These corrections accompany the November 2026 implementation and book review. The numbered R vignettes remain the canonical curriculum. Positive-degree partial moments retain their directional severity meaning; event probabilities, uncertainty, and predictive performance require their corresponding definitions and evidence.
| Topic | Correct interpretation or revised procedure |
|---|---|
| Multivariate fitted contrasts | Describe conditional associations and heterogeneity. Resample sampling units and refit the model to include fitting uncertainty; do not compare a fixed-fit contrast bootstrap to an OLS coefficient interval as if they had the same target. The editable workflow now refits. |
| Time-series interval comparison | Tune on past data and score an explicit future horizon. Match forecast origins and available lags. The invalid historical ranking is withdrawn; the corrected report links executable source. Three interval summary points do not identify a native predictive density. |
| MNIST and machine-learning claims | The supplied PDF's NNS score is one reported run; it does not include a matched kNN fit establishing a 75% error reduction. Those comparative and general dimensionality claims are withdrawn pending common splits, preprocessing, tuning, metrics, and repeated runs. |
| Five-way ranking | Normalized degree-one UPMs are shares of above-target deviation severity. They are not probabilities of winning a race. Sampling winners using those same shares merely reproduces the supplied weights; it does not validate their event-frequency interpretation. |
| Partitional estimation | For rnorm(n) + runif(n), mean noise is 0.5. The
conditional mean is f(x) + 0.5, or use
runif(n) - 0.5 to target f(x). Compare all
fitted methods on a common held-out set; random-forest OOB error and NNS
training error are different assessments. |
| Dependence permutation output | Shuffled-statistic quantiles are a null reference distribution under exchangeability, not a confidence interval for population dependence. Specify dependence-preserving resampling for serial or clustered data. |
| Independent 3-D copula output | Page 9 of
Nonlinear_Correlation_and_Dependence_Using_NNS.pdf prints
0.09571785 for
set.seed(123); x <- rnorm(1000); y <- rnorm(1000); z <- rnorm(1000); NNS.copula(cbind(x,y,z)).
The reviewed implementation returns 0.197845241162; the
revised companion recomputes it. This descriptive finite-sample score is
not a p-value or an independence threshold. The population event mass
2 * (1/2)^3 at independent marginal medians is a different
quantity. |
| Copula sampling | Degree-zero marginal transforms or an explicitly randomized probability transform at atoms target frequency. Positive-degree inverses change the marginal law; their directional geometry does not imply preservation of the original marginal distribution. |
| Iris classification | Rows 141:150 are all virginica. Vignette08 now uses a seeded stratified split, removes historical perfect-accuracy output, and makes class decoding explicit. |
| Derivative summaries | Quantiles across fitted derivatives describe heterogeneity over evaluation locations. Sampling uncertainty requires resampling and refitting with a design appropriate to the observations. Smoothness and admissible local steps still matter. |
| Normalization | Positive constant multiplication of columns preserves Pearson correlation; a negative multiplier reverses its sign. Changing portfolio weights is a separate operation. Check scaling-factor signs before interpreting transformed associations. |
| Forecast period selection | Periods selected using a test window make that window validation. Vignette09 now selects periods within the training prefix, reserves the final observations for testing, and computes current outputs rather than printing stale values. |
| Noisy differentiation | Adding frozen real noise to a complex-valued evaluation leaves the imaginary part unchanged. Tiny complex-step error in that construction is not evidence of generic noise robustness. A reproducible noise generator, complex-input contract, and repeated evaluation protocol are required. |
Issue #72 remains open for the six archived HTML pages below. The documentation workflow copies their existing content and adds a visible historical-output notice to the staged copies; it does not rerun their analyses. The original HTML files remain unchanged. Neither a successful site build nor the revised PDF companions should be read as regeneration of these outputs.
For each page, add editable source, data provenance, package versions, and a repeatable build command. Retain only the current results or explicitly withdraw unsupported historical comparisons. The exact book text corrections, the independent three-variable copula erratum, and current executable vignette sources are addressed separately; they do not close this remaining checklist.
The twelve published PDFs are retained as historical documents. Matching original editable sources were not present in the audited repository. The new revised-companions.Rmd is an explicitly revised, executable companion, not a reconstruction of the original authors' source. It provides corrected demonstrations and clear limits for each PDF below.
| Historical PDF | Revised companion section |
|---|---|
Bayes' Theorem From Partial Moments.pdf |
Conditional event probabilities |
Continuous_CDFs_and_ANOVA_with_NNS.pdf |
Frequency, severity, and comparison targets |
Nonlinear_Correlation_and_Dependence_Using_NNS.pdf |
Directional dependence and null calibration |
Normalization.pdf |
Positive scaling and association |
5_way_rank_probabilities_using_R.pdf |
Severity shares and winner frequencies |
Partitional_Estimation_Using_Partial_Moments.pdf |
A correctly centered regression target |
Machine_Learning.pdf |
Held-out classification |
NNS vs KNN MNIST dataset.pdf |
A reproducible MNIST comparison contract |
Spiral Matching Example.pdf |
Geometric matching |
Nonlinear_FWL.pdf |
Residualization and conditional targets |
Causal_Inference_with_NNS_stack.pdf |
Predictive contrasts and identification |
NNS ARMA.pdf |
Training, validation, and test chronology |
The rendered companion is rebuilt by the documentation workflow. Build it locally with:
Rscript examples/build-revised-companions.RThe build records sessionInfo() and the NNS version.
Original source recovery remains a provenance task: when those sources
are available, add them with their assets and build instructions rather
than representing revised material as an exact source restoration.
Historical numerical tables are not carried forward as current
validation results.