Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Quasicyclic Principal Component Analysis

Susanna E. Rumsey, Stark C. Draper, Frank R. Kschischang

TL;DR

QPCA addresses the limitation of standard PCA for cyclostationary data by constraining components to families of shift-orthogonal signals and formulating an optimization that decomposes into $N$ PCA problems in the frequency domain when $n=Ns$. The authors derive an explicit algorithm, analyze its computational complexity, and demonstrate applications to carrier-pulse recovery, symbol-period estimation, and oversampling-rate adjustment. They discuss non-uniqueness due to phase degrees of freedom and provide practical resampling strategies for non-integer oversampling. Overall, QPCA offers a principled approach to exploiting cyclostationarity in communications and other domains, enabling more interpretable representations of periodic or repeating structures.

Abstract

We present quasicyclic principal component analysis (QPCA), a generalization of principal component analysis (PCA), that determines an optimized basis for a dataset in terms of families of shift-orthogonal principal vectors. This is of particular interest when analyzing cyclostationary data, whose cyclic structure is not exploited by the standard PCA algorithm. We first formulate QPCA as an optimization problem, which we show may be decomposed into a series of PCA problems in the frequency domain. We then formalize our solution as an explicit algorithm and analyze its computational complexity. Finally, we provide some examples of applications of QPCA to cyclostationary signal processing data, including an investigation of carrier pulse recovery, a presentation of methods for estimating an unknown oversampling rate, and a discussion of an appropriate approach for pre-processing data with a non-integer oversampling rate in order to better apply the QPCA algorithm.

Quasicyclic Principal Component Analysis

TL;DR

QPCA addresses the limitation of standard PCA for cyclostationary data by constraining components to families of shift-orthogonal signals and formulating an optimization that decomposes into PCA problems in the frequency domain when . The authors derive an explicit algorithm, analyze its computational complexity, and demonstrate applications to carrier-pulse recovery, symbol-period estimation, and oversampling-rate adjustment. They discuss non-uniqueness due to phase degrees of freedom and provide practical resampling strategies for non-integer oversampling. Overall, QPCA offers a principled approach to exploiting cyclostationarity in communications and other domains, enabling more interpretable representations of periodic or repeating structures.

Abstract

We present quasicyclic principal component analysis (QPCA), a generalization of principal component analysis (PCA), that determines an optimized basis for a dataset in terms of families of shift-orthogonal principal vectors. This is of particular interest when analyzing cyclostationary data, whose cyclic structure is not exploited by the standard PCA algorithm. We first formulate QPCA as an optimization problem, which we show may be decomposed into a series of PCA problems in the frequency domain. We then formalize our solution as an explicit algorithm and analyze its computational complexity. Finally, we provide some examples of applications of QPCA to cyclostationary signal processing data, including an investigation of carrier pulse recovery, a presentation of methods for estimating an unknown oversampling rate, and a discussion of an appropriate approach for pre-processing data with a non-integer oversampling rate in order to better apply the QPCA algorithm.

Paper Structure

This paper contains 16 sections, 1 theorem, 32 equations, 5 figures.

Table of Contents

  1. Introduction
  2. Principal Component Analysis
  3. Problem Statement
  4. Mathematical Background
  5. The QPCA Problem
  6. Solution
  7. Definitions
  8. Derivation of QPCA
  9. Non-integer Oversampling Rate
  10. Lack of Uniqueness
  11. Algorithmic Description
  12. Examples and Applications
  13. Recovering a Dominant Pulse
  14. Symbol Period Estimation
  15. Recomputing the Oversampling Factor
  16. ...and 1 more sections

Key Result

Corollary 1

The signal $x$ is $s$-quasicyclic shift-orthonormal if and only if every component of the orthogonal decomposition in (eq:orthdecomp) has the same energy $1/N$, i.e., if and only if for $0 \le t < N$.

Figures (5)

  • Figure 1: An illustration of QPCA. In (a), a shift-orthogonal pulse $\phi$ is used to create data consisting of $n=100$ runs of signals containing $N=6$ modulated symbols each, such as that shown in (b). The first $N$ orthogonal PCA components are shown in (c), and have a complicated form that tells the user little about the structure of the data. In contrast, the QPCA results in (d) give the first shift-orthogonal component (along with its orthogonal family of shifts), which accurately corresponds to the underlying pulse $\phi$.
  • Figure 2: Absolute values of output spectra of first two QPCA components for a variety of mixtures of pulses, listed as $(P_1, P_2)$.
  • Figure 3: Ratio of first to second eigenvalues for each candidate value of $s$.
  • Figure 4: Primary pulses returned by QPCA for different values of $s$ in Example 2.
  • Figure 5: Comparison of pulses obtained when sampling RRC data with a non-integer oversampling rate, and the same data resampled at an appropriate integer rate.

Theorems & Definitions (10)

  • Definition 1: Autocorrelation
  • Definition 2: Shift-orthonormal signals
  • Definition 3
  • Definition 4: Orthogonal Projection
  • Definition 5: Discrete Fourier Transform
  • Claim 1: Unitarity of the DFT
  • Definition 6: Energy Spectrum
  • Claim 2
  • proof
  • Corollary 1