Updates
-Weekly research briefing
+A critical weekly selection of worthwhile papers, preprints, and technical reports in numerical optimization, numerical linear algebra, nonlinear least squares, and nonlinear equations. Each issue explains why the work matters and where to begin reading.
-Nothing posted yet — check back soon.
+Four selected papers, led by a decentralized objective-function-free trust-region method with centralized-order complexity guarantees.
+ + Read issue 4 → +Four core selections and one strategic read on optimization-driven progress in matrix multiplication.
+ + Read issue 2 → +The inaugural shortlist: five methodologically substantial contributions across Krylov methods, optimization, nonlinear least squares, and structured KKT systems.
+ + Read issue 1 → +Computational Mathematics Research Digest · Issue 1
Coverage: substantive postings from 2–12 August 2026, ranked by relevance and potential research value.
+Why it matters
For Hermitian positive-definite A, the authors show that Lanczos approximations to f(A)b are within an explicit condition-number-dependent factor of the best approximation in the same Krylov space. The result covers Stieltjes-related functions and the constant is proved optimal.
Read first
Section 1.2, especially Theorems 1–2 and Figure 2, followed by Remark 3 for the sharpened Euclidean-norm interpretation for conjugate gradients.
Why it matters
The method combines adaptively scaled proximal subproblems with an inner Newton–line-search solver, including a reduced Newton system, merit function, diagonal scaling, stopping rules, and implementation-oriented pseudocode.
Read first
Algorithms 1–2 and Sections 2.0.1 and 3.1 for the outer regularization and inner solver, then Section 4’s tests on 100 benchmark problems.
Why it matters
The framework separates basis learning from coefficient fitting, handling the latter through Levenberg–Marquardt and exposing more exploitable least-squares structure than ordinary end-to-end PINN training. It includes convergence results and reports errors up to two orders of magnitude below selected baselines.
Read first
Sections 3.3–3.4 and Theorem 1 for the alternating LM framework and convergence conditions, followed by Sections 4.3–4.5 on the high-frequency heat, cavity-flow, and Burgers problems.
Why it matters
The exact tree-contraction solver has O(N) work and storage but O(log N) parallel span, independently of tree height or balance. It reconstructs Riccati coefficients, primal variables, and multipliers, and is accompanied by JAX implementations.
Read first
Sections IV–V for the contraction construction, Theorem 3 for equivalence with the KKT system, and Section VII for parallel complexity.
Why it matters
The principal structural result converts a monotone linear-quadratic variational GNE problem into one convex quadratic program. Regularized and accelerated variants obtain O(1/k²) approximate-equilibrium convergence, while an invertible game Jacobian permits a smaller dual-space QP.
Read first
The QP-equivalence theorem and lower-dimensional reduction, followed by the game-theoretic MPC experiment and comparisons with extragradient-type methods.
Selection and commentary are editorial. Dates and bibliographic details link to the cited primary sources.
Computational Mathematics Research Digest · Issue 2
Coverage: substantive postings from 13–20 August 2026. Four core selections and one strategic cross-disciplinary read.
+Why it matters
A rescaling of Gaussian or hashing Johnson–Lindenstrauss matrices improves evaluation complexity from O(pn²ε−2) to O(pnε−2) without redesigning the underlying random-subspace model-based DFO method.
Read first
Section 2.2, particularly Corollary 2.4 and Remark 2.5, for why Q=Sᵀ/√n removes an entire factor of n and why further scaling does not improve the asymptotic result.
Why it matters
The infeasible interior-point method avoids nested outer–inner solution for constrained quadratic min–max problems, following a single central path and targeting nondegenerate stationary points.
Read first
Sections 2.3–2.5 for stationarity, Algorithm 1 in Section 3.2, and Sections 4.3–4.5 for scaling, coupled-constraint, and min-cost-flow tests.
Why it matters
PANDA combines proximal quasi-Newton iterations with adaptive steps and a matrix-free sensitivity calculation using implicit differentiation, automatic-differentiation products, and MINRES. Reported trailer-control tests show substantial forward, backward, and memory improvements.
Read first
Algorithm 1 and Section 3.2 for the forward/backward construction, then Figure 3 and Table 1. Broader benchmarking beyond the demonstrated NMPC family remains desirable.
Why it matters
The method targets ill-conditioned nonlinear parameter estimation with confounded parameters and heteroscedastic or serially correlated residuals. It updates the ridge parameter automatically within iterative generalized least squares.
Read first
Section 5 for the BIC/AICc hyperparameter fixed-point formulas, Section 6 for the full procedure, and Section 7’s battery state-of-health simulation.
Why it matters
A differentiable reformulation, Sinkhorn iterations, implicit differentiation, and GPU optimization lead to the certified bound ω<2.371177, illustrating numerical optimization’s contribution to linear-algebra complexity theory.
Read first
Section 3 for the optimization machinery and Section 4 for the exact-rational certification that converts a floating-point candidate into a rigorous result.
Selection and commentary are editorial. Dates and bibliographic details link to the cited primary sources.
Computational Mathematics Research Digest · Issue 3
Coverage: substantive postings from 21–27 August 2026, ranked by reading priority.
+Why it matters
This resolves a long-standing quasi-Newton question negatively. Classical DFP can fail even for a C² uniformly convex two-dimensional objective with global Hessian condition number at most three, while every accepted step satisfies strong Wolfe conditions.
Read first
Theorem 1 and Corollary 2 in Section 2, Section 3.2 for the construction’s mechanism, and Section 6.4 for the relation to existing convergence results.
Why it matters
A direct accelerated Newton scheme achieves a global O(k−3) functional-residual rate for convex functions with Lipschitz Hessian using only primal variables and one regularized linear solve per iteration. It admits Hessian-free inexact solves.
Read first
Equation (3) and Table 1 for the method and comparison, then Section 4 for the accelerated-rate proof and Section 5 for composite objectives and non-Euclidean geometry.
Why it matters
A diagonal-plus-rank-one factorization reduces a difficult scaled proximal map to a two-dimensional monotone system. The resulting oracle makes DFP/BFGS-type metrics usable with structured nonsmooth regularizers and has certified polylogarithmic accuracy dependence.
Read first
Algorithm 1 and Sections 3.1–3.2 for the oracle, Sections 3.3–3.4 for its Broyden interpretation, and Sections 5.2–5.4 for SLOPE and group-lasso experiments.
Why it matters
This extends asymptotic conditioning theory from scalar to fixed-block Toeplitz sequences, distinguishing bounded, polynomial, and exponential growth. It links the theory directly to stability of space-time isogeometric discretizations.
Read first
Section 3 for general bounds, Section 4 for banded block matrices, and Sections 5.1 and 5.3 for the wave and Schrödinger applications.
Why it matters
For Ax−|x|=b, CPPC replaces backtracking with an exact coordinate proximal predictor and a full-residual correction. Each iteration needs one new matrix–vector product, with whole-sequence, sublinear residual, and strong-monotonicity linear convergence guarantees.
Read first
Proposition 2.1, Algorithm 1 in Section 3.1, the rate results in Sections 4.1–4.2, and Section 5’s dimension and matrix-geometry experiments.
Selection and commentary are editorial. Dates and bibliographic details link to the cited primary sources.
Computational Mathematics Research Digest · Issue 4
Coverage: substantive postings from 28 August–3 September 2026. Four papers met the quality threshold.
+Why it matters
This brings adaptive trust-region globalization to decentralized, potentially nonconvex optimization without requiring shared objective values or gradient vectors. Agents communicate mainly scalar gradient norms, reducing privacy exposure and communication cost. Certain variants match the iteration-complexity order of centralized trust-region methods.
Read first
Section 3, especially Algorithms 3.1 and 3.2, for selective flooding and the DOFFO_TR iteration. Then examine Section 4’s complexity results and Section 5.3’s comparison with the Gradient Tracking Algorithm.
Why it matters
Gradient norm is a computable stationarity certificate and can be more informative than objective residual. The paper develops near-optimal oracle-complexity methods for Hölder-smooth convex objectives across the full range of ℓp geometries, closing previously open complexity gaps.
Read first
The main complexity table and theorem statements in the introduction, followed by the first algorithm family based on mirror duality and controlled inexactness. Then inspect the matching lower bounds.
Why it matters
The authors extend a d-th-order Newton method to problems with SOS-convex polynomial constraints. Each regularized Taylor subproblem can be represented by a semidefinite program. Under strong convexity and Lipschitz continuity of the d-th derivative tensor, the method has local convergence order d and identifies the optimal active set locally in one iteration.
Read first
The constrained SDP formulation, the theorem proving order-d local convergence, and the active-set identification theorem. Then examine the performance-estimation analysis of the unconstrained third-order method.
Why it matters
The paper proves that macroscopic density and microscopic stochastic-trajectory adjoints agree in the continuum. Although their gradients differ after discretization, both consistently approximate the same continuum gradient with explicit rates.
Read first
The continuum adjoint-correspondence theorem, followed by the two discrete-gradient convergence theorems and the numerical convergence plots comparing Eulerian and particle formulations.
Selection and commentary are editorial. Dates and bibliographic details link to the cited primary sources.
Computational Mathematics Research Digest · Issue NUMBER
+Coverage: REPLACE WITH DATE RANGE AND SHORT EDITORIAL SUMMARY.
+ +Why it matters
REPLACE WITH EDITORIAL ASSESSMENT.
Read first
REPLACE WITH SPECIFIC SECTION, THEOREM, ALGORITHM, OR EXPERIMENT.
Selection and commentary are editorial. Dates and bibliographic details link to the cited primary sources.
+ +