So I am studying IIR filters and I came to read about limit-cycle oscillations in IIR filters. From what I understand that are two type of oscillations

  • granularity limit-cycle: when signal levels become very low or constant for a period of time, the filter can lock into unstable modes and generate low-level output oscillations. This is causes by quantization errors.

  • overflow limit-cycle: oscillations can appear (as large as the maximum signal range) due to overflow.

Now my question is. Knowing the transfer function of the IIR Filter is there any way to predict if this errors can occur knowing the coefficients of my filter. My guess is that this has to do with the denominator coefficients since this behaviour is exclusive of IIR filters and not FIR filters. Can someone please help me? Thank you!


1 Answer 1


granularity limit-cycle:

Correct. However, limit cycles only occur in fixed point processing. It's not a practical issues for floating point processing.

overflow limit-cycle:

Let's ignore these for now. If you have overflows, you need to fix them anyway, limit cycles or not.

is there any way to predict if this errors can occur knowing the coefficients of my filter

Yes and no. Limit cycles depend A LOT on the exact implementation of the filter. Things that matter the most is the topology of the filter (Direct Form I, Transposed Form I, etc.), the bit depths of the state variables, coefficients and data, and most importantly on the exact way that rounding is implemented. Let's say you multiply two signed 8-bit numbers. The result has 15 bits so you need to throw 7 bits away to store the result in an 8-bit location again. That's typically called "rounding".

There are an astonishing number of ways to implement rounding. For a full overview see https://en.wikipedia.org/wiki/Rounding

The most common ones are

  1. Truncate of "floor": just discard the lower 7 bits. Choose the next smaller number. This is the "default" behavior in many programming languages.
  2. Nearest neighbor: round to the nearest number
  3. Nearest neighbor with zero bias: see Wikipedia, it's a bit complicated.
  4. Round towards zero: Always make sure that the magnitude of the rounded number is smaller than the actual number.

"Round towards zero" is guaranteed to eliminate limit cycles. However it also generates 6 dB more quantization noise then the "nearest neighbor" options. So it's a tradeoff.

Limit cycles are a combination of the exact implementation of the filter PLUS the filter coefficients and (to a smaller extend) the input signal as well. Filter coefficients alone are not enough to predict limit cycles, you will also need a bit-exact simulation of your implementation.


Your Answer

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

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