{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,7,11]],"date-time":"2026-07-11T02:08:41Z","timestamp":1783735721360,"version":"3.55.0"},"reference-count":95,"publisher":"Cambridge University Press (CUP)","license":[{"start":{"date-parts":[[2021,8,4]],"date-time":"2021-08-04T00:00:00Z","timestamp":1628035200000},"content-version":"unspecified","delay-in-days":95,"URL":"https:\/\/2.zoppoz.workers.dev:443\/http\/creativecommons.org\/licenses\/by\/4.0\/"}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":["Acta Numerica"],"published-print":{"date-parts":[[2021,5]]},"abstract":"<jats:p>Neural networks (NNs) are the method of choice for building learning algorithms. They are now being investigated for other numerical tasks such as solving high-dimensional partial differential equations. Their popularity stems from their empirical success on several challenging learning problems (computer chess\/Go, autonomous navigation, face recognition). However, most scholars agree that a convincing theoretical explanation for this success is still lacking. Since these applications revolve around approximating an unknown function from data observations, part of the answer must involve the ability of NNs to produce accurate approximations.<\/jats:p><jats:p>This article surveys the known approximation properties of the outputs of NNs with the aim of uncovering the properties that are not present in the more traditional methods of approximation used in numerical analysis, such as approximations using polynomials, wavelets, rational functions and splines. Comparisons are made with traditional approximation methods from the viewpoint of rate distortion, <jats:italic>i.e.<\/jats:italic> error versus the number of parameters used to create the approximant. Another major component in the analysis of numerical approximation is the computational time needed to construct the approximation, and this in turn is intimately connected with the stability of the approximation algorithm. So the stability of numerical approximation using NNs is a large part of the analysis put forward.<\/jats:p><jats:p>The survey, for the most part, is concerned with NNs using the popular ReLU activation function. In this case the outputs of the NNs are piecewise linear functions on rather complicated partitions of the domain of <jats:italic>f<\/jats:italic> into cells that are convex polytopes. When the architecture of the NN is fixed and the parameters are allowed to vary, the set of output functions of the NN is a parametrized nonlinear manifold. It is shown that this manifold has certain space-filling properties leading to an increased ability to approximate (better rate distortion) but at the expense of numerical stability. The space filling creates the challenge to the numerical method of finding best or good parameter choices when trying to approximate.<\/jats:p>","DOI":"10.1017\/s0962492921000052","type":"journal-article","created":{"date-parts":[[2021,8,4]],"date-time":"2021-08-04T09:15:21Z","timestamp":1628068521000},"page":"327-444","source":"Crossref","is-referenced-by-count":185,"title":["Neural network approximation"],"prefix":"10.1017","volume":"30","author":[{"given":"Ronald","family":"DeVore","sequence":"first","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Boris","family":"Hanin","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Guergana","family":"Petrova","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"56","published-online":{"date-parts":[[2021,8,4]]},"reference":[{"key":"S0962492921000052_r78","doi-asserted-by":"publisher","DOI":"10.1037\/h0042519"},{"key":"S0962492921000052_r6","first-page":"1","article-title":"Breaking the curse of dimensionality with convex neural networks","volume":"18","author":"Bach","year":"2017","journal-title":"J. Mach. Learn. Res."},{"key":"S0962492921000052_r24","doi-asserted-by":"crossref","unstructured":"Cohen, A. , DeVore, R. , Petrova, G. and Wojtaszczyk, P. (2020), Optimal stable nonlinear approximation. Available at arXiv:2009.09907.","DOI":"10.1007\/s10208-021-09494-z"},{"key":"S0962492921000052_r73","doi-asserted-by":"publisher","DOI":"10.1016\/j.neunet.2018.08.019"},{"key":"S0962492921000052_r55","doi-asserted-by":"publisher","DOI":"10.1016\/0893-6080(89)90020-8"},{"key":"S0962492921000052_r90","volume-title":"Statistical Learning Theory","author":"Vapnik","year":"1989"},{"key":"S0962492921000052_r93","first-page":"639","volume-title":"31st Conference on Learning Theory (COLT 2018)","volume":"75","author":"Yarotsky","year":"2018"},{"key":"S0962492921000052_r84","doi-asserted-by":"publisher","DOI":"10.1038\/nature24270"},{"key":"S0962492921000052_r4","unstructured":"Arora, R. , Basu, A. , Mianjy, P. and Mukherjee, A. (2018a), Understanding deep neural networks with rectified linear units, in 6th International Conference on Learning Representations (ICLR 2018). Available at https:\/\/2.zoppoz.workers.dev:443\/https\/openreview.net\/forum?id=B1J_rgWRW."},{"key":"S0962492921000052_r54","volume-title":"The Organization of Behavior: A Neuropsychological Theory","author":"Hebb","year":"1949"},{"key":"S0962492921000052_r51","first-page":"361","volume-title":"Advances in Neural Information Processing Systems 32 (NeurIPS 2019)","author":"Hanin","year":"2019"},{"key":"S0962492921000052_r85","doi-asserted-by":"publisher","DOI":"10.1090\/pcms\/013\/08"},{"key":"S0962492921000052_r8","doi-asserted-by":"publisher","DOI":"10.1109\/18.256500"},{"key":"S0962492921000052_r16","doi-asserted-by":"publisher","DOI":"10.1137\/100795772"},{"key":"S0962492921000052_r95","unstructured":"Zhang, C. , Bengio, S. , Hardt, M. , Recht, B. and Vinyals, O. (2017), Understanding deep learning requires rethinking generalization, in 5th International Conference on Learning Representations (ICLR 2017). Available at https:\/\/2.zoppoz.workers.dev:443\/https\/openreview.net\/forum?id=Sy8gdB9xx."},{"key":"S0962492921000052_r20","doi-asserted-by":"publisher","DOI":"10.1109\/MSP.2017.2693418"},{"key":"S0962492921000052_r92","doi-asserted-by":"publisher","DOI":"10.1016\/j.neunet.2017.07.002"},{"key":"S0962492921000052_r74","doi-asserted-by":"publisher","DOI":"10.1007\/BFb0078887"},{"key":"S0962492921000052_r29","first-page":"51","volume-title":"Acta Numerica","volume":"7","author":"DeVore","year":"1998"},{"key":"S0962492921000052_r30","doi-asserted-by":"publisher","DOI":"10.1090\/S0002-9947-1988-0920166-3"},{"key":"S0962492921000052_r81","doi-asserted-by":"publisher","DOI":"10.1016\/j.neunet.2019.07.011"},{"key":"S0962492921000052_r63","doi-asserted-by":"publisher","DOI":"10.1006\/jath.1998.3304"},{"key":"S0962492921000052_r50","doi-asserted-by":"publisher","DOI":"10.3390\/math7100992"},{"key":"S0962492921000052_r10","first-page":"6240","volume-title":"Advances in Neural Information Processing Systems 30 (NIPS 2017)","author":"Bartlett","year":"2017"},{"key":"S0962492921000052_r64","doi-asserted-by":"publisher","DOI":"10.1006\/jath.1996.0031"},{"key":"S0962492921000052_r70","unstructured":"Parhi, R. and Nowak, R. D. (2020), Banach space representer theorems for neural networks and ridge splines. Available at arXiv:2006.05626v2."},{"key":"S0962492921000052_r41","doi-asserted-by":"publisher","DOI":"10.1109\/JSAIT.2020.2991422"},{"key":"S0962492921000052_r52","doi-asserted-by":"publisher","DOI":"10.1016\/S0168-2024(08)70133-4"},{"key":"S0962492921000052_r23","first-page":"2937","volume-title":"Advances in Neural Information Processing Systems 32 (NeurIPS 2019)","author":"Chizat","year":"2019"},{"key":"S0962492921000052_r71","unstructured":"Peetre, J. (1976), New Thoughts on Besov Spaces, Vol. 1 of Duke University Mathematics Series, Mathematics Department, Duke University."},{"key":"S0962492921000052_r49","unstructured":"G\u00fchring, I. , Raslan, M. and Kutyniok, G. (2020), Expressivity of deep neural networks. Available at arXiv:2007.04759."},{"key":"S0962492921000052_r46","doi-asserted-by":"publisher","DOI":"10.1090\/cbms\/079"},{"key":"S0962492921000052_r82","doi-asserted-by":"crossref","unstructured":"Siegel, J. W. and Xu, J. (2020), High order approximation rates for neural networks with ReLU k activation functions. Available at arXiv:2012.07205.","DOI":"10.1016\/j.neunet.2020.05.019"},{"key":"S0962492921000052_r69","unstructured":"Opschoor, J. , Schwab, C. and Zech, J. (2019b), Exponential ReLU DNN expression of holomorphic maps in high dimension. SAM Research Report, ETH Z\u00fcrich."},{"key":"S0962492921000052_r61","doi-asserted-by":"publisher","DOI":"10.1007\/978-3-642-60932-9"},{"key":"S0962492921000052_r58","first-page":"1097","volume-title":"Advances in Neural Information Processing Systems 25 (NIPS 2012)","author":"Krizhevsky","year":"2012"},{"key":"S0962492921000052_r53","doi-asserted-by":"crossref","first-page":"502","DOI":"10.4208\/jcm.1901-m2018-0160","article-title":"ReLU deep neural networks and linear finite elements","volume":"38","author":"He","year":"2020","journal-title":"Comput. Math"},{"key":"S0962492921000052_r77","unstructured":"Pinkus, A. (2012), N-widths in Approximation Theory, Vol. 7 of A Series of Modern Surveys in Mathematics, Springer Science & Business Media."},{"key":"S0962492921000052_r32","first-page":"843","article-title":"Besov spaces on domains in RD","volume":"335","author":"DeVore","year":"1993","journal-title":"Trans. Amer. Math. Soc."},{"key":"S0962492921000052_r17","doi-asserted-by":"publisher","DOI":"10.1137\/15M1025384"},{"key":"S0962492921000052_r62","unstructured":"Lu, J. , Shen, Z. , Yang, H. and Zhang, S. (2020), Deep network approximation for smooth functions. Available at arXiv:2001.03040."},{"key":"S0962492921000052_r65","doi-asserted-by":"publisher","DOI":"10.3934\/cpaa.2020181"},{"key":"S0962492921000052_r67","unstructured":"Ongie, G. , Willett, R. , Soudry, D. and Srebro, N. (2020), A function space view of bounded norm infinite width ReLU nets: The multivariate case, in 8th International Conference on Learning Representations (ICLR 2020). Available at https:\/\/2.zoppoz.workers.dev:443\/https\/openreview.net\/forum?id=H1lNPxHKDH."},{"key":"S0962492921000052_r18","doi-asserted-by":"publisher","DOI":"10.1137\/18M118709X"},{"key":"S0962492921000052_r88","volume-title":"A General Theory of Optimal Algorithms","author":"Traub","year":"1980"},{"key":"S0962492921000052_r13","volume-title":"Interpolation of Operators","author":"Bennett","year":"1990"},{"key":"S0962492921000052_r43","doi-asserted-by":"publisher","DOI":"10.1007\/s11425-018-9387-x"},{"key":"S0962492921000052_r94","volume-title":"Facing Up To Arrangements: Face-Count Formulas for Partitions of Space by Hyper-Planes","author":"Zaslavsky","year":"1975"},{"key":"S0962492921000052_r28","doi-asserted-by":"publisher","DOI":"10.1007\/978-1-4612-6333-3"},{"key":"S0962492921000052_r38","doi-asserted-by":"publisher","DOI":"10.1007\/s00365-013-9186-2"},{"key":"S0962492921000052_r68","doi-asserted-by":"publisher","DOI":"10.1142\/S0219530519410136"},{"key":"S0962492921000052_r66","first-page":"2924","volume-title":"Advances in Neural Information Processing Systems 27 (NIPS 2014)","author":"Montufar","year":"2014"},{"key":"S0962492921000052_r33","doi-asserted-by":"publisher","DOI":"10.1007\/BF02124742"},{"key":"S0962492921000052_r34","doi-asserted-by":"publisher","DOI":"10.1007\/BF01171759"},{"key":"S0962492921000052_r1","volume-title":"Sobolev Spaces","author":"Adams","year":"2003"},{"key":"S0962492921000052_r45","unstructured":"Elbr\u00e4chter, D. , Perekrestenko, D. , Grohs, P. and B\u00f6lcskei, H. (2019), Deep neural network approximation theory. Available at arXiv:1901.02220."},{"key":"S0962492921000052_r3","first-page":"242","volume-title":"Proceedings of the 36th International Conference on Machine Learning (ICML 2019)","volume":"97","author":"Allen-Zhu","year":"2019"},{"key":"S0962492921000052_r80","first-page":"1875","article-title":"Nonparametric regression using deep neural networks with ReLU activation","volume":"48","author":"Schmidt-Hieber","year":"2020","journal-title":"Ann. Statist."},{"key":"S0962492921000052_r37","doi-asserted-by":"publisher","DOI":"10.1007\/s00365-010-9105-8"},{"key":"S0962492921000052_r22","doi-asserted-by":"publisher","DOI":"10.1016\/0022-1236(81)90076-8"},{"key":"S0962492921000052_r2","unstructured":"Ali, M. and Nouy, A. (2020), Approximation of smoothness classes by deep ReLU networks. Available at arXiv:2007.15645v1."},{"key":"S0962492921000052_r91","unstructured":"Wu, Y. , Schuster, M. , Chen, Z. , Le, Q. V. , Norouzi, M. , Macherey, W. , Krikun, M. , Cao, Y. , Gao, Q. , Macherey, K. et al. (2016), Google\u2019s neural machine translation system: Bridging the gap between human and machine translation. Available at arXiv:1609.08144."},{"key":"S0962492921000052_r76","first-page":"143","volume-title":"Acta Numerica","volume":"8","author":"Pinkus","year":"1999"},{"key":"S0962492921000052_r39","first-page":"1675","volume-title":"Proceedings of the 36th International Conference on Machine Learning (ICML 2019)","volume":"97","author":"Du","year":"2019a"},{"key":"S0962492921000052_r36","first-page":"261","article-title":"Approximation by feed-forward neural networks","volume":"4","author":"DeVore","year":"1997","journal-title":"Ann. Numer. Math."},{"key":"S0962492921000052_r59","doi-asserted-by":"publisher","DOI":"10.1038\/nature14539"},{"key":"S0962492921000052_r14","doi-asserted-by":"publisher","DOI":"10.1090\/coll\/048"},{"key":"S0962492921000052_r12","doi-asserted-by":"publisher","DOI":"10.1073\/pnas.1907378117"},{"key":"S0962492921000052_r75","doi-asserted-by":"publisher","DOI":"10.1137\/S0036141097322959"},{"key":"S0962492921000052_r60","unstructured":"Liu, C. , Zhu, L. and Belkin, M. (2020), Toward a theory of optimization for over-parameterized systems of non-linear equations: The lessons of deep learning. Available at arXiv:2003.00307."},{"key":"S0962492921000052_r35","doi-asserted-by":"publisher","DOI":"10.1007\/BF02071385"},{"key":"S0962492921000052_r56","first-page":"8571","volume-title":"Advances in Neural Information Processing Systems 31 (NeurIPS 2018)","author":"Jacot","year":"2018"},{"key":"S0962492921000052_r87","first-page":"1","article-title":"Representation benefits of deep feedforward networks","volume":"49","author":"Telgarsky","year":"2016","journal-title":"JMLR: Workshop and Conference Proceedings"},{"key":"S0962492921000052_r5","first-page":"254","volume-title":"Proceedings of the 35th International Conference on Machine Learning (ICML 2018)","volume":"80","author":"Arora","year":"2018b"},{"key":"S0962492921000052_r79","first-page":"2667","volume-title":"Proceedings of the 32nd Conference on Learning Theory (COLT 2019)","volume":"99","author":"Savarese","year":"2019"},{"key":"S0962492921000052_r25","first-page":"1","article-title":"Tight lower bounds on the VC-dimension of geometric set systems","volume":"20","author":"Csiskos","year":"2019","journal-title":"J. Mach. Learn. Res."},{"key":"S0962492921000052_r89","doi-asserted-by":"publisher","DOI":"10.1007\/s10208-020-09472-x"},{"key":"S0962492921000052_r86","volume-title":"Singular Integrals and Differentiability Properties of Functions","author":"Stein","year":"1970"},{"key":"S0962492921000052_r42","unstructured":"Dziugaite, G. K. and Roy, D. M. (2017), Computing nonvacuous generalization bounds for deep (stochastic) neural networks with many more parameters than training data, in Workshop on Principled Approaches to Deep Learning (ICML 2017)."},{"key":"S0962492921000052_r21","doi-asserted-by":"publisher","DOI":"10.1051\/m2an\/2011056"},{"key":"S0962492921000052_r31","doi-asserted-by":"publisher","DOI":"10.2307\/1971227"},{"key":"S0962492921000052_r11","first-page":"1","article-title":"Nearly-tight VC-dimension and pseudodimension bounds for piecewise linear neural networks","volume":"20","author":"Bartlett","year":"2019","journal-title":"J. Mach. Learn. Res."},{"key":"S0962492921000052_r48","unstructured":"Gribonval, R. , Kutyniok, G. , Nielsen, M. and Voigtlaender, F. (2019), Approximation spaces of deep neural networks. Available at arXiv:1905.01208."},{"key":"S0962492921000052_r15","doi-asserted-by":"publisher","DOI":"10.1007\/978-3-642-66451-9"},{"key":"S0962492921000052_r27","unstructured":"Daubechies, I. , DeVore, R. , Foucart, S. , Hanin, B. and Petrova, G. (2019), Nonlinear approximation and (deep) ReLU networks. Available at arXiv:1905.02199 (to appear in Constr. Approx.)."},{"key":"S0962492921000052_r44","unstructured":", W. E , Ma, C. and Wu, L. (2019), The Barron space and the flow-induced function spaces for neural network models. Available at arXiv:1906.08039."},{"key":"S0962492921000052_r40","unstructured":"Du, S. S. , Zhai, X. , Poczos, B. and Singh, A. (2019b), Gradient descent provably optimizes over-parameterized neural networks, in 7th International Conference on Learning Representations (ICLR 2019). Available at https:\/\/2.zoppoz.workers.dev:443\/https\/openreview.net\/forum?id=S1eK3i09YQ."},{"key":"S0962492921000052_r72","unstructured":"Petersen, P. (2020), Neural network theory. Available at https:\/\/2.zoppoz.workers.dev:443\/http\/pc-petersen.eu\/Neural_Network_Theory.pdf."},{"key":"S0962492921000052_r57","doi-asserted-by":"publisher","DOI":"10.1109\/TIT.2018.2874447"},{"key":"S0962492921000052_r7","doi-asserted-by":"publisher","DOI":"10.1109\/JPROC.2020.3042100"},{"key":"S0962492921000052_r83","doi-asserted-by":"publisher","DOI":"10.1038\/nature16961"},{"key":"S0962492921000052_r9","doi-asserted-by":"publisher","DOI":"10.1007\/BF00993164"},{"key":"S0962492921000052_r19","doi-asserted-by":"crossref","first-page":"323","DOI":"10.1051\/ps:2005018","article-title":"Theory of classification: A survey of some recent advances","volume":"9","author":"Bousquet","year":"2005","journal-title":"ESAIM Probab. Statist."},{"key":"S0962492921000052_r26","doi-asserted-by":"publisher","DOI":"10.1007\/BF02551274"},{"key":"S0962492921000052_r47","doi-asserted-by":"publisher","DOI":"10.1214\/20-AOS1990"}],"container-title":["Acta Numerica"],"original-title":[],"language":"en","link":[{"URL":"https:\/\/2.zoppoz.workers.dev:443\/https\/www.cambridge.org\/core\/services\/aop-cambridge-core\/content\/view\/S0962492921000052","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2022,3,29]],"date-time":"2022-03-29T10:37:09Z","timestamp":1648550229000},"score":1,"resource":{"primary":{"URL":"https:\/\/2.zoppoz.workers.dev:443\/https\/www.cambridge.org\/core\/product\/identifier\/S0962492921000052\/type\/journal_article"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2021,5]]},"references-count":95,"alternative-id":["S0962492921000052"],"URL":"https:\/\/2.zoppoz.workers.dev:443\/https\/doi.org\/10.1017\/s0962492921000052","relation":{},"ISSN":["0962-4929","1474-0508"],"issn-type":[{"value":"0962-4929","type":"print"},{"value":"1474-0508","type":"electronic"}],"subject":[],"published":{"date-parts":[[2021,5]]}}}