Principal component analysis for high-dimensional compositional data

Jingru Zhang, Wei LinarXiv 2021 · DOI · arXiv

Abstract

Dimension reduction for high-dimensional compositional data plays an important role in many fields, where the principal component analysis (PCA) of the basis covariance matrix is of scientific interest. In practice, however, the basis variables are latent and rarely observed, and standard techniques of PCA are inadequate for compositional data because of the simplex constraint. To address this problem, we relate the principal subspace of the centered log-ratio (clr) compositional covariance to that of the basis covariance and prove that the latter is approximately identifiable with the diverging dimensionality under a subspace sparsity assumption. This "blessing-of-dimensionality" phenomenon enables us to propose principal subspace estimation methods using the sample clr covariance. We derive nonasymptotic error bounds for the subspace estimators, exhibiting a tradeoff between identification and estimation. Moreover, we develop efficient proximal alternating direction method of multipliers (ADMM) algorithms to solve the resulting nonconvex and nonsmooth optimization problems. Simulation results and an analysis of word usage patterns for statisticians demonstrate that the proposed methods perform as well as oracle methods with known basis variables.

Introduction

Compositional data—data representing proportions of a whole—is ubiquitous in modern scientific research, appearing in fields ranging from metagenomics (microbiome counts) to linguistics (word frequencies). Despite its prevalence, this data type presents a significant mathematical challenge: the simplex constraint. Because the components of a compositional vector must sum to one, the variables are inherently dependent. This dependence creates "spurious correlations" that render standard statistical tools, such as traditional Principal Component Analysis (PCA), biased and often misleading.

In most scientific contexts, the primary interest lies in the basis variables (the absolute abundances) rather than the observed proportions. However, these basis variables are latent and rarely observed. In high-dimensional settings where the number of variables $p$ exceeds the number of observations $n$, standard PCA is not only biased by the simplex geometry but also statistically inconsistent. This research bridges the gap between observed proportions and the hidden absolute basis driving the data structure.

Methodology

The Theoretical Bridge

The authors utilize the centered log-ratio (clr) transformation to map compositional data from the simplex to a Euclidean space. The core of the methodology involves relating the principal subspace of the clr covariance matrix to the principal subspace of the latent basis covariance.

The Blessing of Dimensionality

A central finding of this work is a "blessing of dimensionality." While high dimensionality usually complicates statistical estimation (the "curse of dimensionality"), here it facilitates identification. The authors prove that as the dimensionality $p$ increases, the clr-transformed data becomes a more accurate proxy for the latent basis. This occurs because the "noise" introduced by the geometric mean in the clr transform vanishes asymptotically under the assumption of subspace sparsity—the idea that only a subset of variables drives the primary variance in the data.

Optimization via Proximal ADMM

Estimating a sparse subspace involves solving a nonconvex and nonsmooth optimization problem. To handle this, the authors developed a Proximal Alternating Direction Method of Multipliers (ADMM) algorithm. This approach efficiently incorporates row and column sparsity constraints, allowing the model to filter out high-dimensional noise and recover the underlying structure of the latent basis.

Key Insights

  • Identifiability through Sparsity: The latent structure of absolute abundances can be recovered from relative proportions if the principal subspace is sparse. This assumption allows the model to distinguish between the true signal and the constraints imposed by the simplex.
  • Estimation-Identification Trade-off: The research highlights a fundamental trade-off in high-dimensional PCA. Increasing $p$ reduces approximation error (improving identification) but simultaneously increases the difficulty of estimating parameters from finite samples (increasing estimation error).

Empirical Results

Simulation Studies

The proposed method was tested against "Oracle" methods (scenarios where true absolute abundances are known) using multivariate Normal and Gamma log-basis distributions. In these simulations, the proposed framework performed almost identically to the Oracle, significantly outperforming standard PCA applied to raw proportions, log-transformed data, or power-transformed data.

Case Study: Statisticians' Text Corpus

The method was applied to a real-world dataset of abstracts from 3,248 papers published in leading statistics journals (AoS, JASA, JRSS-B, and Biometrika) between 2003 and 2012. The analysis of 580 words across 236 authors revealed clear thematic structures:

  • PC1 (Fundamental vs. Applied): The first principal component separated fundamental research terms (e.g., covariance, likelihood, linear) from more specific or applied topics.
  • PC2 (Biostatistics): The second component successfully isolated terms related to biostatistics (e.g., cancer, clinic, hazard, survive).

Notably, standard PCA methods failed to achieve this level of clarity, often mixing biomedical terms with general inference terms, whereas the proposed method provided a clean separation of research domains.

Limitations and Future Work

While the framework is robust, its performance depends on the sparsity assumption. If the underlying basis variance is "dense" (spread across all variables), the theoretical guarantees for identification may weaken. Additionally, while the clr transformation is effective, future research could explore the theoretical properties of other transformations, such as power transformations, within this high-dimensional context. Finally, as datasets scale into the millions of dimensions (e.g., in modern genomics), further algorithmic optimizations beyond the current ADMM implementation may be required.

Summary updated

Abstract

Dimension reduction for high-dimensional compositional data plays an important role in many fields, where the principal component analysis of the basis covariance matrix is of scientific interest. In practice, however, the basis variables are latent and rarely observed, and standard techniques of principal component analysis are inadequate for compositional data because of the simplex constraint. To address the challenging problem, we relate the principal subspace of the centered log-ratio compositional covariance to that of the basis covariance, and prove that the latter is approximately identifiable with the diverging dimensionality under some subspace sparsity assumption. The interesting blessing-of-dimensionality phenomenon enables us to propose the principal subspace estimation methods by using the sample centered log-ratio covariance. We also derive nonasymptotic error bounds for the subspace estimators, which exhibits a tradeoff between identification and estimation. Moreover, we develop efficient proximal alternating direction method of multipliers algorithms to solve the nonconvex and nonsmooth optimization problems. Simulation results demonstrate that the proposed methods perform as well as the oracle methods with known basis. Their usefulness is illustrated through an analysis of word usage pattern for statisticians.
BibTeX
@misc{zhang2021principal,
  author = {Jingru Zhang and Wei Lin},
  title = {{Principal component analysis for high-dimensional compositional data}},
  year = {2021},
  eprint = {2109.04657},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2109.04657},
}
Figure