Codex @codex  0

Brigitte Le Roux studies geometric data analysis, including correspondence analysis and principal component analysis. Her research account describes investigations of how principal axes change when observations, weights or the underlying Euclidean geometry are modified. This stability question asks which apparent structures survive a change in the analysis, rather than treating every fitted axis as intrinsic to the data.^[ref]

Her joint book with Henry Rouanet, Geometric Data Analysis by Brigitte Le Roux and Henry Rouanet, develops the weighted-cloud approach through correspondence analysis, Euclidean clouds and principal component analysis. It connects the linear algebra of projections with explicit choices about representation of multivariate observations.^[ref]

The centering matrix subtracts the arithmetic mean from every component of a real vector. For a positive integer $d$, let $1\in {\mathbb{R}}^d$ have every component equal to one and defineFor $x\in {\mathbb{R}}^d$, write $\bar{x}=d^{-1}{1}^{\mathsf{T}}x$. Then $Px=x-\bar{x}1$, so $1^{\mathsf{T}}Px=0$. The output lies in the subspace of vectors whose components sum to zero.

Its orthogonal projection properties follow directly. Set $J=11^{\mathsf{T}}/d$. Since $1^{\mathsf{T}}1=d$, we have $J^2=J$ and henceThus $P$ is a symmetric idempotent matrix, the orthogonal projector onto $1^{\perp }$. It kills constant vectors and leaves zero-sum vectors unchanged. Its kernel of a linear map is $\mathrm{span}\{1\}$ and its rank is $d-1$; for $d=1$ it is the zero matrix. The eigenvalues are zero on the constant direction and one on its orthogonal complement.

For an original example, $x=(1,2,4,9{)}^{\mathsf{T}}$ has mean $4$, giving $Px=(-3,-2,0,5{)}^{\mathsf{T}}$. The centered sum of squares isFor these four observations, the usual sample variance is $38/(4-1)=38/3$. Centering removes the mean; it does not itself divide by a standard deviation.

For a data matrix $X$ with observations in its $d$ rows, $PX$ centers each column. When $d>1$, its sample covariance matrix is $X^{\mathsf{T}}PX/(d-1)$. Also $P(x+c1)=Px$ for any scalar $c$, which proves invariance to adding the same offset to every observation. These identities make the centering matrix useful in least squares and principal component analysis.

Chemometrics applies statistics and mathematical modeling to chemical measurements and the design of chemical experiments. It connects recorded signals with chemical composition or other properties, while accounting for experimental variation. IUPAC's definition includes both analysis of chemical data and design of experiments and simulations.^[ref]

An original two-component example illustrates multivariate calibration. At two chosen wavelengths, suppose absorbance follows a calibrated linear modelwith fixed optical path and concentration units chosen so the displayed coefficients apply. This is the additive-response model suggested by the Beer-Lambert law, with no interactions or other absorbers. Readings $A_1=1.00$, $A_2=0.56$ giveNeither wavelength alone distinguishes both concentrations; their different response ratios make the pair informative.

If the second response row were instead twice the first, the matrix would be singular. More precise repeated readings would not restore the missing chemical selectivity. Nearly proportional response rows instead make the inverse sensitive to noise, connecting experimental wavelength selection with the condition number of the response matrix.

Principal component analysis can summarize patterns in many-channel measurements, while regression can predict an independently measured chemical quantity. These are different tasks: a direction of large signal variance need not identify a particular analyte. Cross-validation should keep correlated preparations or production batches together when the intended prediction is for new batches; splitting repeated readings of one preparation across training and validation can overstate performance.

A useful experiment measures known mixtures spanning the intended concentration range, reagent blanks and separately prepared validation samples. Inspecting calibration residuals tests departures from the model, while varying preparation conditions tests whether a fitted relation transfers. A mathematical fit gains chemical meaning from the materials, reference measurements and experimental design behind its data.

For a random vector with covariance $\Sigma$ and positive coordinate variances, let $D=\mathrm{diag}(\Sigma )$. Its correlation matrix is the covariance of $D^{-1/2}(X-\mathbb{E}X)$, so it is positive semidefinite with unit diagonal. Each entry is the correlation of the corresponding pair of variables. In particular $\mathrm{tr}(C)=p$, explaining why an eigenvalue in correlation-based principal component analysis represents that eigenvalue divided by $p$ of the total variance. The empirical version is the sample correlation matrix.

For a real random vector $X=(X_1,\dots ,X_p{)}^T$ with finite second moments, its covariance matrix collects all pairwise covariances:The diagonal entries are variances. The covariance matrix is a symmetric matrix and a positive semidefinite matrix, because for every real vector $a$,Thus one matrix describes the variance of every linear combination of the coordinates. It is a positive-definite matrix exactly when no nonzero linear combination is constant almost surely. A deterministic linear relation produces a zero eigenvalue.

For example,has unit coordinate variances, covariance $0.8$, and eigenvalues $1.8$ and $0.2$. Its unit eigenvectors $(1,1{)}^T/\sqrt{2}$ and $(1,-1{)}^T/\sqrt{2}$ describe directions of greatest and least projected variance. For a unit vector $a=(\cos \theta ,\sin \theta {)}^T$, direct multiplication gives $a^T\Sigma a=1+0.8\sin (2\theta )$: the variation in the sum direction is nine times that in the difference direction.

The spectral theorem supplies an orthonormal basis of these directions. This connects covariance matrices to principal component analysis: its leading eigenvector identifies the unit-direction projection with the largest variance. For a linear transformation $A$, the transformed random vector has covariance matrix $A\Sigma A^T$, so uncertainties propagate through a linear model by matrix multiplication.

David Weenink co-created Praat with Paul Boersma, providing tools for measuring and modeling speech. His 2006 thesis investigated speaker-adaptive vowel identification: how recognition can remain useful when acoustic patterns vary between speakers.^[ref][ref]

The thesis describes recordings from 30 speakers, each producing 12 Dutch vowels in controlled sentence frames. It relates these data to formants, principal component analysis, discriminant analysis and adaptive classification methods. This makes the recognition problem experimentally concrete: a classifier must separate vowel identity from speaker-related variation rather than memorize a single speaker's frequency spectrum.^[ref]

Homepage: www.fon.hum.uva.nl/david/

The general intelligence factor, conventionally $g$, is a latent variable used to account for shared variation among performances on different cognitive tasks. Charles Spearman's 1904 analysis distinguished a common component from task-specific contributions, helping establish the historical basis of this approach.^[ref]

A simple one-factor measurement model writes a standardized test score aswhere ${\lambda }_j$ is a factor loading. Suppose all variables have zero mean, $\mathrm{Var}(g)=1$, each ${ϵ}_j$ has zero covariance with $g$, and distinct residuals have zero covariance with one another. Then for different tests,Thus a single shared component can explain correlated scores even when the tests contain different content. The covariance equation follows from the stated assumptions; it does not identify a unique biological cause of the correlations.

Factor analysis estimates such models and evaluates their fit. The general intelligence factor should also be distinguished from a first component obtained by principal component analysis: one is a proposed latent structure with residual assumptions, while the other is a variance-maximizing summary of observed scores. Interpretation requires the tasks, population and measurement assumptions, rather than a number alone.

Geometric data analysis represents multivariate observations as weighted clouds of points, then studies their Euclidean geometry through distances, orthogonal projections and principal components. Principal component analysis applies this approach to numerical variables; correspondence analysis adapts it to a table of counts with its own weights and metric. Choosing the weights and metric is part of the analysis, since these choices decide which differences between observations count as large.^[ref]

For a column-centered data matrix with singular value decomposition $X=U\Sigma V^T$, the first $k$ principal components give coordinates $X{V}_k=U_k{\Sigma }_k$. They are the orthogonal projection onto the chosen principal directions. Rescaling one variable changes the distances and can change those directions: measurements in metres and millimetres should not receive different influence merely because the units were changed. The geometry must therefore be interpreted together with the scaling convention, rather than treating the resulting plot as an intrinsic picture of the observations.

Henry Rouanet worked on statistics for researchers in psychology and the social sciences, emphasizing interpretable methods for analyzing actual data. His own account describes that aim and organizes his work by statistical problem rather than by an abstract sequence of prerequisites.^[ref]

With Brigitte Le Roux, he wrote Geometric Data Analysis by Brigitte Le Roux and Henry Rouanet. Its treatment of correspondence analysis and principal component analysis expresses multivariate data as weighted clouds in Euclidean geometry. The choice of weights and metric specifies which departures count as large; the resulting picture is part of the statistical model, not simply an illustration added afterward.^[ref]

In principal component analysis, the matrix whose columns are the chosen orthonormal eigenvectors of a covariance matrix, ordered by decreasing eigenvalues. Under this convention the principal component scores of centered data are $Z=X_{c}V$. Some literature instead scales the vectors by square roots of eigenvalues; the convention should be stated explicitly.

The latent factor vector has mean zero and covariance $I_q$, the specific errors have mean zero and diagonal covariance $\Psi$, and the two are uncorrelated. Thus the off-diagonal covariance is explained by shared factors. Normality additionally gives a likelihood model, but the covariance identity needs only these second-moment assumptions. Positive diagonal entries of $\Psi$ describe variation not explained by the factors. Unlike principal component analysis, the model separates shared and specific variation rather than only maximizing total projected variance.

A reduced state space is useful when it preserves the structure needed for the question being asked; low dimension and high explained variance do not by themselves establish a sufficient neural code. Start with a clear observation model. A vector of binned spike counts from $N$ neurons may be approximated through $d<N$ latent coordinates, but the choices of bins, transformations, noise model and preserved criterion all affect the result.

For centered observations, principal component analysis chooses an orthonormal projection $U\in {\mathbb{R}}^{N\times d}$ to minimize expected squared reconstruction error,where $\Sigma =E[X{X}^T]$. The identity follows by expanding the square and using $U^{T}U=I$. Diagonalize $\Sigma =V\mathrm{diag}({\lambda }_i)V^T$. The trace term is ${\sum }_i{\lambda }_i{\alpha }_i$, with ${\alpha }_i=\Vert U^T{v}_i{\Vert }^2$ for the columns $v_i$ of $V$. Here $0\le {\alpha }_i\le 1$ and ${\sum }_i{\alpha }_i=d$ because the columns of $U$ are orthonormal. Allocating this total weight to the largest eigenvalues maximizes the sum. This is an optimality statement about squared error in the chosen coordinates; PCA does not require a Gaussian distribution to satisfy it and does not optimize every behavioral question.

An exact counterexample makes the limitation decisive. Let $W,S$ be independent balanced signs, define $X=(10W,S{)}^T$, and let the remembered class be $Y=S$. Zero means and independence giveThe first component retains $100/101\approx 99.01\%$ of total variance. Yet $P(Y=1\mid W)=1/2$: no classifier using that component alone exceeds one-half expected accuracy. The second component determines the class exactly. Positive rates $(20+10W,5+S)$ give the same example after centering. These are constructed rates, not experimental recordings.

This example also exposes a scaling assumption. Covariance-based PCA emphasizes large-variance coordinates; standardization changes that objective and may increase the influence of noisy low-rate cells. Neither preprocessing choice is automatically correct. Behaviorally relevant information may lie in a small-amplitude direction, in precise spike timing lost by binning, or in correlations not retained by a rate-only summary. Dimensionality must therefore be evaluated against decoding, reconstruction and the scientific question separately.

Byron M. Yu and colleagues' 2009 Gaussian-process factor analysis combined temporal smoothing and dimensionality reduction in a probabilistic framework. Its demonstration used 61 simultaneously recorded macaque premotor and motor neurons during reaching. It shows a method for estimating single-trial trajectories, not direct evidence that episodic memories have the same dimensional structure.^[ref]

In a linear observation model $y_t=C{z}_t+b+{\epsilon }_t$, the noise and latent temporal assumptions matter. A Gaussian observation approximation can be imperfect for sparse counts; independent residual errors can misassign shared fluctuations; a smooth temporal prior can suppress abrupt transitions. Trial averaging can manufacture a smooth path while concealing different single-trial paths. A useful low-dimensional description must also generalize beyond the sampled cells, session and task: unrecorded populations may encode variables absent from the fitted coordinates.

Latent axes require care as well. In $y=Cz+b$, an orthogonal rotation $Q$ gives the identical reconstruction through $z^{\prime }=Qz$ and $C^{\prime }=C{Q}^T$, because $Q^{T}Q=I$. Reconstruction alone therefore cannot label one axis a unique psychological process. Model priors and constraints can restrict this ambiguity. Similarly, attractive geometry does not establish intrinsic dynamics: unobserved inputs, behavior or the analysis itself may shape the apparent trajectory. If a true dynamical state is claimed, the retained coordinates should predict relevant future activity under a specified model, not merely provide a convenient picture.

Validation should choose dimension and preprocessing within training data, then assess held-out trials or sessions and, where appropriate, held-out neurons. Behavioral labels must be kept out of unsupervised fitting decisions used to claim independent prediction. Comparing multiple dimensions and reporting uncertainty is more informative than selecting the most compelling two-dimensional plot. A reduced representation can support a useful decoder without being a complete causal explanation of the circuit.

Treat the forty-person convenience sample as a pilot, then develop and validate the measure in its intended population. Start by defining whose organizational ability the questionnaire measures and who answers it: lecturer self-reports and student ratings are different measurements. Specify the relevant activities and context. Discuss the candidate items with lecturers and appropriate users, seek expert review, and use cognitive interviewing to find ambiguous wording, double-barreled items and irrelevant content. This establishes content validity rather than assuming that correlations alone identify the intended construct.

Recruit a larger, better-targeted sample covering relevant roles, disciplines and institutions. A convenience sample of adults who cannot meaningfully judge lecturers' organization is not made suitable simply by increasing its size. Plan the sample size using plausible factor loadings, communalities, number of dimensions and indicators, response distributions, missingness and required estimation stability, preferably with simulation. Forty observations for sixteen items offer limited grounds for a stable generalizable structure; no universal participants-per-item cutoff makes this analysis automatically impossible or automatically adequate. If students rate the same lecturers, recognize that the ratings are clustered rather than independent lecturer-level observations.

Check missing responses, item distributions, scoring direction and correlations, identifying redundant or poorly functioning items. If the underlying dimensions are not already well specified, use exploratory factor analysis to model shared variation among items. Principal component analysis instead forms combinations explaining total observed variation; it is not the same latent measurement model. Choose a suitable common-factor estimator, such as principal axis factoring, or maximum likelihood estimation when its assumptions are credible. For ordered response categories, consider an ordinal estimator and polychoric correlations, while checking whether category counts support their estimation. Do not treat those correlations as assumption-free.^[ref]

Choose the number of factors using parallel analysis, the scree plot, substantive interpretation and the quality of the resulting model, rather than an eigenvalue-one rule alone. Parallel analysis compares the observed eigenvalues with a reference generated from uncorrelated data of comparable dimensions; its implementation must match the chosen correlation and extraction approach.^[ref] Use oblique factor rotation if organizational dimensions can plausibly correlate. Inspect factor loadings, cross-loadings, communalities, factor correlations and model discrepancies, and retain enough substantively relevant indicators for each dimension. Do not repeatedly delete items solely to improve numerical fit at the expense of coverage.

After developing the structure, test it using confirmatory factor analysis in an independently recruited validation sample, with the proposed structure fixed beforehand. Splitting the existing forty participants would not solve the shortage of information. Evaluate model fit and uncertainty, measurement reliability, test-retest reliability where stability is intended, and evidence for construct validity using related, distinct and relevant behavioral measures. Check measurement invariance before comparing scores across groups. Report alternative plausible models, all item decisions, sampling limitations and the distinction between an initial exploratory instrument and an independently validated measure.

Use factor analysis when several observed measurements may reflect fewer unobserved dimensions, such as related questionnaire items about planning, prioritization and record keeping. A common-factor model separates shared variation represented by latent factors from item-specific variation. Factor loadings describe how measurements relate to those latent factors; communalities summarize their shared-variance contributions.

Exploratory factor analysis investigates a structure not yet fixed; confirmatory factor analysis evaluates prespecified measurement restrictions, preferably in independent data. Extraction, factor count and factor rotation affect interpretation, so report them and justify the sample and measurement assumptions. A statistically convenient factor solution does not alone establish construct validity or demonstrate the existence of a literal psychological mechanism. The distinction from principal component analysis is the common-versus-total-variation objective, and from cluster analysis the modeling of measurement relationships rather than the grouping of cases.

There is evidence for particular structures of social intelligence, but no single established factor structure independent of definition, measurement method and population. “Not known” should therefore be interpreted as unresolved construct organization, not an absence of factor analyses or proof that social understanding cannot be measured. A factor is a proposed source of shared variation among selected measurements; it does not acquire a psychological meaning simply because a statistical extraction produces it.

The first difficulty is defining what belongs in the construct. Understanding another person's intentions, remembering social information, knowing social conventions and behaving effectively in an interaction are distinguishable achievements. They may draw on partly different abilities. A questionnaire asking whether one feels socially competent also measures self-perception and typical behavior, whereas a scored performance task asks what one can do under specified conditions. Related constructs such as general intelligence, personality and theory of mind overlap with some definitions but are not interchangeable with every proposed social ability. Different definitions therefore generate different item domains before factor analysis even begins.

Historical negative findings demonstrate the problem of discriminant evidence. Daniel P. Keating studied 117 college students using three academic-intelligence measures and three proposed social-intelligence measures. Correlations within the social set were not higher than correlations across the social and academic sets, and factor analysis did not identify a separate social factor. This challenged the coherence and distinctiveness of that test battery. It did not logically refute every possible social-intelligence construct: the chosen measures, restricted student sample and criteria determine what the negative result can exclude.^[ref]

Conversely, a positive structure within one instrument does not settle the whole domain. David H. Silvera and colleagues developed the Tromsø Social Intelligence Scale through studies of expert interpretations, item selection and stability, obtaining a 21-item self-report instrument with three dimensions: social information processing, social skills and social awareness. This provides a usable operationalization and evidence for internal organization. But the structure of self-descriptions is not automatically the structure of maximal social performance. Shared response style, self-esteem, item wording and perceived competence may generate covariation that a performance test would not reproduce.^[ref]

A different strategy separates content from method. Susanne Weis and colleagues tested a performance model distinguishing social understanding, memory and knowledge using verbal, pictorial and video-based tasks in a multitrait-multimethod design. In 118 students, confirmatory analyses supported multidimensional organization and provided evidence of differentiation from academic intelligence, although social-memory measures showed overlap. This is positive evidence against the claim that no structure has ever been found. The factor labels differ from the self-report scale because the proposed processes and tasks differ; the fact that both studies obtained three dimensions does not make the dimensions equivalent.^[ref]

These studies expose the need to distinguish construct validity from attractive factor labels. A social story task may mainly depend on reading comprehension; a facial or video judgment may introduce perceptual and memory demands. If all putative measures of one trait use the same format, a “trait factor” can actually be a method factor. Multiple methods for each proposed trait allow stronger tests of convergent validity and discriminant validity, although even an MTMM model depends on identifying restrictions and adequate indicators. A good model should explain cross-method covariance as well as within-format similarity.

Measurement quality also changes the observed structure. Unreliable tasks attenuate correlations, potentially hiding a shared ability; overlapping items or shared response biases can inflate them. Social performance depends on cultural conventions and context, and a scoring key based on majority agreement need not establish objectively superior understanding in every setting. Small or homogeneous samples can produce unstable loadings or restrict ability variance. Evidence for measurement invariance across groups and contexts is therefore needed before interpreting different scores or factor solutions as differences in the same underlying construct.

Statistical choices contribute additional uncertainty. Exploratory factor analysis models common variation, whereas principal component analysis summarizes total variance; their dimensions answer different questions. Factor retention, residual correlations and oblique factor rotation affect interpretation. In a general factor model,where $\Phi$ is factor covariance and $\Psi$ residual covariance. For invertible $T$, replacing $\Lambda$ with $\Lambda T$ and $\Phi$ with $T^{-1}\Phi T^{-T}$ leaves $\Sigma$ unchanged. Hence covariance fit alone cannot uniquely name unrestricted factors. Identifying restrictions and substantive predictions are needed. This rotational issue concerns orientation; it does not by itself imply that every possible number of factors fits equally well.

A stronger research program would specify the proposed abilities before choosing indicators, obtain multiple reliable measures across methods and social contexts, and measure competing general-intelligence and personality constructs. Exploratory work and parallel analysis can guide model development; a prespecified confirmatory factor analysis in an independent, adequately varied sample provides a separate test. Competing correlated-factor, hierarchical and method-factor models should be compared, with appropriate attention to identification and out-of-sample performance. Predicting consequential social outcomes beyond competing constructs adds evidence that internal consistency alone cannot supply.

The unresolved structure reflects construct and measurement choices as well as statistical uncertainty. Existing studies support specific, testable organizations, but universal agreement requires convergence across performance and report methods, populations and external criteria. The productive conclusion is to improve competing models and their tests, rather than treating one questionnaire's factors as a final inventory of social intelligence or interpreting an early null result as proof of impossibility.

Yes, the data are centred. Both coordinate sample means vanish:Thus principal component analysis can be applied without further data centering.

The one-dimensional principal component analysis reconstruction lies along $v_1=(4/5,-3/5{)}^T$ through the zero sample mean. Its discarded principal component coordinate is $P{C}_2=0$, soAt $x=6$, $\boxed{\hat{y}=-4.5}$. This intersects the retained principal component line with the vertical line at the supplied $x$. It is a reconstruction from that line, rather than an ordinary linear regression of $y$ on $x$.

For centred data, classical multidimensional scaling from Euclidean distances constructs the observation Gram matrixwhere $D_{ij}^{(2)}=\Vert x_i-x_j{\Vert }^2$. By part (i), its nonzero eigenvalues agree with those of $X^{T}X$, the matrix used for principal component analysis before dividing by a covariance normalization.

If $u_i=X{v}_i/\sqrt{{\lambda }_i}$ is the unit observation-space eigenvector, the corresponding classical multidimensional scaling coordinate column isThat is exactly the column of principal component scores. Consequently classical multidimensional scaling of Euclidean pairwise distances recovers the same reduced coordinates as principal component analysis, up to signs and arbitrary orthogonal transformations within equal-eigenvalue subspaces. Distances alone cannot fix translation or orientation. The equivalence presumes centred data and Euclidean distances; general dissimilarities need not yield a positive semidefinite Gram matrix or a corresponding data covariance matrix.

The total variance is the trace $\mathrm{tr}\Sigma =10$. Retaining two original genes preserves the sum of their marginal variances; the largest choice includes the third gene and either of the first two, giving $4+3=7$. Thus the largest proportion is 70%. This refers to retaining measured coordinates themselves, rather than replacing them by linear combinations as in principal component analysis.

Principal component analysis replaces correlated measurements by orthogonal linear coordinates ordered by their variance. For centered observations collected in an $n\times p$ matrix $Z$, write the sample covariance matrix as $S=Z^{T}Z/(n-1)$. A unit loading vector $a$ gives score vector $Za$ with sample variance $a^{T}Sa$. The Rayleigh quotient is maximized by a unit eigenvector of $S$ for its largest eigenvalue.

The subsequent principal components maximize the same quadratic form subject to orthogonality to the earlier loading vectors. By the spectral theorem, if $S=Q\Lambda Q^T$ with descending eigenvalues, the scores $ZQ$ have diagonal covariance $\Lambda$. Their fractions of explained variance are ${\lambda }_j/{\sum }_i{\lambda }_i$. Repeated eigenvalues identify an eigenspace rather than a unique axis.

Keeping the first $q$ principal components gives the rank-$q$ reconstruction $\hat{Z}=Z{Q}_q{Q}_q^T$. Its squared reconstruction error isFor any rank-$q$ orthogonal projection $P$, retained variance is $\mathrm{tr}(SP)={\sum }_j{\lambda }_j{q}_j^{T}P{q}_j$. The weights $q_j^{T}P{q}_j$ lie in $[0,1]$ and sum to $q$, so this is at most the sum of the largest $q$ eigenvalues. Projecting onto any candidate reconstruction subspace is its best least-squares reconstruction; this proves the minimum-error property and agrees with the singular value decomposition. Thus PCA finds the linear subspace preserving the greatest variance, equivalently minimizing squared reconstruction error. The upper-left sketch shows the first axis aligned with the elongated cloud; the second is perpendicular. A scree plot and substantive interpretability help choose $q$.

These methods depend on measurement scale. A variable measured in large numerical units can dominate covariance-based principal component analysis; standardizing variables gives principal component analysis on a correlation matrix. Scores summarize patterns, but the method neither establishes causation nor guarantees that the largest-variance directions best predict a separate response.