Jun 152014

Everything that happens in the world can be described in some way. Our descriptions range from informal and causal to precise and scientific, yet ultimately they all share one underlying characteristic: they carry an abstract idea known as “information” about what is being described. In building complex systems, whether out of people or machines, information sharing is central for building cooperative solutions. However, in any system, the rate at which information can be shared is limited. For example, on Twitter, you’re limited to 140 characters per message. With 802.11g you’re limited to 54 Mbps in ideal conditions. In mobile devices, the constraints go even further: transmitting data on the network requires some of our limited bandwidth and some of our limited energy from the battery.

Obviously this means that we want to transmit our information as efficiently as possible, or, in other words, we want to transmit a representation of the information that consumes the smallest amount of resources, such that the recipient can convert this representation back into a useful or meaningful form without losing any of the information. Luckily, the problem has been studied pretty extensively over the past 60-70 years and the solution is well known.

First, it’s important to realize that compression only matters if we don’t know exactly what we’re sending or receiving beforehand. If I knew exactly what was going to be broadcast on the news, I wouldn’t need to watch it to find out what happened around the world today, so nothing would need to be transmitted in the first place. This means that in some sense, information is a property of things we don’t know or can’t predict fully, and it represents the portion that is unknown. In order to quantify it, we’re going to need some math.

Let’s say I want to tell you what color my car is, in a world where there are only four colors: red, blue, yellow, and green. I could send you the color as an English word with one byte per letter, which would require 3, 4, 5, or 6 bytes, or we could be cleverer about it. Using a pre-arranged scheme for all situations where colors need to be shared, we agree on the convention that the binary values 00, 01, 10, and 11 map to our four possible colors. Suddenly, I can use only two bits (0.25 bytes, far more efficient) to tell you what color my car is, a huge improvement. Generalizing, this suggests that for any set \chi of abstract symbols (colors, names, numbers, whatever), by assigning each a unique binary value, we can transmit a description of some value from the set using \log_2(|\chi|) bits on average, if we have a pre-shared mapping. As long as we use the mapping multiple times it amortizes the initial cost of sharing the mapping, so we’re going to ignore it from here out. It’s also worthwhile to keep this limit in mind as a max threshold for “reasonable;” we could easily create an encoding that is worse than this, which means that we’ve failed quite spectacularly at our job.

But, if there are additional constraints on which symbols appear, we should probably be able to do better. Consider the extreme situation where 95% of cars produced are red, 3% blue, and only 1% each for yellow and green. If I needed to transmit color descriptions for my factory’s production of 10,000 vehicles, using the earlier scheme I’d need exactly 20,000 bits to do so by stringing together all of the colors in a single sequence. But, given that by the law of large numbers, I can expect roughly 9,500 cars to be red, so what if I use a different code, where red is assigned the bit string 0, blue is assigned 10, yellow is assigned 110, and green 111? Even though the representation for two of the colors is a bit longer in this scheme, the total average encoding length for a lot of 10,000 cars decreases to 10,700 bits (1*9500 + 2*300 + 3*100 + 3*100), almost an improvement of 50%! This suggests that the probabilities for each symbol should impact the compression mapping, because if some symbols are more common than others, we can make them shorter in exchange for making less common symbols longer and expect the average length of a message made from many symbols to decrease.

So, with that in mind, the next logical question is, how well can we do by adapting our compression scheme to the probability distribution for our set of symbols? And how do we find the mapping that achieves this best amount of compression? Consider a sequence of n independent, identically distributed symbols taken from some source with known probability mass function p(X=x), with S total symbols for which the PMF is nonzero. If n_i is the number of times that the i th symbol in the alphabet appears in the sequence, then by the law of large numbers we know that for large n it converges almost surely to a specific value: \Pr(n_i=np_i)\xrightarrow{n\to \infty}1 .

In order to obtain an estimate of the best possible compression rate, we will use the threshold for reasonable compression identified earlier: it should, on average, take no more than approximately \log_2(|\chi|) bits to represent a value from a set \chi , so by finding the number of possible sequences, we can bound how many bits it would take to describe them. A further consequence of the law of large numbers is that because \Pr(n_i=np_i)\xrightarrow{n\to \infty}1 we also have \Pr(n_i\neq np_i)\xrightarrow{n\to \infty}0 . This means that we can expect the set of possible sequences to contain only the possible permutations of a sequence containing n_i realizations of each symbol. The probability of a specific sequence X^n=x_1 x_2 \ldots x_{n-1} x_n can be expanded using the independence of each position and simplified by grouping like symbols in the resulting product:

P(x^n)=\prod_{k=1}^{n}p(x_k)=\prod_{i=1}^{S} p_i^{n_i}=\prod_{i=1}^{S} p_i^{np_i}

We still need to find the size of the set \chi in order to find out how many bits we need. However, the probability we found above doesn’t depend on the specific permutation, so it is the same for every element of the set and thus the distribution of sequences within the set is uniform. For a uniform distribution over a set of size |\chi| , the probability of a specific element is \frac{1}{|\chi|} , so we can substitute the above probability for any element and expand in order to find out how many bits we need for a string of length n:


Frequently, we’re concerned with the number of bits required per symbol in the source sequence, so we divide B(n) by n to find H(X) , a quantity known as the entropy of the source X , which has PMF P(X=x_i)=p_i :

H(X) = -\sum_{i=1}^Sp_i\log_2(p_i)

The entropy, H(X) , is important because it establishes the lower bound on the number of bits that is required, on average, to accurately represent a symbol taken from the corresponding source X when encoding a large number of symbols. H(X) is non-negative, but it is not restricted to integers only; however, achieving less than one bit per symbol requires multiple neighboring symbols to be combined and encoded in groups, similarly to the method used above to obtain the expected bit rate. Unfortunately, that process cannot be used in practice for compression, because it requires enumerating an exponential number of strings (as a function of a variable tending towards infinity) in order to assign each sequence to a bit representation. Luckily, two very common, practical methods exist, Huffman Coding and Arithmetic Coding, that are guaranteed to achieve optimal performance.

For the car example mentioned earlier, the entropy works out to about 0.35 bits, which means there is significant room for improvement over the symbol-wise mapping I suggested, which only achieved a rate of 1.07 bits per symbol, but it would require grouping multiple car colors into a compound symbol, which quickly becomes tedious when working by hand. It is kind of amazing that using only ~3,500 bits, we could communicate the car colors that naively required 246,400 bits (=30,800 bytes) by encoding each letter of the English word with a single byte.

H(X) also has other applications, including gambling, investing, lossy compression, communications systems, and signal processing, where it is generally used to establish the bounds for best- or worst-case performance. If you’re interested in a more rigorous definition of entropy and a more formal derivation of the bounds on lossless compression, plus some applications, I’d recommend reading Claude Shannon’s original paper on the subject, which effectively created the field of information theory.

Jul 062013

The Wiener filter is well known as the optimal solution to the problem of estimating a random process when it is corrupted by another additive process, using only a linear combination of values of the measured process. Mathematically, this means that the Wiener filter constructs an estimator of some original signal x(t) given z(t)=x(t)+n(t) with the property that \|\hat{x}(t)-x(t)\| is minimized among all such linear estimators, assuming only that both x(t) and n(t) are stationary and have known statistics (mean, variance, power spectral density, etc.). When more information about the structure of x(t) is known, different estimators may be easier to implement (such as a Kalman filter for signals with a recursive structure).

Such a filter is very powerful—it is optimal, after all—when the necessary statistics are available and the input signals meet the requirements, but in practice, signals of interest are never stationary (rarely even wide sense stationary, although it is a useful approximation), and their statistics change frequently. Rather than going through the derivation of the filter, which is relatively straightforward and available on Wikipedia (linked above), I’d like to talk about how to adapt it to situations that do not meet the filter’s criteria and still obtain high quality results, and then provide a demonstration on one such signal.

The first problem to deal with is the assumption that a signal is stationary. True to form for engineering, the solution is to look at only a brief portion of the signal and approximate it as stationary. This has the unfortunate consequence of preventing us from defining the filter once and reusing it; instead, as the measured signal is sliced into approximately stationary segments, we must estimate the relevant statistics and construct an appropriate filter for each segment. If we do the filtering in the frequency domain, then for segments of length N we are able to do the entire operation with two length N FFTs (one forward and one reverse) and O(N) arithmetic operations (mostly multiplication and division). This is comparable to other frequency domain filters and much faster than the O(N^2) number of operations required for a time domain filter.

This approach creates a tradeoff. Because the signal is not stationary, we want to use short time slices to minimize changes. However, the filter operates by adjusting the amplitude of each bin after a transformation to the frequency domain. Therefore, we want as many bins as possible to afford the filter high resolution. Adjusting the sampling rate does not change the frequency resolution for a given amount of time, because the total time duration of any given buffer is f_{s}N. So, for fixed time duration, the length of the buffer will scale inversely with the sampling rate, and the bin spacing in an FFT will remain constant. The tradeoff, then, exists between how long each time slice will be and how much change in signal parameters we wish to tolerate. A longer time slice weakens the stationary approximation, but it also produces better frequency resolution. Both of these affect the quality of the resulting filtered signal.

The second problem is the assumption that the statistics are known beforehand. If we’re trying to do general signal identification, or simply “de-noising” of arbitrary incoming data (say, for sample, cleaning up voice recorded from a cell phone’s microphone in windy area, or reducing the effects of thermal noise in a data acquisition circuit), then we don’t know what the signal will look like beforehand. The solution here is a little bit more subtle. The normal formulation of the Wiener filter, in the Laplace domain, is

G(s)= \frac{S_{z,x}(s)}{S_{z}(s)}
\hat{X}(s)=G(s) Z(s)

In this case we assume that the cross-power spectral density, S_{z,x}(s), between the measured process z(t) and the true process x(t) is known, and we assume that the power spectral density, S_{z}(s), of the measured process z(t) is known. In practice, we will estimate S_{z}(s) from measured data, but as the statistics of x(t) are unknown, we don’t know what S_{z,x}(s) is (and can’t measure it directly). But, we do know the statistics of the noise. And, by (reasonable) assumption, the noise and the signal of interest are independent. Therefore, we can calculate several related spectra and make some substitutions into the definition of the original filter.


If we substitute these into the filter definition to eliminate S_x(s), then we are able to construct and approximation of the filter based on the (known) noise PSD and an estimate of the signal PSD (if the signal PSD were known, it’d be exact, but as our PSD estimate contains errors, the filter definition will also contain errors).


You may ask: if we don’t know the signal PSD, how can we know the noise PSD? Realistically, we can’t. But, because the noise is stationary, we can construct an experiment to measure it once and then use it later. Simply identify a time when it is known that there is no signal present (i.e. ask the user to be silent for a few seconds), measure the noise, and store it as the noise PSD for future use. Adaptive methods can be used to further refine this approach (but are a topic for another day). It is also worth noting that the noise does not need to be Gaussian, nor does it have any other restrictions on its PSD. It only needs to be stationary, additive, and independent of the signal being estimated. You can exploit this to remove other types of interference as well.

One last thing before the demo. Using the PSD to construct the filter like this is subject to a number of caveats. The first is that the variance of each bin in a single PSD estimate is not zero. This is an important result whose consequences merit more detailed study, but the short of it is that the variance of each bin is essentially the same as the variance of each sample from which the PSD was constructed. A remedy for this is to use a more sophisticated method for estimating the PSD by combining multiple more-or-less independent estimates, generally using a windowing function. This reduces the variance and therefore improves the quality of the resulting filter. This, however, has consequences related to the trade-off between time slice length and stationary approximation. Because you must average PSDs computed from (some) different samples in order to reduce the variance, you are effectively using a longer time slice.

Based on the assigned final project in ECE 4110 at Cornell University, which was to use a Wiener filter to de-noise a recording of Einstein explaining the mass-energy equivalence with added Gaussian white noise of unknown power, I’ve put together a short clip comparing the measured (corrupted) signal, the result after filtering with a single un-windowed PSD estimate to construct the filter, and the result after filtering using two PSD estimates with 50% overlap (and thus an effective length of 1.5x the no-overlap condition) combined with a Hann window to construct the filter. There is a clear improvement in noise rejection using the overlapping PSD estimates, but some of the short vocal transitions are also much more subdued, illustrating the tradeoff very well.

Be warned, the first segment (unfiltered) is quite loud as the noise adds a lot of output power.

Here is the complete MATLAB code used to implement the non-overlapping filter

% Assumes einsteindistort.wav has been loaded with
[d,r] = wavread('EinsteinDistort.wav');

% Anything that can divide the total number of samples evenly
sampleSize = 512;

% Delete old variables
% clf;
clear input;
clear inputSpectrum;
clear inputPSD;
clear noisePSD;
clear sampleNoise;
clear output;
clear outputSpectrum;
clear weinerCoefficients;

% These regions indicate where I have decided there is a large amount of
% silence, so we can extract the noise parameters here.
noiseRegions = [1 10000;
                81000 94000;
                149000 160000;
                240000 257500;
                347500 360000;
                485000 499000;
                632000 645000;
                835000 855000;
                917500 937500;
                1010000 1025000;
                1150000 116500];

% Now iterate over the noise regions and create noise start offsets for
% each one to extract all the possible noise PSDs
noiseStarts = zeros(length(noiseRegions(1,:)), 1);
z = 1;
for k = 1:length(noiseRegions(:,1))
    for t = noiseRegions(k,1):sampleSize:noiseRegions(k,2)-sampleSize
        noiseStarts(z) = t;
        z = z + 1;

% In an effort to improve the PSD estimate of the noise, average the FFT of
% silent noisy sections in multiple parts of the recording.
noisePSD = zeros(sampleSize, 1);
for n = 1:length(noiseStarts)
    sampleNoise = d(noiseStarts(n):noiseStarts(n)+sampleSize-1);
    noisePSD = noisePSD + (2/length(noiseStarts)) * abs(fft(sampleNoise)).^2 / sampleSize;

% Force the PSD to be flat like white noise, for comparison
% noisePSD = ones(size(noisePSD))*mean(noisePSD);

% Now, break the signal into segments and try to denoise it with a
% noncausal weiner filter.
output = zeros(1, length(d));
for k = 1:length(d)/sampleSize
    input = d(1+sampleSize*(k-1):sampleSize*k);
    inputSpectrum = fft(input);
    inputPSD = abs(inputSpectrum).^2/length(input);
    weinerCoefficients = (inputPSD - noisePSD) ./ inputPSD;
    weinerCoefficients(weinerCoefficients < 0) = 0;
    outputSpectrum = inputSpectrum .* weinerCoefficients;

    % Sometimes for small outputs ifft includes an imaginary value
    output(1+sampleSize*(k-1):sampleSize*k) = real(ifft(outputSpectrum, 'symmetric'));

% Renormalize and write to a file
output = output/max(abs(output));
wavwrite(output, r, 'clean.wav');

To convert this implementation to use 50% overlapping filters, replace the filtering loop (below "Now, break the signal into segments…") with this snippet:

output = zeros(1, length(d));
windowFunc = hann(sampleSize);
while sampleSize*(k-1)/2 + sampleSize < length(d)
    input = d(1+sampleSize*(k-1)/2:sampleSize*(k-1)/2 + sampleSize);
    inputSpectrum = fft(input .* windowFunc);
    inputPSD = abs(inputSpectrum).^2/length(input);
    weinerCoefficients = (inputPSD - noisePSD) ./ inputPSD;
    weinerCoefficients(weinerCoefficients < 0) = 0;
    outputSpectrum = inputSpectrum .* weinerCoefficients;

    % Sometimes for small outputs ifft includes an imaginary value
    output(1+sampleSize*(k-1)/2:sampleSize*(k-1)/2 + sampleSize) = output(1+sampleSize*(k-1)/2:sampleSize*(k-1)/2 + sampleSize) + ifft(outputSpectrum, 'symmetric')';
    k = k +1;

The corrupted source file used for the project can be downloaded here for educational use.

This can be adapted to work with pretty much any signal simply by modifying the noiseRegions matrix, which is used to denote the limits of "no signal" areas to use for constructing a noise estimate.