Riemann distinguishes validated input representations, available geometry primitives, algorithm contracts, and statistical calibration. A wrapped point or available distance does not establish every algorithm's assumptions.
Details
The installed file method-contracts.csv contains one disposition
for every exported function and registered S3 method, including its domain,
assumptions, restrictions, and evidence scope. Read it with the example below.
The four possible dispositions are:
- core
The consolidated input, geometry, summary, clustering, tangent PCA, or scalar regression contract, within its documented capabilities.
- validated_restricted
A specific legacy contract checked against independent formulas or reference fixtures; the stated restrictions apply.
- experimental
A retained legacy interface whose full numerical, mathematical, or statistical contract has not been independently verified.
- disabled
An unavailable interface with a documented incompatible implementation. Current geometry restrictions can disable combinations without disabling an entire exported function.
A test-file reference is an evidence pointer, not branch coverage or a proof. No disposition guarantees a global optimum, an identifiable model, finite sample calibration, or validity for every manifold. Experimental methods are retained for compatibility and exploration; their source references and examples should not be read as completed validation.
Common distinctions are particularly relevant: nonnegative graph affinities or regression weights need not form a positive-semidefinite kernel; a metric dissimilarity need not be Euclidean; a local tangent representation need not solve exact nonlinear principal geodesic analysis; and a permutation test requires exchangeability under its specified null. Rayleigh/Bingham and Frechet mean/variance tests need not detect all distributional alternatives.
riem.capabilities() records primitive availability and geometry status.
The table here separately records method status. The package maintenance
inventory records source and dependency ownership; the release/check logs
supply actual execution evidence for each environment.
Examples
contracts <- read.csv(system.file("method-contracts.csv", package = "Riemann"))
contracts[contracts$name %in% c("riem.kpca", "riem.fanova"),
c("name", "disposition", "limitations")]
#> name disposition
#> 44 riem.fanova validated_restricted
#> 56 riem.kpca validated_restricted
#> limitations
#> 44 Not an omnibus all-distribution test; paired/clustered data require another resampling design. Empirical nondegeneracy and solver convergence do not prove population assumptions.
#> 56 Empirical negative spectrum is reported, warns and is explicitly truncated or rejected. Truncation is an approximation; an empirically PSD sample does not prove global positive definiteness.