Using it with e.g. 16MHz crystal:

• use T2 in Programmable ClockOut mode down to 16MHz/4/65536 = cca 62Hz; for the worst case the error is:
  the real frequency f depends on real divider n (f=16MHz/4/n); as f<=1kHz; n<=4000. n is a rounding of ideal divider N=16MHz/4/F (F is the requred frequency) so it differs from it by abs(n-N)<=0.5; we won't bother with abs assuming n<N (it makes no difference, we could make the proof for the other case in the same way), so the maximum possible relative error is (F-f)/f=(16MHz/4/n-16MHz/4/N)/(16MHz/4/n)=(1-n/N)<=1-n/(n+0.5)=0.5/(n+0.5)<0.000125=125ppm.
  The real maximum occurs at 989Hz, has relative value of 121ppm which means absolute error of cca 0.12Hz. Note that the common xxMHz crystals have error (inaccuracy+drift) of 50-100ppm.
  
  
• use an interrupt-driven counter (any, T2 will do) for frequencies below 62Hz. The error given by interrupt latency will be in the order of a couple of us, well below 1 ppm...
  
  
• stop toggling output pin for Fout=0Hz :-)

Of course any other crystal value will do, the higher the better, recalculate as needed.

Am I missing something?

Jan Waclawek