Abstract:
A key problem of low-density parity-check (LDPC) codes analysis is estimation
of an extremely low error floor that occurs at a high level of the signal-to-noise ratio (SNR).
The importance sampling (IS) method is a popular approach to address this problem. Existing
works typically use a normal sampling probability density function (PDF) with shifted mean,
which yields a large variance of the estimate. In contrast, uniform distribution has equally
probable samples on the entire range and thus should reduce the variance, but results in a biased
estimation. This paper proposes a modified IS approach (IS-U) that allows considering the
uniform distribution as a sampling PDF, and shows that this estimation is better than the
traditional one. Also, this paper demonstrates that the existing criteria cannot be applied to
evaluate the accuracy of the IS-U on the whole SNR range. To address this issue, a new metric
is proposed, which uses only the convergence rate and does not depend on the true data.
Fulltext:
PDF file (230 kB)
First page:
PDF file