Chapter Four · failure evidence
What Dimensionality Reduction & PCA got wrong, from 75 dissertations
The records evaluate dimensionality reduction and principal component analysis across diverse machine learning and scientific domains. These methods frequently fail due to information loss, non-linear manifold structures, noise sensitivity, and degradation of interpretability compared to unreduced baselines. These records come from PhD theses at 25 institutions, 2021 to 2026. Each links to its thesis. They were extracted by language models reading the full text, so treat each as a lead to read, not a verdict.
Linear projections fail to capture complex non-linear structures and manifold geometries
Linear transformations failed to resolve clusters, phenotypes, and continuous transition states across biological, physical, and engineering datasets. These methods captured insufficient variance and produced excessively large latent representations compared to non-linear alternatives.
Tried and failed
PCA dimensionality reduction applied to single-cell expression of disease states. Outcome: no signal. Reason: linear principal components failed to separate healthy from diseased phenotypes
Tried and failed
principal component analysis for dimensionality reduction applied to patient survival subgroup stratification. Outcome: worse than baseline. Reason: linear projection failed to capture complex non-linear patterns compared to autoencoders
Integrated systems approach for mechanistic understanding of human cancers · UT Austin
Tried and failed
traditional dimensionality reduction applied to single-cell RNA sequencing data visualization. Outcome: no signal. Reason: failed to preserve local density information and clustering structure in high-dimensional biological data
Similarity Metrics for Biological Data: Algorithmic developments for high-dimensional datasets · MIT
Considered and rejected
Considered and rejected: Standard PCA for dimensionality reduction of multi-omics data, rejected due to failure to cluster survival-distinguishable glioma subtypes compared to deep autoencoders.
Integrated systems approach for mechanistic understanding of human cancers · UT Austin
Considered and rejected
Considered and rejected: Rejected Principal Component Analysis (PCA) for dimensionality reduction of power flow observations because it performs linear mapping, cannot represent complex polynomial relationships, and fails to preserve local cluster structure unlike T-SNE
Operational moving target defences for improved power system cyber-physical security · Imperial
Tried and failed
principal component analysis for dimensionality reduction applied to high-dimensional molecular descriptor space. Outcome: no signal. Reason: linear projection failed to capture complex nonlinear variance, yielding poor separation on early components
Tried and failed
principal component analysis on raw diffraction patterns applied to polycrystalline grain segmentation. Outcome: no signal. Reason: dynamical diffraction effects introduced strong nonlinearities that PCA could not separate
Transforming 4D-STEM Data: From Diffraction to Cepstra & Aberrations to Brightness · Cornell
Tried and failed
2D principal component analysis projection applied to spectroscopic tissue classification. Outcome: no signal. Reason: First two principal components captured insufficient variance to separate classes linearly.
Characterization of cervid skin tissues with chronic wasting disease by Raman spectroscopy and machine learning · Iowa State
Tried and failed
standard dimensionality reduction and manifold learning applied to phase-change material optical transmission spectra. Outcome: no signal. Reason: linear and non-linear unsupervised methods failed to separate continuous intermediate crystallization states in 2D latent space
Machine Learning for Knowledge Discovery in Engineering and Science: From Nanophotonic Design to Medical Diagnosis · Georgia Tech
Tried and failed
principal component analysis for 2D clustering applied to high-dimensional audio feature vectors. Reason: Non-linear relationships across features prevented distinct cluster separation in linear projection
Real-Time Sound Visualization · Texas Tech
Tried and failed
principal component analysis for trajectory clustering applied to intrinsically disordered protein conformational ensembles. Outcome: no signal. Reason: linear components failed to resolve distinct clusters and captured under half of trajectory variance
Considered and rejected
Considered and rejected: Rejected PCA (Principal Component Analysis) for feature selection due to limitations with nonlinear correlation structures
Artificial intelligence driven anomaly detection for big data systems · Imperial
Considered and rejected
Considered and rejected: Linear dimensionality reduction (SVD/PCA) for mobility time-series due to failure to capture intrinsic nonlinear manifold geometry
Applications of Optimization and Machine Learning to Healthcare · ResearchWorks
Considered and rejected
Considered and rejected: Rejected linear POD/SVD dimensionality reduction for full workflow due to inability to capture nonlinear spatial features across varying parameter sets.
Enabling computer-aided food engineering: mechanistic models and their deep learning surrogates · Cornell
Lost to a baseline
Linear PCA dimensionality reduction on agent-based stochastic data found a latent space size of n - 1 (25x larger than non-linear DR), resulting in poor reducibility compared to non-linear methods.
Reduced Order Non-INtrusive (RONIN) Modeling for Strategic Defense Planning · Georgia Tech
Considered and rejected
Considered and rejected: Discarded Cartesian and dihedral PCA as dimensionality reduction techniques because of poor explained variance and unclusterable landscapes.
Considered and rejected
Considered and rejected: Rejected Proper Orthogonal Decomposition (POD) / Principal Component Analysis (PCA) for reduced-order modeling of cislunar tracking uncertainty due to poor handling of non-linear behavior and scalability limitations.
Uncertainty-Based Methodology for the Development of Space Domain Awareness Architectures in Three-Body Regimes · Georgia Tech
Unsupervised variance maximization discards critical low-variance discriminative signals
Relying on overall variance caused models to discard subtle components and directions essential for distinguishing classes, faults, and target phenotypes. Because unsupervised components ignore target labels, critical predictive information was dropped prior to downstream classification.
Tried and failed
PCA dimensionality reduction before one-class SVM applied to anomaly detection from sensor data. Outcome: worse than baseline. Reason: Nominal-only PCA bases discarded variance directions critical for distinguishing faulty signals
Applications of Probabilistic Machine Learning Models to Semiconductor Fabrication · MIT
Tried and failed
PCA preprocessing before independent component analysis applied to sparse low-signal genomic data. Outcome: no signal. Reason: signal was discarded during the PCA dimension reduction step prior to ICA
Any-way and Sparse Analyses for Multimodal Fusion and Imaging Genomics · Georgia Tech
Tried and failed
PCA dimensionality reduction for feature selection applied to brain connectivity network classification. Outcome: worse than baseline. Reason: Unsupervised variance maximization discarded domain-specific discriminative connectivity features captured by hypothesis-driven network-based statistics
Tried and failed
PCA dimensionality reduction before neural network training applied to multivariate sensor feature classification. Outcome: worse than baseline. Reason: low-variance components retained critical discriminatory information for the target task
Wet gas flow metering with pattern recognition techniques · Cranfield
Tried and failed
principal component regression applied to phenotype prediction from transcriptomic data. Outcome: worse than baseline. Reason: unsupervised variance maximization discarded components predictive of the target compared to partial least squares
Tried and failed
PCA dimensionality reduction before linear SVM classification applied to spectroscopic image classification. Outcome: worse than baseline. Reason: Discarded components contained variance relevant to class separation
Advancing Biological BCARS Imaging: Simulation-Based Optimization, and Machine Learning Analysis · Georgia Tech
Tried and failed
PCA dimensionality reduction before linear discriminant analysis applied to mass spectrometry profiles. Outcome: did not generalise. Reason: PCA discarded low-intensity features that contained critical discriminative signal
Validating the ability of REIMS to differentiate lamb flavor performance based on consumer preference · Texas Tech
Considered and rejected
Considered and rejected: Data reconstruction by removing principal components (e.g. PC18) was rejected because discarding components causes irrecoverable test information loss
Coping with the distribution change in soil classification with LIBS · oURspace
Considered and rejected
Considered and rejected: Rejected using Principal Component Analysis prior to logistic regression (PC-LR) because PCA ignored class labels and destroyed group structure
Statistical Analysis of Mid-Infrared Spectroscopy Data to support Dairy Cow Management · Research Repository UCD
Tried and failed
PCA dimensionality reduction before supervised classification applied to spectroscopy classification of mixed polymers. Outcome: worse than baseline. Reason: High variance within heterogeneous classes caused loss of discriminative spectral variance during unsupervised dimensionality reduction
Considered and rejected
Considered and rejected: Rejected pixel-level PCA dimensionality reduction preprocessing for hyperspectral detection due to loss of subtle discriminative spectral information in faint signals.
Tried and failed
PCA dimensionality reduction before SVM classification applied to molecular descriptor property prediction. Outcome: worse than baseline. Reason: Unsupervised variance maximization discarded non-linear features predictive of class boundaries compared to supervised selection
Tried and failed
PCA dimensionality reduction on fiducial covariance applied to summary statistics for simulation-based inference. Outcome: worse than baseline. Reason: compresses along dominant noise variance modes rather than parameter-sensitive variations
Considered and rejected
Considered and rejected: Rejected using PCA dimensionality reduction (83 components) because it discards low-variance class-discriminative features and degraded F1 score by 0.11.
Tried and failed
PCA variance-based dimensionality reduction applied to surface EMG signal classification. Outcome: worse than baseline. Reason: Discarded low-variance components that contained critical discriminative motor unit action potential features
Non-invasive neural interfacing for wearable electromyographic systems · Imperial
Dimensionality reduction underperforms raw feature sets and simpler selection baselines
Compressed representations degraded predictive accuracy, cross-domain generalization, and correlation matching compared to retaining the original full feature sets. In multiple evaluations, principal component transformations performed worse than direct regression, full channels, or random feature selection.
Tried and failed
Principal component analysis for feature extraction applied to microstructural image representations. Outcome: worse than baseline. Reason: overcomplicated data representation, degrading downstream surrogate model accuracy compared to simple summary metrics
Exploration of the Additive Manufacturing Process Development Space using High-throughput Mechanical Property Assays · Georgia Tech
Tried and failed
PCA dimensionality reduction on temporal sequence features applied to hidden Markov model state prediction. Outcome: worse than baseline. Reason: Component reduction discarded informative variance and dynamic correlations present in raw sequential features
Development of A Trajectory Population Data and its Application in CAV Research · Virginia Tech
Lost to a baseline
PCA dimensionality reduction performed comparably to random sets and worse than curated supersets for average indication accuracy.
Comprehensive Elucidation of Small Molecule Therapeutic Behavior Using Multitarget Theory · DSpace at SUNY Buffalo
Tried and failed
RBF kernel PCA dimensionality reduction applied to dense document embeddings for classification. Outcome: worse than baseline. Reason: Nonlinear projection degraded feature representation compared to linear PCA and the original embedding space
Enabling Context-Aware Natural Language Processing: From Dense Vector Representations to Contextual Features · Texas Tech
Tried and failed
PCA dimensionality reduction before regression applied to tabular physical test data. Outcome: worse than baseline. Reason: PCA feature extraction did not systematically improve predictive performance compared to direct regression
Compacted Snow Testing Methodology and Instrumentation · Virginia Tech
Lost to a baseline
PCA-based feature selection (Jaccard 0.9624) was beaten by random feature selection (Jaccard 0.985) in K-Means segmentation dimensionality reduction
MORPHOLOGICAL CLASSIFICATION OF SUBTYPES OF VOLUMETRIC PROJECTION NEURONS FROM MOUSE BRAIN SCANS · JScholarship
Lost to a baseline
Two-dimensional Principal Component Analysis (PCA) score projections failed to separate distinct ink groups that were clearly resolved by visual spectral inspection.
DEVELOPMENT AND VALIDATION OF A TECHNIQUE FOR ANALYSIS OF TONER PRINTED DOCUMENTS USING MAGNETIC FLUX MEASUREMENTS · DSpace at SHSU
Considered and rejected
Considered and rejected: Rejected feature selection dimensionality reduction (PCA, ANOVA, Fisher Score) for EEG ErrPs because unreduced covariance and correlation features performed better.
Tried and failed
PCA feature reduction for tabular signal features applied to time-series signal classification. Outcome: worse than baseline. Reason: linear dimensionality reduction on mixed statistical and domain features produced inconsistent and degraded classification performance
Tried and failed
PCA projection for cross-correlation alignment applied to multimodal time-series synchronization. Outcome: worse than baseline. Reason: dimensionality reduction degraded cross-correlation accuracy without downstream confidence weighting mechanisms
Analyzing Health-related Behaviors Using First-person Vision · Georgia Tech
Tried and failed
PCA preprocessing before regression applied to open circuit potential sensor signals. Reason: dimensionality reduction did not improve multivariate calibration accuracy compared to raw inputs
Hardware and Software Interfaces Design for Multi-Panel Electrochemical Sensors · EPFL
Lost to a baseline
PCA dimensionality reduction to 100 components (18.70% accuracy) lost drastically to the 20,000-channel baseline (93.15%).
Electromagnetic Side-Channel Analysis Methods for Digital Forensics on Internet of Things · Research Repository UCD
Considered and rejected
Considered and rejected: PCA-based dimensionality reduction for filtered K-means feature selection was rejected in favor of 700 random features due to small sample size vs high dimensionality
MORPHOLOGICAL CLASSIFICATION OF SUBTYPES OF VOLUMETRIC PROJECTION NEURONS FROM MOUSE BRAIN SCANS · JScholarship
Tried and failed
PCA dimensionality reduction after dynamic time warping applied to time series neural network classification. Outcome: did not generalise. Reason: PCA subspace projection did not consistently improve accuracy and hurt cross-domain generalization compared to raw aligned inputs
Physical Side-Channel Vulnerability Assessment of Implementations of Cryptographic Algorithms · Georgia Tech
Tried and failed
feature selection and dimensionality reduction applied to error-related potential classification from EEG. Outcome: worse than baseline. Reason: yielded no significant performance increase over using the full feature set
Principal components are dominated by noise, scale imbalances, and technical confounders
Unsupervised components frequently aligned with dominant technical noise, inter-subject baseline shifts, or unstandardized high-variance features rather than meaningful signals. High-order components also introduced sample-specific noise that degraded downstream clustering, factor estimation, and out-of-sample generalization.
Tried and failed
canonical correlation analysis on principal components applied to single-cell reference mapping. Outcome: did not generalise. Reason: projections strongly biased query data toward a narrow, unrepresentative subset of reference samples
Tried and failed
adjusted covariate principal component analysis applied to multimodal neuroimaging pattern extraction. Outcome: no signal. Reason: severe loss of primary phenotype information leading to low correlations and poor prediction
Statistical Methods for Extracting and Comparing Patterns in Multimodal Neuroimaging Studies · Penn
Considered and rejected
Considered and rejected: Excluded first principal component (PC1) in initial droplet single-cell RNA-seq clustering because it primarily reflected capture efficiency (number of genes expressed).
Transcriptional and Epigenomic Landscapes of Vascular Endothelial Cells · JScholarship
Tried and failed
Simultaneous spectroscopy acquisition using CCD detectors applied to hyperspectral data processing with PCA. Reason: Detector introduced correlated channel noise that disrupted downstream dimensionality reduction algorithms.
Tried and failed
PCA using unstandardized covariance matrix applied to multivariate time series with varying units. Reason: features with the largest raw variance completely dominated the extracted principal components
Macroeconomic Forecasting: Statistically Adequate, Temporal Principal Components · Virginia Tech
Tried and failed
functional principal component analysis without phase alignment applied to unaligned curve data. Reason: fails to capture peak heights and creates nonlinear dependence across principal component scores
Heteroscedastic Functional Data Models · Cornell
Tried and failed
PCA for coordinate axis alignment applied to non-uniformly distributed point clouds. Outcome: did not generalise. Reason: Principal components biased toward high-density spatial clusters rather than true structural axes
Behavior-Coupled Neural Circuit Analysis of Chemosensory Responses in C. elegans · Georgia Tech
Tried and failed
principal component analysis on untargeted lipidomics applied to phenotypic group separation in metabolomics. Outcome: no signal. Reason: unsupervised variance was dominated by factors other than the disease phenotype
Mapping the metabolome in the developing gut · Imperial
Tried and failed
retaining high-order principal components for weak signals applied to linear factor asset pricing models. Outcome: did not generalise. Reason: high-order components captured primarily sample-specific noise rather than true weak latent factors, causing severe out-of-sample degradation
Asset pricing with unsystematic risk · Imperial
Considered and rejected
Considered and rejected: Rejected dimensionality reduction via PCA because it introduces unwanted noise, diminishes important attributes, and adds computational overhead.
Efficient Multi-GPU K-means Clustering · TXST Digital Repository
Considered and rejected
Considered and rejected: Rejected Principal Component Analysis (PCA) for joint word-document space reduction due to disproportionate vector magnitude variance between words and documents and edge skew from high-frequency terms.
Temporal Topic Embeddings with a Compass · Virginia Tech
Tried and failed
principal component analysis applied to multi-subject neural time series. Outcome: no signal. Reason: inter-individual baseline activation differences masked the shared underlying dynamical structure
Tried and failed
k-means clustering on higher-order principal components applied to functional connectivity feature representations. Outcome: worse than baseline. Reason: higher-order principal components introduced noise that progressively degraded cluster separation quality
Subspace projections obscure feature semantics, physical units, and model interpretability
Projecting variables into composite linear axes obscured the physical, biological, and spatial meanings of original measurements. This transformation hindered feature-level interpretation methods like SHAP and prevented discrete band selection or accurate centroid reconstruction.
Tried and failed
PCA dimensionality reduction before classification applied to high-dimensional tabular genomic data. Outcome: worse than baseline. Reason: Reduced classification performance compared to raw features and obstructed feature-level SHAP interpretability
Applications of Machine Learning in Source Attribution and Gene Function Prediction · Virginia Tech
Considered and rejected
Considered and rejected: Rejected using PCA dimensionality reduction for the final SHAP interpretation pipeline because mapping principal components back to distinct genes introduces loss of biological specificity
Applications of Machine Learning in Source Attribution and Gene Function Prediction · Virginia Tech
Considered and rejected
Considered and rejected: Rejected using Principal Component Analysis (PCA) for dimensionality reduction across connectivity models because transforming data out of pairwise connectivity space hindered the interpretability of neuroanatomical feature weights.
Considered and rejected
Considered and rejected: Rejected using Principal Component Analysis (PCA) on all derived variables for dimensionality reduction because composite PCA axes obscure specific ecological and physical interpretations of individual terrain drivers.
Considered and rejected
Considered and rejected: Rejected dimensionality reduction via PCA for experimental subset selection because PCA projections do not preserve original feature semantics.
Metabolic phenotypes of marine heterotrophic bacteria · OpenBU
Considered and rejected
Considered and rejected: Rejected standard Principal Component Analysis (PCA) for sensor design because it creates linear combinations of all bands rather than selecting discrete physical spectral bands.
Towards Transferable Pollution Detection Methods in Aquatic Environments Using Hyperspectral Technology · accedaCRIS
Considered and rejected
Considered and rejected: Rejected using Principal Component Analysis (PCA) for ROI directional analysis because frequency domain 2D Fourier transformation preserves spatial frequency components.
Considered and rejected
Considered and rejected: Rejected Principal Component Analysis (PCA) for clustering GM biomarkers because PCA components are linear signed combinations causing difficult ROI interpretation across multiple clusters.
In vivo MRI perfusion and structural correlates of pathology in frontotemporal lobar degeneration · ScholarlyCommons at Penn
Considered and rejected
Considered and rejected: Rejected using Principal Component Analysis (PCA) or ordination-based data reduction axes as continuous morphospaces for threshold models because the axes lack mechanistic generative meaning and fail under trait correlation.
Considered and rejected
Considered and rejected: Principal Component Analysis (PCA) for index construction was rejected because it reduces dimensionality of large datasets, which was unsuitable for measuring full disclosure items here
Impact of reserve and decommissioning disclosures on value and performance of listed oil companies in the UK · University of Nottingham Repository
Considered and rejected
Considered and rejected: Rejected PCA dimensionality reduction prior to k-means because it requires loading the full dataset and projecting centroids back to original space yields inaccurate centroids and lower cluster quality.
Approaches to extract, characterize, and interpret dynamic functional network connectivity in fMRI data · Georgia Tech
Left open by the authors
Problems the authors named and did not get to.
Left open
Apply principal component analysis to the consensus matrix to analyze latent subspaces and branching structures in single-cell developmental trajectories. Blocker: None
Ensemble Methods for Latent Structure Detection from Heterogeneous Genomic and Phenotypic Data · Harvard
Left open
Quantify cell-type differences using multiple dimensionality reduction runs and XGBoost models trained on gene subsets to reduce multicollinearity. Blocker: None
Developing Graph-based Computational Algorithms for Single-cell Data Science · Georgia Tech
Left open
Apply Principal Component Analysis and tensor decomposition to the combination dose-response L1000 gene expression dataset to identify emergent transcriptional changes. Blocker: None
A Framework for The Study of Compound Interactions in L1000 · Harvard
Left open
Implement and evaluate feature exchange augmentation using higher-order feature statistics like principal components instead of mean and standard deviation. Blocker: None
Left open
Apply Principal Component Analysis to decorrelate fitted thermal model parameters for irregular air-gap defect reconstruction in pulsed thermography. Blocker: None
Three-dimensional subsurface defect reconstruction for industrial components using pulsed thermography · Cranfield
Left open
Apply hierarchical dimensionality reduction to public molecular omics datasets to test if principal component scaling trends persist across biological layers. Blocker: None
Left open
Implement dimensionality reduction (e.g., PCA) for shape latent code projections in the differentiable corrector to avoid linear memory scaling with object count. Blocker: None
Left open
Apply principal component analysis to user response data to reduce dimensionality to orthogonal dimensions and develop targeted chatbot priming techniques. Blocker: Requires the private user survey and interaction response dataset from the thesis experiment.
Left open
Implement PCA-based dimensionality reduction before Heterogeneous Compound Symmetry modeling in the MAGMa proteomics pipeline for large sample size datasets. Blocker: None
Left open
Explore spatial EEG features and hybridize dimensionality reduction methods with high-performing classifiers for cognitive state detection. Blocker: Lacks specific implementation details and relies on proprietary flight crew psychophysiological dataset
Checking a claim in this area?
We can run the same search on any method or claim. If nothing turns up, we will say so, and that proves nothing on its own.