MATHEMATICS
Rank or tolerance? What librla changes about matrix compression
The 2026 librla preprint puts rank and tolerance choices into a common factorization library. Understanding those choices takes more than counting stored values.
Research checked
Newest cited publication
A matrix illustration provides context for the distinction between a requested approximation rank and a requested error tolerance.
Maschen / Wikimedia Commons. Wikimedia-generated PNG rendering of the SVG; no editorial changes. · Image source · Public domain (released by the creator)
Matrix compression is often presented as a request for fewer numbers. The more important question is which information those numbers must preserve. A scientist may know the storage budget before knowing the acceptable error, or may know the acceptable error before knowing how much storage it requires. Those are different numerical problems. Adrianna Gillman and Zydrunas Gimbutas address both in librla, described in a July 22, 2026 preprint revised on August 4. The library offers randomized low rank factorizations in Python, Julia, and MATLAB. Its practical choices provide a useful occasion to examine what rank, tolerance, and validation actually mean when a large matrix is replaced by a smaller representation.
Two requests that can produce different answers
Consider a diagonal matrix whose entries are one hundred, ten, one, and one tenth. Keeping the largest two singular directions gives a rank two approximation. In the spectral norm, which measures the largest amplification error on a unit vector, the discarded part has norm one. If the request is instead a relative spectral error of one thousandth, the allowed absolute error is one tenth because the original norm is one hundred. A rank three approximation reaches that threshold in this exact example. A storage request and an accuracy request have led to different ranks.
The same approximation has a different error in the Frobenius norm, which combines squared entry errors. Discarding one and one tenth gives a Frobenius error of the square root of 1.01, approximately 1.005. Neither number is wrong. They answer different questions. Before interpreting a tolerance field, a reader must establish the norm, whether the tolerance is absolute or relative, and how the implementation estimates it. A short decimal value without those details is an incomplete accuracy specification.
Random samples identify an important space
A matrix maps input directions to output directions. If most of its action concentrates near a small output space, one can seek a basis Q for that space and approximate A by Q(QᵀA) in the real case. The large matrix is then represented through two thinner factors. Random sampling explores this action by applying A to several random input vectors. Directions that the matrix amplifies strongly tend to appear in the resulting samples. The older framework developed by Nathan Halko, Per-Gunnar Martinsson, and Joel Tropp supplies the foundational analysis behind this general strategy.
One sample is easy to misinterpret. For A = diag(100, 1), multiplying by the vector (1, 0) identifies only the first axis and completely misses the second. A random vector is less deliberately unfortunate, but a guarantee still needs a probability statement and assumptions about the sample distribution. Extra samples provide room to capture directions that a minimal sketch might miss. They also increase work. Randomization replaces an exhaustive search with a carefully analyzed experiment; it does not make the approximation error disappear.
The reported 2026 software scope
The librla paper offers QR, singular value, and interpolative factorizations with a rank or tolerance mode. Its stated target is matrices of intermediate size, roughly up to dimension ten thousand depending on resources. The paper explicitly distinguishes this scope from matrices too large to fit in memory. It also supports operator based access when both the matrix and its conjugate transpose can be applied. These are the reported capabilities of a particular research library, not a claim that all matrix compression problems now share one optimal implementation.
The official repository provides the implementation and usage material through which those claims can be checked. Code matters because a mathematical description leaves practical questions unresolved: what shapes are accepted, what defaults are selected, what is returned, and how a failure is signaled. For an actual computational study, the repository revision and dependencies should be pinned. A changing default can alter the sample count or stopping behavior even when an application continues calling the same public function name.
Count the representation before promising savings
A dense ten thousand by ten thousand real matrix contains one hundred million entries. At eight bytes per entry, its values occupy eight hundred million bytes, or eight hundred decimal megabytes, before object overhead. A rank one hundred singular value representation stores two ten thousand by one hundred factors plus one hundred singular values. Those 2,000,100 values require about sixteen million bytes. This arithmetic suggests a large reduction in the final representation when the approximation is suitable.
It does not predict peak memory. Constructing the factors may require the original matrix, random samples, temporary products, and work arrays at the same time. Nor does it predict speed: a factorized application can be cheaper while the one time construction is expensive. If a matrix will be applied once, preparation may dominate. If it will be applied a million times, preparation can be amortized. These are independent accounting examples, not measured librla benchmarks, and they explain why a compression ratio should be reported alongside construction and usage costs.
Two ways to improve a sketch
Oversampling and power iteration spend additional resources in different ways. Oversampling asks for more candidate directions than the final rank requires. Power iteration repeatedly emphasizes dominant singular directions before the small factorization is completed. If two singular values are ten and nine, their ratio is only about 1.11. Applying the operator AAᵀA transforms their magnitudes to one thousand and seven hundred twenty nine, increasing the ratio to about 1.37. This small calculation illustrates spectral separation; actual algorithms also need orthogonalization to prevent numerical collapse.
The benefit depends on the spectrum. If a matrix has a sharp gap after the desired rank, a modest sketch may suffice. If many singular values are similar, more work may be needed and the compressed representation may deliver little benefit. The librla paper exposes oversampling and power iteration choices and describes a tolerance stopping criterion based on diagonal information from the sampled QR factorization. Such an indicator should be understood through the algorithm and its assumptions, rather than silently reinterpreted as an exact deterministic certificate for every global error norm.
Why keeping actual columns can matter
A singular value decomposition creates abstract directions, usually combining all original columns. An interpolative decomposition instead represents the matrix using selected columns and coefficients that reconstruct the remaining columns approximately. Suppose columns correspond to measured spectra. Retaining actual columns can make the compressed representation easier to inspect because each retained spectrum came from a real measurement. That interpretability is useful, but it does not mean the chosen columns reveal a unique physical mechanism or a causal explanation.
The quality of the reconstruction coefficients also matters. If two selected columns are almost parallel, fitting another column may require large coefficients with opposing signs. Small measurement noise can then become a large reconstruction change. A useful assessment therefore considers the residual and the conditioning of the chosen representation. This issue is visible even in a two vector example and does not depend on a fashionable application. Different factorizations solve related storage problems while preserving different structural advantages.
Validate the question that the matrix will answer
An approximation can have a small overall error and still be unsuitable for a particular task. If the scientific signal lies in a weak singular direction, a compression optimized to preserve dominant energy may remove it. A sensible evaluation includes representative downstream operations alongside norm estimates. For a solver, check residuals in the original equations. For repeated matrix applications, compare selected outputs against the original operator. For a data product, preserve a route to the uncompressed source and record the numerical settings used to derive the compact version.
As of October 18, 2026, librla is an openly available research implementation described by a current preprint. The strongest general lesson is to make the optimization target explicit. Fixed rank asks what can be retained within a budget. Tolerance asks what budget is needed for a specified numerical goal. Random sampling makes both approachable, while spectra, conditioning, and application requirements determine whether the result is useful. The smaller representation earns trust through those checks, rather than through its smallness alone.
Sources & further reading
Primary papers, project documentation, and research releases used in this article. Publication dates below refer to the sources themselves.
- Algorithm librla: A library of randomized linear algebra routinesarXivPublished Checked
- librla: Randomized Linear Algebra LibraryGillman and Gimbutas official repositoryChecked
- Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositionsarXivPublished Checked
FOLLOW THE QUESTION
Explore mathematics