Does L1 mean the direction has L1 length one?

Not in the common formulation here: the projection objective uses L1, while the direction has Euclidean unit length.

Is L1-PCA guaranteed to remove outliers?

No. It changes sensitivity; it is not an automatic outlier-removal procedure.

Are all L1-PCA formulations equivalent?

No. Projection and reconstruction objectives can give different results.

AI Glossary / Data science and analytics / Advanced

L1-norm principal component analysis

Definition

L1-norm principal component analysis refers to PCA-related methods that use absolute-value criteria instead of the squared-value criteria associated with ordinary PCA. A common formulation selects unit-length projection directions that maximize the sum of absolute projected data values. Other methods minimize absolute reconstruction errors, and these formulations are not generally equivalent. The precise objective must therefore accompany the name.

What it means

Ordinary PCA finds directions that capture large squared variation in centered data. Squaring gives extreme values substantial influence. An absolute-value objective changes that weighting and can reduce sensitivity to some outlying observations. It does not make the analysis immune to contamination or remove the need to decide how data are centered and scaled. The components are still mathematical summaries whose practical meaning requires interpretation and validation.

How it works

In the projection-maximization version, start with a numerical data matrix and a specified centering and scaling procedure. For one component, search for a direction w with Euclidean length one that maximizes the sum of absolute dot products between w and the data vectors. The unit-length condition prevents the score from increasing simply by enlarging w. Multi-component versions impose appropriate orthonormality constraints. Unlike conventional PCA, the answer is not generally obtained by simply taking covariance eigenvectors. Algorithms differ in whether they find a local solution, an approximation, or a global optimum under particular conditions.

Three practical examples

Sensor analysis: researchers explore lower-dimensional representations when occasional large measurement errors may distort ordinary PCA.
Image analysis: a team compares absolute-value and squared-value component methods under a specified corruption scenario.
Signal processing: analysts investigate subspace estimates for data containing atypical observations, while measuring the effect on the downstream task.

Think of it this way

Think of comparing two scoring rules for deviations: one counts a deviation in direct proportion to its size, while the other squares it. A very large deviation receives disproportionately more weight under the second rule. This explains a motivation for L1 methods, but it does not establish that every unusual observation is erroneous or that an absolute-value method always finds the desired structure.

Worked example

Consider six already-centered two-dimensional observations: two copies of (2, 0), two copies of (-2, 0), one (0, 3), and one (0, -3). Along the horizontal unit direction, the absolute projections sum to 8, while squared projections sum to 16. Along the vertical direction, absolute projections sum to 6 and squared projections sum to 18. The two objectives therefore prefer different axes when comparing only those axes.
The L1 search need not select an axis. For the unit direction (0.8, 0.6), the four horizontal observations project to values of magnitude 1.6 and the two vertical observations to values of magnitude 1.8. Their absolute sum is 4 times 1.6 + 2 times 1.8 = 10. More generally the L1 score is 8|w1| + 6|w2| under w1 squared plus w2 squared equal to one, whose maximum is 10. The squared-projection objective instead equals 16w1 squared + 18w2 squared, maximized at a vertical direction. This constructed example shows different optimization criteria; it does not represent an outlier-detection experiment.

Illustrative case study

In a fictional inspection project, a few sensor spikes visibly influence an ordinary PCA display. A team compares an L1 projection method with the existing approach. Before concluding that one is superior, it checks whether the spikes are errors or important rare events, documents centering, and evaluates both representations on held-out inspection tasks. It also records which optimization algorithm was used and whether repeated initializations produce different solutions. The case illustrates an evaluation process rather than claiming automatic robustness or improved defect detection.

Common misconceptions and mistakes

L1-PCA is not automatically sparse PCA. Using an absolute-value objective on projections differs from penalizing component coefficients to encourage zeros. It is also not identical to every method called robust PCA, some of which decompose a matrix into low-rank and sparse parts. Absolute projection maximization and absolute reconstruction minimization must not be treated as interchangeable. A local optimum is not a global guarantee, and a component sign change may represent the same direction for interpretation.

Business use

Consider this family when a project has a concrete reason to question squared-error sensitivity. State the exact objective and implementation, the preprocessing choices, and the criterion used to judge usefulness. Compare with ordinary PCA and other appropriate alternatives on realistic data issues, not only a favorable synthetic example. Assess computational cost and stability as well as visualization quality. For decision-making, explain what the projected values retain and what they discard, and avoid interpreting a component as a causal factor merely because it summarizes several features.

Check your understanding

Why is it incomplete to compare only horizontal and vertical directions in the worked example?

Answer: The optimization permits other unit directions. The direction (0.8, 0.6) achieves an absolute-projection sum of 10, exceeding either axis's score.

Related terms

Learn next

Review dot products and ordinary PCA before tracing the absolute-projection objective on a small dataset. Compare it explicitly with an absolute reconstruction-error objective. Then inspect an implementation's convergence guarantee and examine how centering, scaling, and atypical observations influence the result.

Sources

Nojun Kwak, original PCA-L1 paper abstract: https://pubmed.ncbi.nlm.nih.gov/18617723/ Markopoulos, Karystinos, and Pados, original formulation and distinctions: https://arxiv.org/pdf/1309.1194 pcaL1 authors, PCA-L1 implementation documentation: https://search.r-project.org/CRAN/refmans/pcaL1/html/pcal1.html

Previous: L1-based feature selection Next: Label