Umeå University's logo

umu.sePublikasjoner
Endre søk
RefereraExporteraLink to record
Permanent link

Direct link
Referera
Referensformat
  • apa
  • ieee
  • vancouver
  • Annet format
Fler format
Språk
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Annet språk
Fler språk
Utmatningsformat
  • html
  • text
  • asciidoc
  • rtf
Revisiting Frank-Wolfe for structured nonconvex optimization
Umeå universitet, Teknisk-naturvetenskapliga fakulteten, Institutionen för matematik och matematisk statistik.ORCID-id: 0000-0001-8251-2605
Umeå universitet, Teknisk-naturvetenskapliga fakulteten, Institutionen för matematik och matematisk statistik.
Technical University of Munich, Germany.
Umeå universitet, Teknisk-naturvetenskapliga fakulteten, Institutionen för matematik och matematisk statistik.ORCID-id: 0000-0001-7320-1506
2025 (engelsk)Inngår i: NeuriPS 2025: Downloads, 2025Konferansepaper, Poster (with or without abstract) (Fagfellevurdert)
Abstract [en]

We introduce a new projection-free (Frank-Wolfe) method for optimizing structured nonconvex functions that are expressed as a difference of two convex functions. This problem class subsumes smooth nonconvex minimization, positioning our method as a promising alternative to the classical Frank-Wolfe algorithm. DC decompositions are not unique; by carefully selecting a decomposition, we can better exploit the problem structure, improve computational efficiency, and adapt to the underlying problem geometry to find better local solutions. We prove that the proposed method achieves a first-order stationary point in  iterations, matching the complexity of the standard Frank-Wolfe algorithm for smooth non-convex minimization in general. Specific decompositions can, for instance, yield a gradient-efficient variant that requires only  calls to the gradient oracle by reusing computed gradients over multiple iterations. Finally, we present numerical experiments demonstrating the effectiveness of the proposed method compared to other projection-free algorithms.

sted, utgiver, år, opplag, sider
2025.
HSV kategori
Forskningsprogram
matematik
Identifikatorer
URN: urn:nbn:se:umu:diva-248742OAI: oai:DiVA.org:umu-248742DiVA, id: diva2:2030515
Konferanse
NeuriPS 2025, The Thirty-Ninth Annual Conference on Neural Information Processing Systems, San Diego, USA, December 1, 2025
Tilgjengelig fra: 2026-01-20 Laget: 2026-01-20 Sist oppdatert: 2026-01-22bibliografisk kontrollert
Inngår i avhandling
1. From accelerated first-order methods to structured nonconvex optimization: analysis and perspectives
Åpne denne publikasjonen i ny fane eller vindu >>From accelerated first-order methods to structured nonconvex optimization: analysis and perspectives
2026 (engelsk)Doktoravhandling, med artikler (Annet vitenskapelig)
Alternativ tittel[sv]
Från accelererade första ordningens metoder till strukturerad icke-konvex optimering : analys och perspektiv
Abstract [en]

Modern technologies like machine learning and data-driven decision systems, depend on solving large and often highly complex optimization problems. These problems rarely come with simple shapes or smooth surfaces; instead, they can twist and bend in ways that make finding good solutions surprisingly difficult. This thesis explores two ideas that help us navigate such complexity more effectively. 

The first idea focuses on acceleration and investigates how optimization algorithms can reach good solutions faster while using only basic information such as function values and gradients. By studying these algorithms through a continuous-time analysis, we show that many of the fastest methods behave like carefully designed dynamical systems. This perspective not only clarifies why acceleration happens, but also allows us to design fast algorithms and understand how they behave in the presence of noisy information. 

The second idea focuses on a mathematical structural assumption called difference-of- convex (DC) programming, which captures a remarkably wide range of nonconvex problems. By leveraging this structure, we develop practical algorithms that avoid costly operations like projections or high dimensional gradient evaluations, making them efficient for large-scale applications. Through the connection between DC programming and the classical Expectation–Maximization (EM) algorithm, we develop more scalable EM variants and provide the first general performance guarantees for them. 

sted, utgiver, år, opplag, sider
Umeå: Umeå University, 2026. s. 43
Serie
Research report in mathematical statistics, ISSN 1653-0829 ; 80/26
Emneord
accelerated methods, convex optimization, nonconvex optimization, DC pro- gramming, EM algorithm, randomized DC algorithm
HSV kategori
Identifikatorer
urn:nbn:se:umu:diva-248866 (URN)978-91-8070-910-1 (ISBN)978-91-8070-911-8 (ISBN)
Disputas
2026-02-24, BIO.E.203 - Aula Biologica + Zoom, 08:30 (engelsk)
Opponent
Veileder
Forskningsfinansiär
Wallenberg AI, Autonomous Systems and Software Program (WASP)
Merknad

Link to participate via Zoom: https://umu.zoom.us/j/63834231447

Tilgjengelig fra: 2026-02-03 Laget: 2026-01-22 Sist oppdatert: 2026-01-22bibliografisk kontrollert

Open Access i DiVA

fulltext(2623 kB)38 nedlastinger
Filinformasjon
Fil FULLTEXT01.pdfFilstørrelse 2623 kBChecksum SHA-512
b23448ce7431dd6d8e2c7327d3ad09ab61c303953411a941086c4f542cc7c5250def14ac6b15e1e385b65eace51a9b300dd7c86bdec72273ae78c7ef7d81225c
Type fulltextMimetype application/pdf

Andre lenker

Presentation

Person

Maskan, HoomaanHou, YikunYurtsever, Alp

Søk i DiVA

Av forfatter/redaktør
Maskan, HoomaanHou, YikunYurtsever, Alp
Av organisasjonen

Søk utenfor DiVA

GoogleGoogle Scholar
Antall nedlastinger er summen av alle nedlastinger av alle fulltekster. Det kan for eksempel være tidligere versjoner som er ikke lenger tilgjengelige

urn-nbn

Altmetric

urn-nbn
Totalt: 2045 treff
RefereraExporteraLink to record
Permanent link

Direct link
Referera
Referensformat
  • apa
  • ieee
  • vancouver
  • Annet format
Fler format
Språk
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Annet språk
Fler språk
Utmatningsformat
  • html
  • text
  • asciidoc
  • rtf