schemaxisNumpy operation reference
Generated from the current declarations. Do not edit this file by hand.
Types, limits and defaults below describe resolved values. Runtime expressions must produce these types. Unknown options are rejected; defaults apply only when omitted. Constructor options must be constant after document composition.
Index#
principalComponents#
Find the dominant correlated patterns in captured multi-channel measurements. The route centres each channel at its mean, scales it as declared, and computes principal components by singular value decomposition in isolated NumPy. It publishes the spectrum with explained variance, labelled loadings and scores (or a retained-subspace projector) and a reconstruction in the channels' own units, each independently checked by a standard-library verifier.
Common mistakes and assumptions:
- Repeated or near-repeated eigenvalues never publish an arbitrary per-component basis; use subspace output.
- Accuracy is relative to the total variance, not to each small eigenvalue; no statistical significance or sampling-uncertainty claim is made.
- Requirements are decided by the host verifier; unmet or indeterminate results are retained evidence, not accepted deliverables.
Native binding: numpy.linalg.svd(full_matrices=False) on the centred, scaled data (LAPACK gesdd). Output kind: ["table","text"].
Inputs: none.
| Option | Meaning | Type / constraints | Default |
|---|---|---|---|
source |
Required absolute path to a regular ASCII .csv file of at most 8 MiB. The header must equal the optional row key followed by the channels, in order. Cells are plain decimal numbers; quotes, blank lines, missing values, nan/inf, hex, digit separators and whitespace are refused with their line and column. The bytes are captured for standalone verification, and changed bytes invalidate workflow reuse. | {"type":"string","minLength":1,"maxLength":4096} |
required |
rowKey |
Optional numeric column that identifies rows, such as time or sample number. It must increase strictly, is excluded from the analysis and is carried into scores and reconstruction. Omit it to number rows 1..n in file order. | {"type":"object","properties":{"name":{"type":"string","minLength":1,"maxLength":40,"required":true},"unit":{"type":"string","minLength":1,"maxLength":32,"required":true}}} |
omitted |
channels |
The 2-32 analysed channels in CSV header order (after the optional row key), at most 160000 analysed cells and 10000 rows. Every listed channel is analysed; there are no ignored columns. | {"type":"array","items":{"type":"object","properties":{"name":{"type":"string","minLength":1,"maxLength":40,"required":true},"unit":{"type":"string","minLength":1,"maxLength":32,"required":true},"scale":{"type":"number","minimum":0,"exclusiveMinimum":true}}},"minItems":2,"maxItems":32} |
required |
scaling |
How channels are scaled after centring at their mean. none: covariance PCA in the shared unit (eigenvalues in unit squared). standardise: divide by each n-1 sample deviation, giving correlation PCA. declared: divide by each channel's declared scale. Choose standardise when channels have different units or very different spreads you do not want to dominate. | {"type":"enum","values":["none","standardise","declared"]} |
required |
retention |
How many components to keep: exactly one of components or minimumExplainedVariance. Kept components must be numerically resolved from the rest of the spectrum. | {"type":"object","properties":{"components":{"type":"integer","minimum":1,"maximum":10},"minimumExplainedVariance":{"type":"number","minimum":0,"maximum":1,"exclusiveMinimum":true,"exclusiveMaximum":true}}} |
required |
output |
components publishes per-component loadings and scores and requires every kept component to be separated from its neighbours. subspace publishes only the projector onto the kept subspace and its reconstruction, so repeated eigenvalues inside the kept set are allowed. | {"type":"enum","values":["components","subspace"]} |
"components" |
loadingTolerance |
Validated accuracy eta in [1e-9, 1e-2] (default 1e-6): the sine of the angle between each published loading (or the projector) and the exact one, bounded a posteriori in each run. After sign alignment the loading 2-norm distance is at most sqrt(2)*eta. It also sets the resolution threshold and each channel's reconstruction certification (uncertainty at most eta times the channel deviation). | {"type":"number","minimum":1e-09,"maximum":0.01} |
1e-06 |
requirements |
Optional acceptance requirements decided by the independent host verifier as met, unmet or indeterminate using an effective tolerance published in pca-verification.json. Only met requirements give an accepted run; unmet and indeterminate runs keep their evidence but fail. | {"type":"object","properties":{"minimumExplainedVariance":{"type":"number","minimum":0,"maximum":1,"exclusiveMinimum":true,"exclusiveMaximum":true},"reconstructionRMS":{"type":"array","items":{"type":"object","properties":{"channel":{"type":"string","minLength":1,"maxLength":40,"required":true},"maximum":{"type":"number","minimum":0,"exclusiveMinimum":true,"maximum":1e+100,"required":true}}},"maxItems":32}}} |
omitted |