Recommended Free Tools
PCA ranks directions in your data by how much variance they explain. When you keep fewer components, the reduced data no longer represents variation along the directions you leave out. Those are the lowest-variance directions among the components PCA found—but “low variance” does not mean “useless” for every analysis.
What PCA removes when you keep fewer components
Principal component analysis (PCA) finds orthogonal directions through multivariate data, ordering them by explained variance. The first component captures the greatest variance possible along one direction; each successive component captures the greatest remaining variance subject to being orthogonal to the earlier ones. Scikit-learn’s decomposition guide and PCA API documentation describe this behavior.
As an Amazon Associate I earn from qualifying purchases.
Keeping only the first few components projects observations into the subspace they span. Variation represented by the omitted directions is no longer present in that lower-dimensional representation. In that sense, PCA discards the quieter directions first: “quiet” means lower variance relative to the retained components, not necessarily random noise.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Does PCA always remove the least useful information?
No. PCA optimizes variance representation, not relevance to a particular prediction, diagnosis, or scientific question. A direction with little overall variation can still matter to a downstream goal; its usefulness depends on the task and evidence, not its position in PCA’s variance ranking. Conversely, a high-variance direction is not automatically the most useful one for every task.
#1 Best Overall
So a component cutoff answers a specific question: how many orthogonal directions will represent the data? It does not, by itself, establish that everything omitted is irrelevant. Evaluate the reduced representation against the goal for which you intend to use it.
Choosing how many components to retain in scikit-learn
In scikit-learn, PCA learns its directions during fit and projects observations onto them during transform. The n_components setting controls how many directions are retained. The API also documents a variance-threshold option: with the full solver, a fractional threshold selects enough components to exceed the requested explained-variance percentage. See the API documentation for the current parameter behavior.
- Fixed component count: Choose a set number of directions when the target dimensionality is part of your workflow. The representation has that many dimensions, and variation outside their subspace is omitted.
- Variance threshold: Choose a target share of explained variance when you want the number of retained components to follow that criterion. This preserves a chosen amount of variance, not necessarily a chosen amount of task-relevant information.
Neither choice makes the cutoff universally correct. The first fixes dimensionality; the second uses variance as its selection rule. The downstream objective still determines whether the resulting representation is suitable.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesWhy feature scaling changes what PCA sees
Scikit-learn’s PCA centers input features before applying singular value decomposition (SVD), but it does not scale each feature to unit variance. As a result, features with larger scales can contribute more to the variance PCA ranks. The scikit-learn preprocessing guide explains feature standardization and its relevance to estimators.
If features are measured in unlike units or have substantially different scales, decide whether to standardize them before PCA. That preprocessing changes the variation the analysis sees, so it can change the component directions and their ranking. Centering alone and standardizing are distinct operations; scikit-learn PCA performs the former, not the latter.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Whitening is different from dropping components
Whitening changes the scale of retained component scores: in scikit-learn, it scales outputs to unit component-wise variance. The API notes that this removes the relative variance scales of the components and can sometimes help downstream estimators with particular assumptions. The PCA API reference documents the option.
Truncation and whitening therefore do different things. Truncation omits selected directions from the representation; whitening rescales the retained component outputs. Whitening is not another name for discarding low-variance components.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Quick Recap
A practical way to interpret a PCA cutoff
- Set the purpose. Identify what the reduced data will be used for; variance preservation alone cannot answer whether the representation serves that purpose.
- Choose preprocessing deliberately. Decide whether centering alone is appropriate or whether feature standardization is needed for the scales and units in your data.
- Select a retention rule. Use a fixed component count for a predetermined dimensionality, or a variance threshold when the target is a share of explained variance.
- Consider whitening separately. Use it only if unit-variance component outputs fit the downstream method’s assumptions, recognizing that relative variance scales are removed.
- Assess the result against the goal. Treat the variance ranking as PCA’s compression criterion, not as proof that omitted directions contain no useful task-specific information.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
