D.2 Discrete Fourier transform
7.1 Theoretical background
In strongly scattering samples light is scattered many times before it leaves the sample again. This results in a turbid and soft appearance, like e.g. that of a cloud in the sky. If light is scattered often enough in a disordered material, each light path can be regarded as a random walk. One can say that light is diffusing.
Mathematically the transport of light can in this case be described by the diffusion equation, which makes the theoretical treatment of those systems possible.
The foundations of DWS go back to the eighties, when Maret and Wolf [107] investigated multiply scat-tered light in dense samples of randomly distributed interacting particles under Brownian motion. They showed that the diffusion approximation is well suited to describe the temporal correlations measured in the fluctuating intensity of speckles originating from a coherent light beam going through the sample.
Since then, DWS has been established and has become a standard tool to retrieve information about the dynamics of many optically dense systems like particle suspensions, emulsions, foams, polymers, gran-ular matter etc. [108]. Some of the theoretical ideas behind DWS are given in the following. We will partly follow the explanations in [109] and [110], which are good reviews including detailed derivations together with important applications.
7.1.1 General form of the autocorrelation function
In DWS we want to see a signal from the motion of the scatterers in a highly scattering turbid sample. Like in single dynamic light scattering (DLS), it is retrieved from measuring the auto correlation functionπ1(π‘) of the electric fieldπΈ(π‘) =βπ
π=1πΈππβππ(π‘), which is produced by allπinterfering plain waves registered in the detector:
π1(π‘)β‘ β¨πΈβ(π‘)πΈ(0)β©
β¨|πΈ2|β© = 1
β¨πΌβ©
β¨ π
β
π,π=1
πΈπβπβπππ(π‘)πΈπππππ(0)
β©
=
βπ π=1
β¨πΌπβ©
β¨πΌβ©
β¨πβπΞππ(π‘)β©
(7.1) Each of theπwaves corresponds to one of the light paths through the sample. For the last equality we defineπΌπ =πΈπβπΈπand use that the field vectorsπΈπand the phaseΞ¦π are not correlated. Their ensemble averages β¨β¦β©can thus be computed individually. Further we assume that the fields of different light paths are uncorrelated, so that only summands withπ =πremain. This approximation obviously works well for independent particles. But it was also proven to be valid for interacting particles, as long as their interactions are short-ranged compared to mean free path of light [111]. For single scattering, the phase shift for each wave due to the movement of the scattering particle isΞππ(π‘) =πββ Ξβππ. In the case of multiple scattering, many scatterers contribute to the phase shift of each light path. As indicated in Figure7.1(to a very good approximation [109]) the total phase shift for one path is the following sum of the phase shifts from all scattering sites:
Ξππ(π‘) =
βπ(π) π=1
β
ππβ Ξβππ(π‘) (7.2)
Hereπ(π)is the number of scatterers contributing to path numberπ. Theπβπare the individual scattering vectors for each scattering particle andΞβππis the displacement of that particle. In order to simplify our
7.1 Theoretical background
Figure 7.1:A selection of detected light paths in a multiply scattering medium. Light comes in from the left and is detected on the right hand side. Paths in blue correspond toπ‘= 0and paths in light blue to a later timeπ‘. The total phase shift for each light path due to the motion of the scatterers can be computed from the individual phase shiftπβπβ Ξβππ(π‘)at each scatterer. While some correlation remains after timeπ‘ for each individual path, among themselves the paths are uncorrelated.
expression forπ1(π‘)further, we rewrite the sum over allπdifferent light paths as a sum over the possible number of scatterersπ: which is the relative contribution of all light paths withπscatterers to the total intensity. Now the average is not only an ensemble average over particle positions but also an average over the possible individual scattering vectorsπβπtaking part in paths withπscatterers. To obtain the second equality in7.3, we assume that the successive phase shiftsπβπΞππ(π‘)are uncorrelated. This is a safe assumption in most disordered samples. Hence the average over allπβπis reduced to the averaging over the scattering vector in a single scattering event.
In the following we consider a sample that consists of particles in a solvent or a gas that does not contribute to the scattering. To go further we need to assume that the particle displacementsΞβππobey a Gaussian distribution. As discussed in AppendixBthis holds for free Brownian motion and also for Brownian motion in a harmonic potential. In the previous chapter (6.2.4) we have seen that there is some deviation from the Gaussian distribution in glassy systems. However, this deviation is relatively small, so the Gaussian approximation should be sufficiently good. The scalar productπβπ β Ξβππ then equally obeys a Gaussian distribution. By using that only moments of even order are nonzero, we obtain
β¨
For the last equality we need the assumption that scattering vectorsπβπand corresponding displacements Ξβππ are not correlated, the factor 13 comes in because we are in 3D. As a consequence, hydrodynamic interactions are neglected from here on. The ensemble average over particle positions is restricted to
β¨Ξπ2(π‘)β©, so that it gives the mean squared displacement (MSD). An average over absolute values of the scattering vectorπ remains to be done. As long as the MSD is small compared toπβ2, which is always true for small lag timesπ‘, one can restrict the calculation to the first term of a cumulant expansion [110]:
β¨ Now theπaverage is restricted to the termβ¨π2β©. As mentioned above, we only need the distribution ofπ
for a single scattering event. This is given by the productπ(π)β π(π)of the form factor and the structure factor, which would be the result of a static single light scattering experiment (cf. Equ.1.26). Then we obtain Here we use that2π0πef fis the largest possible scattering vector that occurs for backscattering. The wave vector π0 = (2π)βπcorresponds to the wavelength of lightπoutside the sample,πef f is the refractive index in the sample. For the second equality in 7.6we use the relation π = 2π0πef fsin(πβ2)between scattering vectorπand scattering angleπ.
The average π = β¨1 β cosπβ©is known as the anisotropy factor. For π = 1the scattering is isotropic:
The same intensity is scattered toπand toπβπ, so that the correlation between incoming and outgoing direction is lost. This is the case if the structure factor can be neglected, i.e. π(π) β 1, and the diameterπ of the particles is small compared to the wavelengthπ(Rayleigh scattering forπ βͺ πin dilute systems).
Anisotropic scattering is indicated byπ <1. It occurs in dense particle systems whereπ(π) β 1and/or for particles withπ β³ πwhich must be described as Mie scatterers [112]. They usually scatter significant proportion of the light in forward direction to small anglesπ.
To transfer the description into a picture of diffusing light waves, we need know the average distance between two scatterers in a light path. It is given as the mean free pathπ.
π = 1 It decreases with an increasing density of the scatterersπand an increasing scattering cross sectionπ. A calculation ofπ[112] involves the form factorπΉ(π, π)orπΉ(π)and additionally the structure factorπ(π) in dense systems. Since the structure peaks ofπ(π)increase with increasing densityπ, the mean free pathπ is no more inversely proportional toπin dense systems. For systems with anisotropic scattering π < 1it makes sense to define an average lengthπβ, the so called transport mean free path, after which light has travelled long enough so that the correlation of its direction with the incoming direction is lost.
A natural definition is to divideπby the anisotropy factorπ[110]:
πβ = π For the second equality one uses Equation7.6. This makes clear thatπβdepends primarily on the values ofπΉ(π)andπ(π)for largeπ close to the backscattering directionπ = 2π0πef f.
Now we can rewrite the autocorrelation function π1(π‘) in terms of path lengthsπ = ππ instead of the number of scatterersπ. We combine Equation7.3with7.4,7.5,7.6and7.8to get:
π1(π‘) =β
Here, eventually we go to the continuum limit and write the sum over all occurring path lengthsπ as an integral over the distributionπ(π ). One should note that this integration implies anincoherentsum in a sense that each light path contributes independently to the decay ofπ1(π‘) [109]. Hence the correlation function in Equ.7.9is comparable to the incoherent intermediate scattering functionπΉπ (π, π‘)for single light scattering, which is equally a function of the MSD (in the Gaussian approximation, cf. Equ.1.23).
7.1 Theoretical background It has been found that coherent parts make up for a fraction ofπβ1withπβ (π0πβ)(π0πΏ)[109], whereπΏis the minimum distance that light must travel through the sample. Thus forπβ β« πcoherent contributions to the decay ofπ1(π‘)are negligible.
It is also possible to rewriteπ1(π‘) in terms of a diffusion coefficientπ· = β¨Ξπ2(π‘)β©β(6π‘). Now one can also include the dependence ofπ·on particle interactions by takingπ·(π) =π·π β(π)βπ(π), whereβ(π) = π»(π)βπ»(β)describes the hydrodynamic interactions andπ·π is the short time diffusion coefficient. In this case one can omit the approximation applied in Equ.7.4, that requiredΞβπandπβto be independent:
π1(π‘) = With the use of visible light, the latter approximation is valid in systems with big particles (π >2Β΅m) [109]. For those one can equally use Equation7.9, that directly connectsπ1(π‘)to the MSD. For particles of a sizeπ= 412nm the approximation leads to a deviation of about20 %[109] in the decay time. Thus, one should have in mind that the error might be even larger for smaller particles.
It is convenient to hide the specialties of the particle system in a characteristic decay timeπ0: π0 =[
(π0πef f)2π·π ]β1
= 6π‘β [
(π0πef f)2β¨Ξπ2(π‘)β©]β1
, (7.11)
Then one obtains a more clear and catchy form for the autocorrelation function:
π1(π‘) = the correlation decreases by a factor ofexp(β2π‘βπ0). Note thatexp(β2π‘βπ0) is exactly the correlation function for single scattering (DLS) at an angle of90β¦(π=β2π0πef f).
Summary of approximations
Most important to obtain an explicit expression forπ1(π‘)is the assumption of independent light paths, which should be well-fulfilled forπβ β« π. Somehow connected to this is the assumed independence of successive scattering events in each light path. Ifπ β« πand if the correlation between particle positions (seen e.g. inπ(π)) has decayed at a distance of π, this should be valid as well. Possibly problematic for glassy systems could be the assumption of a Gaussian distribution for the displacementsΞβππof the scatterers. To obtain Equation7.9, the scattering vectorπin a single scattering event must be uncorrelated to the displacement of that scatterer. Collective movements, that occur e.g. in hydrodynamic interactions, are therefore ignored. By usage of Equation7.10one could at least approximately include hydrodynamic interactions, but this goes beyond the scope of this thesis. Another approximation is done in Equation7.5, where the lag timeπ‘must to be small enough, so that β¨Ξπ2(π‘)β© βͺ π2. However, in most of the cases, the average path length is much larger thanπβ so that the factorπ βπβ in the exponent of Equations7.9 and7.10leads to a decay ofπ1(π‘)long before the MSD becomes comparable to the wavelengthπ.
7.1.2 Diffusion equation and path length distributionπ(π )
In the diffusion approximation (for length scales larger thanπβ) one can describe the energy densityπ(βπ, π‘) of light with the diffusion equation
ππ(βπ, π‘)
ππ‘ =π·πβ2π(βπ, π‘) with π·π= ππβ
3 . (7.13)
The diffusion coefficientπ·πcontains the speed of light in the mediumπ =π0βπef fand the transport mean free pathπβ, the distance after which the direction of light is randomized, the division by3accounts for the number of space dimensions. For a solution of Equ.7.13one also needs initial and boundary conditions.
The idea is to compute the outgoing fluxπ½βout(βπout, π‘)at a positionβπout at the boundary of the sample after an instantaneous light pulse has started to diffuse in the sample at time π‘ = 0. Since light needs the timeπ‘=π βπto reach the position of the detector atβπout, the distributionπ(π ) is directly proportional to π½βout(βπout, π‘=π βπ). Because the light path is only randomized after a distance ofπβthe starting point for the light pulse is not directly at the border of the sample. Instead one defines it to start at a distanceπ§0 βπβ inside the sample. Therefore, as an example, the initial condition for a sample with a planar boundary at π§= 0and a parallel beam coming fromπ§ <0is given by:
π(π§, π‘= 0) =π0πΏ(π§βπ§0, π‘) (7.14)
Boundary conditions are usually obtained by requiring that the incoming net flux into the sample is zero [110]. This results in the condition
( π+ 2
3πβ Μβπβ βπ) |
||βπβππ = 0, (7.15)
whereππ means the surface of the sample andΜβπis a unit vector which is always orthogonal to that surface and directing out of the sample. Other boundary conditions, e.g. absorption of light at the surface, were investigated, but Equ.7.15gave the best agreement with experiments [109]. With a solutionπ(βπ, π‘)for the full diffusion problem, the path length distribution is given by
π(π ) β|||π½β(βπout, π‘)|||=π·π|||Μβπβ βπ(βπout, π‘)|||= π
2π(βπout, π‘=π βπ) (7.16) Here one makes use of Fickβs law of diffusion and the boundary condition in Equation7.15. Since the autocorrelation functionπ1(π‘)is basically a Laplace transform ofπ(π )(cf. equs.7.9and7.10) it is often convenient to solve the diffusion problem in Laplace space and obtain directly an expression forπ1(π‘) [109].
7.1.3 Transmission geometry
The detailed derivation of the path length distributions for arbitrary geometries and the final computation ofπ1 uses nontrivial mathematics and goes beyond the scope of this description. Only the formulas for the two most important scattering geometries (see Fig.7.2), that were also used in this work, are presented here including comments.
In transmission geometry, a parallel beam of coherent laser light uniformly illuminates one side of a slab with a thicknessπΏinπ§direction and an infinite extent in the other two directions. In experiments this idealized situation is well approximated using a beam width much larger thanπβand a sample that is larger than the beam width by more than10πβ. As mentioned in the preceding paragraph, light is considered to
7.1 Theoretical background
Figure 7.2: Idealized transmission and backscattering geometry for a slab of thicknessπΏwith infinite extent perpendicular to the incoming beam. The coherent parallel light is considered to be homogeneous with an infinite width, like an idealized plane wave. Precise angles of the detector with respect to the incoming beam are arbitrary. Sketches of typically detected light paths are given in red.
start diffusing from a plane atπ§0βπβ. However, the exact value ofπ§0is not very important sinceπΏ β« πβ is required anyway. At leastπΏ >10πβis necessary for the diffusion approximation to be valid. Using the characteristic decay time of the particle systemπ0, one obtains the autocorrelation function [109]:
π1(π‘) =
As the approximation suggests, the shape of this function somewhat resembles an exponential decay. But particularly for largeπ‘, deviations from exponential behaviour are observed. The characteristic decay time isπ0(πββπΏ)2. Recalling thatπ1(π‘) decreases by a factorexp(β2π‘βπ0)for each step of the random walk, one infers that the mean number of steps isβ¨πββ©= 1
2(πΏβπβ)2. This corresponds to an average path length ofβ¨π β© = πΏ2β(2πβ). An easy way to derive this is to consider the growing mean squared displacement 6π·ππ‘ of the photons diffusing in the sample: In order to diffuse an average distance ofπΏ, which is the sample thickness, light needs the timeπΏ2β(6π·π) =πΏ2β(2π πβ). A division by the speed of lightπyields the average lengthβ¨π β©of a light path. Since all light paths must travel the distanceπΏinπ§direction, the distributionπ(π )has a sharp maximum at the average path lengthβ¨π β©.
From the definition ofπ0in Equation7.11one can see that each scatterer moves an average distance of Ξπ0 =β transmis-sion geometry is very sensitive to the dynamics in the sample. An average translation of the scatterers by only a fractionπββπΏof the wavelength is enough to destroy the correlation. By adjusting the thick-ness of the sample cell the sensitivity can be adjusted, for instance starting from a decay for an average displacement ofβΌ 50nm in a thin cell down to several nanometers in a thick cell.
7.1.4 Backscattering geometry
The path length distributionπ(π )for the backscattering geometry is a lot broader than that for the trans-mission case. Both very short light paths with a length of a fewπβ and very long paths are possible.
Incoming beam and detector are on the same side of the sample (see Fig.7.2). The exact angle at which light is collected is not important. But it is substantial that the incoming parallel beam is expanded much
wider thanπβ. The detector needs to focus at a spot near the centre of the illuminated area. If the latter two conditions are not fulfilled, the path length distribution is qualitatively different. For instance, if the detector is focused to the border of the illuminated area, the contribution of short paths becomes smaller while that of large paths becomes longer.
For a computation ofπ1(π‘)one again assumes that the diffusing light starts from a plane atπ§=π§0 =πΎπβ in the sample. The exact value ofπΎ β 2(see [110]) depends on the sample and the sample cell (e.g.
on the particle sizes, thickness of glass walls, reflectivities etc.). Additionally, there is a dependence ofπΎ on the polarization of incoming and detected light, which is measurable by putting a polarizer in front of the detector. If the polarizations of incoming and detected light are perpendicular the very short paths coming from one single backscattering are not detected, because single scattering preserves the polarization. This yields an increase inπΎ. On the other hand, if incoming and detected light have mainly parallel polarizations,πΎdecreases compared to a measurement without a polarizer.
For the longer pathsπ β« πβ, one can show thatπ(π ) βπ β3β2. Therefore,π(π )only decays algebraically forπ ββ. This means that also longer paths contribute to the decay ofπ1, not only the more frequent short paths. Since many scatterers contribute to the phase shift in long paths, very small displacements Ξβππare enough to destroy their correlation. Thus, the long paths are responsible for the decay ofπ1(π‘) at small lag times π‘. On the other hand, the longer lag times are dominated by the correlations of the short paths.
For a sample of thicknessπΏ, the exact solution of the diffusion equation yields [109]:
π1(π‘) = accu-rate normalizationπ1(π‘ = 0) = 1. Since paths of many different lengthsπ contribute, the decay upon increasing the lag timeπ‘is slower than exponential. By comparison ofexp(βπΎβ
6π‘βπ0)with the decay exp(β2π‘βπ0)for90β¦ single scattering, one comes to the conclusion that forπ‘ >3β2πΎ2π0the correlation for backscattering DWS is actually larger than that for single scattering. It is however clear that correla-tions of multiply scattered light must decay faster. Therefore, the result given in Equation7.19can only be valid for small lag timesπ‘ βͺ π0. This is not only true for the approximation, but also for the exact solution of the diffusion equation. The reason is that forπ‘ > π0 the correlations mainly stem from the shortest paths. But for short paths, many of the approximations mentioned above (see7.1.1), including the diffusion approximation, are not valid.
Still DWS in backscattering geometry is a very useful tool especially for glassy samples. It allows to probe the dynamics at many length scales at once. Small movements of the particles contribute to the decay ofπ1(π‘)at short lag times. But even quite large displacements do not yield a total decay, because the dominating short light paths are still correlated. Using the definition ofπ0in Equation7.11together with the approximatedπ1(π‘)from above, we see thatπ1(π‘)has decayed to1βπwith an average displacement of the scatterers by:
Ξπ0= (πΎπ0πef f)β1 =πβ (2ππef fπΎ)β1 (7.20) This is only a factorβΌπΎearlier than in a single scattering experiment. A total decay requires an average displacement by about one wavelengthπ.