frequency divider

Question

How would one implement a frequency divider? I know I can resample up and then treat it as if I'm at the same frequency, however I want to divide the frequency of my signal by 2 and keep the same number of samples.

I've tried multiplying by a cosine wave at half the frequency then low pass filtering. However the problem with this approach is the timing of the output wave matches the signal mixed in and not the original signal. I need the timing to match the original signal, just be at half the frequency.

Answer 1

Going by your last comment, I assume I've hit the task you were trying to accomplish.

This is pretty easy to do in software. We'll call you input signal I and your output signal C

Do the following:

1. C starts out at value zero.
2. Watch the values for I. When I goes from negative to positive ( or crosses some other threshold going up) set C to 1
3. Watch the values for I. When I goes from negative to positive ( or crosses some other threshold going up) set C to 0
4. Repeat steps 2 and 3.

For every sample of I, you get a C that is either 0 or 1. You can send the C value to a DAC or process it further.

This will generate a square wave at half of the frequency of the signal from I.

In an analog system "sharpening the rising edge" (see my earlier comment) would be accomplished by amplifying the signal. Digitally, you can either set a fixed threshold and multiply the incoming values, or lower the threshhold.

The accuracy of the synchronization depends on the sharpness of the rising edge. The sharper the better. With a perfectly vertical edge, the generated clock will be one sample behind the real signal. The slower the rising edge goes up, the more the generated C will lag. At worst you will be 1/2 of a cyle behind, but that takes a really flat rising edge.

Answer 2

The approach described by @JRE is identical to how a digital frequency divider works and has the limitation that it is only edge sensitive (so works well in high SNR conditions but not optimum for lower SNR's where we may have multiple zero crossings as the horizontal axis is transitioned). Here is an approach that can work with every sample and thus be more optimized for lower SNR conditions:

Measure the phase of the tone, determine the sample to sample phase difference (which is proportional to the frequency) and divide that phase difference by two (simple bit-shift) and accumulate that divided phase result to recreate the divide by two tone.

Phase measurements can be done with real or complex tones, and the created divide-by-two tone can be created with a CORDIC Rotator or Numerically Controlled Oscillator (NCO), (or any equivalent process to compute the cosine of the phase argument).

If a phase measurement with complex tones is desired, a Hilbert Transform can be used to convert the real tone in the form $x[n] = A\cos(\omega n +\phi)$ to the complex form $x[n] = Ae^{j(\omega n + \phi)} = I[n]+jQ[n]$).

Create the divided signal by tracking the phase change from sample to sample and dividing that phase change by two. This requires a phase computation of which can be done with $atan2(Q,I)$ or any of multiple optimized phase recovery methods as detailed here, where the phase difference can be detected directly (Note that the difference in phase is directly proportional to frequency), or if cycle time is not a premium the CORDIC algorithm:

phase difference detection

Phase synchronization in BPSK

CORDIC, What is it?

Dividing by two digitally is simply a bit shift. Once the new phase difference is determined, accumulate that to create phase versus time and then use a look-up table to produce the divide by two frequency output. (Note that this structure is done with a Numerically Controlled Oscillator, where optimized look-up tables are used with quarter-cycle storage, etc, or the CORDIC Rotator if cycle time is not a premium). Note how the CORDIC can be used for both determining the original phase step and rotating the divide by two signal at half the rate.

The original signal can be synchronized to the divide by two signal with an additional absolute phase measurement against an arbitrary reference, and therefore an approach that computed the phase difference from the absolute measurements would be preferred.