Pith. sign in

REVIEW 2 major objections 5 minor 45 references

Efficient Sparse Flow Decomposition Methods for RNA Multi-Assembly

T0 review · 2 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper proposes that sparse flow decomposition is best solved as a convex data-fitting problem on the flow polytope with Frank-Wolfe methods, avoiding explicit path-count minimization.

desk verdict Solid empirical contribution to sparse flow decomposition, but the linear-convergence proof for the least-squares loss is invalid as stated. read the letter →

arxiv 2501.14662 v5 pith:HCOWBKYV submitted 2025-01-24 math.OC cs.DM

Mathieu Besançon This is my paper · ORCID
classification math.OCcs.DM MSC 90C2590C3505C85
keywords sparseflowdecompositionFrank-WolfealgorithmRNAtranscriptassemblypolytopeconditionalgradientPoissonregressionintegerprogrammingconvergenceanalysis
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

This paper claims that the sparse flow decomposition problem—decomposing a flow on a DAG into few weighted s-t paths, as in RNA transcript multi-assembly—can be solved as a convex regression problem over the flow polytope, solved by Frank-Wolfe methods, without explicitly minimizing the number of paths. The paper argues this formulation subsumes earlier exact and inexact flow formulations and can fit different error models through least-squares or Poisson log-likelihood losses. If correct, it means the NP-hard cardinality-minimizing integer formulations are not required for practical multi-assembly, and a projection-free first-order method can produce sparse, high-quality transcript reconstructions in far less runtime. The paper supports this with convergence analysis for both losses and with experiments on tens of thousands of splice graphs.

What carries the argument

The central object is the flow polytope $\mathrm{conv}(\mathcal{X})$, the convex hull of s-t path incidence vectors in a DAG, together with the linear minimization oracle that computes a shortest path. BPCG maintains a sparse active set of vertices and uses away-steps to stay in the optimal face; the paper leverages active-set identification bounds to argue the iterates are sparse. The least-squares analysis rides on a quadratic growth property with constant $\frac{1}{2}$, and the Poisson objective is shown to be the composition of a logarithmically homogeneous self-concordant barrier with an affine map, which activates known linear-convergence results for Frank-Wolfe variants.

What would settle it

Compute, on a two-path DAG with a reference flow $r$ not proportional to a path, the least-squares objective gap $f(x)-f^*$ and the squared distance to the constrained minimizer set $\mathrm{conv}(\mathcal{X})$; if the ratio drops below $1/2$ at any feasible $x$, the quadratic growth constant claimed in Proposition 2 is false, and the stated linear-rate proof collapses. Alternatively, run BPCG on such an instance and check whether the primal gap decays linearly with a rate consistent with the theorem.

Watch

Extended reading notes

Core claim

The central discovery is that minimizing a convex loss over the convex hull of s-t paths, with sparsity arising from the algorithm rather than the formulation, solves sparse flow decomposition effectively. For the least-squares loss, the problem is $\min_{x\in \mathrm{conv}(\mathcal{X})} \frac{1}{2}\lVert x - \frac{\langle x,r\rangle}{\lVert r\rVert^2} r \rVert^2$, whose optimal step can be computed in closed form; for the Poisson model, the log-likelihood is written as a logarithmically homogeneous self-concordant barrier composed with an affine map. Blended pairwise conditional gradient (BPCG) is applied, with each iteration requiring one shortest-path computation on the DAG; the paper claims linear convergence for both losses despite lack of strong convexity, based on a quadratic growth property for least squares and on self-concordant barrier theory for Poisson. Experiments on 81,439 gene graphs show the FW methods match or improve reconstruction quality of recent integer optimization approaches while being orders of magnitude faster, and sparser solutions are produced without minimizing cardinality.

Load-bearing premise

The linear-convergence claim for the least-squares objective depends on the assumption that the objective grows at least quadratically with distance to the minimizer set inside the flow polytope, but the proof only checks distance to the unconstrained minimizer ray, and a two-path DAG shows that inequality can fail.

Editorial extensions

If this is right

  • On 81,439 gene graphs from four species, least-squares Frank-Wolfe achieves sub-millisecond mean runtimes, several orders of magnitude faster than the integer and robust-integer baselines.
  • The least-squares and Poisson formulations cover exact flows, inexact flows, and interval uncertainty from previous work, so the same solver can be reused across modeling choices.
  • The BPCG active set produces sparse path decompositions, and active-set identification bounds the number of vertices by the dimension of the optimal face plus one after finitely many iterations.
  • Linear convergence is claimed for both losses despite the absence of strong convexity, so the method remains scalable to large splice graphs.
  • On noisy data, the least-squares model has the lowest flow error under both Poisson and binomial perturbations, while the robust integer model often wins on path-error counts.

Reading between the lines

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

  • If these results generalize, minimum-cardinality flow decomposition may be viewed as a diagnostic ideal rather than the operational objective for transcript assembly, with statistical fit plus algorithmic sparsity as the working target.
  • The same convex-over-polytope template suggests testing other error distributions, such as negative binomial or zero-inflated Poisson, within the Frank-Wolfe machinery.
  • A natural stress test is to run the method on very large splice graphs and check whether the active set size stays within the predicted face-dimension bound or grows with graph size.
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.

Referee Report

2 major / 5 minor

Summary. The paper proposes a convex optimization approach to sparse flow decomposition (SFD) on DAGs, modeling the problem as a least-squares or Poisson log-likelihood fit over the flow polytope, solved by the Blended Pairwise Conditional Gradient (BPCG) method. Sparsity is achieved through the active-set structure of Frank-Wolfe rather than explicit cardinality constraints. The authors claim linear convergence for both losses, report an early-termination heuristic, and provide extensive experiments on RNA multi-assembly datasets showing runtime and reconstruction quality competitive with mixed-integer formulations.

Significance. If the convergence guarantees held, the paper would make a strong case that NP-hard cardinality-constrained formulations can be replaced by a convex projection plus a projection-free algorithm, with empirical sparsity arising from active-set identification. The experimental section is a clear strength: it uses a large public dataset (81,439 DAGs), compares against recent integer methods, reports source code and an archived version, and includes shifted geometric means and instance-level counts. The derived step-size formula in Proposition 1 is correct and useful. However, the theoretical support for the main algorithmic claim is currently not established, which tempers the significance until the proof is repaired.

major comments (2)
  1. [Section 3, Proposition 2] Proposition 2 claims that the least-squares objective satisfies quadratic growth with µ = 1/2, but the proof computes the distance to the unconstrained minimizer ray {αr : α ≥ 0}, whereas Definition 1 and the convergence theorem [41, Thm 3.6] require the distance to the set of minimizers over the constraint set conv(X). These sets generally differ. For a two-path DAG with vertices v1 (k-edge path) and v2 (single-edge path), r = K v1 + v2, the constrained minimizer is x* = (K v1 + v2)/(K+1) ∈ conv{v1,v2}, so f(x*) = 0. At x = v1, f(v1) - f* = k/[2(K^2 k + 1)] but (1/2) dist²(v1, X*_con) = (k+1)/[2(K+1)²]. The ratio tends to 1/[2(k+1)] as K → ∞, so the claimed inequality with µ = 1/2 fails and the quadratic-growth constant can be arbitrarily small. Consequently, the linear convergence guarantee for BPCG stated immediately after Proposition 2 is unsupported. The authors should either prove the growth condition with respect to the constrained minimizer set (possibly with an instance-dependent constant) or revise the theoretical claim.
  2. [Section 2.3 and Algorithm 1] The Poisson objective is a log-barrier: it is finite only when (Ax)_u > 0 for every node u with positive observed inflow c_u = Σ_{e∈δ^in_u} r_e, and it tends to +∞ on the boundary of this domain. However, Algorithm 1 initializes x0 as a single path pG(1). In a DAG with branching, such a path has zero inflow at nodes not on that path, so f(x0) is infinite and the BPCG algorithm cannot start from this point. The paper does not describe an alternative initialization or a domain-regularization for the Poisson variant, and the convergence framework of [44,20] cited in Section 3 relies on the barrier property that the objective is finite on the interior only. The authors need to specify a feasible starting point (e.g., a strictly positive convex combination covering all nodes with c_u > 0) or justify why the implemented algorithm avoids this issue.
minor comments (5)
  1. [Section 2.3, Proposition 3 proof] The logarithmic homogeneity coefficient θ is written as (Σ_{e∈δ^in_u} r_e), but the sum over u is missing; it should be θ = Σ_{u∈V} Σ_{e∈δ^in_u} r_e.
  2. [Section 2.1, formulation] The notation in the integer formulation mixes w_s ∈ Z_+ and the support constraint ∥w∥_0 ≤ k with the index s ∈ [[k]]; the intended range of s should be clarified (it seems to be the number of paths in the decomposition, not the bound k).
  3. [Section 4.1, paragraph after Table 2] The sentence 'The robust integer model yields the best path error performance at the cost of an increased runtime' is consistent with Table 2, but the runtime comparison in Figure 2 would be easier to read if the y-axis were labeled with the exact time unit and if the note about shifted geometric means were repeated in the caption.
  4. [Section 4.1, Figure 3] The text attributes the stalling of the 64-bit Poisson trajectory to numerical errors, but it would be helpful to state explicitly that 'PF' uses standard double precision and 'PE' uses BigFloat, and to report the number of iterations at which the gap stalls.
  5. [References] Reference [18] is missing a publication venue (Citeseer), and reference [34] is an arXiv preprint with no version number; please update these entries.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the formulation, algorithm, and experiments are self-contained; the quadratic-growth proof gap is a correctness issue, not a circular dependency.

full rationale

I find no load-bearing circular step. The central formulation is a direct data-fitting problem: min_{x in conv(X)} 1/2 ||x - (⟨x,r⟩/||r||^2) r||^2 for least squares and the Poisson log-likelihood over scaled paths; these objectives are not defined in terms of the claimed output (sparse transcript sets), and no constant is fitted to the benchmark decompositions. The elimination of the scaling variable τ is a closed-form algebra step, not a smuggled conclusion. Sparsity is not inserted as a constraint; it is inherited from the BPCG active-set identification (Section 2.1), with citations to [7,42]. Although [42] and [20] are co-authored by the present author, they are stated theorems about Frank-Wolfe algorithms whose assumptions do not include this paper's claims, so they are independent support rather than a self-citation chain. The early termination criterion (Section 3.1) explicitly uses an integrality heuristic, is disclosed as such, and does not change the model. The only substantive concern is Proposition 2: its proof computes distance to the unconstrained minimizer ray α r, whereas Definition 1 and [41, Thm 3.6] require distance to the constrained minimizer set inside conv(X); the claimed μ=1/2 quadratic growth is therefore not established. That is a correctness gap, not a circularity: the result is not made true by definition of the objective, and the downstream convergence theorem is an external result.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard flow decomposition theory, external Frank-Wolfe convergence results, and the correctness of the quadratic growth proof. No data-fitted parameters are introduced. The main gap is the quadratic growth argument, which is used to invoke an external theorem but actually measures distance to the wrong set.

assumptions (4)
  • standard math Every flow on a DAG can be decomposed into at most |E|-|V|+2 s-t paths (flow decomposition theorem).
    Invoked in Section 2.1 to justify the upper bound on the number of paths in the original formulation.
  • standard math The active set identification and vertex-count bounds for BPCG from Bomze et al. (2020) and Wirth et al. (2025) apply to the flow polytope.
    Used in Section 2.1 to argue that iterates are sparse and eventually lie on the optimal face.
  • standard math The linear convergence theorems for BPCG on polytopes (Wirth et al. 2024, Hendrych et al. 2025, Zhao 2025) apply to the least-squares and Poisson objectives.
    Used in Section 3 to claim linear convergence rates, but the least-squares application relies on the flawed quadratic growth proof.
  • domain assumption The observed flow data are generated from ground-truth transcripts with independent error distributions.
    The benchmark uses splice graphs with known ground truth and simulates Poisson or binomial noise; the statistical models assume such error structures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Sparse Flow Decomposition Methods for RNA Multi-Assembly." pith.science (2026). https://pith.science/paper/HCOWBKYV

@misc{pith2026250114662,
  author       = {Pith},
  title        = {Pith review of: Efficient Sparse Flow Decomposition Methods for RNA Multi-Assembly},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HCOWBKYV}},
  note         = {Machine review of arXiv:2501.14662}
}
read the original abstract

Decomposing a flow on a Directed Acyclic Graph (DAG) into a weighted sum of a small number of paths is an essential task in operations research and bioinformatics. This problem, referred to as Sparse Flow Decomposition (SFD), has gained significant interest, in particular for its application in RNA transcript multi-assembly, the identification of the multiple transcripts corresponding to a given gene and their relative abundance. Several recent approaches cast SFD variants as integer optimization problems, motivated by the NP-hardness of the formulations they consider. We propose an alternative formulation of SFD as a data fitting problem on the conic hull of the flow polytope. By reformulating the problem on the flow polytope for compactness and solving it using specific variants of the Frank-Wolfe algorithm, we obtain a method converging rapidly to the minimizer of the chosen loss function while producing a parsimonious decomposition. Our approach subsumes previous formulations of SFD with exact and inexact flows and can model different priors on the error distributions. Computational experiments show that our method outperforms recent integer optimization approaches in runtime, but is also highly competitive in terms of reconstruction of the underlying transcripts, despite not explicitly minimizing the solution cardinality.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 39 canonical work pages

  1. [42]

    Fast convergence of Frank-Wolfe algorithms on polytopes

    E. Wirth, J. Pena, and S. Pokutta. Fast convergence of Frank-Wolfe algorithms on polytopes. arXiv preprint arXiv:2406.18789, 2024

    work page Pith review arXiv 2024

  2. [1]

    Bernard, L

    E. Bernard, L. Jacob, J. Mairal, and J.-P. Vert. Efficient RNA iso- form identification and quantification from RNA-Seq data with network flows. Bioinformatics, 30(17):2447–2455, 2014

    work page 2014

  3. [2]

    Besançon, A

    M. Besançon, A. Carderera, and S. Pokutta. FrankWolfe.jl: A high- performance and flexible toolbox for Frank-Wolfe algorithms and con- ditional gradients. INFORMS Journal on Computing, 34(5):2611–2620, 2022

    work page 2022

  4. [3]

    Besançon, S

    M. Besançon, S. Designolle, J. Halbey, D. Hendrych, D. Kuzinowicz, S. Pokutta, H. Troppens, D. V . Herrmannsdoerfer, and E. Wirth. Im- proved algorithms and novel applications of the FrankWolfe.jl library. arXiv preprint arXiv:2501.14613, 2025

    arXiv 2025

  5. [4]

    Besançon

    M. Besançon. Code for efficient sparse flow decomposition methods for RNA multi-assembly, July 2025. URL https://doi.org/10.5281/zenodo. 16033781

    doi:10.5281/zenodo 2025

  6. [5]

    Besançon

    M. Besançon. Repository for efficient sparse flow decomposition meth- ods for RNA multi-assembly, July 2025. https://github.com/inria-ghost/ sparse-flow-decomposition

    work page 2025

  7. [6]

    Bolusani, M

    S. Bolusani, M. Besançon, K. Bestuzheva, A. Chmiela, J. Dioní- sio, T. Donkiewicz, J. van Doornmalen, L. Eifler, M. Ghannam, A. Gleixner, et al. The SCIP optimization suite 9.0. arXiv preprint arXiv:2402.17702, 2024

    arXiv 2024

  8. [7]

    I. M. Bomze, F. Rinaldi, and D. Zeffiro. Active set complexity of the away-step Frank-Wolfe algorithm. SIAM Journal on Optimization , 30 (3):2470–2500, 2020

    work page 2020

Show all 45 references
  1. [8]

    I. M. Bomze, F. Rinaldi, and D. Zeffiro. Frank-Wolfe and friends: a journey into projection-free first-order optimization methods. 4OR, 19: 313–345, 2021

    2021

  2. [9]

    Braun, A

    G. Braun, A. Carderera, C. W. Combettes, H. Hassani, A. Karbasi, A. Mokhtari, and S. Pokutta. Conditional gradient methods. arXiv preprint arXiv:2211.14103, 2022

    2022 arXiv

  3. [10]

    Carderera, M

    A. Carderera, M. Besançon, and S. Pokutta. Simple steps are all you need: Frank-Wolfe and generalized self-concordant functions. Advances in Neural Information Processing Systems , 34:5390–5401, 2021

    2021

  4. [11]

    Carderera, M

    A. Carderera, M. Besançon, and S. Pokutta. Scalable Frank-Wolfe on generalized self-concordant functions via simple steps. SIAM Journal on Optimization, 34(3):2231–2258, 2024

    2024

  5. [12]

    B. V . Cherkassky, A. V . Goldberg, and T. Radzik. Shortest paths algo- rithms: Theory and experimental evaluation. Mathematical program- ming, 73(2):129–174, 1996

    1996

  6. [13]

    F. H. Dias and A. I. Tomescu. Accurate flow decomposition via ro- bust integer linear programming. IEEE/ACM Transactions on Compu- tational Biology and Bioinformatics, 2024

    2024

  7. [14]

    F. H. Dias and A. I. Tomescu. Dataset for: Accurate flow decom- position via robust integer linear programming, Mar. 2024. Zenodo, https://doi.org/10.5281/zenodo.10775004

    2024 doi

  8. [15]

    F. H. Dias, L. Williams, B. Mumey, and A. I. Tomescu. Efficient min- imum flow decomposition via integer linear programming. Journal of Computational Biology, 29(11):1252–1267, 2022

    2022

  9. [16]

    Dvurechensky, K

    P. Dvurechensky, K. Safin, S. Shtern, and M. Staudigl. Generalized self-concordant analysis of Frank-Wolfe algorithms.Mathematical Pro- gramming, 198(1):255–323, 2023

    2023

  10. [17]

    Frank, P

    M. Frank, P. Wolfe, et al. An algorithm for quadratic programming. Naval research logistics quarterly, 3(1-2):95–110, 1956

    1956

  11. [18]

    A. V . Goldberg and T. Radzik. A heuristic improvement of the Bellman- Ford algorithm. Citeseer, 1993

    1993

  12. [19]

    Grigorjew, F

    A. Grigorjew, F. H. C. Dias, A. Cracco, R. Rizzi, and A. I. Tomescu. Accelerating ILP solvers for minimum flow decompositions through search space and dimensionality reductions. In L. Liberti, edi- tor, 22nd International Symposium on Experimental Algorithms (SEA 2024), volume...

    2024

  13. [20]

    URL https://drops.dagstuhl.de/ entities/document/10.4230/LIPIcs.SEA.2024.14

    doi: 10.4230/LIPIcs.SEA.2024.14. URL https://drops.dagstuhl.de/ entities/document/10.4230/LIPIcs.SEA.2024.14

    2024 doi

  14. [21]

    Hendrych, M

    D. Hendrych, M. Besançon, and S. Pokutta. Solving the optimal ex- periment design problem with mixed-integer convex methods. arXiv preprint arXiv:2312.11200, 2025

    2025

  15. [22]

    Hendrych, S

    D. Hendrych, S. Pokutta, M. Besançon, and D. Martínez-Rubio. Secant line search for Frank-Wolfe algorithms. In Forty-second International Conference on Machine Learning, 2025

    2025

  16. [23]

    M. Jaggi. Revisiting Frank-Wolfe: Projection-free sparse convex opti- mization. In International conference on machine learning, pages 427–

  17. [24]

    Karimi, J

    H. Karimi, J. Nutini, and M. Schmidt. Linear convergence of gradient and proximal-gradient methods under the Polyak-łojasiewicz condition. In Machine Learning and Knowledge Discovery in Databases: Euro- pean Conference, ECML PKDD 2016, Riva del Garda, Italy, September 19-23, 2...

    2016

  18. [25]

    Kloster, P

    K. Kloster, P. Kuinke, M. P. O’Brien, F. Reidl, F. S. Villaamil, B. D. Sullivan, and A. van der Poel. A practical FPT algorithm for flow de- composition and transcript assembly. In 2018 Proceedings of the Twen- tieth Workshop on Algorithm Engineering and Experiments (ALENEX), ...

    2018

  19. [26]

    T. Koch, T. Berthold, J. Pedersen, and C. Vanaret. Progress in math- ematical programming solvers from 2001 to 2020. EURO Journal on Computational Optimization, 10:100031, 2022

    2001

  20. [27]

    E. S. Levitin and B. T. Polyak. Constrained minimization methods. USSR Computational mathematics and mathematical physics , 6(5):1– 50, 1966

    1966

  21. [28]

    J. J. Li, C.-R. Jiang, J. B. Brown, H. Huang, and P. J. Bickel. Sparse linear modeling of next-generation mRNA sequencing (RNA-Seq) data for isoform discovery and abundance estimation. Proceedings of the National Academy of Sciences, 108(50):19867–19872, 2011

    2011

  22. [29]

    W. Li, J. Feng, and T. Jiang. Isolasso: a LASSO regression approach to RNA-Seq based transcriptome assembly. In Research in Computa- tional Molecular Biology: 15th Annual International Conference, RE- COMB 2011, Vancouver, BC, Canada, March 28-31, 2011. Proceedings 15, pages 1...

    2011

  23. [30]

    Lubin, O

    M. Lubin, O. Dowson, J. D. Garcia, J. Huchette, B. Legat, and J. P. Vielma. JuMP 1.0: Recent improvements to a modeling language for mathematical optimization. Mathematical Programming Computation, 15(3):581–589, 2023

    2023

  24. [31]

    Pedregosa, G

    F. Pedregosa, G. Negiar, A. Askari, and M. Jaggi. Linearly convergent Frank-Wolfe with backtracking line-search. InInternational conference on artificial intelligence and statistics, pages 1–10. PMLR, 2020

    2020

  25. [32]

    S. Pokutta. The Frank-Wolfe algorithm: a short introduction. Jahres- bericht der Deutschen Mathematiker-Vereinigung, 126(1):3–35, 2024

    2024

  26. [33]

    Sena and A

    F. Sena and A. I. Tomescu. Safe paths and sequences for scalable ILPs in RNA transcript assembly problems. arXiv preprint arXiv:2411.03871, 2024

    2024 arXiv

  27. [34]

    F. Sena, E. Ingervo, S. Khan, A. Prjibelski, S. Schmidt, and A. Tomescu. Flowtigs: safety in flow decompositions for assembly graphs. iScience, 27(12), 2024

    2024

  28. [35]

    Z. Song, Z. Xu, Y . Yang, and L. Zhang. Accelerating Frank-Wolfe algo- rithm using low-dimensional and adaptive data structures, 2022. URL https://arxiv.org/abs/2207.09002

    2022 arXiv

  29. [36]

    Sun and Q

    T. Sun and Q. Tran-Dinh. Generalized self-concordant functions: a recipe for Newton-type methods. Mathematical Programming, 178(1): 145–213, 2019

    2019

  30. [37]

    A. I. Tomescu, A. Kuosmanen, R. Rizzi, and V . Mäkinen. A novel min- cost flow method for estimating transcript expression with RNA-Seq. In BMC bioinformatics, volume 14, pages 1–10. Springer, 2013

    2013

  31. [38]

    A. I. Tomescu, T. Gagie, A. Popa, R. Rizzi, A. Kuosmanen, and V . Mäkinen. Explaining a weighted DAG with few paths for solving genome-guided multi-assembly. IEEE/ACM transactions on computa- tional biology and bioinformatics, 12(6):1345–1354, 2015

    2015

  32. [39]

    K. K. Tsuji, K. Tanaka, and S. Pokutta. Pairwise conditional gradients without swap steps and sparser kernel herding. InInternational Confer- ence on Machine Learning, pages 21864–21883. PMLR, 2022

    2022

  33. [40]

    Vatinlen, F

    B. Vatinlen, F. Chauvet, P. Chrétienne, and P. Mahey. Simple bounds and greedy algorithms for decomposing a flow into a minimal set of paths. European Journal of Operational Research, 185(3):1390–1401, 2008

    2008

  34. [41]

    Williams, G

    L. Williams, G. Reynolds, and B. Mumey. RNA transcript assembly using inexact flows. In 2019 IEEE International Conference on Bioin- formatics and Biomedicine (BIBM), pages 1907–1914. IEEE, 2019

    2019

  35. [43]

    Wirth, M

    E. Wirth, M. Besançon, and S. Pokutta. The pivoting framework: Frank- Wolfe algorithms with active set size control. In The 28th International Conference on Artificial Intelligence and Statistics, 2025

    2025

  36. [44]

    Z. Xu, Z. Song, and A. Shrivastava. Breaking the linear iteration cost barrier for some well-known conditional gradient methods using MaxIP data-structures. Advances in Neural Information Processing Systems , 34:5576–5589, 2021

    2021

  37. [45]

    R. Zhao. New analysis of an away-step Frank-Wolfe method for min- imizing log-homogeneous barriers. Mathematics of Operations Re- search, 2025

    2025

Pith tools

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