{"references": ["D.D. Lee, H.S. Seung, Learning the parts of objects by non-negative\nmatrix factorization, Nature 401 (6755) (1999) 788-791.", "E. F. Gonzales, Y. Zhang, Accelerating the Lee-Seung algorithm for nonnegative\nmatrix factorization, Dept. Comput. Appl. Math., Rice Univ.,\nHouston, TX, Tech. Rep. (2005).", "M. Berry, M. Browne, A. Langville, P. Pauca, R. Plemmons, Algorithms\nand applications for approximate nonnegative matrix factorization,\nComputational Statistics and Data Analysis 52 (2006) 155-173.", "P. Paatero, U. Tapper, Positive matrix factorization: a non-negative\nfactor model with optimal utilization of error estimates of data values,\nEnvironmetrics, 5 (1994) 111-126.", "M. Chu, F. Diele, R. Plemmons, S. Ragni, Optimality, computation, and\ninterpretation of nonnegative matrix factorizations, unpublished report\n(2004).", "P. Paatero and U. Tapper, Positive matrix factorization: A non-negative\nfactor model with optimal utilization of error, Environmetrics, 5 (1994)\n111-126.", "L.K. Saul, F. Sha, D.D. Lee, Statistical signal processing with\nnonnegativity constraints, Proceedings of EuroSpeech 2 (2003) 1001-\n1004.", "F. Shahnaz, M.W. Berry, V.P. Pauca, R. Plemmons, Document clusting\nusing nonnegative matrix factorization, Information Processing and\nManagement 42 (2006) 373-386.", "P.M. Kim, B. Tidor, Subsystem identification through dimensionality\nreduction of large-scale gene expression data Genome Research, 13\n(2003) 1706C1718.\n[10] V.P. Pauca, J. Piper, R.J. Plemmons, Nonnegative matrix factorization for\nspectral data analysis, Linear Algebra and its Applications 416 (2006)\n29-47.\n[11] F. Sha, L.K. Saul, Real-time pitch determination of one or more voices\nby nonnegative matrix factorization, Advances in Neural Information\nProcessing Systems (2005). online papers.\n[12] P. Smaragdis, J.C. Brown, Non-negative matrix factorization for\npolyphonic music transcription, in: Proceedings of IEEE Workshop on\nApplications of Signal Processing to Audio and Acoustics (2003) 177-\n180.\n[13] L. Lu, Alternative gradient algorithms with applications to nonnegative\nmatrix factorizations, Applied Mathematics and Computation 216\n(2010) 1763-1770.\n[14] A. Cichocki, R. Zdunek, S. Amari, New algorithms for non-negative\nmatrix factorization in applications to blind source separation, ICASSP\n2006, Toulouse, France, 621-625.\n[15] C.J. LIN, Projected gradient methods for non-negative matrix\nfactorization, Tech. Report Information and Support Service ISSTECH-\n95-013 (2005).\n[16] S.M. Yang, Z. Yi, Convergence analysis of non-negative matrix\nfactorization for BSS algorithm, Neural Process Lett (2010) 45-64.\n[17] A. Cichocki, R. Zdunek, Multilayer non-negative matrix factorization\nusing project gradient approaches, International Journal of Neural\nSystems (2007) 431-446.\n[18] R. Salakhutdinov, S.T. Roweis, Z. Ghahramani, The convergence of\nbound optimization algorithms, In: The 19th International Conference\non Uncertainty in Artificial Intelligence, (2003).\n[19] C.J. Lin, On the convergence of multiplicative update algorithms for\nnonnegative matrix factorization, Transactions on Neural Networks\n(2007) 18.\n[20] V.L. Kocic, G. Ladas, Global behavior of nonlinear difference equations\nof higher order, Kluwer Academic Publishers (1953).\n[21] J.K. Hale, Theory of functional differential equations, Springer,\nHeidelberg, New York (1977).\n[22] J. Dai, S. Meyn, Stabilitity and convergence of moments for multiclass\nqueueing networks via fluid limit models, Translation on Automatic\nControl (1995) 1889-1904."]}

Non-negative matrix factorization (NMF) is a useful computational method to find basis information of multivariate nonnegative data. A popular approach to solve the NMF problem is the multiplicative update (MU) algorithm. But, it has some defects. So the columnwisely alternating gradient (cAG) algorithm was proposed. In this paper, we analyze convergence of the cAG algorithm and show advantages over the MU algorithm. The stability of the equilibrium point is used to prove the convergence of the cAG algorithm. A classic model is used to obtain the equilibrium point and the invariant sets are constructed to guarantee the integrity of the stability. Finally, the convergence conditions of the cAG algorithm are obtained, which help reducing the evaluation time and is confirmed in the experiments. By using the same method, the MU algorithm has zero divisor and is convergent at zero has been verified. In addition, the convergence conditions of the MU algorithm at zero are similar to that of the cAG algorithm at non-zero. However, it is meaningless to discuss the convergence at zero, which is not always the result that we want for NMF. Thus, we theoretically illustrate the advantages of the cAG algorithm.