Explainer

Maximum-Entropy Spectral Analysis: What Burg's Method Estimates

Compare an FFT-based periodogram with a Burg autoregressive spectrum, work through a small prediction example, and understand model-order and resolution limits.

On this page

A short recording may contain a recurring pattern that is hard to distinguish in a noisy Fourier spectrum. One response is to fit a model of how each sample depends on previous samples, and inspect the spectrum implied by that model. Burg’s method is one way to fit that model.

The result can show sharp features even from a short record. That sharpness is useful only if the model is a reasonable description of the data. It is not a way to manufacture missing measurements.

First separate the FFT from a spectrum estimate

The discrete Fourier transform (DFT) converts a finite sequence into frequency-domain coefficients. The FFT is an efficient way to compute that transform. A periodogram is a power-spectrum estimate built from the squared magnitude of a windowed DFT, with a chosen normalization.

If a recording has NN samples at sampling frequency FsF_s, the unpadded DFT grid has spacing Fs/NF_s/N. Zero-padding gives more points on the same finite-record transform; it does not add observations. A different window changes leakage and the width of spectral features. Neither a smooth plot nor closely spaced displayed points alone proves that two nearby components can be resolved.

The Signals and Spectra experiment shows how the sampled waveform and its FFT spectrum change together. Burg’s method answers a different question: what spectrum follows from a fitted autoregressive model?

Model the next sample from previous ones

For a real, approximately stationary sequence, an autoregressive model of order pp can be written as

x[n]+∑k=1pakx[n−k]=e[n].x[n] + \sum_{k=1}^{p} a_k x[n-k] = e[n].

Here aka_k are model coefficients and e[n]e[n] is the prediction error, idealized as zero-mean white noise with variance σe2\sigma_e^2. With this sign convention, the corresponding two-sided discrete-time power spectral density is

S(ω)=σe2∣1+∑k=1pake−iωk∣2,−π≤ω≤π.S(\omega) = \frac{\sigma_e^2} {\left|1 + \sum_{k=1}^{p} a_k e^{-i\omega k}\right|^2}, \qquad -\pi \leq \omega \leq \pi.

This convention measures ω\omega in radians per sample and takes variance to be the integral of S(ω)S(\omega) divided by 2π2\pi. Per-hertz and one-sided conventions use different scaling, so compare units and normalization before comparing amplitudes.

The denominator determines the shape of the modeled spectrum. Its peaks are implications of the estimated coefficients, not extra samples of the original signal.

What Burg’s recursion fits

Burg’s method estimates autoregressive coefficients by considering prediction errors in both the forward and backward directions. At each order it finds a reflection coefficient from the current errors, then updates the prediction filter and error sequences. An order-pp fit therefore depends on a deliberate choice of pp as well as the data.

For a short, concrete first step, take the mean-zero real sequence x=[1,2,0,−3]x=[1,2,0,-3]. At order one the forward errors start as x[n]x[n] and backward errors as x[n−1]x[n-1], for n=1,2,3n=1,2,3. With the sign convention above, the first reflection coefficient is

k1=−2∑n=13x[n]x[n−1]∑n=13(x[n]2+x[n−1]2)=−418=−29.k_1 = -\frac{2\sum_{n=1}^{3}x[n]x[n-1]} {\sum_{n=1}^{3}\bigl(x[n]^2+x[n-1]^2\bigr)} = -\frac{4}{18} = -\frac{2}{9}.

The numerator’s unscaled cross-product sum is 22; the denominator is 1818. For this order-one example the filter coefficient is a1=k1a_1=k_1, so its forward prediction is −a1x[n−1]-a_1x[n-1]. The following plain Python snippet checks the arithmetic:

x = [1.0, 2.0, 0.0, -3.0]
cross = sum(x[n] * x[n - 1] for n in range(1, len(x)))
power = sum(x[n] ** 2 + x[n - 1] ** 2 for n in range(1, len(x)))
k1 = -2 * cross / power
print(cross, power, round(k1, 6))  # 2.0 18.0 -0.222222

This four-sample example illustrates one algebraic update only. It is far too short to support a meaningful scientific spectral conclusion, and a real analysis must also estimate the error power. The Python spectrum documentation describes a complete Burg implementation and its input/order requirements.

Why “maximum entropy”?

The maximum-entropy formulation selects, among spectra compatible with specified autocorrelation constraints, the one that maximizes spectral entropy under that formulation. The resulting all-pole autoregressive representation provides a way to extend an estimated correlation structure beyond the lags used as constraints.

This is a statement about a constrained model, not a guarantee that an inferred peak exists in the physical source. Burg’s finite-sample forward/backward error procedure is a practical estimator in this family; its estimates should not be mistaken for exact population autocorrelations.

The order is a modeling decision

ChoiceWhat it can doWhat to watch
Small ppGive a compact, smooth modelMiss structure that needs more parameters
Larger ppFit more complicated short-range dependenceFit noise, split peaks, or produce unstable-looking estimates
Longer or repeated recordsProvide more evidence about persistent featuresRequire comparable conditions and correct sampling

Model-selection criteria such as AIC can help compare candidate orders; they do not certify a uniquely correct model. Check whether a feature persists across plausible orders, independent recordings and a non-parametric estimate such as a windowed periodogram or Welch estimate. Remove an unwanted mean or trend deliberately, and document the sampling rate.

A peak from an AR fit may be much narrower than an FFT periodogram peak when the signal agrees with the model. The opposite is also possible: a narrow model peak can be sensitive to order, noise or nonstationarity. Do not call it a newly resolved physical component without checking those alternatives.

Which one should I use?

Start with a windowed FFT-based estimate when you want a relatively direct view of the observed finite record and its leakage. Try Burg when an autoregressive model is plausible, a short record makes a parametric comparison useful, and you can test the effect of changing model order.

They are not competing definitions of truth. The best practice is to show the data and assumptions behind each estimate, compare like units, and treat agreement across methods and independent records as stronger evidence than a single sharp curve.

Sources