i
i
i
i
i
i
i
i
206 9. Signal Processing
overshoot
overshoot
Figure 9.26. A filter with negative lobes will
always produce some overshoot when filtering
or reconstructing a sharp discontinuity.
A filter that takes on negative
values has ringing or overshoot:it
will produce extra oscillations in the
value around sharp changes in the
value of the function being filtered.
For instance, the Catmull-Rom
filter has negative lobes on either
side, and if you filter a step function
with it, it will exaggerate the step a bit, resulting in function values that under-
shoot 0 and overshoot 1 (Figure 9.26).
A continuous filter is ripple free if, when used as a reconstruction filter, it
will reconstruct a constant sequence as a constant function (Figure 9.27). This is
equivalent to the requirement that the filter sum to one on any integer-spaced grid:
i
f(x + i)=1 for all x.
A continuous filter has a degree of continuity, which is the highest-order
derivative that is defined everywhere. A filter, like the box filter, that has sud-
den jumps in its value is not continuous at all. A filter that is continuous but
has sharp corners (discontinuities in the first derivative), such as the tent filter,
has order of continuity zero, and we say it is C
0
.Afilter that has a continuous
derivative (no sharp corners), such as the piecewise cubic filters in the previous
section, is C
1
; if its second derivative is also continuous, as is true of the B-spline
filter, it is C
2
. The order of continuity of a filter is particularly important for a
reconstruction filter because the reconstructed function inherits the continuity of
the filter.
ripple-free
not ripple-free
Σ = 1
Σ ≠ 1
Figure 9.27. The tent filter of radius 1 is a ripple-free reconstruction filter; the Gaussian
filter with standard deviation 1/2 is not.