## What is a practical tolerance

**for convergence?## Cannot be less than round-off error.

## For single-precision (float) accuracy, typically take

**10-6 in__fractional__error.- i.e.
- where
*x*= numerical solution,*xr*= actual root.

## When

*f*(*xr*) = 0 this fails, so use**10-6 as__absolute__error (or perhaps use**(|a| + |b|)/2).