I started from the original PaltaniPianZhang Fortran code and
made a simplified IDL version which uses the same structure
with a triple loop (on tau, i and j) as a reference.
I then made a faster IDL version which removes one loop
exploiting IDL array notation and runs 16 times faster with the same results.
The procedure is called as :
sfu4,t,lc,f,nk,binsize,binwidth,tmin,tmax,tt
where :
The simplification with respect to Zhang's full code is that
Also I do not fully understand the roles of binsize and binwidth
which apparently can be assigned arbitrarily irrespectively of the bin size of
the light curve. In practice I tried varying them (one or both) to values different from the actual bin size of the light curve, with confusing results, so all simulations are reported with binsize=binwidth=light curve binsize = 600 s. 
Let us first assess the effect of the fluctuations (since we do not compute errors on single points we can't call it the noise) generating uniform light curves with the same plateau=1000 and different level (since the same random seed is used, the fluctuations are exactly the same, just scaled differently).  
generated with makecurve,100,1000.,x*sqrt(1000.),600.,t,lc,lastt,a2,sdev  Structure function generated with
sfu4,t,lc,f,nk,600.,600.,0.,lastt,tt plot,tt,f/nk  
Reference light curve A level=sqrt(1000)  Comparison of the 4 structure functions
The colours correspond to the 4 light curves plotted on the left. The SF scales exactly as the square of the fluctuations and is of course flat, pure white noise, so that one can reasonably compute its mean.


double fluctuation amplitude B : level=2*sqrt(1000)  
quadruple fluctuation amplitude C : level=4*sqrt(1000)  
half fluctuation amplitude D : level=0.5*sqrt(1000)  
Let us then assess the effect of intensity generating uniform light curves with different plateau (1000 and 2000) and the same percentage level of fluctuations.  
Reference light curve I : plateau=1000. level=30  Comparison of the 2 structure functions
The colours correspond to the 2 light curves plotted on the left.
The SF scales this time exactly as the square of intensity


double intensity J : plateau=2000 level=60  
Let us then examine the SF of a periodic (sinusoidal) modulation, superimposed on the previous white noise light curve.  
Sinusoid of amplitude 100, P=3230 s superimposed on (white)
reference light curve A (lc) makesin,t,cc,100.,3230. lcs=lc+cc
The SF of a periodically modulated white noise light curve (in red on the
right) is also periodic, with equal height maxima at P/2+nP and minima
tangent to the SF of the reference light curve The example shown is with a period reasonably large and not multiple of the bin size (600 s), i.e. not so well sampled. Tests have been made with other periods, with similar results (although for periods smaller than 600 s the periodicity is poorly determined, and for submultiples of 600 s (or 600 s itself) there is no contribution to the SF (the modulation adds to fixed bins). 
Comparison of the structure functions 

The next case is the addition to the reference white noise light curve of finite wavetrains derived from the same sinusoidal light curves (for a noninteger small number of periods)  
generated with
u=findgen(i+n)+n lcs=lc lcs(u)=lcs(u)+cc(u)  Structure function generated with
sfu4,t,lcs,fs,nks,600.,600.,0.,lastt,tt plot,tt,fs/nks  
wavetrain i=57 n=39 (3.2P from 23400 to 33600 s)  Comparison of the 4 structure functions with reference
The colours correspond to the 4 light curves plotted on the left, while the white line is the SF of the reference light curve.
The first three light curves (red, magenta and yellow) are all wavetrains
ending at the same moment but covering a different number of cycles
(the number of peaks in the SF match the number of cycles). All SF show similar features, that is, a number of peaks of decreasing amplitude (matching the number of cycles in the wavetrain) and a sharp drop which connects exactly with the reference light curve. The location of the drop seems somewhat correlated with the duration of the wavetrain.
The cyan SF however is strange since the drop is very close to the end of
the tau axis, and there is an intermediate dip in the SF. 

wavetrain i=57 n=15 (7.6P from 9000 to 33600 s)  
wavetrain i=57 n=43 (2.4P from 25800 to 33600 s)  
wavetrain i=90 n=76 (2.4P from 45600 to 53400 s)  
Since sinusoidal variations are quite unlikely in the sources we are interested in, a simpler example superimposes onto the white noise light curve a single flare derived from the positive part of a single sinusoidal cycle (with the period and resolution chosen this is just less than 4 time bins).  
generated with
flare=[cc(0:2),0.0] lcs=lc lcs(i+3)=lcs(i+3)+flare  Structure function generated with
sfu4,t,lcs,fs,nks,600.,600.,0.,lastt,tt plot,tt,fs/nks  
Example of 5 single flares
The colours correspond to the 5 SF plotted on the right, while the white line is the reference light curve (or its SF). Each analysed light curve differs from the reference only in the single flare highlighted in colour.
Each flare is located at bin i and lasts 4 bins (the last one is forced to
be zero, i.e. equal to the reference curve, instead of the negative value
in the original sinusoid, compare IDL expression above), i.e. 1800 s. The two blue Lshaped lines in the SF plots extend respectively :
The features of the SF can be described as follows : For the first two light curves (red and magenta), which correspond to flares located in the first half of the light curve, the SF shows a shallow peak, followed by a little drop in correspondence of tend, followed by a plateau or shallow dip, followed by a sharp drop in correspondence of lastttend after which one reaches exactly the reference SF.
For a flare located in the second half of the light curve (like the third, in
yellow) the location of the two drops (the minor one and the sharp one) is
reversed. The terminal, sharpest drop corresoponds to the end time of the
flare tend
A flare located at the very beginning of the light curve (fourth, in cyan)
shows a steady rise without final drop, and also the minor drop at tend
is almost invisible. 
i=20, flare from 12000 to 13800 s  
i=40, flare from 24000 to 25800 s  
i=75, flare from 45000 to 46800 s  
i=0, flare from 0 to 1800 s  
Let us examine the effect of the intensity of a single flare.  
The figure on the right compares the SF of two light curves with a single
flare at the same time (the light curve is not shown). The red curve is the same reported above for a flare at i=20 (1200013800 s), while the green curve corresponds to a flare of double intensity occurring at the same time. One can see that the two SF have approximately the same shape, with the same features at the same positions, but with the one corresponding to the more intense flare being scaled up. 

Let us examine the effect of multiple flares.  
The figure on the right compares the SF of three light curves, two with a single
flare at two different times, and the third (not shown) with both flares present. The red and magenta curves are the same reported above for a flare at i=20 and i=40, while the green curve corresponds to a light curve with two flares of the same intensity occurring at both i=20 and i=40. One can see that the green SF stays up until the time of the drop of the red curve (end of first flare) where it has a dip. then stays up again mimicking the magenta curve until the time of the final drop of the magenta curve : at this point it drops until it superimposes exactly to the red curve, which then follows until the red (hidden) and green drop together to the reference white noise value. 

The figure on the right compares the SF of four light curves, two with a single
flare at two different times, and two with two flares, one of which is of double
intensity with respecto to the other.
The red and magenta dottedcurves are the same reported above for a flare
at i=20 and for a flare at i=40,
while the yellow and cyan curves corresponds to a light curve with two flares
occurring at both i=20 and i=40 : for the yellow curve the first flare is the
most intense, while for the cyan curve it is the second flare to be stronger.
(the light curves can be viewed clicking here One can see that the main features occur as sharp drops at characteristics times, connecting otherwise almost flat portions of the SF curve. 

I suspended writing these pages on Oct 15. I restart now on Oct 30 adding some results obtained 2 weeks ago and not yet inserted here. Let us examine now the behaviour of red noise (or something similar to it), which I generate coadding several sinusoids of different periods and amplitude proportional to a power of period using the following IDL code :  
lcr1=lc lcc1=lcr1lcr1+1000 seed=123456789L for i=P1,P2 do begin makesin,t+randomu(seed)*i,cc,float(i),float(i) lcr1=lcr1+cc/(P2P1) endfor seed=123456789L for i=P1,P2 do begin makesin,t+randomu(seed)*i,cc,float(i),float(i) lcc1=lcc1+cc/(P2P1) endfor 
use white noise light curve A (lc) as base for the "r" light curve use a flat light curve at intensity 1000 as base for the "c" light curve assign a random seed loop from period P1 to period P2 generate a sinusoid of amplitude proportional to period and coadd it to the "r" light curve, normalize by number of periods used reassign same random seed to have same noise loop from period P1 to period P2 generate same sinusoid as above and coadd it to the "c" light curve, normalize by number of periods used  
Note the trick by which each sinusoid is random shifted by a fraction of its period
randomu(seed)*i in order to break coherence. The "r" light curve is the superposition of white noise with red noise from P1 to P2, while the "c" light curve should be pure red noise.  
"r" (red) and "c" (green) light curves for P1=3000, P2=10000 s
On the right their SF compared with the white noise one (in white).
One can note that the SF of the "c" curve is smoother, and closer to the
tau^{2} slope in its beginning. 

"c" red noise light curves for the following period ranges (same start)
On the right their SF with the respective tau^{2} (dashed) trends
which is confirmed. I am however not sure at all that the "red noise" simulations are by any mean realistic (but a periodic function is simpler than shot noise). 

I have also tried a different "power index " in the red noise using an
amplitude proportional to the square of the period, replacing the statement
above for the generation of sinusoidal components with the one below (where the
prescaling by 10^{4} is used to preserve the same approximate total
amplitude of the resulting light curve. makesin,t+randomu(seed)*i,cc,1.E4*float(i)^2,float(i) The red curve below has been generated with such formula in the period range i=P1,P2 with P1=3000 and P2=10000 (same as the reference green curve, which has been generated with the default formula (amplitude proportional to period).
The relevant SFs are reported on the right. They scale in a similar way, with just minor shifts of the peak, and the initial rise in both cases follows a tau^{2} trend (despite the different power index used in the simulation) 