As an alternative to the method of random projections, this article introduces an efficient information-theoretic method for dimensionality reduction. The efficacy of this method for choosing informative dimensions relies upon the fact that the Von Neumann entropy of a normalised Pearson Correlation matrix is well-defined.

Another important use of this method is to verify that all the dimensions in an empirically-derived orthogonal basis for a dynamical system are informative.

The Von Neumann entropy of a Pearson Correlation matrix:

Let’s suppose we have an ergodic dynamical system with \(N\) observables \(x_i(t) \in \mathbb{R}\) which are sampled using a sequence of \(n\) measurements so we have a dataset \(X \in \mathbb{R}^{N \times n}\). Given \(X\), we may compute the statistics \(X_i = x_i - \langle x_i \rangle\) and \(\sigma_i^2 = \langle X_i^2 \rangle\) which allows us to define the Pearson Correlation Matrix with entries:

\begin{equation} R_{i,j} = \frac{X_i \cdot X_j}{\sigma_i \cdot \sigma_j} \end{equation}

Given \(R \in \mathbb{R}^{N \times N}\), we may define the density matrix \(\rho = \frac{R}{N}\) which is positive semi-definite, Hermitian and has unit-trace. Thus, we may calculate the entropy of \(R\) using the Von Neumann entropy:

\begin{equation} S(\rho) = -\text{tr}(\rho \cdot \ln \rho) = - \sum_{i=1}^N \lambda_i \cdot \ln \lambda_i \end{equation}

where \(\lambda_i\) are the eigenvalues of \(\rho\).

Using the Von Neumann entropy for dimensionality reduction:

If \(N\) is large, in order to compress the dataset \(X\) so that we keep the dimensions that contain \(95\%\) of the statistical information it is sufficient to find the discrete subset \(S \subset [1,N]\) that maximises:

\begin{equation} -\sum_{i \in S} \lambda_i \cdot \ln \lambda_i \end{equation}

subject to the constraint \(\frac{- \sum_{i \in S} \lambda_i \cdot \ln \lambda_i}{- \sum_{i=1}^N \lambda_i \cdot \ln \lambda_i} \leq \frac{95}{100}\) which may be done using sorting algorithms such as Quick Sort.


The method of random projections implicitly relies upon the Johnson-Lindenstrauss lemma which implies that in high-dimensional Euclidean spaces most dimensions aren’t informative. For datasets where this assumption is valid, the cardinality \(\lvert S \rvert\) as calculated in (3) provides us with an upper-bound on the intrinsic dimension of an ergodic dynamical system.


  1. von Neumann, John (1932). Mathematische Grundlagen der Quantenmechanik (Mathematical Foundations of Quantum Mechanics) Princeton University Press., . ISBN 978-0-691-02893-4.

  2. H. Felippe et al. The von Neumann entropy for the Pearson correlation matrix: A test of the entropic brain hypothesis. Arxiv. 2021.

  3. E.T. Jaynes. Information Theory and Statistical Mechanics. The Physical Review. 1957.

  4. Sharpee, Tatyana, Nicole C. Rust, and William Bialek. Maximally informative dimensions: analyzing neural responses to natural signals. Advances in Neural Information Processing Systems (2003): 277-284.

  5. Edward Witten. A Mini-Introduction To Information Theory. Arxiv. 2019.

  6. Karl Friston. The free-energy principle: a rough guide to the brain? Cell Press. 2009.

  7. Ping Li, Trevor J. Hastie & Kenneth W. Church. Very Sparse Random Projections. 2006.