0
$\begingroup$

Suppose I have discrete noisy signal $X = (0.096, -0.0632, 0.351, 0.531, 0.360, 0.006, -0.320)$ sampled at discrete time points $T = (1, 2, 3, 4, 5, 6, 7)$. Filtering (zero-padded) $X$ with symmetrical FIR $G = (\frac{1}{4}, \frac{1}{2}, \frac{1}{4})$ gives me a good approximation of underlying $f(t)$ but only at discrete time points $T$. Is it possible to use filter coefficients to approximate the values of $f(t)$ between discrete values, e.g. estimate $f(\sqrt 2)$?

In Interpolation with an FIR filter user Hilmar mentions Whittaker–Shannon interpolation formula, but it is sensitive to noise. Is it possible to construct smooth but more noise-robust interpolation by using FIR coefficients somehow?

Or, separate interpolation step (e.g. spline or polynomial interpolation) after FIR application is the only practical way to go?

Improve this question
$\endgroup$

2 Answers 2

Reset to default
1
$\begingroup$

From Hilmar's answer to the linked question, the Whittaker–Shannon interpolation formula is:

$$x(t)=\sum_{n=-\infty }^{\infty }x[n]\cdot \operatorname{sinc}\left(\frac{t-nT}{T}\right).$$

If the signal is low-pass and the noise is high-pass, to filter high-pass noise, we can make a low-pass version of the interpolation formula:

$$x(t)=\sum_{n=-\infty }^{\infty }x[n]\cdot 2f_c/f_s\cdot \operatorname{sinc}\left(\frac{t-2f_c/f_s\cdot nT}{T}\right),$$

where $f_c$ is the cutoff frequency and $f_s$ is the sampling frequency. The highest signal-to-noise ratio is obtained by setting the cutoff frequency to a frequency for which the signal and the noise have equal power. Actually the Whittaker–Shannon interpolation formula is low-pass as well, with $f_c=f_s/2$.

Improve this answer
$\endgroup$
3
  • $\begingroup$ Unfortunately, I don't know $f_c$, only filter coefficients $G$. Is it possible to generalize the formula for arbitrary symmetric $G$? $\endgroup$
    – Andrey Paramonov
    Mar 31, 2016 at 13:58
  • $\begingroup$ You could the -3dB cutoff frequency of your filter with coefficients $G$: $\omega_c = 1.14371774040242049375067,$ solved numerically from $\left|\frac{1}{4}\exp(- i\omega) + \frac{1}{2} + \frac{1}{4}\exp(i \omega)\right| = \sqrt{0.5}.$ Then $f_c/f_s = \omega_c/(2\pi)$ $\endgroup$
    – Olli Niemitalo
    Mar 31, 2016 at 14:33
  • $\begingroup$ ...assuming that your filter (with coefficients $G$) has the same sampling rate as your signal. $\endgroup$
    – Olli Niemitalo
    Mar 31, 2016 at 21:24
0
$\begingroup$

The filter you are applying is a lowpass. So noise is high frequency? You could just do an upsampling of your signal and set the cutoff frequency of the lowpass, that is required for upsampling, to the lowest frequency of the noise. Then you have two tasks in one step: de-noising and interpolating.

Improve this answer
$\endgroup$
1
  • $\begingroup$ Upsampling wouldn't enable me to get $f(t)$ for any real $t$, unfortunately. Also, upsampling would require a lot of memory, while the time points $t$ I'm interested in are sparse. $\endgroup$
    – Andrey Paramonov
    Mar 31, 2016 at 11:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.