0

I have a radio spectrum (converted from the time domain with FFT). For each bin (frequency), I have 100 samples, taken a couple of seconds apart. These samples are power (e.g. -47.5 dBm).

I am testing for normality using the technique seen here. Presumably channels (man-made radio signals) will be "less random" than the noise floor, which is (supposed to be) Gaussian noise.

When I run normaltest on each frequency array, it returns p < 0.055 the majority of the time (which according to the reference above, means "probably not normal"). This includes many, many frequencies that are part of the noise floor.

Why doesn't this test work well with my setup?

Ken - Enough about Monica
  • 1,595
  • 4
  • 21
  • 50
  • Could you add some data, i.e. some spectrum you are getting? But according to what your are saying I do not know why you think you should get a higher p, since your spectrum samples are far enough from being a pure gaussian. – Paradox Jul 20 '17 at 12:01
  • @Paradox - the (radio) noise floor is Gaussian. – Ken - Enough about Monica Jul 20 '17 at 13:39
  • Yes, but since you did not give the SNR or any other metric and just said "_Presumably channels (man-made radio signals) will be "less random" than the noise floor, which is (supposed to be) Gaussian noise_", it's hard to say what you are getting either 'real-life radio signal' + noise or juste noise... Anyway, the p-value gives you the probability of it being a gaussian-like process, and nothing else. If your distribution is too far from a gaussian process, of course the p-value will be lower than what you are expecting. Maybe you can give a picture of your samples distribution? – Paradox Jul 20 '17 at 13:51

0 Answers0