Quadratic manifold fitting provides a flexible framework for modeling nonlinear low-dimensional structures embedded in high-dimensional spaces. However, existing methods typically employ prescribed noise models or fixed loss functions, limiting their adaptability to varying noise characteristics. To address this limitation, we propose an alternating maximum likelihood estimation method for subspace-constrained quadratic matrix factorization, termed AMLE-SQMF, which jointly estimates the quadratic manifold and the underlying noise distribution. Specifically, we model the residuals using a generalized Gaussian distribution (GGD), whose shape parameter controls the tail behavior and determines the associated fitting loss. By exploiting the closed-form maximum likelihood estimator of the scale parameter, we eliminate the scale variable and derive a profile negative log-likelihood. Based on this formulation, we develop an alternating optimization algorithm that iteratively updates the manifold parameters and estimates the noise shape parameter. This enables adaptive loss selection within a unified likelihood-based framework, without requiring prior specification of the noise characteristics. Experiments on synthetic and real-world datasets demonstrate the effectiveness of AMLE-SQMF in quadratic manifold fitting and its robustness to diverse noise distributions.
Fig. 1: Illustration of local quadratic surface fitting to noisy sphere samples using IRGU. The blue sphere shows the original unit sphere, and the colored patches show the fitted quadratic surfaces. Circles and crosses represent noisy samples and their Euclidean projections onto the fitted surfaces, respectively. Diamonds and stars represent the noisy neighborhood centers and their projections, respectively.
Fig. 2: Objective function value F versus iteration number for GD and IRGU when fitting a local quadratic surface to noisy sphere samples. Panels (a)–(d) use p=1.25,1.5,1.75,2 , respectively, for both the noise shape and the fitting loss exponent.
Methods / p
1.25
1.50
1.75
2.00
GD
13.91
3.01
12.41
0.81
IRGU
2.32
2.24
2.42
0.75
TABLE I: Optimization time (seconds) on the sphere example.
p
1.25
1.50
1.75
2.00
Mean
0.046
0.032
0.033
0.036
SD
0.038
0.027
0.029
0.038
TABLE II: Mean and sample standard deviation of the minimum Euclidean distances from 30 noisy neighborhood centers to their fitted quadratic surfaces for each p .
Fig. 3: Shape-estimation MSE across the (D,n) configurations.
Fig. 4: Effects of sample size and ambient dimension on shape-parameter estimation. The top row shows the effects of varying sample size n at fixed ambient dimension D , whereas the bottom row shows the effects of varying ambient dimension D at fixed sample size n .
Fig. 5: Mean estimates of p for 30 Frey Face images versus (a) the noise scale α⋆ and (b) the corresponding noise standard deviation. Shaded regions indicate one sample standard deviation; dashed lines indicate the true values of p .
Fig. 6: Shape (left) and scale (right) estimates over 30 repeats at different generating scales. Dots show individual estimates; boxes show the interquartile ranges with medians, and diamonds indicate the generating parameters. Scale estimates are multiplied by 102 .
p⋆
α⋆ = 0.05
α⋆ = 0.10
α⋆ = 0.15
α⋆ = 0.20
Shape p
1.00
1.066 ± 0.024
1.050 ± 0.020
1.027 ± 0.015
1.021 ± 0.012
1.25
1.142 ± 0.066
1.273 ± 0.021
1.264 ± 0.021
1.251 ± 0.018
1.50
1.150 ± 0.092
1.461 ± 0.043
1.513 ± 0.019
1.516 ± 0.018
1.75
1.129 ± 0.098
1.564 ± 0.073
1.693 ± 0.031
1.726 ± 0.019
2.00
1.112 ± 0.098
1.627 ± 0.104
1.848 ± 0.050
1.930 ± 0.033
TABLE III: Estimates of p and α across 30 Frey Face images under different noise scales. Entries are reported as means ± sample standard deviations.
Fig. 7: Parameter estimates for one image as the nested neighborhood grows: (a) shape and (b) scale.
p⋆
n = 30
n = 60
n = 90
Shape p
1.00
0.583 ± 0.066
0.940 ± 0.016
1.007 ± 0.012
1.25
0.817 ± 0.058
1.135 ± 0.019
1.235 ± 0.019
1.50
0.981 ± 0.029
1.450 ± 0.041
1.506 ± 0.022
1.75
1.130 ± 0.188
1.695 ± 0.030
1.723 ± 0.021
2.00
1.748 ± 0.254
1.911 ± 0.029
1.930 ± 0.027
TABLE IV: Estimates of p and α across 20 Frey Face neighborhood centers with varying sample sizes, where the true noise scale is fixed at α⋆=0.2 . Entries are reported as the mean ± sample standard deviation.
Method
GGD p⋆ = 1.25
GGD p⋆ = 1.50
GGD p⋆ = 1.75
Logistic
Student- t 10
AMLE-SQMF (Ours)
8.902 ± 2.226
8.595 ± 1.949
9.038 ± 2.042
8.828 ± 1.901
8.967 ± 1.989
SQMF ( p = 1)
9.857 ± 2.465
10.443 ± 4.573
10.247 ± 1.971
9.642 ± 1.676
9.756 ± 1.551
SQMF ( p = 2)
8.895 ± 2.007
9.033 ± 2.756
9.066 ± 1.923
8.884 ± 1.884
8.998 ± 1.928
LPCA
12.121 ± 2.637
12.049 ± 2.333
12.163 ± 2.532
11.815 ± 2.394
11.914 ± 2.299
SPH
14.889 ± 5.137
14.457 ± 4.230
14.565 ± 4.823
14.583 ± 4.272
14.485 ± 3.832
Q-LS
11.923 ± 2.399
11.825 ± 2.357
11.816 ± 2.251
11.614 ± 2.162
11.660 ± 2.137
TABLE V: Reconstruction MSE for 30 images from the Isomap Face dataset. Entries are reported as the means ± sample standard deviations.
Fig. 8: Isomap Face reconstructions (top) and pixelwise squared errors (bottom) under generalized Gaussian noise with p⋆=1.25 .
Fig. 9: Frey Face reconstruction with α⋆=0.20 and p⋆=2 : (a) clean image, (b) noisy image, and (c) fitted image using the estimated shape parameter.
Fig. 10: Latent-space visualization for 40 images of one subject. (a) Images generated on a 10×10 latent grid. (b) Fitted coordinates; blue and red indicate negative and positive lighting azimuths, respectively.
Zheng Zhai is currently a faculty member in the Department of Statistics, Faculty of Arts and Sciences at Beijing Normal University, Zhuhai. He received his Ph.D. degree from the School of Mathematical Sciences at Zhejiang University and worked as a postdoctoral researcher at the University of Toronto and the National University of Singapore. His current research interests include diffusion model, matrix computation, and manifold learning.
Table 18
Zhuan Liang is currently a graduate student in the Department of Statistics, Faculty of Arts and Sciences at Beijing Normal University, Zhuhai. He received his B.S. degree from the School of Mathematics and Statistics at Shandong University. His current research interests include diffusion models and matrix computation.
Table 19
Sujun Guo is currently a graduate student in the Department of Statistics, Faculty of Arts and Sciences, Beijing Normal University, Zhuhai, China. She received the B.S. degree from the School of Mathematics and Statistics, Northeast Normal University. Her research interests include statistical methods and their applications in machine learning.
Department of Statistics, Faculty of Arts and Sciences at Beijing Normal University, Zhuhai. · School of Mathematics and Information Sciences, Yantai University.