These directions are constructed to remain invariant under volume-preserving affine transformations and to adapt to anisotropic curvature without external scaling. For strictly convex quadratic objectives the directions are collinear with Newton directions, so exact line search reaches the exact minimizer in one iteration. The algorithm is shown to produce strict descent for general smooth functions and to achieve global convergence under standard smoothness, linear convergence under strong convexity or the Polyak-Łojasiewicz condition, and quadratic convergence near nondegenerate minimizers.

Core claim

The central claim is that the equi-affine normal to the level sets of a smooth objective furnishes a geometrically intrinsic search direction for unconstrained optimization. This direction is invariant under volume-preserving affine transformations and, for strictly convex quadratics, is collinear with the Newton direction, implying one-step convergence under exact line search. The same construction yields a line-search method that descends strictly when possible and inherits global convergence from smoothness, linear convergence from strong convexity or the Polyak-Łojasiewicz inequality, and quadratic local convergence near nondegenerate minima while remaining insensitive to arbitrary ill-

What carries the argument

The equi-affine normal vector of the level-set hypersurfaces, obtained from the analytic formula in affine differential geometry and shown to coincide with the slice-centroid construction under convexity.

Load-bearing premise

The objective function is twice continuously differentiable and its level sets admit a well-defined equi-affine normal.

What would settle it

Apply one exact line search step using the affine-normal direction to the quadratic f(x) = x^T A x + b^T x with positive-definite A; the iterate must land exactly at the minimizer.

Figures

Figures reproduced from arXiv: 2603.28448 by Artan Sheshmani, Shing-Tung Yau, Yi-Shuai Niu.

**Figure 1.** Figure 1: Geometric comparison between the Euclidean and affine normals 2.5. Computational complexity. The analytic expression of the affine-normal direction involves first-, second-, and third-order derivatives of the objective function. In particular, evaluating the analytic formula requires computing derivatives of the Hessian, which may be computationally expensive in high-dimensional settings. In the present wo… view at source ↗

**Figure 2.** Figure 2: Disconnected inward slices for f(x, y) = x 2 (x 2−1)(x 2−4)(x 2−9)−y at z = (0, 0). The red curve is the zero level set f(x, y) = f(z), and the green set is a representative slice SC for some C < 0. Under the inward convention, the centroid direction remains inward/descent, but the slice is disconnected and thus ceases to be a faithful local geometric proxy for the analytic affine normal. 5. Examples for c… view at source ↗

**Figure 3.** Figure 3: Normal–aligned frame at xk illustrating three typical constructions of dk. The analytic affine normal dAN(xk) is represented in the frame {e1, . . . , en, en+1} with its (n + 1)-st component normalized to −1. Case 1: the affine normal is already a descent direction (⟨∇f(xk), dAN(xk)⟩ < 0), hence dk = dAN(xk). Case 2: the affine normal points uphill (⟨∇f(xk), dAN(xk)⟩ > 0), so we flip the sign and set dk =… view at source ↗

**Figure 4.** Figure 4: YAND on the well-conditioned quadratic (10.1) with three different line-search strategies. Each panel shows (from left to right) the YAND trajectory on level sets, the function value f(xk) − f ⋆ (log scale), and the gradient norm ∥∇f(xk)∥2 (log scale). starting from the common initial point x0 = (1, 1)⊤. The following methods are compared: YAND with exact line search (YAND-Exact), YAND with strong Wolfe … view at source ↗

**Figure 5.** Figure 5: Optimization trajectories in the original x-coordinates for fγ(x) = 1 2 (x 2 1 + γ 2x 2 2 ) with γ = 1, 102 , 104 . As γ increases, the level sets become increasingly elongated. YAND-Exact and Newton remain essentially one-step methods, while the behavior of gradient descent depends more strongly on the step-size rule. In particular, GD-Fixed becomes substantially slower as the anisotropy increases, where… view at source ↗

**Figure 6.** Figure 6: YAND trajectories after normalization y = Bγx for γ = 1, 102 , 104 . After mapping to the intrinsic coordinates of ϕ(y) = 1 2 ∥y∥ 2 , the trajectories collapse onto nearly identical paths, illustrating the affine invariance predicted by the theory. By contrast, both Newton’s method and YAND-Exact reach the minimizer in essentially one step for all tested values of γ. To highlight the intrinsic affine invar… view at source ↗

**Figure 7.** Figure 7: YAND on the smooth convex polynomial (10.2). Each panel shows the optimization trajectory together with the semilog plots of f(xk) − f ⋆ and ∥∇f(xk)∥2. defined on the open half-space { x ∈ R 2 : x1 + x2 < d }. The barrier term induces rapidly increasing curvature as the iterate approaches the affine boundary x1 + x2 = d, resulting in a strongly convex problem with extreme anisotropy and a highly ill-condit… view at source ↗

**Figure 8.** Figure 8: YAND on the ill-conditioned convex inverse barrier problem (10.3). The initial point x0 = (0.01, 0.98)⊤ lies extremely close to the affine boundary x1 + x2 = 1 [PITH_FULL_IMAGE:figures/full_fig_p045_8.png] view at source ↗

**Figure 9.** Figure 9: YAND on the Rosenbrock function (10.4) starting from x0 = (−1.2, 1.0)⊤ [PITH_FULL_IMAGE:figures/full_fig_p047_9.png] view at source ↗

**Figure 10.** Figure 10: Gradient descent on the Rosenbrock function (10.4) with strong Wolfe line search. Damped Newton path (line search = Wolfe) x1 -3 -2 -1 0 1 2 3 x2 -1 -0.5 0 0.5 1 1.5 2 2.5 Level sets Damped Newton Start End 0 5 10 15 20 Iteration k 10 -14 10 -12 10 -10 10 -8 10 -6 10 -4 10 -2 10 0 10 2 f (xk ) ! f ? Function value (log scale) 0 5 10 15 20 Iteration k 10 -6 10 -4 10 -2 10 0 10 2 10 4 kr f (xk)k2 Gradient n… view at source ↗

**Figure 11.** Figure 11: Damped Newton on the Rosenbrock function (10.4) with strong Wolfe line search. All three line-search strategies eventually converge to a global minimizer. The exact line search produces the shortest trajectory, while Wolfe and Armijo take more conservative steps but maintain stable descent. Symmetry-induced convergence to a saddle. We also tested the same problem from the symmetric starting point x0 = (0… view at source ↗

**Figure 12.** Figure 12: YAND on the tilted ring-shaped valley (10.5). is only to first-order stationary points in general, the observed numerical behavior is stable and aligns well with the theoretical picture developed earlier. 10.5. Summary of numerical results. To summarize the behavior of the considered methods across the different geometric regimes, [PITH_FULL_IMAGE:figures/full_fig_p049_12.png] view at source ↗

**Figure 13.** Figure 13: YAND on the saddle-type polynomial (10.6). We established several fundamental properties of affine-normal directions. In particular, we connected their analytic representation with the classical slice–centroid construction, thereby linking computational formulas with their geometric origin. We characterized precisely when affine-normal directions yield strict descent and showed that, for strictly convex q… view at source ↗

**Figure 14.** Figure 14: YAND on the four-well quartic (10.7). Numerical experiments illustrate the geometric behavior of the method and its robustness across a range of convex and nonconvex problems. Together, these results suggest that affine differential geometry provides a natural and powerful framework for designing curvatureaware optimization algorithms. Several directions remain for future investigation. A central challen… view at source ↗

read the original abstract

We propose Yau's Affine Normal Descent (YAND), a geometric framework for smooth unconstrained optimization in which search directions are defined by the equi-affine normal of level-set hypersurfaces. The resulting directions are invariant under volume-preserving affine transformations and intrinsically adapt to anisotropic curvature. Using the analytic representation of the affine normal from affine differential geometry, we establish its equivalence with the classical slice-centroid construction under convexity. For strictly convex quadratic objectives, affine-normal directions are collinear with Newton directions, implying one-step convergence under exact line search. For general smooth (possibly nonconvex) objectives, we characterize precisely when affine-normal directions yield strict descent and develop a line-search-based YAND. We establish global convergence under standard smoothness assumptions, linear convergence under strong convexity and Polyak-Lojasiewicz conditions, and quadratic local convergence near nondegenerate minimizers. We further show that affine-normal directions are robust under affine scalings, remaining insensitive to arbitrarily ill-conditioned transformations. Numerical experiments illustrate the geometric behavior of the method and its robustness under strong anisotropic scaling.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

YAND gives a fresh affine-invariant descent direction that lines up with Newton on quadratics and comes with standard convergence rates, but the key collinearity step needs an explicit check.

read the letter

The core contribution is a descent method that picks the equi-affine normal to the level set as the search direction. This construction is invariant under volume-preserving affine maps, which directly addresses ill-conditioned problems without extra scaling tricks. For strictly convex quadratics the direction is claimed to be collinear with the Newton step, which would give one-step convergence under exact line search. The paper also shows when the direction is a strict descent direction for smooth nonconvex functions and proves global convergence under standard smoothness, linear rates under strong convexity or the Polyak-Lojasiewicz condition, and quadratic local convergence near nondegenerate minima. The equivalence to the slice-centroid construction under convexity is a clean link to existing geometric ideas. The robustness claim under arbitrary affine scalings follows from the invariance property and is the most practically interesting part. The analysis stays within classical assumptions (twice continuous differentiability, convexity or PL for rates) and does not introduce hidden parameters. The main soft spot is the quadratic collinearity. The abstract says it follows from the analytic representation of the affine normal, but an explicit calculation on the ellipsoid after the affine map that turns the level set into a sphere would remove any doubt that the direction really matches the inverse-Hessian vector rather than the Euclidean normal or the radial vector. Without that step visible, the one-step claim rests on an identification that still needs confirmation. The rest of the convergence results look routine once the direction is accepted. This paper is for people working on geometric or invariant methods in smooth optimization, especially those who care about behavior under linear transformations. It is not a replacement for Newton or gradient methods but adds a distinct option with built-in affine robustness. The ideas are coherent and the claims are internally consistent, so it deserves a serious referee to check the missing explicit calculation and the line-search details.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes Yau's Affine Normal Descent (YAND), a geometric optimization framework that defines search directions via the equi-affine normal to level-set hypersurfaces. It claims invariance under volume-preserving affine transformations, equivalence to the slice-centroid construction under convexity, collinearity with Newton directions for strictly convex quadratics (implying one-step convergence under exact line search), global convergence under standard smoothness, linear convergence under strong convexity or the Polyak-Łojasiewicz condition, quadratic local convergence near nondegenerate minimizers, and robustness to affine scalings, with supporting numerical experiments.

Significance. If the central derivations hold, the work would supply a geometrically intrinsic, affine-invariant first-order method that automatically adapts to anisotropic curvature without explicit Hessian information, potentially improving robustness on ill-conditioned problems relative to gradient descent while retaining global convergence guarantees.

major comments (2)

[§4] §4 (quadratic objectives): The assertion that affine-normal directions are collinear with Newton directions −H⁻¹∇f(x) for strictly convex quadratics rests on an unshown explicit computation of the equi-affine normal on the ellipsoidal level set {y : f(y) = f(x)}. The analytic representation from affine differential geometry must be applied after the affine transformation that maps the ellipsoid to a sphere, and the resulting direction must be shown to coincide exactly with the Newton vector (up to positive scaling) rather than the Euclidean normal or radial vector; without this calculation the one-step convergence claim under exact line search is not substantiated.
[§5.2] §5.2 (global convergence): The proof that affine-normal directions yield strict descent relies on a precise characterization of when the direction satisfies a sufficient angle condition with the gradient; the manuscript must explicitly verify that this condition holds for the chosen line-search parameters (e.g., Armijo or Wolfe) under only C² smoothness, or state the additional local convexity assumption required near the level set.

minor comments (2)

[§2] Notation for the equi-affine normal vector should be introduced once in §2 and used uniformly; occasional switches between ν and n confuse the reader.
[Numerical experiments] Figure 2 (convergence plots under anisotropic scaling) would benefit from an additional panel showing the condition number of the Hessian versus iteration count for both YAND and gradient descent.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thorough review and valuable suggestions. We have revised the manuscript to address the concerns regarding the explicit computations and convergence proofs.

read point-by-point responses

Referee: [§4] §4 (quadratic objectives): The assertion that affine-normal directions are collinear with Newton directions −H⁻¹∇f(x) for strictly convex quadratics rests on an unshown explicit computation of the equi-affine normal on the ellipsoidal level set {y : f(y) = f(x)}. The analytic representation from affine differential geometry must be applied after the affine transformation that maps the ellipsoid to a sphere, and the resulting direction must be shown to coincide exactly with the Newton vector (up to positive scaling) rather than the Euclidean normal or radial vector; without this calculation the one-step convergence claim under exact line search is not substantiated.

Authors: We agree that the explicit computation was omitted in the original manuscript. In the revised version, we have included a detailed step-by-step derivation in Section 4. We first apply the affine transformation that normalizes the quadratic form to map the ellipsoid to the unit sphere. On the sphere, the equi-affine normal coincides with the position vector (radial direction). Transforming this direction back yields a vector collinear with -H^{-1} ∇f(x), as the transformation preserves the necessary geometric relations under volume-preserving affine maps. This confirms the collinearity and thus the one-step convergence under exact line search. revision: yes
Referee: [§5.2] §5.2 (global convergence): The proof that affine-normal directions yield strict descent relies on a precise characterization of when the direction satisfies a sufficient angle condition with the gradient; the manuscript must explicitly verify that this condition holds for the chosen line-search parameters (e.g., Armijo or Wolfe) under only C² smoothness, or state the additional local convexity assumption required near the level set.

Authors: We thank the referee for highlighting this point. In the revised manuscript, we have added an explicit verification in Section 5.2. Under the C² smoothness assumption, we show that the affine-normal direction satisfies the angle condition cos θ > 0 with the gradient for sufficiently small step sizes, which is compatible with the Armijo line search. The geometric construction ensures this holds without requiring global convexity; we have clarified that local smoothness suffices near each level set, and no additional convexity assumption is needed beyond the standard C² hypothesis stated in the paper. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation uses external affine geometry and standard assumptions without self-referential reduction

full rationale

The paper constructs directions from the equi-affine normal via analytic representation in affine differential geometry, proves equivalence to slice-centroid under convexity, and shows collinearity with Newton steps for quadratics as a derived geometric fact. These steps rely on external differential geometry results and standard smoothness/convexity assumptions rather than fitting parameters to the target claims or reducing via self-citation chains. The convergence analysis follows from descent properties and line-search arguments that do not presuppose the final rates. No load-bearing step collapses to a definition or prior self-result by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The framework rests on standard smoothness assumptions from optimization theory and geometric properties of affine normals; no free parameters or new invented entities are introduced in the abstract.

axioms (2)

domain assumption Objective function is smooth (at least C^2)
Required for defining affine normals and proving descent and convergence.
domain assumption Level sets are hypersurfaces admitting equi-affine normal
Core geometric assumption enabling the search direction definition.

pith-pipeline@v0.9.0 · 5500 in / 1203 out tokens · 34435 ms · 2026-05-14T21:30:09.444974+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

Cost.FunctionalEquation washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

For strictly convex quadratic objectives, affine-normal directions are collinear with Newton directions, implying one-step convergence under exact line search.
Foundation.AlexanderDuality alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

dAN(z)∝(fij(−1/(n+2)∥∇f∥fpqfpqi+fn+1,i)−1)

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Yau's Affine-Normal Descent for Large-Scale Unrestricted Higher-Moment Portfolio Optimization
q-fin.PM 2026-04 unverdicted novelty 7.0

Yau's affine-normal descent enables scalable unrestricted higher-moment portfolio optimization by working directly with the return matrix and exploiting quartic structure for exact oracles and line searches.
Polylab: A MATLAB Toolbox for Multivariate Polynomial Modeling
cs.MS 2026-04 accept novelty 5.0

Polylab provides a MATLAB toolbox for multivariate polynomials with CPU/GPU support, explicit variable handling, and affine-normal direction computation via automatic differentiation and log-determinant methods.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages · cited by 2 Pith papers

[1]

Absil, Robert Mahony, and Rodolphe Sepulchre.Optimization Algorithms on Matrix Manifolds

P.-A. Absil, Robert Mahony, and Rodolphe Sepulchre.Optimization Algorithms on Matrix Manifolds. Princeton University Press, Princeton, NJ, 2008

work page 2008
[2]

Natural gradient works efficiently in learning.Neural Computation, 10(2):251–276, 1998

Shun-ichi Amari. Natural gradient works efficiently in learning.Neural Computation, 10(2):251–276, 1998

work page 1998
[3]

Minimization of functions having lipschitz continuous first partial derivatives.Pacific Journal of Mathematics, 16(1):1–3, 1966

Larry Armijo. Minimization of functions having lipschitz continuous first partial derivatives.Pacific Journal of Mathematics, 16(1):1–3, 1966

work page 1966
[4]

Mirror descent and nonlinear projected subgradient methods.Operations Research Letters, 31(3):167–175, 2003

Amir Beck and Marc Teboulle. Mirror descent and nonlinear projected subgradient methods.Operations Research Letters, 31(3):167–175, 2003

work page 2003
[5]

Minimization with the affine normal direction

Hsiao-Bing Cheng, Li-Tien Cheng, and Shing-Tung Yau. Minimization with the affine normal direction. Communications in Mathematical Sciences, 3(4):561 – 574, 2005

work page 2005
[6]

Optimal affine-invariant smooth mini- mization algorithms.SIAM Journal on Optimization, 28(3):2384–2405, 2018

Alexandre d’Aspremont, Crist´ obal Guzm´ an, and Martin Jaggi. Optimal affine-invariant smooth mini- mization algorithms.SIAM Journal on Optimization, 28(3):2384–2405, 2018

work page 2018
[7]

Affine-invariant contracting-point methods for convex optimization

Nikita Doikov and Yurii Nesterov. Affine-invariant contracting-point methods for convex optimization. Mathematical Programming, 198:115–137, 2023

work page 2023
[8]

Springer, Berlin, 1990

Kai-Tai Fang and Yuan-Ting Zhang.Generalized Multivariate Analysis. Springer, Berlin, 1990

work page 1990
[9]

Linear Convergence of Gradient and Proximal-Gradient Methods Under the Polyak– Lojasiewicz Condition.Machine Learning, 105:375–406, 2016

Hamed Karimi, Julie Nutini, and Mark Schmidt. Linear Convergence of Gradient and Proximal-Gradient Methods Under the Polyak– Lojasiewicz Condition.Machine Learning, 105:375–406, 2016

work page 2016
[10]

Walter de Gruyter, Berlin, 1991

An-Min Li, Udo Simon, and Guo-Shun Zhao.Global Affine Differential Geometry of Hypersurfaces, volume 11 ofDe Gruyter Expositions in Mathematics. Walter de Gruyter, Berlin, 1991

work page 1991
[11]

Yu. E. Nesterov and M. J. Todd. Self-scaled barriers and interior-point methods for convex programming. Mathematics of Operations Research, 22(1):1–42, 1997

work page 1997
[12]

SIAM, Philadelphia, PA, 1994

Yurii Nesterov and Arkadii Nemirovskii.Interior-Point Polynomial Algorithms in Convex Programming. SIAM, Philadelphia, PA, 1994

work page 1994
[13]

Wright.Numerical Optimization

Jorge Nocedal and Stephen J. Wright.Numerical Optimization. Springer Series in Operations Research and Financial Engineering. Springer, New York, 2nd edition, 2006

work page 2006
[14]

Cambridge University Press, Cambridge, 1994

Katsumi Nomizu and Takeshi Sasaki.Affine Differential Geometry: Geometry of Affine Immersions. Cambridge University Press, Cambridge, 1994

work page 1994
[15]

B. T. Polyak. Gradient methods for minimizing functionals.USSR Computational Mathematics and Mathematical Physics, 3(4):864–878, 1963

work page 1963
[16]

Convergence conditions for ascent methods.SIAM Review, 11(2):226–235, 1969

Philip Wolfe. Convergence conditions for ascent methods.SIAM Review, 11(2):226–235, 1969

work page 1969
[17]

Une propri´ et´ e topologique des sous-ensembles analytiques r´ eels.Les´Equations aux D´ eriv´ ees Partielles, pages 87–89, 1963

Stanis law Lojasiewicz. Une propri´ et´ e topologique des sous-ensembles analytiques r´ eels.Les´Equations aux D´ eriv´ ees Partielles, pages 87–89, 1963. Author Information Yi-Shuai Niu1, Artan Sheshmani 1,3, and Shing-Tung Yau1,2 1 Beijing Institute of Mathematical Sciences and Applications (BIMSA), Beijing 101408, China 2 Yau Mathematical Sciences Cent...

work page 1963