In the ideal vibrating string, the only restoring force for transverse displacement comes from the string tension (§C.1 above); specifically, the transverse restoring force is equal the net transverse component of the axial string tension. Consider in place of the ideal string a bundle of ideal strings, such as a stranded cable. When the cable is bent, there is now a new restoring force arising from some of the fibers being compressed and others being stretched by the bending. This force sums with that due to string tension. Thus, stiffness in a vibrating string introduces a new restoring force proportional to bending angle. It is important to note that string stiffness is a linear phenomenon resulting from the finite diameter of the string.
In typical treatments,C.3bending stiffness adds a new term to the wave equation that is proportional to the fourth spatial derivative of string displacement:
To solve the stiff wave equation Eq.(C.32), we may set to get
At very low frequencies, or when stiffness is negligible in comparison with , we obtain again the non-stiff string: .
At very high frequencies, or when the tension is negligible relative to , we obtain the ideal bar (or rod) approximation:
In an ideal bar, the only restoring force is due to bending stiffness. Setting gives solutions and . In the first case, the wave velocity becomes proportional to . That is, waves travel faster along the ideal bar as oscillation frequency increases, going up as the square root of frequency. The second solution corresponds to a change in the wave shape which prevents sharp corners from forming due to stiffness [95,118].
At intermediate frequencies, between the ideal string and the ideal bar,
the stiffness contribution can be treated as a correction term
. This is the region of most practical interest because
it is the principal operating region for strings, such as piano strings,
whose stiffness has audible consequences (an inharmonic, stretched overtone
Setting as before, corresponding to driving the medium sinusoidally over time at frequency , the medium response is
Because the effective wave velocity depends on , we cannot use Fourier's theorem to construct arbitrary traveling shapes by superposition. At , we can construct any function of time, but the waveshape disperses as it propagates away from . The higher-frequency Fourier components travel faster than the lower-frequency components.
Since the temporal and spatial sampling intervals are related by , this must generalize to , where is the size of a unit delay in the absence of stiffness. Thus, a unit delay may be replaced by
That is, each delay element becomes an allpass filter which approximates the required delay versus frequency. A diagram appears in Fig.C.8, where denotes the allpass filter which provides a rational approximation to .
The general, order , allpass filter is given by 
and the roots of must all have modulus less than . That is, the numerator polynomial is just the reverse of the denominator polynomial. This implies each pole is gain-compensated by a zero at .
For computability of the string simulation in the presence of scattering junctions, there must be at least one sample of pure delay along each uniform section of string. This means for at least one allpass filter in Fig.C.8, we must have which implies can be factored as . In a systolic VLSI implementation, it is desirable to have at least one real delay from the input to the output of every allpass filter, in order to be able to pipeline the computation of all of the allpass filters in parallel. Computability can be arranged in practice by deciding on a minimum delay, (e.g., corresponding to the wave velocity at a maximum frequency), and using an allpass filter to provide excess delay beyond the minimum.
Because allpass filters are linear and time invariant, they commute like gain factors with other linear, time-invariant components. Fig.C.9 shows a diagram equivalent to Fig.C.8 in which the allpass filters have been commuted and consolidated at two points. For computability in all possible contexts (e.g., when looped on itself), a single sample of delay is pulled out along each rail. The remaining transfer function, in the example of Fig.C.9, can be approximated using any allpass filter design technique [1,2,269,274,553]. Alternatively, both gain and dispersion for a stretch of waveguide can be provided by a single filter which can be designed using any general-purpose filter design method which is sensitive to frequency-response phase as well as magnitude; examples include equation error methods (such as used in the matlab invfreqz function (§8.6.4), and Hankel norm methods [178,432,36].
In the case of a lossless, stiff string, if denotes the consolidated allpass transfer function, it can be argued that the filter design technique used should minimize the phase-delay error, where phase delay is defined by 
Minimizing the Chebyshev norm of the phase-delay error,
approximates minimization of the error in mode tuning for the freely vibrating string [432, pp. 182-184]. Since the stretching of the overtone series is typically what we hear most in a stiff, vibrating string, the worst-case phase-delay error is a good choice in such a case.
Alternatively, a lumped allpass filter can be designed by minimizing group delay,
The group delay of a filter gives the delay experienced by the amplitude envelope of a narrow frequency band centered at , while the phase delay applies to the ``carrier'' at , or a sinusoidal component at frequency . As a result, for proper tuning of overtones, phase delay is what matters, while for precisely estimating (or controlling) the decay time in a lossy waveguide, group delay gives the effective filter delay ``seen'' by the exponential decay envelope.
See §9.4.1 for designing allpass filters with a prescribed delay versus frequency. To model stiff strings, the allpass filter must supply a phase delay which decreases as frequency increases. A good approximation may require a fairly high-order filter, adding significantly to the cost of simulation. (For low-pitched piano strings, order 8 allpass filters work well perceptually .) To a large extent, the allpass order required for a given error tolerance increases as the number of lumped frequency-dependent delays is increased. Therefore, increased dispersion consolidation is accompanied by larger required allpass filters, unlike the case of resistive losses.
The function piano_dispersion_filter in the Faust distribution (in effect.lib) designs and implements an allpass filter modeling the dispersion due to stiffness in a piano string [155,171,371].