An Example
I give here an experimental example of a canonical heterogeneous system that helps understanding the concept introduced earlier. It shows what are the effects of a finite tracking depth on the mean-squared displacement distribution, and illustrates the strength and limitations of multiple particle tracking measurements in heterogeneous systems.
The model system
The model system is a binary viscous fluid. One half of the fluid is composed of water, and the other half is composed of a 55% volume fraction solution of glycerol. Both parts are equally represented purely viscous fluids, the water being the "fast" system with viscosity η1 ≈ 9×10-4 Pa.s and the glycerol is the "slow" fluid with viscosity η2 ≈ 9×10-3 Pa.s. I created such a fluid by performing two separate experiments, one in water and one in the solution of glycerol, both with the same concentration of probes, and I merged the results. The tracking was performed for 30 s with 0.5 μm-diameter fluorescent probes at a concentration of 2.9×109 particles per ml. By construction none of the particle travel from one fluid to the other, although this rare event doesn't limit the validity of our artificial example. This canonical heterogeneous system is schematized on the figure on the right. The same number of particle is visible in each fluid at all time. That means that besides using the same concentration of beads in both fluids, the same tracking parameters were also used (a detailed discussion of this can be found in Savin & Doyle (2008), referenced herePointer). This ensures that the same tracking depth is observed in both fluids, as if the tracking was performed simultaneously on a real bimodal material. At the end, about 180 particles were trackable at all time, 90 particles in each fluid. Knowing that the field of view is 140 μm × 95 μm, the tracking depth can be calculated from the known concentration and we find it to be 4.5 μm (this method of determining the tracking depth is very accurate, see Savin & Doyle (2008) listed herePointer). Also, note again that this does not mean that I have a total of 180 trajectories at the end of the tracking. As explained in the previous sectionPointer, that would be true if the particles would remain observable the whole duration of the movie. Here, particles are coming in and out of focus, every time ending or beginning a new trajectory. The resulting tracking contains a total of about 5300 trajectories, 3600 of them are obtained in the "fast" fluid whereas 1700 are extracted from the "slow" side of the fluid. This unbalance is a consequence of the finite depth of tracking as explained in the previous sectionPointer. It poses some immediate problems when looking at the distribution of msd.
The msd distribution
Out of these 5300 trajectories, I calculated 5300 values of msd at a lag time τ=0.1 s. I plot here the distribution of these msd. Since we know exactly the nature of the fluid, where two different viscosities η1 and η2 are equally represented, the distribution of viscosities in our canonical heterogeneous fluid is known: it is two peaks at η1 and η2 of equal heights. A correct mapping of this distribution in term of msd should also gives us two peaks at msd1=4kTτ/(6πη1a) and msd2=4kTτ/(6πη2a), also of equal height (these two peaks are shown in blue on the plot on the right-hand side). Because of the finite depth of tracking, the measured distribution of msd is very different. First, we have much more msd calculated from the "fast" fluid than from the "slow", as previously explained. Second, since the msd calculated in the "fast" fluid are mostly calculated from short trajectories, they are in general not precisely estimated. This explains the wide spread of msd in the "fast" fluid around their supposed value (the asymetry of the peak is a feature of squared quantities distributions). On the other hand longer trajectories are more frequent in the slow fluid and lead to better estimates of their respective msd, so that the peak is sharper. In the end, the distribution of msd calculated in the bimodal fluid is significantly deformed in two ways: the peaks are unbalanced, and they are estimated with different accuracies. This distribution cannot be used directly to measure mean and variance of the msd, as we'll see in the next part.
Calculating msd and heterogeneity
As explained in the previous sectionPointer, it is still possible to calculate unbiased estimators of the msd distribution moments. Weighting each msd in the sample with the duration of the corresponding trajectories circumvents the issues presented above. In the figure on the right I show the calculation of the ensemble mean and variance of the msd vs. lag time using the formula of M1 and M2 given herePointer. Since the structure of the fluid is known, we can compare the results with their exact expected values. The red curves correspond to the "fast" and "slow" regions taken individually, and the blue curve is for the composite bimodal fluid. The squares are obtained using the formula M1 and M2 of the previous sectionPointer, and the solid lines are the expected results. For the bimodal fluid (blue plots), we see that the formula are returning good values up until a lag time of 0.5 s. Above this lag time, a fundamental limit of particle tracking is reached. We explain this limit in the next part. Notably, the estimator M1 and M2 are returning very accurate values at τ=0.1 s, for which the distribution shown above was found. Of interest, I also reported other possible estimates of the msd moments. In green, the dash dotted line is the non-weighted sample mean and variance of the msd population. As expected these estimates are failing to report correct values. More interesting is the calculated mean msd given by the dashed green line in the top plot: this is obtained by a weighted sample mean where the weight is proportional to the number ni of displacements extracted in each trajectory. This is an important quantity because it is the one everybody uses (see the website "Particle tracking using IDL" linked in the referencesPointer). This estimates performs well, although it is slightly biased and its accuracy start to degrade at smaller lag times as compared to M1 (as seen by comparing the green dashed line with the blue squares on the top plot). This tells us that at large lag times, weighting by the displacements number is less robust than by the duration.
Fundamental limit
The plot above shows that significant discrepancies are observed at large lag time, even if the mean msd is calculated with the proper weighting. Above a certain lag time the particles in the "fast" fluid will undergo displacements bigger then the depth of tracking. At these large lag times, only few displacements will be extracted from the amputated trajectories. Eventually, the displacements calculated in the "fast" fluid are rare and the latter will not be represented and becomes undetectable by the multiple particle tracking measurement. Only the "slow" fluid will be detected, and that is why the averaged msd falls on the "slow" fluid results, as seen in the plots of ensemble msd's mean and variance. This is a fundamental limitation of the technique, which cannot be overcome. The assumption of constant density only tells that as many positions are detected in one fluid or the other. But the number of displacements can be different from one fluid to the other because of the finite tracking depth. We found that the detectability of a fluid can be quantified by the following ratio that we called degree of sampling θ:

θ(τ) = Σi:Ti>τ Ti / Σi Ti

where the sum in the numerator runs only on trajectories longer than τ. This factor takes values between 0 and 1, and the limits are reached when on the one hand, no trajectory are longer than τ, thus no displacement and hence no msd can be calculated (θ=0, and the material cannot be assessed) or when all trajectory are used to calculate the msd (θ=1). As a rule of thumb, we found that in heterogeneous systems, values of M1 and M2 were unbiased for θ>80%. In the next sectionPointer we explain how to use the IDL procedures given herePointer in order to obtain all the information presented here, including the value of the degree of sampling θ.