MATLAB's filtfilt() Algorithm

Question

I'm trying to reproduce a lengthy piece of MATLAB code in another language which has no built-in equivalent of the phaseless filter, filtfilt(). I'm trying to re-cast the function in terms of simple filtering (or convolution) operations so I can easily reproduce it. I understand that this filtering operation is equivalent to forward filtering, followed by reverse filtering, but I'm seeing small differences at the edges of the data. Specifically:

data = [1 1 1 2 2 3 5 7 1 1 1 1 1];
ker = [2 1 1];

a = filtfilt(ker,1,data)

b = fliplr( filter( ker, 1, fliplr( filter(ker, 1, data) ) ) )

% a =
%
%   16   18   21   29   39   57   66   68   42   28   16   16   16
%
% b =
%
%   11   16   21   29   39   57   66   68   42   28   16   12    8

I've tried padding the data with zeros at one or both ends, before one or both of the filtering operations. I think I'm likely missing something obvious, but can't spot it.

I had the same issue with `scipy.signal.filtfilt`. There are two important things `filtfilt` does, in addition to running the filter forward-backward: 1. apply 'odd' padding to at start and end of the signal; 2. determine the initial state `zi` and pass it on to the actual filter. Details can be found in this [simple filtfilt implementation](https://stackoverflow.com/a/52742099). — djvg, Oct 11 '18 at 07:45

score 24 · Answer 1 · edited Nov 25 '16 at 16:56

This is my personal implementation of Matlab's filtfilt function. I have tested it with both FIR and IIR filter coefficients and this implementation gives the same results at MATLAB implementation. I used Eigen library to calculate the initial conditions of the filter exactly as done in MATLAB but if you are willing to implement matrix multiplications and inverse operations, you can get rid of it and the code will just depend on STL's <vector> and <algorithm> headers.

#include <vector>
#include <exception>
#include <algorithm>
#include <Eigen/Dense>

typedef std::vector<int> vectori;
typedef std::vector<double> vectord;

void add_index_range(vectori &indices, int beg, int end, int inc = 1)
{
    for (int i = beg; i <= end; i += inc)
    {
       indices.push_back(i);
    }
}

void add_index_const(vectori &indices, int value, size_t numel)
{
    while (numel--)
    {
        indices.push_back(value);
    }
}

void append_vector(vectord &vec, const vectord &tail)
{
    vec.insert(vec.end(), tail.begin(), tail.end());
}

vectord subvector_reverse(const vectord &vec, int idx_end, int idx_start)
{
    vectord result(&vec[idx_start], &vec[idx_end+1]);
    std::reverse(result.begin(), result.end());
    return result;
}

inline int max_val(const vectori& vec)
{
    return std::max_element(vec.begin(), vec.end())[0];
}

void filter(vectord B, vectord A, const vectord &X, vectord &Y, vectord &Zi)
{
    if (A.empty())
    {
        throw std::domain_error("The feedback filter coefficients are empty.");
    }
    if (std::all_of(A.begin(), A.end(), [](double coef){ return coef == 0; }))
    {
        throw std::domain_error("At least one of the feedback filter coefficients has to be non-zero.");
    }
    if (A[0] == 0)
    {
        throw std::domain_error("First feedback coefficient has to be non-zero.");
    }

    // Normalize feedback coefficients if a[0] != 1;
    auto a0 = A[0];
    if (a0 != 1.0)
    {       
        std::transform(A.begin(), A.end(), A.begin(), [a0](double v) { return v / a0; });
        std::transform(B.begin(), B.end(), B.begin(), [a0](double v) { return v / a0; });
    }

    size_t input_size = X.size();
    size_t filter_order = std::max(A.size(), B.size());
    B.resize(filter_order, 0);
    A.resize(filter_order, 0);  
    Zi.resize(filter_order, 0);
    Y.resize(input_size);

    const double *x = &X[0];
    const double *b = &B[0];  
    const double *a = &A[0];
    double *z = &Zi[0];
    double *y = &Y[0];

    for (size_t i = 0; i < input_size; ++i)
    {
        size_t order = filter_order - 1;
        while (order)
        {
            if (i >= order)
            {
                z[order - 1] = b[order] * x[i - order] - a[order] * y[i - order] + z[order];
            }
            --order;
        }
        y[i] = b[0] * x[i] + z[0];
    }
    Zi.resize(filter_order - 1);
}

void filtfilt(vectord B, vectord A, const vectord &X, vectord &Y)
{
    using namespace Eigen;

    int len = X.size();     // length of input
    int na = A.size();
    int nb = B.size();
    int nfilt = (nb > na) ? nb : na;
    int nfact = 3 * (nfilt - 1); // length of edge transients

    if (len <= nfact)
    {
        throw std::domain_error("Input data too short! Data must have length more than 3 times filter order.");
    }

    // set up filter's initial conditions to remove DC offset problems at the
    // beginning and end of the sequence
    B.resize(nfilt, 0);
    A.resize(nfilt, 0);

    vectori rows, cols;
    //rows = [1:nfilt-1           2:nfilt-1             1:nfilt-2];
    add_index_range(rows, 0, nfilt - 2);
    if (nfilt > 2)
    {
        add_index_range(rows, 1, nfilt - 2);
        add_index_range(rows, 0, nfilt - 3);
    }
    //cols = [ones(1,nfilt-1)         2:nfilt-1          2:nfilt-1];
    add_index_const(cols, 0, nfilt - 1);
    if (nfilt > 2)
    {       
        add_index_range(cols, 1, nfilt - 2);
        add_index_range(cols, 1, nfilt - 2);
    }
    // data = [1+a(2)         a(3:nfilt)        ones(1,nfilt-2)    -ones(1,nfilt-2)];

    auto klen = rows.size();
    vectord data;
    data.resize(klen);
    data[0] = 1 + A[1];  int j = 1;
    if (nfilt > 2)
    {
        for (int i = 2; i < nfilt; i++)
            data[j++] = A[i];
        for (int i = 0; i < nfilt - 2; i++)
            data[j++] = 1.0;
        for (int i = 0; i < nfilt - 2; i++)
            data[j++] = -1.0;
    }

    vectord leftpad = subvector_reverse(X, nfact, 1);
    double _2x0 = 2 * X[0];
    std::transform(leftpad.begin(), leftpad.end(), leftpad.begin(), [_2x0](double val) {return _2x0 - val; });

    vectord rightpad = subvector_reverse(X, len - 2, len - nfact - 1);
    double _2xl = 2 * X[len-1];
    std::transform(rightpad.begin(), rightpad.end(), rightpad.begin(), [_2xl](double val) {return _2xl - val; });

    double y0;
    vectord signal1, signal2, zi;

    signal1.reserve(leftpad.size() + X.size() + rightpad.size());
    append_vector(signal1, leftpad);
    append_vector(signal1, X);
    append_vector(signal1, rightpad);

    // Calculate initial conditions
    MatrixXd sp = MatrixXd::Zero(max_val(rows) + 1, max_val(cols) + 1);
    for (size_t k = 0; k < klen; ++k)
    {
        sp(rows[k], cols[k]) = data[k];
    }
    auto bb = VectorXd::Map(B.data(), B.size());
    auto aa = VectorXd::Map(A.data(), A.size());
    MatrixXd zzi = (sp.inverse() * (bb.segment(1, nfilt - 1) - (bb(0) * aa.segment(1, nfilt - 1))));
    zi.resize(zzi.size());

    // Do the forward and backward filtering
    y0 = signal1[0];
    std::transform(zzi.data(), zzi.data() + zzi.size(), zi.begin(), [y0](double val){ return val*y0; });
    filter(B, A, signal1, signal2, zi);
    std::reverse(signal2.begin(), signal2.end());   
    y0 = signal2[0];
    std::transform(zzi.data(), zzi.data() + zzi.size(), zi.begin(), [y0](double val){ return val*y0; });
    filter(B, A, signal2, signal1, zi);
    Y = subvector_reverse(signal1, signal1.size() - nfact - 1, nfact);
}

To test it you could use the following code:

TEST_METHOD(TestFilterAndFiltfilt)
{
    vectord b_coeff = { /* Initialise your feedforward coefficients here */ };
    vectord a_coeff = { /* Initialise your feedback coefficients here */ };

    vectord input_signal = { /* Some input data to be filtered */ };
    vectord y_filter_ori = { /* MATLAB output from calling y_filter_ori = filter(b_coeff, a_coeff, input_signal); */ };
    vectord y_filtfilt_ori = { /* MATLAB output from calling y_filtfilt_ori = filtfilt(b_coeff, a_coeff, input_signal); */ };

    vectord y_filter_out; vectord zi = { 0 };
    filter(b_coeff, a_coeff, input_signal, y_filter_out, zi);
    Assert::IsTrue(compare(y_filter_out, y_filter_ori, 0.0001));

    vectord y_filtfilt_out;
    filtfilt(b_coeff, a_coeff, input_signal, y_filtfilt_out);
    Assert::IsTrue(compare(y_filtfilt_out, y_filtfilt_ori, 0.0001));
}

bool compare(const vectord &original, const vectord &expected, double tolerance = DBL_EPSILON)
{
    if (original.size() != expected.size())
    {
        return false;
    }
    size_t size = original.size();

    for (size_t i = 0; i < size; ++i)
    {
        auto diff = abs(original[i] - expected[i]);
        if (diff >= tolerance)
        {
            return false;
        }
    }
    return true;
}

The result of C++ algorithm is not the same as result of MATLAB. — 劉大為, Jan 04 '22 at 10:53

score 8 · Accepted Answer · answered Jul 16 '13 at 11:22

8

The filtfilt algorithm matches the initial conditions on the filter to minimise start and end transients (from the doc filtfilt). If you type edit filtfilt you can see the code - there is a function getCoeffsAndInitialConditions(b,a,Npts) that will show you the details of this.

answered Jul 16 '13 at 11:22

Hugh Nolan

2,508
13
15

The result of C++ algorithm and MATLAB are not the same. Result of C++ algorithm is as follows. – 劉大為 Jan 04 '22 at 10:47
https://github.com/hdw09/filtfilt/issues/6 (Result of MATLAB and C++ filtfilt() are not the same) – 劉大為 Jan 11 '22 at 13:04
The result of C++ algorithm and MATLAB are not the same. Please refer to https://gitlab.com/libeigen/eigen/-/issues/2450 – 劉大為 Mar 03 '22 at 03:26

MATLAB's filtfilt() Algorithm

2 Answers2

Linked