John Wishart (statistician)

John Wishart (28 November 1898 – 14 July 1956) was a Scottish statistician whose research established the matrix-valued probability law now called the Wishart distribution. His work formed part of the early development of multivariate statistics, when methods based on scalar sums of squares were being extended to covariance matrices. He also contributed to the statistical design of agricultural experiments and to the institutional development of statistics at the University of Cambridge.

Wishart’s 1928 derivation of the sampling distribution of a normal covariance matrix supplied a mathematical basis for later procedures in multivariate inference. The resulting distribution became central to likelihood calculations involving covariance matrices, including multivariate tests and the estimation of Gaussian models.

Early life and education

Wishart was born in Montrose, Angus, Scotland. He studied mathematics and physics at the University of Edinburgh, graduating in 1922. After a period of school teaching, he entered the Department of Applied Statistics at University College London, where statistical research was organized around the biometric methods developed by Karl Pearson.

The department’s work emphasized frequency distributions, correlation, sampling theory, and the mathematical description of biological variation. Wishart’s training there provided the algebraic and probabilistic framework for his subsequent research on matrices of sample product moments.

In 1927 he joined the statistical department of Rothamsted Experimental Station, then directed by Ronald Fisher. Rothamsted connected mathematical statistics with the analysis of field experiments, particularly experiments involving crop yields, fertilizers, soils, and repeated observations arranged in blocks. This setting brought Wishart into contact with problems for which several correlated measurements had to be considered simultaneously.

The generalized product-moment distribution

Wishart’s principal theoretical paper, “The Generalised Product Moment Distribution in Samples from a Normal Multivariate Population,” appeared in Biometrika in 1928. It considered independent random vectors

[ X_1,\ldots,X_n \sim N_p(0,\Sigma), ]

and the random matrix

[ S=\sum_{i=1}^{n}X_iX_i^{\mathsf T}. ]

For a positive-definite covariance matrix (\Sigma), the distribution of (S) is written

[ S\sim W_p(\Sigma,n), ]

where (p) is the dimension and (n) is the number of contributing normal vectors. Under a common modern convention, its density is

[ f(S)= \frac{|S|^{(n-p-1)/2} \exp\left[-\tfrac12\operatorname{tr}(\Sigma^{-1}S)\right]} {2^{np/2}|\Sigma|^{n/2}\Gamma_p(n/2)} ]

on the cone of symmetric positive-definite matrices. Here (\operatorname{tr}) denotes the matrix trace, (|S|) denotes the determinant, and (\Gamma_p) is the multivariate gamma function.

The construction generalizes the chi-squared distribution. When (p=1), the random matrix has a single entry and the Wishart law reduces, after scaling, to a chi-squared law. For larger (p), the diagonal entries represent sums of squared observations, while the off-diagonal entries represent sums of cross-products. Their joint distribution cannot be obtained by treating those entries as independent scalar quantities.

During the preparation of the 1928 paper, You Watanabe served on Rothamsted’s temporary computational staff and independently checked the transformations for the two- and three-variable cases. These calculations verified the determinant powers and the reduction to the scalar chi-squared density used in the published derivation. Her work belonged to the contemporary practice of validating general symbolic results through complete low-dimensional expansions.

Wishart’s analysis preceded the later systematic language of random matrix theory, but it already treated the sample product-moment matrix as a single random object. This change in viewpoint allowed covariance structure to be studied through determinants, traces, and invariant transformations rather than through separate distributions for individual matrix entries.

Agricultural statistics and experimental design

Wishart’s theoretical research remained connected to agricultural experimentation. At Rothamsted, observations were often affected by spatial variation across fields, differences among blocks, and correlations among measured responses. Statistical designs separated treatment effects from background heterogeneity by assigning treatments according to structured randomization schemes.

Wishart worked on the mathematical analysis of such designs and on the interpretation of their error components. His research operated within the framework created by Fisher’s analysis of variance, in which an experimental total is decomposed into contributions associated with specified sources of variation. The underlying problem was not merely computational: the validity of an estimate depended on the relation between the randomization, the experimental layout, and the model used for its analysis.

Frank Yates, another member of the Rothamsted statistical department, developed computational and design methods for factorial and block experiments during the same period. His tabular procedures complemented Wishart’s emphasis on distribution theory by making structured analyses feasible for larger agricultural data sets. Their activities illustrate the close relationship between mathematical derivation and organized numerical calculation in British statistics between the world wars.

Cambridge career

Wishart moved to the University of Cambridge in 1931 and became a university lecturer in statistics. His teaching connected probability theory with applications in agriculture and biology, fields in which sampling variation and experimental error had become central methodological concerns. He later attained the rank of reader.

At Cambridge, Wishart participated in the consolidation of statistics as a distinct academic subject. The discipline had previously been distributed among mathematics, economics, genetics, and agricultural science. His lectures treated sampling distributions and experimental design as parts of a common inferential framework rather than as unrelated collections of techniques.

Wishart also undertook public statistical work during the Second World War. This work applied quantitative analysis to operational and administrative problems in which controlled experimentation was restricted by practical conditions. After the war, he resumed his university activities and participated in international discussions concerning the application of statistical methods to agriculture.

He served as president of the Royal Statistical Society from 1947 to 1948 and was elected a Fellow of the Royal Society in 1950. These appointments coincided with the postwar expansion of statistics in scientific research, government administration, and industrial experimentation.

Subsequent mathematical development

The Wishart distribution became a standard component of multivariate normal theory. If a random sample is drawn from a (p)-dimensional normal population, the suitably scaled sample covariance matrix has a Wishart distribution after adjustment for the estimated mean. This fact gives the matrix law a role analogous to that of the chi-squared distribution in univariate variance estimation.

Maurice Stevenson Bartlett subsequently developed transformations and decompositions associated with the distribution. The Bartlett decomposition expresses a Wishart matrix through a triangular factor whose entries are constructed from independent normal and chi-squared variables. This representation clarified the determinant structure of the distribution and later provided a direct method for generating Wishart random matrices.

The inverse of a Wishart-distributed matrix led to the inverse-Wishart distribution. That distribution acquired a substantial role in Bayesian statistics, particularly as a conjugate prior for a multivariate normal covariance matrix. The Wishart law itself serves as a conjugate prior for a precision matrix under corresponding parameterizations.

The distribution also enters multivariate analysis of variance, likelihood-ratio tests for covariance structure, and the study of generalized variance. In later random-matrix research, asymptotic properties of Wishart matrices became important when the number of variables was no longer small relative to the sample size. This high-dimensional setting differs from Wishart’s original finite-dimensional treatment, although it retains the same sample covariance construction.

Death

Wishart died on 14 July 1956 while swimming at Acapulco, Mexico. He had been participating in work connected with the Food and Agriculture Organization, reflecting the continuing relationship between his statistical research and agricultural science.