A matrix-free interior point continuous trajectory for linearly constrained convex programming

Jie Sun; Li-Zhi Liao; Xun Qian

A matrix-free interior point continuous trajectory for linearly constrained convex programming

T0 review · 0 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper establishes that a matrix-free interior point augmented Lagrangian ODE, starting from any interior feasible point, converges to an optimal solution of a linearly constrained convex program under only the assumption that an…

desk verdict The convergence theorem is real: the ODE system avoids the projection matrix, and the proof that every cluster point is optimal and that x(t) converges to a maximal-support optimum is essentially sound; the gaps are expositional. read the letter →

arxiv 2412.20141 v1 pith:UOM7BQ4A submitted 2024-12-28 math.OC

Xun Qian , Li-Zhi Liao , Jie Sun This is my paper · ORCID

classification math.OC MSC 90C2590C5134A12

keywords matrix-freeinteriorpointmethodaugmentedLagrangiancontinuoustrajectoryordinarydifferentialequationlinearlyconstrainedconvexprogrammingpotentialfunctionstrictcomplementarityconvergenceanalysis

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Interior point methods for linearly constrained convex programming normally maintain the equality constraint at every step, which forces the formation of a projection matrix that can become ill-conditioned near the boundary. This paper proposes a continuous trajectory defined by an ordinary differential equation that keeps the nonnegativity-constrained variables positive and relaxes the equality constraint into an augmented Lagrangian term, so no projection matrix is needed. The central claim is that, starting from any interior feasible point, the trajectory always converges and its limit is an optimal solution of the original problem, under just the assumption that an optimal solution exists, plus mild smoothness and full row rank of the constraint matrix. With an additional strict complementarity condition, the corresponding Lagrange multipliers converge to an optimal solution of the Lagrangian dual problem. The result matters because it gives a matrix-free, projection-free continuous route to optimality and a foundation for discrete algorithms that avoid the ill-conditioning failure mode of classical interior point methods.

What carries the argument

The central object is the potential function $V(x,x',y,y')$ of (4)–(7), an extension of the Losert–Akin potential: for $\gamma=1/2$ it uses terms $x_i - x_i'\ln(x_i/x_i')$, and for $1/2<\gamma<1$ it uses power terms $(x_i^{2-2\gamma}-(x_i')^{2-2\gamma})/(2-2\gamma)$ minus a reciprocal-power term, with value $+\infty$ when the support condition $B(x')\subseteq B(x)$ fails. Choosing $x'=x^*$, an optimal point, and $y'=y^*$, an associated multiplier, yields a Lyapunov function $V_1$ whose time derivative along (1) satisfies $dV_1/dt \leq -\sigma_1\|Ax-b\|^2$. This single inequality delivers boundedness of $(x(t),y(t))$, global existence, and, via Barbalat's lemma, the limiting optimality statements. The ODE itself is matrix-free because the diagonal scaling matrix $U = \mathrm{diag}(x_1^\gamma,\dots,x_s^\gamma,1,\dots,1)$ replaces the projection matrix $I - XA^T(AX^2A^T)^{-1}AX$; only matrix–vector products with $A$ and $A^T$ are needed.

What would settle it

The central claim would be refuted by a linearly constrained convex program with finite optimum and an interior feasible starting point for which a high-accuracy numerical integration of the ODE (1) either fails to converge or accumulates at a non-optimal point; a direct test is to integrate the system on the paper's $A=0$ example and check that every trajectory lands on the optimal set.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the ODE system (1), namely $\dot x = -U^2[\nabla f(x)+A^T y + \sigma_1 A^T(Ax-b)]$ and $\dot y = \sigma_2(Ax-b)$ with $U_{ii}=x_i^\gamma$ for the nonnegativity-constrained coordinates and $U_{ii}=1$ elsewhere, has a globally defined solution for every interior feasible starting point, and its primal trajectory $x(t)$ converges to an optimal solution of (P). The convergence proof does not require the optimal solution set to be bounded; it only assumes a finite optimum is attained. The argument builds a Lyapunov-style potential function from a Losert–Akin construction, shows its derivative along the trajectory is bounded above by $-\sigma_1\|Ax-b\|^2$, and then uses Barbalat's lemma to force both the feasibility residual and the scaled gradient to zero. A further result identifies the limit as the optimal solution with the maximal number of positive components among optimal solutions, and, under strict complementarity, shows the multiplier pair $(y(t), z_S(t))$ converges to an optimal dual solution.

Load-bearing premise

The proof needs a chosen optimal solution to have an associated Lagrange multiplier vector from which the potential function is built; the paper does not list this as an explicit initial assumption, and without such a multiplier the Lyapunov argument has no starting point.

Editorial extensions

If this is right

If the central claim is correct, every linearly constrained convex program with a finite optimum can be approached from any interior feasible point by a trajectory that never forms or inverts the $m\times m$ matrix $AX^2A^T$.
Only existence of an optimal solution is needed: boundedness of the optimal solution set and strict convexity of $f$ are not required for convergence of $x(t)$.
The limit point has the maximal number of positive components among optimal solutions, so the trajectory naturally selects a maximally nondegenerate optimal solution.
Under strict complementarity, the multiplier trajectory converges to an optimal dual solution, so primal and dual information are obtained simultaneously from the same ODE integration.
The explicit and semi-implicit discretizations discussed in the paper give concrete search directions that inherit the matrix-free property, providing a starting point for discrete algorithms without ill-conditioned projections.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper, a natural next test is to prove global convergence and a rate for the explicit Euler discretization with a line search based on the same potential function, turning the trajectory into a practical algorithm.
The numerical example where the classical central path zig-zags but this trajectory converges suggests, beyond the paper, that this construction may be a continuous interior route that avoids ill-behaved central paths in degenerate convex programs; proving this for a broader class would be a separate result.
The weighted-diagonal variant with $W=\mathrm{diag}(w_i)$ is asserted to work by similar proofs but is not analyzed; a testable extension is to verify convergence numerically and theoretically for block-coordinate or asynchronous discrete updates built from it.
The ODE's matrix-free character depends on the Euclidean augmented Lagrangian penalty; extending the idea to semidefinite or cone constraints would require a different scaling than $U^2$ and is not covered by the paper's arguments.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: the convergence theorem is real. The ODE system (1)—U^2 interior scaling plus augmented Lagrangian multiplier dynamics—is genuinely not in the cited literature, and the proof that every cluster point is optimal and that x(t) converges to a maximal-support optimum is essentially sound. The paper deserves a referee.

What's new: the matrix-free property is the point. The potential function is borrowed from your own [22], but it is applied to a new system, and the chain of inequalities in Theorem 3.2 correctly forces Ax(1)=b and f(x(1))=f(x*). Theorem 4.3(b)'s use of I(x,x(1)) as a separation potential is neat. The numerical example (24) does demonstrate the ill-conditioning failure of affine scaling when AX^2A^T degenerates.

Soft spots, all minor: - The abstract says "by only assuming the existence of an optimal solution," but Assumptions 2.2 and 2.3 (f in C^2, A full row rank) are also in play, and Theorem 2.8 silently asserts KKT multipliers exist for the chosen x*. For linear constraints this is standard polyhedral KKT, so it is an omitted citation, not a break. - Theorem 4.2 states lim L_sigma1(x(t),y(t)) = f(x*) with "not hard to see." It follows from boundedness and the cluster-point result of Theorem 3.2, but the sentence should be in the proof. - The claim lim Ax-b=0 "from Theorem 3.2" needs the same cluster argument; again a missing sentence, not a gap. - Section 6's discrete search directions are mostly speculative and some involve inverses, but the paper says this is future work, so that is fine.

Citation pattern is fine: [22] supplies the potential and two lemmas, and it is credited. The numerical section is illustrative rather than a benchmark, which is appropriate for a continuous-trajectory paper.

Verdict: send to a serious referee. The fixes are expositional; the mathematical core is sound.

Referee Report

0 major / 6 minor

Summary. This paper studies a continuous-time interior point augmented Lagrangian method for convex programs with linear equality constraints and nonnegativity bounds. The method is defined by the ODE system (1), which replaces the projection matrix used in classical affine scaling with a diagonal scaling and an augmented Lagrangian penalty. The main results are: the ODE has a unique solution on [t0,∞); the trajectory is bounded; every accumulation point of x(t) is optimal (Theorem 3.2); x(t) itself converges to an optimal solution (Theorem 4.3(b)); and under strict complementarity, the dual multiplier sequence converges to an optimal dual solution (Theorem 4.3(c)). The paper also presents two numerical examples and sketches several discrete search directions derived from the ODE.

Significance. The result is theoretically clean and would be a useful contribution to the continuous-time optimization literature. The matrix-free property is attractive for large-scale problems where forming A(X^2)A^T inverses is prohibitive. The Lyapunov-function proof is explicit and does not rely on numerical fitting or on assuming boundedness of the feasible set; only existence of an optimal solution is needed. The numerical examples, while small, illustrate the claimed difference from the affine scaling trajectory near degenerate solutions. The strengths of the paper are its careful Lyapunov analysis and the use of Barbalat's lemma to obtain convergence without strong convexity.

minor comments (6)

[Theorem 2.8] The proof asserts the existence of KKT multipliers (y*, z*) for the chosen optimal x* without proof or citation. This existence is standard for linearly constrained convex programs (polyhedral KKT theorem), but since the Lyapunov function V1 is defined using these multipliers, the step should be stated explicitly or referenced.
[Abstract and Theorem 4.3(b)] The abstract says "starting from any interior feasible point," but the ODE system (1) only requires x(t0) in R^n_{s++}; the equality constraint Ax0=b is not needed for the convergence results. Consider rewording to "any interior starting point" or "strictly positive point."
[Theorem 4.2, Eq. (19)] The claim "from Theorems 2.8 and 3.2, it is not hard to see that lim_{t→∞} L~_{σ1}(x(t),y(t)) = f(x*)" is terse. A short justification—all subsequential limits have Ax=b and f(x)=f(x*), with y bounded—would improve readability.
[Theorem 2.9, inequality (12)] The bound |dx_i/dt| ≤ L x_i for i≤s is not derived. It follows from boundedness of the bracket term and x_i^{2γ} ≤ M^{2γ-1} x_i for 2γ ≥ 1, but this step should be included explicitly.
[Section 5, problem (23)] In the first numerical example, the variable y is used both as a primal variable and, in the general ODE formulation, as a dual multiplier. This dual use of notation may confuse some readers; consider renaming the primal variable in the example.
[Section 6] The statement "by similar proofs, the same results of the ODE system (1) hold for this weighted ODE system" is made without proof. Since Section 6 is a discussion of possible search directions, this is acceptable, but the sentence should be phrased as a remark or conjecture rather than a proved assertion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the convergence proof is self-contained modulo standard convex-analysis facts.

full rationale

The derivation chain is not circular. The ODE system (1) is constructed from an augmented-Lagrangian step with an interior-point scaling U^2, but convergence is not built in by definition: the Lyapunov potential V1 in (4)-(10) is a genuine Lyapunov function whose derivative along the ODE is computed, not assumed, and Theorem 3.2 shows every cluster point satisfies Ax=b and f(x)=f(x*) using Barbalat's lemma. The potential function and Lemmas 2.6-2.7 are taken from the authors' earlier paper [22], but those lemmas are elementary scalar inequalities with explicit stated assumptions and are externally checkable; the paper does not cite or rely on [22]'s convergence theorem, so this is a minor self-citation, not load-bearing circularity. The one genuine presentational gap is Theorem 2.8's assertion that for the chosen optimal x* there exist KKT multipliers (y*,z*) satisfying (8); this is standard for linearly constrained convex programs (polyhedral KKT), but the paper neither states nor proves it. Since the omitted fact is true and independent of the ODE, the gap affects exposition rather than circularity. Theorem 4.3(a)-(c) also avoids circularity: the two-circle argument is a genuine geometric fact (internal tangency), and the strict-complementarity assumption is an extra hypothesis rather than an input that forces the conclusion. There are no fitted parameters being renamed as predictions, and the numerical examples are illustrative rather than claims of independent predictive success. I therefore find no circular step and assign score 0.

Assumptions & free parameters 3 free parameters · 8 assumptions · 0 invented entities

The central proof rests on the blanket assumptions (existence, smoothness, full row rank), an unstated KKT multiplier existence, standard lemmas, and, for the dual convergence, strict complementarity. The method itself has three user-chosen continuous parameters (gamma, sigma1, sigma2) but they are not fitted to data.

free parameters (3)

gamma = 0.75 in experiments; any value in [1/2,1) in theory
Interior scaling exponent in U=diag(x_i^gamma). Chosen by hand; convergence proofs hold for the whole interval.
sigma1 = 1 in experiments; any sigma1>0 in theory
Augmented Lagrangian penalty weight on ||Ax-b||^2. Set by user; positivity required for the Lyapunov decrease.
sigma2 = 1 in experiments; any sigma2>0 in theory
Gain for the multiplier update dy/dt=sigma2(Ax-b). Set by user; positivity required for the Lyapunov function.

assumptions (8)

domain assumption Existence of an optimal solution to (P) (Assumption 2.1).
Used to choose x* and define the potential V1 in Theorem 2.8; also gives finiteness of f(x*).
domain assumption f is twice continuously differentiable on R^n_{s+} (Assumption 2.2).
Ensures local Lipschitz right-hand side in Theorem 2.4 and uniform continuity of the Lyapunov derivative for Barbalat.
domain assumption A has full row rank (Assumption 2.3).
Assumed to keep the linear constraint nonredundant; used implicitly in KKT multiplier arguments.
domain assumption KKT multipliers (y*, z*) exist for the chosen optimal x* (Eq. 8).
Asserted in Theorem 2.8 without proof or explicit constraint qualification; standard for polyhedral feasible sets, but load-bearing.
standard math Barbalat's lemma (cited [26]).
Used in Theorems 3.2 and 4.2 to convert finite Lyapunov limits into vanishing derivatives.
standard math Lemma 4.1: gradient of f is constant on the optimal solution set ([16]).
Used in Theorem 4.3(c) to identify multipliers along the segment from y* to y(1).
domain assumption Strict complementarity between some primal and dual optimal solutions (Section 4(c)).
Needed only for the convergence of the Lagrange multiplier to a dual optimal solution.
standard math Properties of the Losert-Akin potential (Lemmas 2.6, 2.7 from [22]).
Used to show the potential is nonnegative, zero only at the anchor, and blows up at the boundary.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A matrix-free interior point continuous trajectory for linearly constrained convex programming." pith.science (2026). https://pith.science/paper/UOM7BQ4A

@misc{pith2026241220141,
  author       = {Pith},
  title        = {Pith review of: A matrix-free interior point continuous trajectory for linearly constrained convex programming},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UOM7BQ4A}},
  note         = {Machine review of arXiv:2412.20141}
}

read the original abstract

Interior point methods for solving linearly constrained convex programming involve a variable projection matrix at each iteration to deal with the linear constraints. This matrix often becomes ill-conditioned near the boundary of the feasible region that results in wrong search directions and extra computational cost. A matrix-free interior point augmented Lagrangian continuous trajectory is therefore proposed and studied for linearly constrained convex programming. A closely related ordinary differential equation (ODE) system is formulated. In this ODE system, the variable projection matrix is no longer needed. By only assuming the existence of an optimal solution, we show that, starting from any interior feasible point, (i) the interior point augmented Lagrangian continuous trajectory is convergent; and (ii) the limit point is indeed an optimal solution of the original optimization problem. Moreover, with the addition of the strictly complementarity condition, we show that the associated Lagrange multiplier converges to an optimal solution of the Lagrangian dual problem. Based on the studied ODE system, several possible search directions for discrete algorithms are proposed and discussed.

Figures

Figures reproduced from arXiv: 2412.20141 by the authors.

**Figure 2.** Trajectories of the solution paths with different in [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗

**Figure 3.** Trajectories of the primal affine scaling path ( [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 34 canonical work pages

[22]

The converg ent generalized central paths for linearly constrained convex programming

Xun Qian, Li-Zhi Liao, Jie Sun, and Hong Zhu. The converg ent generalized central paths for linearly constrained convex programming. SIAM Journal on Optimization , 28(2):1183– 1204, 2018

work page 2018
[1]

Limiting behavior of t he aﬃne scaling continuous trajectories for linear programming problems

Ilan Adler and Renato DC Monteiro. Limiting behavior of t he aﬃne scaling continuous trajectories for linear programming problems. Mathematical Programming, 50:29–51, 1991

work page 1991
[2]

On augmented lagrangian methods with general lower-level c onstraints

Roberto Andreani, Ernesto G Birgin, Jos´ e Mario Mart ´ ın ez, and Mar ´ ıa Laura Schuverdt. On augmented lagrangian methods with general lower-level c onstraints. SIAM Journal on Optimization, 18(4):1286–1309, 2008

work page 2008
[3]

Dynamical systems I: ordinary diﬀerential equations and smooth dynamical syste ms

Dmitrij V Anosov, Vladimir Igorevich Arnold, and DV Anos ov. Dynamical systems I: ordinary diﬀerential equations and smooth dynamical syste ms. Springer, 1988

work page 1988
[4]

Regularity versus degeneracy in dynam ics, games, and optimization: a uniﬁed approach to diﬀerent aspects

Immanuel M Bomze. Regularity versus degeneracy in dynam ics, games, and optimization: a uniﬁed approach to diﬀerent aspects. SIAM review, 44(3):394–414, 2002

work page 2002
[5]

Functions of a real variable

Nicolas Bourbaki. Functions of a real variable . Springer Berlin Heidelber, 2004

work page 2004
[6]

Convex optimization

Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004

2004
[7]

Examples of ill-behaved central paths in convex optimization

J Charles Gilbert, Clovis C Gonzaga, and Elizabeth Karas . Examples of ill-behaved central paths in convex optimization. Mathematical programming, 103(1):63–94, 2005

work page 2005

Show all 37 references

[8]

Iterative solution of problems of linear and qu adratic programming

II Dikin. Iterative solution of problems of linear and qu adratic programming. In Doklady Akademii Nauk , volume 174, pages 747–748. Russian Academy of Sciences, 19 67
[9]

On the convergence of an iterative process

II Dikin. On the convergence of an iterative process. In Upravlyaemye Sistemy , volume 12, pages 54–60, 1974

1974
[10]

A primal aﬃne scaling algorithm for linearly con- strained convex programs

Clovis C Gonzaga and Luiz A Carlos. A primal aﬃne scaling algorithm for linearly con- strained convex programs. preprint, 1990

1990
[11]

On full j acobian decomposition of the augmented lagrangian method for separable convex progr amming

Bingsheng He, Liusheng Hou, and Xiaoming Yuan. On full j acobian decomposition of the augmented lagrangian method for separable convex progr amming. SIAM Journal on Optimization, 25(4):2274–2312, 2015

2015
[12]

Multiplier and gradient methods

Magnus R Hestenes. Multiplier and gradient methods. Journal of optimization theory and applications, 4(5):303–320, 1969

1969
[13]

Augmented lagrangian methods and prox imal point methods for convex optimization

Alfredo N Iusem. Augmented lagrangian methods and prox imal point methods for convex optimization. Investigaci´ on Operativa, 8(11-49):7, 1999

1999
[14]

A study of the dual aﬃne scaling continuous trajectories for linear programming

Li-Zhi Liao. A study of the dual aﬃne scaling continuous trajectories for linear programming. Journal of Optimization Theory and Applications , 163:548–568, 2014

2014
[15]

Dynamics of games and gene s: Discrete versus continuous time

Viktor Losert and Ethen Akin. Dynamics of games and gene s: Discrete versus continuous time. Journal of Mathematical Biology , 17:241–251, 1983

1983
[16]

A simple characterization of solut ion sets of convex programs

Olvi L Mangasarian. A simple characterization of solut ion sets of convex programs. Oper- ations Research Letters , 7(1):21–26, 1988

1988
[17]

Boundary behavior of interior poi nt algorithms for linear pro- gramming,

N Megiddo and M Shub. “Boundary behavior of interior poi nt algorithms for linear pro- gramming,” ibm research report rj5319, 1986

1986
[18]

Convergence and boundary behavior of the projective scaling trajec- tories for linear programming

Renato DC Monteiro. Convergence and boundary behavior of the projective scaling trajec- tories for linear programming. Mathematics of Operations Research , 16(4):842–858, 1991. 18

1991
[19]

Global conver gence of the aﬃne scaling al- gorithm for convex quadratic programming

Renato DC Monteiro and Takashi Tsuchiya. Global conver gence of the aﬃne scaling al- gorithm for convex quadratic programming. SIAM Journal on Optimization , 8(1):26–58, 1998

1998
[20]

A method for nonlinear constraints i n minimization problems

Michael JD Powell. A method for nonlinear constraints i n minimization problems. Opti- mization, pages 283–298, 1969

1969
[21]

A strategy of global c onvergence for the aﬃne scaling algorithm for convex semideﬁnite programming

Xun Qian, Li-Zhi Liao, and Jie Sun. A strategy of global c onvergence for the aﬃne scaling algorithm for convex semideﬁnite programming. Mathematical Programming, 179(1):1–19, 2020

2020
[23]

Augmented lagrangians and appl ications of the proximal point algo- rithm in convex programming

R Tyrrell Rockafellar. Augmented lagrangians and appl ications of the proximal point algo- rithm in convex programming. Mathematics of operations research , 1(2):97–116, 1976

1976
[24]

Monotone operators and the prox imal point algorithm

R Tyrrell Rockafellar. Monotone operators and the prox imal point algorithm. SIAM journal on control and optimization , 14(5):877–898, 1976

1976
[25]

A simple proof of a primal aﬃne scaling me thod

Romesh Saigal. A simple proof of a primal aﬃne scaling me thod. Annals of Operations Research, 62(1):303–324, 1996

1996
[26]

Applied nonlinear control

Jean-Jacques E Slotine. Applied nonlinear control. PRENTICE-HALL google schola , 2:1123–1131, 1991

1991
[27]

The rate of converg ence of the augmented lagrangian method for nonlinear semideﬁnite programming

Defeng Sun, Jie Sun, and Liwei Zhang. The rate of converg ence of the augmented lagrangian method for nonlinear semideﬁnite programming. Mathematical Programming, 114(2):349– 391, 2008

2008
[28]

A convergen t proximal alternating direction method of multipliers for conic programming with 4-block co nstraints

Defeng Sun, Kim-Chuan Toh, and Liuqin Yang. A convergen t proximal alternating direction method of multipliers for conic programming with 4-block co nstraints. arXiv preprint arXiv:1404.5378, 2014

2014 arXiv
[29]

A convergence proof for an aﬃne-scaling algori thm for convex quadratic program- ming without nondegeneracy assumptions

Jie Sun. A convergence proof for an aﬃne-scaling algori thm for convex quadratic program- ming without nondegeneracy assumptions. Mathematical Programming, 60(1):69–79, 1993

1993
[30]

A convergence analysis for a convex version of d ikin’s algorithm

Jie Sun. A convergence analysis for a convex version of d ikin’s algorithm. Annals of Operations Research, 62:357–374, 1996

1996
[31]

On the convergence o f the exponential multiplier method for convex programming

Paul Tseng and Dimitri P Bertsekas. On the convergence o f the exponential multiplier method for convex programming. Mathematical programming, 60(1):1–19, 1993

1993
[32]

A ﬁrst-order interior-point method for linearly constrained smooth optimization

Paul Tseng, Immanuel M Bomze, and Werner Schachinger. A ﬁrst-order interior-point method for linearly constrained smooth optimization. Mathematical Programming, 127:399– 424, 2011

2011
[33]

On the convergence of the aﬃ ne-scaling algorithm

Paul Tseng and Zhi-Quan Luo. On the convergence of the aﬃ ne-scaling algorithm. Mathe- matical Programming, 56(1):301–319, 1992

1992
[34]

Global convergence of the aﬃne scali ng methods for degenerate linear programming problems

Takashi Tsuchiya. Global convergence of the aﬃne scali ng methods for degenerate linear programming problems. Mathematical Programming, 52:377–404, 1991

1991
[35]

Sdpnal+: a m ajorized semismooth newton-cg augmented lagrangian method for semideﬁnite pro gramming with nonnegative constraints

Liuqin Yang, Defeng Sun, and Kim-Chuan Toh. Sdpnal+: a m ajorized semismooth newton-cg augmented lagrangian method for semideﬁnite pro gramming with nonnegative constraints. Mathematical Programming Computation, 7(3):331–366, 2015. 19

2015
[36]

An extension of karmarkar’s pro jective algorithm for convex quadratic programming

Yinyu Ye and Edison Tse. An extension of karmarkar’s pro jective algorithm for convex quadratic programming. Mathematical programming, 44:157–179, 1989

1989
[37]

A newton- cg augmented lagrangian method for semideﬁnite programming

Xin-Yuan Zhao, Defeng Sun, and Kim-Chuan Toh. A newton- cg augmented lagrangian method for semideﬁnite programming. SIAM Journal on Optimization , 20(4):1737–1765, 2010. 20

2010

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [22]

The converg ent generalized central paths for linearly constrained convex programming

Xun Qian, Li-Zhi Liao, Jie Sun, and Hong Zhu. The converg ent generalized central paths for linearly constrained convex programming. SIAM Journal on Optimization , 28(2):1183– 1204, 2018

work page 2018

[2] [1]

Limiting behavior of t he aﬃne scaling continuous trajectories for linear programming problems

Ilan Adler and Renato DC Monteiro. Limiting behavior of t he aﬃne scaling continuous trajectories for linear programming problems. Mathematical Programming, 50:29–51, 1991

work page 1991

[3] [2]

On augmented lagrangian methods with general lower-level c onstraints

Roberto Andreani, Ernesto G Birgin, Jos´ e Mario Mart ´ ın ez, and Mar ´ ıa Laura Schuverdt. On augmented lagrangian methods with general lower-level c onstraints. SIAM Journal on Optimization, 18(4):1286–1309, 2008

work page 2008

[4] [3]

Dynamical systems I: ordinary diﬀerential equations and smooth dynamical syste ms

Dmitrij V Anosov, Vladimir Igorevich Arnold, and DV Anos ov. Dynamical systems I: ordinary diﬀerential equations and smooth dynamical syste ms. Springer, 1988

work page 1988

[5] [4]

Regularity versus degeneracy in dynam ics, games, and optimization: a uniﬁed approach to diﬀerent aspects

Immanuel M Bomze. Regularity versus degeneracy in dynam ics, games, and optimization: a uniﬁed approach to diﬀerent aspects. SIAM review, 44(3):394–414, 2002

work page 2002

[6] [5]

Functions of a real variable

Nicolas Bourbaki. Functions of a real variable . Springer Berlin Heidelber, 2004

work page 2004

[7] [6]

Convex optimization

Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004

2004

[8] [7]

Examples of ill-behaved central paths in convex optimization

J Charles Gilbert, Clovis C Gonzaga, and Elizabeth Karas . Examples of ill-behaved central paths in convex optimization. Mathematical programming, 103(1):63–94, 2005

work page 2005