Signal sawtooth decomposition

Question

I'm interested in analyzing a windowed signal in terms of a weighted sum of sawtooth waves. Being a somewhat DSP newbie, is there an obvious approach to this? I realize that sawtooth waves don't form an orthogonal basis like sinusoids do, so the problem is probably a bit ill-defined. A least squares solution or some other relaxation would probably suffice.

I could imagine constructing a matrix of sawtooth waves of different fundamental frequencies (cf. DFT/DCT matrix) but I'm not sure e.g. how phase would affect the analysis.

I could also consider low-passing + comb filtering the signal for each fundamental frequency, but constructing an array of such non over-lapping filters would seem cumbersome and inefficient.

Also doing a Fourier transform and running an analysis over all fundamentals and their harmonics would seem feasible, but clumsy.

Is there a go-to solution to a problem like this?

Answer 1

Assuming you're doing this in discrete time, you can find the least-squares fit for any collection of functions, orthogonal or not.

The following is straight out of linear algebra:

Let $$\mathbf X = \begin{bmatrix} f_1(k_1) & f_2(k_1) & \cdots & f_N(k_1)\\ f_1(k_2) & f_2(k_2) & \cdots & f_N(k_2) \\ \vdots & \vdots & \ddots & \vdots \\ f_1(k_M) & f_2(k_M) & \cdots & f_N(k_M) \end{bmatrix} \tag 1$$

where $f_n$ is any function of $k$

Then if $\mathbf y = \begin{bmatrix}y(k_1) & y(k_2) & \cdots & y(k_M)\end{bmatrix}^T$ is known and you find the $\mathbf b$ that minimizes the Euclidean distance between $\mathbf y$ and

$$\hat {\mathbf y} = \mathbf X \mathbf b \tag 2$$

you have the weights for your "$y$ is a weighted sum of $f$".

• In the case of the DFT, if you let $k_n = n - 1$, $N = M$ and $f_n(k) = e^{-j 2 \pi \frac{n - 1}{N}k}$, and poof -- you have the DFT.
• In the case of $N \le M$, you really want $\mathbf X$ to have rank $M$. Moreover, you would like it to be well conditioned numerically (i.e., to have eigenvalues that are all roughly the same magnitude). If you choose the same definition for $f_n$ as the DFT case, then you're doing a truncated DFT and (essentially) finding the coefficients for a low-pass filtered version of your $\mathbf y$.

I'd make up a set of $f_n$ corresponding to your triangle wave, it's "90 degree" offset, and all their harmonics. Then I'd see if the resulting $\mathbf X$ matrix was full rank and well conditioned. If it is -- try it out, see how things work.