Extended Low Rank Approximation (xLRA)

• xLRA is an adaptive extension of low-rank approximation that incorporates weighting, missing data, and tensor structures.
• It employs methods like projected-gradient, alternating least squares, and acceleration (Nesterov and Anderson) to achieve efficient convergence.
• xLRA is applied in both matrix and materials contexts, offering high accuracy with controlled adaptive rank corrections.

Searching arXiv for the cited papers to ground the article in the current record. Extended Low Rank Approximation (xLRA) denotes an extension of classical low-rank approximation in which the approximation is made adaptive to weighting, missingness, or higher-order structure rather than being restricted to an unweighted fixed-rank matrix factorization. In the cited literature, this extension appears in two closely related forms: an element-wise weighted low-rank matrix approximation framework with projected-gradient, alternating-least-squares, and acceleration mechanisms, and a tensor-decomposition–based surrogate modeling framework that maps high-dimensional microstructure descriptors to local elastic response fields through adaptive canonical polyadic (CP) terms (Tuzhilina et al., 2021, Karmakar et al., 24 Sep 2025). Taken together, these formulations place xLRA within the broader family of low-rank methods while emphasizing adaptive rank growth, weighted objectives, and efficient training procedures.

1. Scope and relation to classical low-rank approximation

Low-rank matrix approximation is described as one of the central concepts in machine learning, with applications in dimension reduction, de-noising, multivariate statistical methodology, and related settings (Tuzhilina et al., 2021). A recent extension to low-rank matrix approximation is low-rank matrix completion, which solves the problem when some observations are missing and is especially useful for recommender systems. The weighted formulation generalizes this further by introducing an element-wise weight matrix and treating matrix completion as the binary-weight special case (Tuzhilina et al., 2021).

In the materials-science setting, xLRA is formulated not as a matrix problem but as a low-rank approximation of a high-order response map from local microstructural descriptors to local elastic strain tensors. The underlying representation is CP decomposition, in which a high-order tensor is approximated as a sum of rank-1 outer products. The extension from ordinary low-rank approximation arises because rank-1 LRA often fails for elastically heterogeneous materials with strong contrasts, so xLRA adaptively incorporates higher-rank correction terms (Karmakar et al., 24 Sep 2025).

This suggests that xLRA is not a single algorithmic primitive but a family of low-rank constructions whose shared feature is the extension of standard LRA to settings with weighting, missing data, tensor structure, or adaptive rank correction.

Setting	Mathematical object	Extension mechanism
Weighted LRA	$M \in \mathbb{R}^{m \times n}$ with $W \in [0,1]^{m \times n}$	element-wise weighting, rank constraint or nuclear-norm surrogate
Materials xLRA	response tensor / local strain map	CP decomposition with adaptive higher-rank correction terms

2. Weighted matrix formulation

The weighted matrix formulation begins with a data matrix $M \in \mathbb{R}^{m \times n}$, a weight matrix $W \in [0,1]^{m \times n}$, and a target rank $r < \min(m,n)$. The objective is to seek a rank-$r$ approximation $X$ of $M$ minimizing weighted squared error: $$\min_{\substack{X\in\mathbb{R}^{m\times n}\ \operatorname{rank}(X)\le r}} \sum_{i=1}^m \sum_{j=1}^n W_{ij}\bigl(M_{ij}-X_{ij}\bigr)^2 \;=\; \min_{\operatorname{rank}(X)\le r}\;\bigl\|\sqrt{W}\ast(M-X)\bigr\|_F^2.$$ Equivalently, one may factor $X = U V^\top$ with $W \in [0,1]^{m \times n}$0 and $W \in [0,1]^{m \times n}$1 and solve

$W \in [0,1]^{m \times n}$2

Within this formulation, $W \in [0,1]^{m \times n}$3 encodes confidence or missing-data masks, $W \in [0,1]^{m \times n}$4 means “ignore entry,” $W \in [0,1]^{m \times n}$5 is the observed data, and $W \in [0,1]^{m \times n}$6 controls the low-rank subspace dimension (Tuzhilina et al., 2021).

A convex surrogate is also considered: $W \in [0,1]^{m \times n}$7 for which large $W \in [0,1]^{m \times n}$8 forces a low-rank solution. This places the weighted framework adjacent to nuclear-norm regularization and proximal splitting methods while preserving the interpretation of low-rank structure under nonuniform confidence in entries (Tuzhilina et al., 2021).

A common misconception is that weighted low-rank approximation is identical to matrix completion. The cited formulation states more narrowly that weighted low-rank matrix approximation covers low-rank matrix completion as a special case with binary weights. The weighted problem is therefore strictly more general than missing-data masking alone (Tuzhilina et al., 2021).

3. Algorithms, acceleration, and high-dimensional variants

For the nonconvex rank-constrained problem, the objective can be written as

$W \in [0,1]^{m \times n}$9

A projected-gradient step with unit step size yields the update

$M \in \mathbb{R}^{m \times n}$0

where $M \in \mathbb{R}^{m \times n}$1 keeps the top $M \in \mathbb{R}^{m \times n}$2 singular values (Tuzhilina et al., 2021). In factorized form, one may also alternate weighted least squares by solving row-by-row for $M \in \mathbb{R}^{m \times n}$3 with $M \in \mathbb{R}^{m \times n}$4 fixed and then swapping roles.

Two acceleration techniques are described. The first is Nesterov acceleration, with extrapolated iterate

$M \in \mathbb{R}^{m \times n}$5

followed by

$M \in \mathbb{R}^{m \times n}$6

In the convex setting, this improves the worst-case rate from $M \in \mathbb{R}^{m \times n}$7 to $M \in \mathbb{R}^{m \times n}$8 (Tuzhilina et al., 2021).

The second is Anderson acceleration. Writing the fixed-point map as

$M \in \mathbb{R}^{m \times n}$9

the method forms the last $W \in [0,1]^{m \times n}$0 residuals $W \in [0,1]^{m \times n}$1, with $W \in [0,1]^{m \times n}$2, solves

$W \in [0,1]^{m \times n}$3

and sets

$W \in [0,1]^{m \times n}$4

The procedure can be guarded by checking whether the objective decreases and otherwise falling back on the plain update. The theoretical comment given in the source is that Anderson acceleration has been shown, for suitably smooth fixed-point maps, to attain superlinear local convergence, while global convergence can be enforced by “Type-I” safeguards (Tuzhilina et al., 2021).

For very large problems with $W \in [0,1]^{m \times n}$5, even a skinny SVD is costly. A non-SVD variant therefore maintains factors $W \in [0,1]^{m \times n}$6 with $W \in [0,1]^{m \times n}$7, $W \in [0,1]^{m \times n}$8, and alternates two unweighted multi-response regressions on the intermediate matrix

$W \in [0,1]^{m \times n}$9

The updates are

$r < \min(m,n)$0

followed by

$r < \min(m,n)$1

Each step is $r < \min(m,n)$2 if $r < \min(m,n)$3 is sparse, and no SVDs are needed. In the convex-relaxed ALS version, $r < \min(m,n)$4 is replaced by $r < \min(m,n)$5, and similarly for the $r < \min(m,n)$6-update (Tuzhilina et al., 2021).

For the nonconvex rank-constraint problem, the stated guarantee is convergence only to a stationary point; in practice, the algorithms are reported as stable (Tuzhilina et al., 2021).

4. CP-based xLRA for heterogeneous materials

The materials formulation of xLRA targets the local elastic strain tensor field in heterogeneous media discretized on a regular grid. At each material point $r < \min(m,n)$7, the local microstructural state is encoded by descriptors $r < \min(m,n)$8, $r < \min(m,n)$9, for neighboring voxels at relative offsets $r$0. The approximation target is

$r$1

where $r$2 is the imposed macroscopic strain (Karmakar et al., 24 Sep 2025).

In tensor language, the discrete map can be regarded as an $r$3th-order response tensor $r$4, and CP-based low-rank approximation seeks

$r$5

corresponding to

$r$6

The specific xLRA extension is

$r$7

where $r$8 are chosen basis functions, given in the source as primitive hat functions for phases and generalized spherical harmonics for orientation, and $r$9 are trainable coefficients (Karmakar et al., 24 Sep 2025).

The final adaptive rank $X$0 is selected by monitoring the local relative error

$X$1

and requiring it to remain below a user-specified threshold $X$2 everywhere. This makes the extension explicitly local-error driven rather than relying on a fixed a priori rank choice (Karmakar et al., 24 Sep 2025).

Training is carried out in the Fourier domain. The first-rank regression forms the target field

$X$3

builds feature maps

$X$4

and solves the complex-valued linear regression

$X$5

for Fourier coefficients $X$6. After inverse DFT and exponentiation, the residual

$X$7

is computed. If the maximal relative error exceeds $X$8 and $X$9, the rank is incremented and a new least-squares problem is solved in Fourier space using

$M$0

The approximation is then updated by

$M$1

The stated loss is standard mean-squared error or mean-absolute error between predicted $M$2 and $M$3, solved via normal-equations linear regression. No explicit Tikhonov regularization was needed, because the Fourier representation and rank truncation were reported to suffice to control overfitting (Karmakar et al., 24 Sep 2025).

5. Empirical validation and benchmarks

The matrix-based study evaluates baseline, Nesterov, and Anderson schemes in both simulation and real-data settings. In simulation, $M$4 is generated, $M$5 is drawn at random, and the methods are compared on both the nonconvex hard-rank and convex soft-threshold problems for various $M$6 or $M$7. The relative change

$M$8

is plotted against iteration, and the reported result is that both accelerated methods reach a given accuracy in far fewer steps, with Anderson especially effective for the convex surrogate (Tuzhilina et al., 2021).

The real-data application uses a $M$9 MovieLens user–movie rating matrix with $$\min_{\substack{X\in\mathbb{R}^{m\times n}\ \operatorname{rank}(X)\le r}} \sum_{i=1}^m \sum_{j=1}^n W_{ij}\bigl(M_{ij}-X_{ij}\bigr)^2 \;=\; \min_{\operatorname{rank}(X)\le r}\;\bigl\|\sqrt{W}\ast(M-X)\bigr\|_F^2.$$0 observed entries, hence an extremely sparse $$\min_{\substack{X\in\mathbb{R}^{m\times n}\ \operatorname{rank}(X)\le r}} \sum_{i=1}^m \sum_{j=1}^n W_{ij}\bigl(M_{ij}-X_{ij}\bigr)^2 \;=\; \min_{\operatorname{rank}(X)\le r}\;\bigl\|\sqrt{W}\ast(M-X)\bigr\|_F^2.$$1. The non-SVD ALS implementation is warm-started by imputing zeros and applied to the convex-relaxed problem with $$\min_{\substack{X\in\mathbb{R}^{m\times n}\ \operatorname{rank}(X)\le r}} \sum_{i=1}^m \sum_{j=1}^n W_{ij}\bigl(M_{ij}-X_{ij}\bigr)^2 \;=\; \min_{\operatorname{rank}(X)\le r}\;\bigl\|\sqrt{W}\ast(M-X)\bigr\|_F^2.$$2. Objective decrease and actual rank are tracked versus CPU time. The reported conclusion is that Nesterov- and Anderson-accelerated ALS converge much faster than plain ALS; Anderson tends to shorten time to high precision, whereas Nesterov may be preferable for a modest accuracy (Tuzhilina et al., 2021).

The materials study evaluates xLRA on four microstructural classes: 3D two-phase composites with hard-phase volume fraction $$\min_{\substack{X\in\mathbb{R}^{m\times n}\ \operatorname{rank}(X)\le r}} \sum_{i=1}^m \sum_{j=1}^n W_{ij}\bigl(M_{ij}-X_{ij}\bigr)^2 \;=\; \min_{\operatorname{rank}(X)\le r}\;\bigl\|\sqrt{W}\ast(M-X)\bigr\|_F^2.$$3 and elastic contrast $$\min_{\substack{X\in\mathbb{R}^{m\times n}\ \operatorname{rank}(X)\le r}} \sum_{i=1}^m \sum_{j=1}^n W_{ij}\bigl(M_{ij}-X_{ij}\bigr)^2 \;=\; \min_{\operatorname{rank}(X)\le r}\;\bigl\|\sqrt{W}\ast(M-X)\bigr\|_F^2.$$4 varied from $$\min_{\substack{X\in\mathbb{R}^{m\times n}\ \operatorname{rank}(X)\le r}} \sum_{i=1}^m \sum_{j=1}^n W_{ij}\bigl(M_{ij}-X_{ij}\bigr)^2 \;=\; \min_{\operatorname{rank}(X)\le r}\;\bigl\|\sqrt{W}\ast(M-X)\bigr\|_F^2.$$5 to $$\min_{\substack{X\in\mathbb{R}^{m\times n}\ \operatorname{rank}(X)\le r}} \sum_{i=1}^m \sum_{j=1}^n W_{ij}\bigl(M_{ij}-X_{ij}\bigr)^2 \;=\; \min_{\operatorname{rank}(X)\le r}\;\bigl\|\sqrt{W}\ast(M-X)\bigr\|_F^2.$$6, totaling $$\min_{\substack{X\in\mathbb{R}^{m\times n}\ \operatorname{rank}(X)\le r}} \sum_{i=1}^m \sum_{j=1}^n W_{ij}\bigl(M_{ij}-X_{ij}\bigr)^2 \;=\; \min_{\operatorname{rank}(X)\le r}\;\bigl\|\sqrt{W}\ast(M-X)\bigr\|_F^2.$$7 instances; 2D porous media with $$\min_{\substack{X\in\mathbb{R}^{m\times n}\ \operatorname{rank}(X)\le r}} \sum_{i=1}^m \sum_{j=1}^n W_{ij}\bigl(M_{ij}-X_{ij}\bigr)^2 \;=\; \min_{\operatorname{rank}(X)\le r}\;\bigl\|\sqrt{W}\ast(M-X)\bigr\|_F^2.$$8 porosity, $$\min_{\substack{X\in\mathbb{R}^{m\times n}\ \operatorname{rank}(X)\le r}} \sum_{i=1}^m \sum_{j=1}^n W_{ij}\bigl(M_{ij}-X_{ij}\bigr)^2 \;=\; \min_{\operatorname{rank}(X)\le r}\;\bigl\|\sqrt{W}\ast(M-X)\bigr\|_F^2.$$9, and $X = U V^\top$0 instances; 3D polycrystals for face-centered cubic metals with $X = U V^\top$1 grains and Zener ratios $X = U V^\top$2, totaling $X = U V^\top$3 instances across seven materials; and dual-phase steel with $X = U V^\top$4 instances containing ferrite and martensite at $X = U V^\top$5 martensite (Karmakar et al., 24 Sep 2025). The finite-element data are generated with ABAQUS UMAT, periodic boundary conditions, a uniform mesh of $X = U V^\top$6 voxels, and tensile strain $X = U V^\top$7. Training and validation splits vary from $X = U V^\top$8 to $X = U V^\top$9 of the data, and the performance metrics are coefficient of determination $W \in [0,1]^{m \times n}$00, relative $W \in [0,1]^{m \times n}$01 error, mean absolute scaled error (MASE), CPU/GPU training time, and floating-point operations (FLOPs) (Karmakar et al., 24 Sep 2025).

The reported findings are specific. Two-phase composites at $W \in [0,1]^{m \times n}$02 reach $W \in [0,1]^{m \times n}$03 with just $W \in [0,1]^{m \times n}$04 training and $W \in [0,1]^{m \times n}$05. Porous media achieve $W \in [0,1]^{m \times n}$06 with maximum rank $W \in [0,1]^{m \times n}$07 and $W \in [0,1]^{m \times n}$08. Single-phase polycrystals for Ni require at least $W \in [0,1]^{m \times n}$09 generalized spherical harmonics bases, and using $W \in [0,1]^{m \times n}$10 training with $W \in [0,1]^{m \times n}$11 yields $W \in [0,1]^{m \times n}$12 for $W \in [0,1]^{m \times n}$13; the same study states data efficiency of $W \in [0,1]^{m \times n}$14 with only $W \in [0,1]^{m \times n}$15 of data. Dual-phase steels achieve $W \in [0,1]^{m \times n}$16 for all strain components with $W \in [0,1]^{m \times n}$17 instances and $W \in [0,1]^{m \times n}$18 (Karmakar et al., 24 Sep 2025).

Method / dataset	Reported result	Resource profile
xLRA ($W \in [0,1]^{m \times n}$19) / 2D composite ($W \in [0,1]^{m \times n}$20)	$W \in [0,1]^{m \times n}$21 with $W \in [0,1]^{m \times n}$22 data	$W \in [0,1]^{m \times n}$23 min CPU, $W \in [0,1]^{m \times n}$24 FLOPs
FNO / same	$W \in [0,1]^{m \times n}$25	$W \in [0,1]^{m \times n}$26 h GPU, $W \in [0,1]^{m \times n}$27 FLOPs
U-Net / polycrystal classes	intra-class $W \in [0,1]^{m \times n}$28, inter-class $W \in [0,1]^{m \times n}$29	$W \in [0,1]^{m \times n}$30 min GPU, $W \in [0,1]^{m \times n}$31 FLOPs

Additional benchmark statements in the source are that xLRA outperforms Fourier Neural Operators by approximately $W \in [0,1]^{m \times n}$32–$W \in [0,1]^{m \times n}$33 orders of magnitude in FLOPs and trains in minutes on CPU instead of hours on GPU; U-Net–based models suffer generalizability losses when applied cross-class, whereas xLRA remains robust; higher-order MKS of order $W \in [0,1]^{m \times n}$34 yields MASE $W \in [0,1]^{m \times n}$35 versus xLRA $W \in [0,1]^{m \times n}$36; TherINO on a 3D composite with $W \in [0,1]^{m \times n}$37 reports $W \in [0,1]^{m \times n}$38 error $W \in [0,1]^{m \times n}$39; SPiFOL on a 2D composite with $W \in [0,1]^{m \times n}$40 reports max error $W \in [0,1]^{m \times n}$41; Micrometer on a fiber composite reports RMSE $W \in [0,1]^{m \times n}$42; and rank-1 LRA on a composite with $W \in [0,1]^{m \times n}$43 yields error $W \in [0,1]^{m \times n}$44 (Karmakar et al., 24 Sep 2025).

6. Interpretation, limitations, and research context

The common principle across the two formulations is that low-rank structure is retained while the approximation is extended in a way that matches the underlying data geometry. In the weighted matrix case, the extension is entrywise weighting plus accelerated fixed-point or projected-gradient solution. In the materials case, the extension is an adaptive CP expansion with basis-dependent factors and local error control. This suggests that xLRA is best interpreted as an adaptive low-rank methodology rather than a single canonical model class (Tuzhilina et al., 2021, Karmakar et al., 24 Sep 2025).

Several limitations are explicit. For the nonconvex rank-constrained matrix problem, convergence guarantees reach only stationary points, not global optima (Tuzhilina et al., 2021). In the materials formulation, the method requires FE or spectral pre-computed local fields for training and is therefore not strictly unsupervised; the current formulation is limited to linear elasticity, with extension to plasticity described as under active development; and basis selection must match descriptor type, specifically primitive hat functions for phases and generalized spherical harmonics for orientation (Karmakar et al., 24 Sep 2025).

A second misconception is that xLRA is simply a rank-1 surrogate. The materials study states the opposite: rank-1 LRA often fails for elastically heterogeneous materials with strong contrasts, and the reported porous-media study requires a maximum rank of only $W \in [0,1]^{m \times n}$45, while dual-phase steels use $W \in [0,1]^{m \times n}$46 (Karmakar et al., 24 Sep 2025). Likewise, in the matrix setting, the availability of nuclear-norm relaxation, Nesterov acceleration, Anderson acceleration, and non-SVD ALS shows that the framework is not confined to a single low-rank solver architecture (Tuzhilina et al., 2021).

The applications described in the sources are correspondingly broad. Weighted low-rank matrix approximation is presented as an essential component of GLM optimization algorithms when an exponential family models matrix entries and the matrix of natural parameters admits a low-rank structure (Tuzhilina et al., 2021). The materials xLRA is positioned for rapid surrogates in microstructure-aware material design and optimization, integration into multiscale frameworks in place of costly FE inner loops, real-time prediction of localization hotspots for damage initiation studies, and uncertainty quantification when coupled with stochastic microstructure generators (Karmakar et al., 24 Sep 2025). Across both settings, the defining feature is compact approximation with explicit control over rank, weighting, and computational cost.