Assume estimated norm of n-dimension coefficient vector: 2 * n,
where n = m_filter_order.
The absolute values of the real/imaginary part of each coefficient
would not exceed 1.0 under the normal operating condition.
Then the update factor alpha should be proportional to 1/(2 * n).
So let alpha be mu/(2 * n), where 0 < mu < 2.

In paper [2], the old constant alpha = 0.00002 was safe
when n = 401 (200+200 stages) when the IF S/N was 40dB.

We tested mu values of 0.008, 0.004, 0.002, and 0.001
with the transmission result of FM COCOLO 76.5MHz in Osaka.
We concluded that for n = 72 (where m_filter_order is 145):
mu = 0.008: overfitting, which caused long-term increase of the error.
mu = 0.004: steadily fit.
mu = 0.002: steadily fit, resulted in even smaller error than mu = 0.004.
mu = 0.001: underfitting.

[2] Takashi Mochizuki, and Mitsutoshi Hatori, "Automatic Cancelling of FM
Multipath Distortion Using and Adaptive Digital Filter", The Journal of the
Institute of Television Engineers of Japan, Vol. 39, No. 3, pp. 228-234
(1985). https://doi.org/10.3169/itej1978.39.228

