Main Content

The goal of *spectral estimation* is to describe the distribution (over
frequency) of the power contained in a signal, based on a finite set of data. Estimation of
power spectra is useful in a variety of applications, including the detection of signals buried
in wideband noise.

The *power spectral density* (PSD) of a stationary random process *x*(*n*) is mathematically related to the autocorrelation sequence by the discrete-time
Fourier transform. In terms of normalized frequency, this is given by

$${P}_{xx}(\omega )=\frac{1}{2\pi}{\displaystyle \sum _{m=-\infty}^{\infty}{R}_{xx}(m){e}^{-j\omega m}}.$$

This can be written as a function of physical frequency *f* (for example,
in hertz) by using the relation *ω* = 2*πf* /
*f _{s}*, where

$${P}_{xx}(f)=\frac{1}{{f}_{s}}{\displaystyle \sum _{m=-\infty}^{\infty}{R}_{xx}(m){e}^{-j2\pi mf/{f}_{s}}}.$$

The correlation sequence can be derived from the PSD by use of the inverse discrete-time Fourier transform:

$${R}_{xx}(m)={\displaystyle \underset{-\pi}{\overset{\pi}{\int}}{P}_{xx}(\omega )\text{\hspace{0.17em}}{e}^{j\omega m}d\omega}={\displaystyle \underset{-{f}_{s}/2}{\overset{{f}_{s}/2}{\int}}{P}_{xx}(f)\text{\hspace{0.17em}}{e}^{j2\pi mf/{f}_{s}}\text{\hspace{0.05em}}df}.$$

The average power of the sequence *x*(*n*) over the entire Nyquist interval is represented by

$${R}_{xx}(0)={\displaystyle \underset{-\pi}{\overset{\pi}{\int}}{P}_{xx}(\omega )d\omega ={\displaystyle \underset{-{f}_{s}/2}{\overset{{f}_{s}/2}{\int}}{P}_{xx}(f)df}}.$$

The average power of a signal over a particular frequency band [*ω*_{1}, *ω*_{2}], 0 ≤ *ω*_{1} ≤ *ω*_{2} ≤ *π*, can be found by integrating the PSD over that band:

$${\overline{P}}_{[{\omega}_{1},{\omega}_{2}]}={\displaystyle {\int}_{{\omega}_{1}}^{{\omega}_{2}}{P}_{xx}}(\omega )\text{\hspace{0.17em}}d\omega ={\displaystyle {\int}_{-{\omega}_{2}}^{-{\omega}_{1}}{P}_{xx}}(\omega )\text{\hspace{0.17em}}d\omega .$$

You can see from the above expression that *P _{xx}*(

The units of the PSD are power (e.g., watts) per unit of frequency. In the case of *P _{xx}*(

For real–valued signals, the PSD is symmetric about DC, and thus *P _{xx}*(

The one-sided PSD is given by

$${P}_{\text{one-sided}}(\omega )=\{\begin{array}{ll}0,\hfill & -\pi \le \omega <0,\hfill \\ 2{P}_{xx}(\omega ),\hfill & 0\le \omega \le \pi .\hfill \end{array}$$

The average power of a signal over the frequency band, [*ω*_{1},*ω*_{2}] with 0 ≤ *ω*_{1} ≤ *ω*_{2} ≤ *π*, can be computed using the one-sided PSD as

$${\overline{P}}_{[{\omega}_{1},{\omega}_{2}]}={\displaystyle {\int}_{{\omega}_{1}}^{{\omega}_{2}}{P}_{\text{one-sided}}}(\omega )d\omega .$$

The various methods of spectrum estimation available in the toolbox are categorized as follows:

Nonparametric methods

Parametric methods

Subspace methods

*Nonparametric methods* are those in which the PSD is
estimated directly from the signal itself. The simplest such method is the
*periodogram*. Other nonparametric techniques such as *Welch's
method* [8], the *multitaper method
*(*MTM*) reduce the variance of the periodogram.

*Parametric methods* are those in which the PSD is
estimated from a signal that is assumed to be output of a linear system driven by white noise.
Examples are the *Yule-Walker autoregressive
*(*AR*)* method* and the *Burg
method*. These methods estimate the PSD by first estimating the parameters
(coefficients) of the linear system that hypothetically generates the signal. They tend to
produce better results than classical nonparametric methods when the data length of the
available signal is relatively short. Parametric methods also produce smoother estimates of the
PSD than nonparametric methods, but are subject to error from model misspecification.

*Subspace methods*, also known as
*high-resolution methods* or *super-resolution methods*,
generate frequency component estimates for a signal based on an eigenanalysis or
eigendecomposition of the autocorrelation matrix. Examples are the multiple signal
classification (*MUSIC*) *method* or the eigenvector
(*EV*) *method*. These methods are best suited for line
spectra — that is, spectra of sinusoidal signals — and are effective in the
detection of sinusoids buried in noise, especially when the signal to noise ratios are low. The
subspace methods do not yield true PSD estimates: they do not preserve process power between the
time and frequency domains, and the autocorrelation sequence cannot be recovered by taking the
inverse Fourier transform of the frequency estimate.

All three categories of methods are listed in the table below with the corresponding
toolbox function names. More information about each function is on the corresponding function
reference page. See Parametric Modeling for details about
`lpc`

and other parametric estimation functions.

**Spectral Estimation Methods/Functions**

Method | Description | Functions |
---|---|---|

Periodogram | Power spectral density estimate | |

Welch | Averaged periodograms of overlapped, windowed signal sections | |

Multitaper | Spectral estimate from combination of multiple orthogonal windows (or “tapers”) | |

Yule-Walker AR | Autoregressive (AR) spectral estimate of a time-series from its estimated autocorrelation function | |

Burg | Autoregressive (AR) spectral estimation of a time-series by minimization of linear prediction errors | |

Autoregressive (AR) spectral estimation of a time-series by minimization of the forward prediction errors | ||

Modified Covariance | Autoregressive (AR) spectral estimation of a time-series by minimization of the forward and backward prediction errors | |

MUSIC | Multiple signal classification | |

Eigenvector | Pseudospectrum estimate |