// Date of creation: 20 Jan, 2016 function [w,pow] = rootmusic(x,p,varargin) // Frequencies and power of sinusoids using the root MUSIC algorithm // // Calling Sequence // w = rootmusic(x,p) // [w,pow] = rootmusic(x,p) // [f,pow] = rootmusc(...,fs) // [w,pow] = rootmusic(...,'corr') // // Parameters // x - int|double - vector|matrix // Input signal. // If x is a vector, then it reprsenets one realization of the signal. // If x is a matrix, then each row represents a separate observation of // the signal. // p - int|double - scalar|2 element vector // p(1) is the signal subspace dimension and hence the number of // complex exponentials in x. // p(2), if specified, represents a threshold that is multiplied by // the smallest estimated eigenvalue of the signal's correlation // matrix. // fs - int|double - scalar // Sampling frequency (in Hz) // If fs is specified by an empty vector or unspecified, it defaults // to 1 Hz // 'corr' flag // If specified, x is interpreted as a correlation matrix rather than // a matrix of the signal data. For x to be a correlation matrix, // x must be a square matrix and all its eigenvalues must be // nonnegative // // Examples: // 1) 3 complex exponentials: // // n=0:99; // s=exp(1*%i*%pi/2*n)+2*exp(1*%i*%pi/4*n)+exp(1*%i*%pi/3*n)+rand(1,100,"normal"); // [A,R]=corrmtx(s,12,'mod'); // [W,P] = rootmusic(R,3,'corr'); // //2) // n=0:99; // s=exp(1*%i*%pi/2*n)+2*exp(1*%i*%pi/4*n)+exp(1*%i*%pi/3*n); // [A,R]=corrmtx(s,12,'mod'); // [W,P] = rootmusic(R,3,'corr'); //EXPECTED OUTPUT: //W = 0.7738111 1.5690374 1.0426234 //P =377.4255 103.18124 123.86659 // // Author // Ayush // // See also // corrmtx | peig | pmusic | rooteig // // References // 1) Monson H. Hayes, Statistical Digital Signal Processing And Modeling, // Wiley & Sons, Inc, [Section 8.6.3] // // // Output arguments // w - double - vector // Estimated frequencies of the complex sinusoids // pow - double - vector // estimated absolute value squared amplitudes of the sinusoids at // the frequencies w // funcprot(0); // **** checking the number of input and output arguments **** [numOutArgs, numInArgs] = argn(0); if numOutArgs~=1 & numOutArgs~=2 then error(78,"rootmusic"); end if numInArgs<1 | numInArgs>4 then error(77,"rootmusic"); end // **** parsing the input arguments **** isFsSpecified = %F; fs = []; varargLength = length(varargin); // searching for the 'corr' flag isCorrFlag = %F; if varargLength==0 then stringIndices = []; else stringIndices = find(type(varargin(1:varargLength))==10); end if ~isempty(stringIndices) then // ignoring all other strings except the corr flag isCorrFlag = or(strcmpi(varargin(stringIndices),"corr")==0); varargin(stringIndices) = []; end // varargin can have only an entry for fs if length(varargin)==1 then fs = varargin(1); if length(fs)==1 then if ~IsIntOrDouble(fs, %T) then msg = "rootmusic: Wrong type for argument #4 (fs); Positive scalar expected"; error(msg,10084); end fs = double(fs); isFsSpecified = %T; elseif length(fs)>1 then msg = "rootmusic: Wrong type for argument #4 (fs); Positive scalar expected"; error(msg,10084); end elseif length(varargin)>1 then msg = "rootmusic: Wrong type for argument #4 (fs); Positive scalar expected"; error(msg,10084); end // extracting primary input x/R primaryInput = x; if ndims(primaryInput)<1 | ndims(primaryInput)>2 then msg = "rootmusic: Wrong dimension for argument #1; Vector or a matrix expected"; error(msg,10053); end if ~IsIntOrDouble(primaryInput, %F) then msg = "rootmusic: Wrong type for argument #1; Numeric vector or a matrix expected"; error(msg,10053); end // covert to a column vector if ndims(primaryInput)==1 then primaryInput = primaryInput(:); end // casting to double primaryInput = double(primaryInput); //****extracting p**** // p must be either scalar or a 2-element vector if length(p)~=1 & length(p)~=2 then msg = "rootmusic: Wrong type for argument #2 (p); " + ... "A scalar or a 2-element vector expected"; error(msg,10053); end // first argument of p must be an integer if ~IsIntOrDouble(p(1),%T) then msg = "rootmusic: Wrong input argument #2 p(1); " + ... "positive integer expected"; error(msg,10036); return end p(1) = int(p(1)); // TODO: check if positive required // 2nd argument, if exists, must be a positive integer' if length(p)==2 then if ~IsIntOrDouble(p(2),%F) then msg = "rootmusic: Wrong type for argument #2 p(2); must be a scalar"; error(msg,10053); end end isXReal = isreal(x) if ~isCorrFlag then // check that p(1) should be even if x is real if isXReal & modulo(p(1),2)~=0 then msg = "rootmusic: Wrong input argument #2 p(1); " + ... " An even value expected for real input x"; error(msg,10036); end end // **** calling pmusic **** data= struct(); data.x = primaryInput; data.p = p; data.nfft = 256; data.w = []; data.fs = fs; data.isWindowSpecified = %F; data.windowLength = 2*p(1); data.windowVector = []; data.noverlap = []; data.isCorrFlag = isCorrFlag; data.isFsSpecified = isFsSpecified; data.freqrange = "twosided"; [outData,msg] = musicBase(data); if length(msg)~=0 then // throw error msg = "rootmusic: "+msg error(msg); end pEffective = outData.pEffective; eigenvals = outData.eigenvals; w = computeFreqs(outData.noiseEigenvects,pEffective,%f,eigenvals); if isempty(w) then // assign all frequency and powers as -nan w = %nan*(1:pEffective)'; pow = w; return; end // **** Estimating the variance of the noise **** // Estimate is the mean of the eigenvalues belonging to the noise subspace sigma_noise = mean(eigenvals(pEffective+1:$)); pow = computePower(outData.signalEigenvects,eigenvals,w,pEffective,... sigma_noise,isXReal); // is fs is specified, convert normailized frequencies to actual frequencies if isFsSpecified then w = w*fs/(2*%pi); end endfunction function w = computeFreqs(noiseEigenvects,pEffective,EVFlag,eigenvals) // Computes the frequencies of the complex sinusoids using the roots of // the polynomial formed with the noise eigenvectors // // Parameters // noiseEigenvects - // A matrix where noise eigenvectors are represented by each column // pEffective - // The effective dimension of the signal subspace // EVFlag - // Flag to indicate weighting to be used for rooteig // eigenvals - // Eigenvals of the correlation matrix // // Output arguments // w - // A vector with frequencies of the complex sinusoids numOfNoiseEigenvects = size(noiseEigenvects,2); if EVFlag then // weights are the eigenvalues in the noise subspace weights = eigenvals($-numOfNoiseEigenvects+1:$); else weights = ones(numOfNoiseEigenvects,1); end // Form a polynomial consisting of a sum of polynomials given by the // product of the noise subspace eigenvectors and the reversed and // conjugated version. (eq 8.163 from [1]) D = 0; for i=1:numOfNoiseEigenvects eigenvect = noiseEigenvects(:,i); D = D + conv(eigenvect,conj(eigenvect($:-1:1)))./weights(i); end roots = roots(D); // selecting the roots inside the unit circle rootsSelected = roots(abs(roots)<1); // sort the roots in order of increasing distance from the unit circle [dist,indices] = gsort(abs(rootsSelected)-1); sortedRoots = rootsSelected(indices); if isempty(sortedRoots) then w = []; else w = atan(imag(sortedRoots(1:pEffective)),real(sortedRoots(1:pEffective))); end endfunction function power = computePower(signalEigenvects,eigenvals,w,pEffective,... sigma_noise,isXReal) if isXReal then // removing the negative frequencies as sinusoids will be present in // complex conjugate pairs w = w(w>=0); pEffective = length(w); end // Solving eq. 8.160 from [1] (Ap = b) where p is the power matrix A = zeros(length(w),pEffective); for i=1:pEffective A(:,i) = computeFreqResponseByPolyEval(signalEigenvects(:,i), ... w,1,%F); end A = (abs(A).^2)'; b = eigenvals(1:pEffective) - sigma_noise; // Solving Ap=b with the constraint that all elements of p >=0 power = nnls(A,b+A*sqrt(%eps)*ones(pEffective,1)); endfunction function h = computeFreqResponseByPolyEval(b,f,fs,isFsSpecified) // returns the frequency response (h) for a digital filter with numerator b. // The evaluation of the frequency response is done at frequency values f f = f(:); b = b(:); if isFsSpecified then // normalizing the f vector w = f*2*%pi/fs; else w = f; end n = length(b); powerMatrix = zeros(length(f),n); powerMatrix(:,1) = 1; for i=2:n powerMatrix(:,i) = exp(w*(-i+1)*%i); end h = powerMatrix*b; endfunction function result = IsIntOrDouble(inputNum, isPositiveCheck) // Checks if The Input Is Integer Or Double // Also Checks if It Is Greater Than 0 For IsPositiveCheck = True if ~(type(inputNum)==1 | type(inputNum)==8) then result = %F; return end if isPositiveCheck & or(inputNum<=0) then result = %F; return end result = %T; return endfunction