Transactions on Machine Learning Research · 2025

On a Gradient Approach to Chebyshev Center Problems with Applications to Function Learning

Introducing gradOL — the first gradient-based optimization framework for Chebyshev center problems

Abhinav Raghuvanshi1 Mayank Baranwal1,2 Debasish Chatterjee1
1Indian Institute of Technology, Bombay  ·  2Tata Consultancy Services Research, Mumbai

Abstract

A New Frontier in Optimal Function Learning

We introduce gradOL, the first gradient-based optimization framework for solving Chebyshev center problems, a fundamental challenge in optimal function learning and geometric optimization. gradOL hinges on reformulating the semi-infinite problem as a finitary max-min optimization, making it amenable to gradient-based techniques.

By leveraging automatic differentiation for precise numerical gradient computation, gradOL ensures numerical stability and scalability, making it suitable for large-scale settings. Under strong convexity of the ambient norm, gradOL provably recovers optimal Chebyshev centers while directly computing the associated radius. This addresses a key bottleneck in constructing stable optimal interpolants.

Empirically, gradOL achieves significant improvements in accuracy and efficiency on 34 benchmark Chebyshev center problems from a benchmark CSIP library. Moreover, we extend gradOL to general convex semi-infinite programming (CSIP), attaining up to 4000× speedups over the state-of-the-art SIPAMPL solver tested on the indicated CSIP library containing 67 benchmark problems. gradOL thus offers a unified solution framework for Chebyshev centers and broader CSIPs.

4000×
Max Speedup vs SIPAMPL
67
Benchmark Problems Solved
638ms
Avg. Runtime (Cheby. Center)
34/34
Chebyshev Instances Solved

Contributions

What gradOL Brings to the Table

01
Efficient Chebyshev Center Solver
A robust, gradient-based Julia implementation using automatic differentiation libraries for high accuracy and scalability. The optimal value corresponds directly to the Chebyshev radius.
02
Theoretical Foundation
We establish that the map u ↦ G(u) is locally Lipschitz, justifying gradient-based learning and proving convergence to a generalized Goldstein stationary point.
03
Benchmark Superiority
Tested on 34 Chebyshev center and 33 additional CSIP benchmark instances, gradOL consistently outperforms all existing solvers — including SIPAMPL and simulated annealing methods.
04
General CSIP Extension
The gradient-based framework extends naturally to general convex semi-infinite programming, achieving several orders-of-magnitude improvement in speed with solution quality matching or exceeding state-of-the-art.

Presentation

Paper Walkthrough

A short presentation covering the key ideas, formulation, and results of gradOL.

Results

Performance Comparison

gradOL is benchmarked against the industry-standard SIPAMPL, MSA–Simulated Annealing, and Iterative Sampling on all benchmark instances. Runtime measured in milliseconds on an AWS t2.large instance (Intel Xeon, 8 GiB RAM).

Method Chebyshev Center (34 instances) CSIP (33 instances)
Solved Avg. Runtime (ms) Solved Avg. Runtime (ms)
Iterative Sampling 9 3,880.95 20 1,695.37
MSA–Simulated Annealing 19 23,768.12 20 16,417.43
SIPAMPL 32 50,906.00 28 72,336.00
gradOL OURS 34 638.79 33 303.77

Table 1: gradOL solves all benchmark instances while being up to ~80× faster than SIPAMPL on Chebyshev center problems and up to ~238× faster on general CSIPs.

Detailed Benchmark Comparison

Per-problem breakdown of gradOL against best-known values from the literature. Problems marked † are Chebyshev center instances. Values marked * were not in original literature but obtained via SIPAMPL.

Problem Source Reported Value gradOL Value Reported Time (ms) gradOL Time (ms)
coopeLPrice & Coope (1996)3.4310 × 10⁻¹3.3631 × 10⁻¹4.067
coopeMPrice & Coope (1996)1.0000 × 10⁰1.0000 × 10⁰3.052
coopeNPrice & Coope (1996)0.0000 × 10⁰−6.9580 × 10⁻³1.19 × 10³3.209
fang1Fang & Wu (1994)4.7927 × 10⁻¹4.7939 × 10⁻¹2.173 × 10⁴28.737
fang2Fang & Wu (1994)6.9315 × 10⁻¹6.7982 × 10⁻¹2.307 × 10⁴65.094
fang3Fang & Wu (1994)1.7185 × 10⁰1.7183 × 10⁰2.128 × 10⁴33.526
ferris1 †Ferris & Philpott (1989)4.8800 × 10⁻³4.9828 × 10⁻⁴5.99 × 10³2.296 × 10²
ferris2Ferris & Philpott (1989)−1.7869 × 10⁰−1.7865 × 10⁰7.46 × 10³8.177
goerner4 †Goerner (1997)5.3324 × 10⁻²2.6209 × 10⁻²8.25 × 10³2.658 × 10²
goerner5 †Goerner (1997)2.7275 × 10⁻²2.5906 × 10⁻²2.2 × 10⁴1.665 × 10²
goerner6 †Goerner (1997)1.0770 × 10⁻³5.9995 × 10⁻⁵4.658 × 10⁴47.247
honstedelHonstede (1979)1.2124 × 10⁰1.0424 × 10⁰3.984
kortanek1Kortanek & No (1993)3.2212 × 10⁰3.2184 × 10⁰4.802 × 10⁴6.018
kortanek2Kortanek & No (1993)6.8629 × 10⁻¹6.8628 × 10⁻¹1.11 × 10³9.324 × 10²
kortanek3 †Kortanek & No (1993)1.4708 × 10⁻²2.2281 × 10⁻⁴1.5 × 10³95.012
kortanek4 †Kortanek & No (1993)5.2083 × 10⁻³3.7413 × 10⁻⁵2.666 × 10⁴1.443
leon1 †Leon et al. (2000)4.5050 × 10⁻³2.4607 × 10⁻⁴1.29 × 10³7.283
leon2 †Leon et al. (2000)4.1880 × 10⁻⁵1.0002 × 10⁻⁷1.111 × 10⁴12.273
leon3 †Leon et al. (2000)5.2190 × 10⁻⁴1.0002 × 10⁻⁵5.12 × 10³4.778 × 10²
leon4 †Leon et al. (2000)2.6028 × 10⁻³1.3479 × 10⁻⁴1.381 × 10⁴6.175 × 10³
leon5 †Leon et al. (2000)1.4257 × 10⁻²4.8963 × 10⁻⁴4.722 × 10⁴14.99
leon6 †Leon et al. (2000)1.5540 × 10⁻⁴1.1692 × 10⁻⁵4.12 × 10³7.378
leon7 †Leon et al. (2000)2.0997 × 10⁻³2.5904 × 10⁻⁴4.37 × 10³21.41
leon8 †Leon et al. (2000)5.4222 × 10⁻²1.0002 × 10⁻⁷2.139 × 10⁴5.146 × 10³
leon9 †Leon et al. (2000)1.6338 × 10⁻¹1.6338 × 10⁻¹1.696 × 10⁴8.694
leon10 †Leon et al. (2000)5.3825 × 10⁻¹5.3776 × 10⁻¹2.65 × 10³19.612
leon11 †Leon et al. (2000)4.8414 × 10⁻²2.3620 × 10⁻³1.28 × 10³15.384
leon13 †Leon et al. (2000)2.3607 × 10⁻¹2.3531 × 10⁻¹1.26 × 10³3.972
leon14Leon et al. (2000)6.6667 × 10⁻¹6.6640 × 10⁻¹1.4 × 10³5.151
leon15Leon et al. (2000)−6.6667 × 10⁻¹−6.6657 × 10⁻¹9.5 × 10²4.234
leon16Leon et al. (2000)1.7263 × 10⁰1.7187 × 10⁰2.2 × 10²3.04
leon17Leon et al. (2000)−2.0000 × 10⁰−1.9998 × 10⁰1.9 × 10²3.4
leon18 *Leon et al. (2000)−1.7500 × 10⁰−1.7000 × 10⁰3.63 × 10³3.303
leon19 *Leon et al. (2000)7.8584 × 10⁻¹7.7543 × 10⁻¹2.12 × 10³4.003
leon20Leon et al. (2000)3.2380 × 10⁻¹3.2316 × 10⁻¹1.68 × 10³11.713
leon21Leon et al. (2000)−9.9661 × 10¹−9.9670 × 10¹3.73 × 10³4.399 × 10³
leon22Leon et al. (2000)−1.0472 × 10¹−1.0472 × 10¹5.9 × 10²4.211 × 10³
leon23Leon et al. (2000)−3.0857 × 10¹−3.0857 × 10¹5.1 × 10²16.434
leon24 *Leon et al. (2000)−1.1998 × 10¹−1.0370 × 10¹1.54 × 10³7.526
lin1 *Lin et al. (1998)−1.8244 × 10⁰−1.8300 × 10⁰2.98 × 10⁴3.785
reemtsen1 †Reemtsen (1991)1.5249 × 10⁻¹1.5231 × 10⁻¹1.2689 × 10⁵4.974 × 10²
reemtsen2 †Reemtsen (1991)5.8359 × 10⁻²5.6952 × 10⁻²1.0145 × 10⁵17.249
reemtsen3 †Reemtsen (1991)7.3547 × 10⁻¹7.3548 × 10⁻¹1.6633 × 10⁵15.484
reemtsen4 †Reemtsen (1991)1.1401 × 10⁻²1.0001 × 10⁻⁵4.5109 × 10⁵64.185
reemtsen5 †Reemtsen (1991)8.8932 × 10⁻²2.8761 × 10⁻²1.4513 × 10⁵1.950
potchinkov3 †Potchinkov (1997)3.4226 × 10⁻³5.788
potchinkovPL †Potchinkov (1997)9.2940 × 10⁻⁷2.595
powell1Todd (1994)−1.0000 × 10⁰−1.0000 × 10⁰9.2 × 10²3.096
hettich2 †Hettich (1979)5.3800 × 10⁻¹5.3742 × 10⁻¹2.68 × 10³3.98
hettich4 †Hettich (1979)1.0000 × 10⁰1.0001 × 10⁰3.6 × 10²8.137
hettich5 †Hettich (1979)5.3800 × 10⁻¹5.3505 × 10⁻¹1.1995 × 10⁵10.909
hettich6 †Hettich (1979)2.8100 × 10⁻²2.8163 × 10⁻²5.504 × 10⁴1.735
hettich7 †Hettich (1979)1.7800 × 10⁻¹1.7776 × 10⁻¹4.976 × 10⁴7.776
hettich8 †Hettich (1979)2.996 × 10⁻²2.290 × 10³1.139 × 10²
hettich9 †Hettich (1979)3.4700 × 10⁻³3.4791 × 10⁻³8.465 × 10⁴3.837
hettich12 †Hettich (1979)1.0252 × 10⁻³7.844 × 10⁴8.695 × 10³
priceKPrice (1992)−3.0000 × 10⁰−3.0000 × 10⁰5.1 × 10²3.191
still1Still (2001)1.0000 × 10⁰9.9771 × 10⁻¹3.9 × 10²4.888
userman1.2802 × 10⁻⁷4.028
watson4aWatson (1983)6.4904 × 10⁻¹6.5012 × 10⁻¹1.78 × 10³10.116
watson4bWatson (1983)6.1688 × 10⁻¹6.1610 × 10⁻¹2.22 × 10³37.654
watson4cWatson (1983)6.1661 × 10⁻¹6.1564 × 10⁻¹2.82 × 10³10.029
watson5Watson (1983)4.3012 × 10⁰4.2966 × 10⁰8.4 × 10²17.097
watson7Watson (1983)1.0000 × 10⁰9.9777 × 10⁻¹1.23 × 10³7.534
watson8Watson (1983)2.4356 × 10⁰2.4356 × 10⁰2.189 × 10⁴1.627 × 10²
watson10Watson (1983)2.7527 × 10⁻¹−7.0000 × 10⁻⁵6.952
zhou1 †Zhou & Tits (1996)2.3605 × 10⁻¹2.3505 × 10⁻¹3.09 × 10³42.561

Table 2: Per-problem comparison. Teal rows (†) indicate Chebyshev center problems. * denotes values not in original literature, obtained via SIPAMPL. For 65 of 67 problems, gradOL deviates by less than 10⁻² from previously reported results.

Authors

Meet the Team

This work was conducted at the Indian Institute of Technology, Bombay

AR
Abhinav Raghuvanshi
Indian Institute of Technology, Bombay
MB
Mayank Baranwal
TCS Research, Mumbai &
Indian Institute of Technology, Bombay
DC
Debasish Chatterjee
Indian Institute of Technology, Bombay

Citation

BibTeX

@article{raghuvanshi2025gradol,
  title={On a Gradient Approach to {Chebyshev} Center Problems with
         Applications to Function Learning},
  author={Raghuvanshi, Abhinav and Baranwal, Mayank and Chatterjee, Debasish},
  journal={Transactions on Machine Learning Research},
  year={2025},
  url={https://openreview.net/forum?id=lPZVsDhyj3}
}