Pith. sign in

REVIEW 3 major objections 4 minor 72 references

The paper establishes explicit problem-size-dependent lower bounds on the geometric constant γ(ℓ_sub) that governs finite-time exact consistency of gradient-based inverse optimization for integer linear programs, yielding iteration counts a

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 05:17 UTC pith:ASZZYCR5

load-bearing objection A genuinely new explicit lower bound on γ(ℓ_sub) for ILPs, with mostly solid proofs, but the iteration-count conclusion rests on an unproved companion theorem and a couple of technical gaps to patch. the 3 major comments →

arxiv 2607.22263 v1 pith:ASZZYCR5 submitted 2026-07-24 math.OC cs.AIcs.LGstat.ML

Explicit Iteration Complexity of Exact Data-Driven Inverse Optimization for Integer Linear Programs

Akira Kitaoka This is my paper
classification math.OC cs.AIcs.LGstat.ML MSC 90C9090C2590C5290C1190C0568Q25
keywords inverse optimizationinteger linear programmingsuboptimality lossprojected subgradient methodGraver basisM-convexitytotal unimodularityfinite-time convergence
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper seeks to turn a qualitative finite-termination guarantee into a quantitative one for data-driven inverse optimization of integer linear programs. Earlier work showed that projected subgradient descent on the suboptimality loss reaches exact consistency with observed data in finitely many iterations, with the count controlled by an abstract geometric constant γ(ℓ_sub). The paper proves explicit lower bounds on γ(ℓ_sub) in terms of sample size, feature dimension, feature ranges, and, for linear inequality constraints, the Graver-basis norm of the constraint matrix. Substituted into the known iteration bounds, these yield explicit finite iteration budgets that can be computed before running the algorithm. For structured feasible regions such as M-convex and M♮-convex sets, the bounds are polynomial in the dimension and independent of feature ranges; for general ILPs they are exponentially small in dimension, and the paper shows this exponential dependence is inherent.

Core claim

For integer linear programs, the geometric constant γ(ℓ_sub)—the largest margin separating the observed feature sums from all alternative feature tuples—can be lower-bounded explicitly from problem data. Under the probability-simplex weight set, γ ≥ 1/(N^d max(d−1,√2)‖m‖_2^{d−1}) for general ILPs; when the feasible region is M-convex or M♮-convex, γ ≥ 2/(N d(d−1)) (resp. 2/(N d(d+1))) independent of feature ranges; and for linear inequality constraints Ax ≤ b it is Ω(1/(N d^{(d+1)/2}(2C)^{d−1})), with C the ℓ∞ norm of the Graver basis of the slack-augmented matrix [A|I]. These bounds make the previously abstract iteration upper bound T = O(1/γ²) an explicit function of the problem size.

What carries the argument

The carrying object is the geometric constant γ(ℓ_sub), defined as the maximum over weights of the minimum normalized score advantage of the observed features over any other combination of vertices from the sample-wise feature vertex sets. The arguments reduce γ to a distance problem between the observed sum P and the convex hull of alternative sums Conv(W), then use an integral separating hyperplane: because all features are integers, the separating normal is an integer vector whose norm is bounded through Hadamard's inequality and the feature ranges. For structured regions, explicit test sets—single exchanges for M-/M♮-convex sets, Graver basis elements for linear inequality systems—give t

Load-bearing premise

The entire chain rests on every observed sample being generated exactly by a single true weight with a unique optimal feature vector, and on all features being integer-valued; if either fails, the lattice-based lower bounds that make the iteration count explicit do not follow.

What would settle it

Take an explicit ILP instance with N=1, d=2, feature ranges m=(1,1), and a feasible region whose vertex set is, say, {(0,0),(1,0),(0,1)}; compute the min-max on the right side of Equation (4.3) exactly and compare to the lower-bound value 1/(2·2·√2·√2)=1/4. Any instance yielding a value below 1/4 would refute Theorem 6.6 under its stated assumptions.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Practitioners can precompute a sufficient iteration budget before running the algorithm, turning the finite-time guarantee into an a priori stopping criterion.
  • The finite-step exact-consistency guarantee can now be compared fairly with asymptotic regret bounds, both expressed as functions of the problem size.
  • For M-convex and M♮-convex feasible regions, the iteration count is polynomial in dimension and independent of feature ranges, making exact inverse optimization computationally attractive for those structures.
  • For general ILPs, the exponential dependence of the iteration bound on dimension is unavoidable, consistent with the known hardness of inverse optimization under noisy data.
  • The same lower bounds yield explicit iteration counts for attaining zero prediction loss of features (PLF), not just zero suboptimality loss.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural testable extension is whether the same explicit bounds can be derived for mixed-integer programs by applying the lattice separating hyperplane to the integer part; the current result covers pure ILPs.
  • The striking gap between the strong M-convex bound and the weaker general linear-inequality bound for the same region suggests that the choice of structural representation itself changes the complexity estimate by an exponential factor, and algorithm designers should exploit combinatorial type rather than the inequality description.
  • Real-world data noise or tie-breaking suboptimality in observed solutions would violate the uniqueness assumption; measuring how γ degrades under small perturbations would indicate how quickly these explicit budgets erode.
  • The bounds are for exact consistency; for approximate consistency one would expect much smaller budgets, and the explicit form of γ could yield an ε-dependent complexity that interpolates smoothly between asymptotic and exact regimes.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper studies the data-driven inverse optimization problem (DDIOP) for integer linear programs (ILPs), where one estimates objective-function weights from observed optimal solutions. Its central contribution is a set of explicit lower bounds on the geometric constant γ(ℓ_sub) that controls the finite-time exact-convergence guarantees of projected subgradient methods: a general ILP bound under Θ=Δ^{d-1} (Theorem 6.6), polynomial-in-dimension bounds for M-convex and M♮-convex sets (Theorems 6.13/6.14 and 6.18/6.19), a Graver-basis bound for linear-inequality-constrained ILPs (Theorems 6.23/6.24), and a matching exponential-in-dimension tightness example (Proposition 6.34). These lower bounds are then substituted into iteration-complexity results from the author's companion paper (Kitaoka 2024, Corollaries 4.8/4.9) to obtain explicit iteration counts, summarized in Tables 1 and 4.

Significance. If the results are correct, the paper provides a genuinely new quantification of a previously abstract constant: practitioners would be able to precompute an iteration budget that guarantees exact consistency before running the algorithm. The test-set reduction (Lemma 6.9) and the lattice-separating-hyperplane arguments are elegant, and the structural separation between polynomial-in-dimension bounds (M-convex, M♮-convex, totally unimodular) and exponential-in-dimension worst cases for general ILPs is valuable. The tightness construction in Proposition 6.34 is a particular strength because it shows the exponential dependence is not an artifact of the proof. No fitted parameters appear; the lower bounds are derived from test sets, Graver bases, and lattice geometry. However, the final explicit iteration counts are conditional on the author's companion paper, and the scope is restricted by exact-data, uniqueness, and integrality assumptions that are not fully discussed.

major comments (3)
  1. [§4.1, Theorem 4.5 and Corollaries 4.8/4.9] The explicit iteration counts in Table 4 are obtained by substituting the new lower bounds into Q_{L(ℓsub),θ1}(γ(ℓsub)) from the companion paper, but Theorem 4.5 is only quoted, not proved, and its statement here omits the relative-interior-point condition that §2 identifies as essential. Assumptions 4.1/4.2 are conditions on an abstract update map, and the reader is not shown how the PSGD variants satisfy them in this setting. If the companion's γ has a different normalization, or if a hidden hypothesis (e.g., relative interior of argmin_Θ ℓsub) is needed, then every row of Table 4 is unsupported. Please restate the exact companion theorem and corollaries with all hypotheses, and either prove them or provide a self-contained verification for ℓsub under Assumptions 3.1/3.2.
  2. [§6.6, Propositions 6.30/6.33 and Theorems 6.4/6.6/6.23/6.24] The lattice-separation mechanism is the backbone of the general-ILP lower bounds: the integrality of the normal vector a and the inequality a·P−c≥1 require Assumption 6.1 (integer-valued features) and Assumption 3.2 (unique optimal feature). If features are real-valued or the data contain ties, the m_i-based and C-based bounds lose their lattice support and need not hold. The paper states these assumptions but does not discuss their necessity or provide examples showing they are tight. Additionally, Theorems 6.4 and 6.23 require full-dimensionality conditions (dim Conv(W)=d and dim Conv(S+)=d) that are not derived from the ILP data; Remark 6.5 gives only a sufficient condition. Please state explicitly which claims fail without these assumptions and add verification criteria or counterexamples.
  3. [§6.2–6.4, Lemma 6.9 and M-convex/M♮-convex bounds] Lemma 6.9 assumes Y(n)⊆X(s(n)) and identity features, but the inclusion is not justified in the text. For a bounded integer set X, the vertices of Conv(X) are indeed elements of X, so the assertion is true; nonetheless it should be stated and proved explicitly because it is used in all M-convex and M♮-convex theorems. More importantly, the lower bounds in Theorems 6.13/6.14 and 6.18/6.19 depend on choosing a θ† that is ordered relative to the unknown true θ*. The resulting bounds are independent of θ*, so this is not a circularity, but the presentation should clarify that the bound is existential rather than constructed from data.
minor comments (4)
  1. [§6.1, Theorems 6.4/6.13/6.18] The unit ball is consistently written as 'Bd' rather than 'B^d', which is easy to misread as 'B_d'. Please use a consistent notation.
  2. [§5, Tables 2 and 3] The acronyms UPA and RPA are used without definition. Please define these methods or cite the exact references from which the rates are taken.
  3. [§6.4.1, Theorem 6.18] The handling of the degenerate case θ*=0 is embedded in the proof. It would be clearer to state at the start that if θ*=0 then Assumption 3.2 forces every X(s(n)) to be a singleton, in which case γ(ℓsub)=+∞ and the bound is trivial, and otherwise the proof proceeds.
  4. [§6.5, Proposition 6.22] The constant C=g∞(eA) is described as 'determined by A alone', but computing a Graver basis can be expensive. The paper should state explicitly whether C is assumed to be known to the practitioner or whether only its bounds (e.g., Proposition 6.27) are intended to be used in the stopping criterion.

Circularity Check

0 steps flagged

No circular reduction: the gamma lower bounds are derived from test sets and lattice separation, and the explicit iteration counts are a modular composition with a self-cited companion theorem.

full rationale

The paper's substantive new content is the lower bounds on gamma(ell_sub) in Section 6. These are obtained by: (i) a minimax reformulation of gamma(ell_sub) (Propositions 6.2 and 6.3); (ii) test-set decompositions from external results (Murota's M-convex/M♮-convex exchange test sets; Sturmfels/Onn Graver-basis test sets); and (iii) integral separating-hyperplane arguments using Assumption 6.1. None of these steps uses the target iteration count T or the companion's Q_{L,theta_1} as an input; the inequalities are in terms of N, d, m, C, and lattice geometry. The final Table 4 bounds are obtained by substituting these independently proved lower bounds into the finite-termination/iteration bounds of Kitaoka (2024, Corollaries 4.8/4.9). That is a composition with a same-author companion, so the explicit iteration counts are not fully self-contained and inherit any unverified hypothesis of the companion; this is an external-dependency caveat rather than a circular reduction. The companion theorem does not assume the new lower bounds, and the lower bounds do not assume the companion iteration bound. The known weakest assumptions (Assumption 3.2 unique optimal feature; Assumption 6.1 integer-valued features) are explicitly stated as conditions, not hidden restatements of the conclusion. I therefore find no circular step; the score of 2 only flags the load-bearing same-author citation that makes the headline explicit-complexity claim depend on a companion preprint.

Axiom & Free-Parameter Ledger

0 free parameters · 9 axioms · 0 invented entities

The paper rests on standard optimization/integer-programming results plus a set of domain assumptions about clean, uniquely-optimal integer data. The main nonlocal input is the same author's companion theorem that finite termination holds with rate function Q(γ); that is imported rather than proved here. The full-dimensionality/nondegeneracy conditions are explicit but may be easy to overlook in the summary tables.

axioms (9)
  • domain assumption Kitaoka (2024) Theorem 4.5 and Corollaries 4.8/4.9: under Assumptions 3.1, 3.2, 4.1, 4.2, T ≥ Q_{L,θ1}(γ(ℓsub)) implies min_{t≤T} ℓsub(θt) = 0, with explicit Q for PSGD (SRSS/SRSL).
    Imported unchanged from the same author's companion paper; the present paper does not reprove these theorems, and all final iteration bounds are obtained by substitution into them.
  • domain assumption Assumption 3.2: for each n, a*(θ*, s(n)) is uniquely determined.
    Needed for Proposition 4.4 (γ > 0), the strict inequalities in Theorems 6.4/6.6, and Lemma 6.9; guaranteed only almost everywhere by Lemma 3.3.
  • domain assumption Assumption 6.1: X ⊂ Z^{dX} and f(X × S) ⊂ Z^d (integer-valued features).
    The lattice separation argument (Prop 6.30) and all m_i-based bounds rely on integrality of feature values.
  • standard math Minimax theorem (Prop 6.2).
    Used to convert max-min over finite sets into min over convex hulls; standard compact convex/concave minimax.
  • standard math Hadamard's inequality and determinant cofactor bounds (Props 6.31–6.33).
    Bounds the integer separating normal in Theorems 6.4, 6.6, 6.23, and 6.24.
  • standard math Conformal decomposition property of Graver bases and the Graver-basis test set (Prop 6.22; Onn 2010; Sturmfels 1996).
    Provides the test set Tx with ‖g‖∞ ≤ C used in Theorems 6.23 and 6.24.
  • standard math Test-set results for M-convex and M♮-convex sets (Props 6.12 and 6.17; Murota; Murota–Shioura).
    Supplies T = {e_i − e_j} and T = {e_i − e_j, ±e_i} used in the polynomial lower bounds.
  • ad hoc to paper Full-dimensionality / nondegeneracy conditions in Theorems 6.4, 6.6, 6.23, and 6.24 (e.g., dim Conv(W) = d or Y(n0) non-singleton).
    These conditions are stated in the theorem statements but are not always carried into the summary tables; if they fail, the stated lower bounds may not apply.
  • standard math Total unimodularity implies C = 1 (Remark 6.25, citing Onn 2010).
    Used to specialize the linear-inequality bound to totally unimodular matrices.

pith-pipeline@v1.3.0-alltime-deepseek · 31620 in / 20516 out tokens · 162485 ms · 2026-08-01T05:17:52.982380+00:00 · methodology

0 comments
read the original abstract

A data-driven inverse optimization problem (DDIOP) is the problem of estimating the objective-function parameters (weights) that explain observed optimal-solution data, and it arises in many applications, including integer linear programming (ILP). It is known that, by applying gradient-based optimization methods to the suboptimality loss, the inverse optimization of ILPs can be solved exactly within finitely many oracle iterations, and that the required number of iterations is bounded as $T=O(1/\gamma(\ell_{\mathrm{sub}})^2)$ in terms of a problem-dependent geometric constant $\gamma(\ell_{\mathrm{sub}})$. However, no means of bounding $\gamma(\ell_{\mathrm{sub}})$ from below as a function of the problem size has been available, and hence the number of iterations could not be given as an explicit function of the problem size. We therefore give, when the forward problem is an integer linear program (ILP), the number of iterations sufficient for projected subgradient descent applied to the suboptimality loss to achieve exact consistency with the observed data, as a fully explicit function of the number of samples, the dimension of the features, the ranges of the features, and the structure of the constraint coefficient matrix, up to polynomial factors in the basic constants (the diameter of the weight set, the step-size parameter, and the Lipschitz constant of the suboptimality loss).

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

72 extracted references · 2 canonical work pages

  1. [1]

    2001 , publisher =

    Inverse optimization , author =. 2001 , publisher =

    2001

  2. [2]

    2012 , publisher =

    The multiplicative weights update method: a meta-algorithm and applications , author =. 2012 , publisher =

    2012

  3. [3]

    2018 , journal =

    Aswani, Anil and Shen, Zuo-Jun and Siddiq, Auyon , title =. 2018 , journal =

    2018

  4. [4]

    2020 , journal =

    Berthet, Quentin and Blondel, Mathieu and Teboul, Olivier and Cuturi, Marco and Vert, Jean-Philippe and Bach, Francis , title =. 2020 , journal =

    2020

  5. [5]

    Gomory cuts revisited , year =

    Balas, Egon and Ceria, Sebastian and Cornu. Gomory cuts revisited , year =

  6. [6]

    2021 , journal =

    Babier, Aaron and Chan, Timothy CY and Lee, Taewoo and Mahmood, Rafid and Terekhov, Daria , title =. 2021 , journal =

    2021

  7. [7]

    2017 , doi =

    Beck, Amir , title =. 2017 , doi =

    2017

  8. [8]

    2015 , journal =

    Bertsimas, Dimitris and Gupta, Vishal and Paschalidis, Ioannis Ch , title =. 2015 , journal =

    2015

  9. [9]

    2021 , organization =

    Online learning from optimal actions , author =. 2021 , organization =

    2021

  10. [10]

    2025 , publisher =

    Contextual inverse optimization: Offline and online learning , author =. 2025 , publisher =

    2025

  11. [11]

    An online-learning approach to inverse optimization , year =

    B. An online-learning approach to inverse optimization , year =

  12. [12]

    2017 , publisher =

    Inverse optimization for the recovery of market structure from market outcomes: An application to the MISO electricity market , author =. 2017 , publisher =

    2017

  13. [13]

    2003 , url =

    Subgradient methods , author =. 2003 , url =

    2003

  14. [14]

    2019 , journal =

    Chan, Timothy CY and Lee, Taewoo and Terekhov, Daria , title =. 2019 , journal =

    2019

  15. [15]

    2022 , publisher =

    An inverse optimization approach to measuring clinical pathway concordance , author =. 2022 , publisher =

    2022

  16. [16]

    2023 , journal =

    Chan, Timothy CY and Mahmood, Rafid and Zhu, Ian Yihang , title =. 2023 , journal =

    2023

  17. [17]

    Online convex optimization perspective for learning from dynamically revealed preferences , year =

    Chen, Violet Xinying and K. Online convex optimization perspective for learning from dynamically revealed preferences , year =

  18. [18]

    Adaptive subgradient methods for online learning and stochastic optimization , author =

  19. [19]

    2022 , journal =

    Elmachtoub, Adam N and Grigas, Paul , title =. 2022 , journal =

    2022

  20. [20]

    2023 , booktitle =

    Ferber, Aaron M and Huang, Taoan and Zha, Daochen and Schubert, Martin and Steiner, Benoit and Dilkina, Bistra and Tian, Yuandong , title =. 2023 , booktitle =

    2023

  21. [21]

    1977 , journal =

    Flatto, Leopold and Newman, Donald J , title =. 1977 , journal =

    1977

  22. [22]

    2020 , booktitle =

    Ferber, Aaron and Wilder, Bryan and Dilkina, Bistra and Tambe, Milind , title =. 2020 , booktitle =

    2020

  23. [23]

    2023 , note =

    Handbook of convergence theorems for (stochastic) gradient methods , author =. 2023 , note =

    2023

  24. [24]

    Contextual recommendations and low-regret cutting-plane algorithms , author =

  25. [25]

    and Lawler, E.L

    Graham, R.L. and Lawler, E.L. and Lenstra, J.K. and Kan, A.H.G. Rinnooy , title =. 1979 , booktitle =

    1979

  26. [26]

    2021 , journal =

    Ghobadi, Kimia and Mahmoudzadeh, Houra , title =. 2021 , journal =

    2021

  27. [27]

    1960 , address =

    Gomory, Ralph Edward , title =. 1960 , address =

    1960

  28. [28]

    2018 , journal =

    Gunantara, Nyoman , title =. 2018 , journal =

    2018

  29. [29]

    2007 , publisher =

    Logarithmic regret algorithms for online convex optimization , author =. 2007 , publisher =

    2007

  30. [30]

    2019 , note =

    Hazan, Elad , title =. 2019 , note =

    2019

  31. [31]

    Harris, C. R. and Millman, K. J. and van der Walt, S. J. and Gommers, R. and Virtanen, P. and Cournapeau, D. and Wieser, E. and Taylor, J. and Berg, S. and Smith, N. J. and Kern, R. and Picus, M. and Hoyer, S. and van Kerkwijk, M. H. and Brett, M. and Haldane, A. and Del R. 2020 , journal =

    2020

  32. [32]

    2004 , publisher =

    Inverse combinatorial optimization: A survey on problems, methods, and results , author =. 2004 , publisher =

    2004

  33. [33]

    and Smith, Alice E

    Konak, Abdullah and Coit, David W. and Smith, Alice E. , title =. 2006 , issn =

    2006

  34. [34]

    Kitaoka, Akira and Eto, Riki , title =

  35. [35]

    2023 , note =

    Kitaoka, Akira and Eto, Riki , title =. 2023 , note =

    2023

  36. [36]

    2023 , doi =

    Data-driven inverse optimization for marginal offer price recovery in electricity markets , author =. 2023 , doi =

    2023

  37. [37]

    2018 , journal =

    Mohajerin Esfahani, Peyman and Shafieezadeh-Abadeh, Soroosh and Hanasusanto, Grani A and Kuhn, Daniel , title =. 2018 , journal =

    2018

  38. [38]

    1996 , journal =

    Murata, Tadahiko and Ishibuchi, Hisao and Tanaka, Hideo , title =. 1996 , journal =

    1996

  39. [39]

    2014 , journal =

    Mansini, Renata and Ogryczak, Wlodzimierz and Speranza, M Grazia , title =. 2014 , journal =

    2014

  40. [40]

    1983 , publisher =

    Problem Complexity and Method Efficiency in Optimization , author =. 1983 , publisher =

    1983

  41. [41]

    2004 , journal =

    Nemirovski, Arkadi , title =. 2004 , journal =

    2004

  42. [42]

    Algorithms for inverse reinforcement learning , author =

  43. [43]

    2002 , journal =

    Ovchinnikov, Sergei , title =. 2002 , journal =

    2002

  44. [44]

    2023 , url =

    Perron, Laurent and Furnon, Vincent , title =. 2023 , url =

    2023

  45. [45]

    and Saff, E

    Reznikov, A. and Saff, E. B. , title =. 2015 , issn =

    2015

  46. [46]

    1970 , publisher =

    Conjugate convex functions in optimal control and the calculus of variations , author =. 1970 , publisher =

    1970

  47. [47]

    Online Inverse Linear Optimization: Improved Regret Bound, Robustness to Suboptimality, and Toward Tight Regret Analysis , author =

  48. [48]

    2023 , organization =

    Understanding and generalizing contrastive learning from the inverse optimal transport perspective , author =. 2023 , organization =

    2023

  49. [49]

    2023 , booktitle =

    Sun, Chunlin and Liu, Shang and Li, Xiaocheng , title =. 2023 , booktitle =

    2023

  50. [50]

    1967 , volume =

    Stallings, John R and Swarup, G Ananda , title =. 1967 , volume =

    1967

  51. [51]

    2019 , booktitle =

    Suzuki, Yasuhisa and Wee, Werner M and Nishioka, Itaru , title =. 2019 , booktitle =

    2019

  52. [52]

    2005 , pages =

    Tarantola, Albert , title =. 2005 , pages =. doi:10.1137/1.9780898717921 , address =

    doi:10.1137/1.9780898717921 2005

  53. [53]

    1981 , journal =

    Tversky, Amos and Kahneman, Daniel , title =. 1981 , journal =

    1981

  54. [54]

    , title =

    Van Rossum, Guido and Drake, Fred L. , title =. 2009 , isbn =

    2009

  55. [55]

    Metagrad: Multiple learning rates in online learning , author =

  56. [56]

    Metagrad: Adaptation using multiple learning rates in online learning , author =

  57. [57]

    Projection onto the probability simplex: An efficient algorithm with a simple proof, and an application , year =

    Wang, Weiran and Carreira-Perpin. Projection onto the probability simplex: An efficient algorithm with a simple proof, and an application , year =

  58. [58]

    2019 , booktitle =

    Wilder, Bryan and Dilkina, Bistra and Tambe, Milind , title =. 2019 , booktitle =

    2019

  59. [59]

    Steepest Descent Algorithm for

    Oki, Taihei and Shioura, Akiyoshi , year =. Steepest Descent Algorithm for

  60. [60]

    Finite and Corruption-Robust Regret Bounds in Online Inverse Linear Optimization under

    Oki, Taihei and Sakaue, Shinsaku , year =. Finite and Corruption-Robust Regret Bounds in Online Inverse Linear Optimization under

  61. [61]

    2010 , publisher =

    Nonlinear discrete optimization , author =. 2010 , publisher =. doi:10.4171/093 , address =

    work page doi:10.4171/093 2010

  62. [62]

    1998 , issn =

    Minimization of an M-convex function , journal =. 1998 , issn =. doi:10.1016/S0166-218X(97)00140-6 , author =

    work page doi:10.1016/s0166-218x(97)00140-6 1998

  63. [63]

    M-convex function minimization under

    Shioura, Akiyoshi , note =. M-convex function minimization under

  64. [64]

    1996 , publisher =

    Gr\"obner bases and convex polytopes , author =. 1996 , publisher =. doi:10.1090/ulect/008 , address =

    doi:10.1090/ulect/008 1996

  65. [65]

    1996 , journal =

    Murota, Kazuo , title =. 1996 , journal =

    1996

  66. [66]

    1998 , journal =

    Murota, Kazuo , title =. 1998 , journal =

    1998

  67. [67]

    2003 , publisher =

    Murota, Kazuo , title =. 2003 , publisher =

    2003

  68. [68]

    1999 , journal =

    Murota, Kazuo and Shioura, Akiyoshi , title =. 1999 , journal =

    1999

  69. [69]

    Burke, J. V. and Ferris, M. C. , title =. 1993 , journal =

    1993

  70. [70]

    Ferris, M. C. , title =. 1991 , journal =

    1991

  71. [71]

    Polyak, B. T. , title =. 1987 , publisher =

    1987

  72. [72]

    2024 , eprint =

    Kitaoka, Akira , title =. 2024 , eprint =

    2024