Revisiting the method of cumulants for the analysis of

dynamic light-scattering data

Barbara J. Frisken

The method of cumulants is a standard technique used to analyze dynamic light-scattering data mea-
sured for polydisperse samples. These data, from an intensity–intensity autocorrelation function of the
scattered light, can be described in terms of a distribution of decay rates. The method of cumulants
provides information about the cumulants and the moments of this distribution. However, the method
does not permit independent determination of the long-time baseline of the intensity correlation function
and can lead to inconsistent results when different numbers of data points are included in the fit. The
method is reformulated in terms of the moments about the mean to permit more robust and satisfactory
fits. The different versions of the method are compared by analysis of the data for polydisperse-vesicle
samples. © 2001 Optical Society of America
OCIS codes: 120.5820, 290.5870.

1. Introduction The ubiquitous use of nonlinear fitting routines

The method of cumulants1 for the analysis of dynamic makes formulation in terms of a polynomial unnec-
light-scattering 共DLS兲 data was first introduced by essary. Reformulating the method in terms of the
Koppel in 1972. He showed that the logarithm of moments of the distribution rather than of the cumu-
the field–field correlation function is equivalent to the lants results in more satisfactory and robust fits and
cumulant-generating function. Information about permits independent fitting of the long-time baseline.
the cumulants of the distribution of decay rates can Furthermore, it is not necessary to limit the fit to a
thus be obtained from the correlation function mea- restricted range of the data. In this paper data from
sured for polydisperse samples. measurements of polydisperse lipid vesicles are used
The method, as commonly used,2–5 allows the log- to highlight the differences between the traditional
arithm of the field-correlation function to be written and the reformulated versions of the method. Like
in terms of a polynomial in the delay time ␶, which is the original cumulant method, this reformulated mo-
a function that can be fitted easily by use of linear ment method is most reliable for monomodal decay-
least-squares techniques. This function has several rate distributions of finite width.
disadvantages. Most remarkably, parameters ob-
2. Theory
tained in the fits are not invariant as more data
points are included. In addition, fitting this function A. Dynamic Light Scattering
requires that the long-time baseline of the intensity
correlation function be an assumed rather than a DLS measurements involve the analysis of the time
floating parameter. Using the baseline as a floating autocorrelation function of scattered light as per-
parameter makes it possible to detect problems in the formed by a digital correlator. The normalized time
data and to fit data when a little bit of noise is autocorrelation function of the intensity of the scat-
present. tered light g共2兲共␶兲 for a given delay time ␶ is given by2

具I共t兲I共t ⫹ ␶兲典
g 共2兲共␶兲 ⫽ , (1)
The author 共frisken@sfu.ca兲 is with the Department of Physics,
具I共t兲典 2
Simon Fraser University, Burnaby, British Colombia V5A 1S6,
Canada. where I共t兲 and I共t ⫹ ␶兲 are the intensities of the scat-
Received 11 December 2000; revised manuscript received 3 April tered light at times t and t ⫹ ␶, respectively, and the
2001. braces indicate averaging over t.
0003-6935兾01兾244087-05$15.00兾0 In most cases of practical interest the intensity–
© 2001 Optical Society of America intensity time autocorrelation function may also be

20 August 2001 兾 Vol. 40, No. 24 兾 APPLIED OPTICS 4087

expressed in terms of the field–field time autocorre- simplest is the method of cumulants first proposed by
lation function g共1兲共␶兲 as Koppel.1 This method is based on two relations:
one between g共1兲共␶兲 and the moment-generating func-
g 共2兲共␶兲 ⫽ B ⫹ ␤关 g 共1兲共␶兲兴 2, (2) tion of the distribution, and one between the loga-
共1兲
with g 共␶兲 given by rithm of g共1兲共␶兲 and the cumulant-generating function
of the distribution. It is appropriate only for use in
具E共t兲 E*共t ⫹ ␶兲典 cases in which G共⌫兲 is monomodal.
g 共1兲共␶兲 ⫽ , (3) In fact, as was discussed by Koppel,1 the form of
具E共t兲 E*共t兲典 g共1兲共␶兲 as given in Eq. 共5兲 is equivalent to the definition
where E共t兲 and E共t ⫹ ␶兲 are the scattered electric of the moment-generating function M共⫺␶, ⌫兲 of the
fields at times t and t ⫹ ␶, respectively, and ␤ is a distribution G共⌫兲 共Ref. 7兲:
factor that depends on the experimental geometry.

兰
⬁
Equation 共2兲 is known as the Siegert relation.6 The
M共⫺␶, ⌫兲 ⫽ G共⌫兲exp共⫺⌫␶兲d⌫ ⬅ g 共1兲共␶兲. (7)
factor B, commonly referred to as the baseline, is the
long-time value of g共2兲共␶兲. Although the factor B 0

should be equal to one, in practice, a small amount of The mth moment of the distribution mm共⌫兲 is given by
noise in the measurement can result in values that the mth derivative of M共⫺␶, ⌫兲 with respect to ␶:
differ from unity by small 共⬃10⫺4兲 amounts. In this
case assuming that the baseline is one changes the
parameter estimates and increases the deviation of
the fit from the data. Larger deviations of the base-
m m共⌫兲 ⫽
dmM共⫺␶, ⌫兲
d共⫺␶兲 m
冏 ⫺␶⫽0

兰
line from one can indicate that there is a problem ⬁

with the data. ⫽ G共⌫兲⌫ m exp共⫺⌫␶兲d⌫兩 ⫺␶⫽0. (8)

For monodisperse particles in solution the field- 0

correlation function decays exponentially, g共1兲共␶兲 ⫽ Similarly, the logarithm of the field-correlation func-
exp共⫺⌫␶兲, with a decay rate of ⌫ ⫽ Dq2, where D is the tion is equivalent to the definition of the cumulant-
diffusion coefficient of the particles and q is the mag- generating function7 K共⫺␶, ⌫兲
nitude of the scattering wave vector. The scattering
wave vector q is defined as the difference between the K共⫺␶, ⌫兲 ⫽ ln关M共⫺␶, ⌫兲兴 ⬅ ln关 g 共1兲共␶兲兴, (9)
incident and the scattered wave vectors, and its mag-
nitude q is given by where the mth cumulant of the distribution ␬m共⌫兲 is
given by the mth derivative of K共⫺␶, ⌫兲:
q⫽
4␲n
␭0
sin
␪
2
, 冉冊 (4)
␬ m共⌫兲 ⫽
dmK共⫺␶, ⌫兲
d共⫺␶兲 m
冏 . (10)
⫺␶⫽0
where n is the refractive index of the solvent, ␭0 is the
wavelength of the laser in vacuum, and ␪ is the scat- By making use of the fact that the cumulants, except
tering angle. The Stokes–Einstein relation, D ⫽ for the first, are invariant under a change of origin,
kB T兾6␲␩Rh, where kB is Boltzmann’s constant, T is one can write the cumulants in terms of the moments
the temperature, and ␩ is the dynamic viscosity, re- about the mean as
lates the diffusion coefficient to the hydrodynamic

兰
⬁
radius Rh of the particles. For a polydisperse sam-
␬ 1共⌫兲 ⫽ G共⌫兲⌫d⌫ ⬅ ⌫៮ , (11)
ple, g共1兲共␶兲 can no longer be represented as a single
0
exponential and must be represented as a sum or an
integral over a distribution of decay rates G共⌫兲 by ␬ 2共⌫兲 ⫽ ␮ 2, (12)

兰
⬁
␬ 3共⌫兲 ⫽ ␮ 3, (13)
g 共1兲共␶兲 ⫽ G共⌫兲exp共⫺⌫␶兲d⌫, (5)
0
␬ 4共⌫兲 ⫽ ␮ 4 ⫺ 3共␮ 2兲 2 . . . , (14)
where G共⌫兲 is normalized so that where ␮m are the moments about the mean, as de-
fined by
兰
⬁
G共⌫兲d⌫ ⫽ 1. (6)

兰
⬁
0
␮m ⫽ G共⌫兲共⌫ ⫺ ⌫៮ 兲 md⌫. (15)
B. Method of Cumulants 0

Finding the precise functional form for the distribu- The first cumulant describes the average decay rate
tion of decay rates G共⌫兲 is problematic because the of the distribution. The second and the third cumu-
correlation function is measured discretely only over lants correspond directly to the appropriate moments
an incomplete range of ␶ and there is always noise about the mean: The second moment corresponds to
associated with the data.3 There are several ways of the variance, and the third moment provides a mea-
using DLS data to characterize G共⌫兲,3 but one of the sure of the skewness or asymmetry of the distribu-

4088 APPLIED OPTICS 兾 Vol. 40, No. 24 兾 20 August 2001

tion. The first two cumulants must be positive, but lation function can be expressed in terms of the mo-
the third cumulant can be positive or negative. ments about the mean, as defined in Eq. 共15兲:

冉 冊
The basis of the cumulant expansion that is usually
used in the analysis of DLS data lies in expanding ␮2 2 ␮3 3
g 共1兲共␶兲 ⫽ exp共⫺⌫៮ ␶兲 1 ⫹ ␶ ⫺ ␶ ⫹ ... .
the logarithm of g共1兲 in terms of the cumulants of the 2! 3!
distribution. This relation follows from the fact that (22)
the mth cumulant is the coefficient of 共⫺␶兲m兾m! in the
Taylor expansion of K共⫺␶, ⌫兲 about ␶ ⫽ 0, as given by Expression 共22兲 was derived by Pusey et al.,8 but they
went on to expand the logarithm of Eq. 共22兲 to obtain
ln关 g 共1兲共␶兲兴 ⬅ K共⫺␶, ⌫兲 a function for ln关 g共1兲共␶兲兴. This expansion adds an
extra approximation to the derivation that is unnec-
␬2 2 ␬3 3 ␬4 4
⫽ ⫺⌫៮ ␶ ⫹ ␶ ⫺ ␶ ⫹ ␶ .... (16) essary. Instead, the moment-based expression for
2! 3! 4! g共1兲 关Eq. 共22兲兴 and the Seigert relation 关Eq. 共2兲兴 can be
used directly to derive a third expression for g共2兲:

冉 冊
To take advantage of this form and use linear least-
2
squares methods to fit this function to the data re- ␮2 2 ␮3 3
quires that a key assumption be made about the data: g 共2兲
⫽ B ⫹ ␤ exp共⫺2⌫៮ ␶兲 1 ⫹ ␶ ⫺ ␶ ... .
2! 3!
The baseline must be assumed to be exactly one.
Then a fit can be made to (23)
The form of Eq. 共23兲 also permits the direct fitting of
␤ ␬ 2␶ 2 ␬ 3␶ 3 B and has the advantage that it eliminates stability
ln关 g 共2兲共␶兲 ⫺ 1兴 ⫽ ln ⫺ ⌫៮ ␶ ⫹ ⫺ ⫹ .... problems that are inherent to Eq. 共18兲 at large ␶.
2 2! 3!
(17) 3. Testing the Model Functions

Equation 共17兲 is the traditional fitting function that is The three model functions of Eqs. 共17兲, 共18兲, and 共23兲
described in many DLS texts.3–5 with terms up to the second moment about the mean
Although most modern correlators do an excellent ␮2 were fitted to data measured from palmitoyl-oleoyl
job of measuring the baseline, small amounts of noise phosphatidylcholine 共POPC兲 vesicles formed by ex-
can lead to small deviations from unity. Nonlinear trusion through polycarbonate membranes with
fitting routines permit the possibility of fitting the 200-nm pores at an extrusion pressure of 35 psi
data to g共2兲 directly. From Eq. 共17兲, we obtain 共2.4 ⫻ 105 Pa兲.9 The apparatus used for the light-
scattering experiments was a Model ALV DLS兾SLS-

冉
g 共2兲 ⫽ B ⫹ ␤ exp ⫺2⌫៮ ␶ ⫹ ␬ 2␶ 2 ⫺
␬3 3
3
␶ ... . 冊 (18)
5000 共ALV-Laser GmbH, Langen, Germany兲 that
used a He–Ne laser as the light source. Figure 1
shows results of a measurement in which light scat-
tered by the sample was collected at 90° from the
Using the form of Eq. 共18兲 makes it possible to deter- transmitted beam. The model functions were fitted
mine B from the data. to the data by use of nonlinear fitting routines;
One can also express the field-correlation function weights were calculated from standard deviations
in terms of the moments about the mean by first provided by the ALV-Laser software. Fits were
rewriting exp共⫺⌫␶兲 in terms of its mean value:

exp共⫺⌫␶兲 ⫽ exp共⫺⌫៮ ␶兲exp关⫺共⌫ ⫺ ⌫៮ 兲␶兴. (19)

By substituting Eq. 共19兲 into Eq. 共5兲, we obtain

兰
⬁
g 共1兲共␶兲 ⫽ exp共⫺⌫␶兲 G共⌫兲exp关⫺共⌫ ⫺ ⌫៮ 兲␶兴d⌫. (20)
0

The exponential function inside the integral in Eq.

共20兲 can be expressed as a series expansion:

g 共1兲共␶兲 ⫽ exp共⫺⌫៮ ␶兲
兰 0
⬁

冋
G共⌫兲 1 ⫺ 共⌫ ⫺ ⌫៮ 兲␶

⫹
共⌫ ⫺ ⌫៮ 兲 2 2 共⌫ ⫺ ⌫៮ 兲 3 3
2!
␶ ⫺
3!
␶ ⫹ . . . d⌫. 册 Fig. 1. Sample data taken for POPC vesicles formed by extrusion
(21) through polycarbonate membranes. The curve through the data
is a fit of Eq. 共23兲 to the data. The dashed curve shows the
This expansion is exact as long as all terms in the weighted residuals: the difference of the fit from the data divided
expansion are included. After integration the corre- by the uncertainty in each point.

20 August 2001 兾 Vol. 40, No. 24 兾 APPLIED OPTICS 4089

 Table 1. Fit of Eq. 共17兲 to DLS Data for POPC Vesicles Extruded Table 3. Fit of Eq. 共23兲 to DLS Data for POPC Vesicles Extruded
through 200-nm Poresa through 200-nm Poresa

␶max ⌫ ␮2 ␶max ⌫ ␮2
共ms兲 ␤ ␹2 共1兾ms兲 共1兾ms2兲 共ms兲 ␹2 B ␤ 共1兾ms兲 共1兾ms2兲

0.20 2.64 0.37533 0.921 0.21 1.64 0.227 0.9981 0.3781 0.9072 0.121
0.41 2.22 0.37502 0.9134 0.112 3.28 0.211 0.99912 0.37603 0.9081 0.0995
0.82 2.29 0.37501 0.9082 0.0855 6.55 0.253 0.99981 0.37532 0.9071 0.0874
1.64 12.6 0.37403 0.9002 0.0702 13.11 0.265 0.99991 0.37512 0.9071 0.0834
3.27 620 0.3702 0.8605 0.0233 26.2 0.379 1.000077 0.37502 0.9072 0.0804
a
52.8 0.421 1.000046 0.37502 0.9072 0.0814
The subscripts refer to the error in the final digit of the param- 105.3 0.525 1.000015 0.37502 0.9072 0.0825
eter. The parameter B is assumed equal to 1. The fit was made
to data ranging from 6.4 ␮s to a maximum delay time ␶max. a
The subscripts refer to the error in the final digit of the param-
eter. The fit was made to data ranging from 6.4 ␮s to a maximum
delay time ␶max.
made to data corresponding to delay times of 6.4 ␮s to
maximum times ranging from 0.31 to 200 ms. It is
traditional to fit the cumulant function to data to the no satisfactory determination of the polydispersity, in
point at which the amplitude has fallen to 10% of the particular, can be made.
original amplitude, which would be 1.84 ms for these The results for the third model function, as given in
data. The fit shown in Fig. 1 is a fit of Eq. 共23兲 to Eq. 共23兲, are shown in Table 3. Again, the function
data from ␶ ⫽ 6.4 ␮s–13.1 ms. The correlation time is hard to fit at small delay times, but as soon as the
of ␶៮ ⫽ 1兾⌫៮ for these data is approximately 1.23 ms. maximum delay time is greater than a time corre-
The results for the traditional fitting function, as sponding to several correlation times the parameters
given by Eq. 共17兲, are shown in Table 1. Results for are well determined with minimal variation as the
the three parameters ␤, ⌫, and ␬2 ⫽ ␮2 are not stable number of data points fitted increases. As well as
in the sense that the parameters vary as the number doing a better job of determining the parameters, the
of data points included in the fit is changed. The third model function is also much more robust; bad
standard error in the parameters is shown as a sub- guesses of the initial parameters still lead to quick
script in the appropriate decimal place. The table convergence to the solution. Table 4 compares the
also includes results for the usual goodness-of-fit pa- number of iterations required for convergence from
rameter ␹2, defined by different starting parameters by use of Eqs. 共18兲 and
共23兲.
1 N
共 yi ⫺ fi 兲2
兺
The difference in effectiveness among the fitting
␹2 ⫽ , (24) functions is due to the different expansions used.
N⫺m i⫽1 ␴ i2
One derives the expansion in terms of cumulants, Eq.
where N is the number of data points, m is the num- 共16兲, by making an expansion about ␶ ⫽ 0. Thus Eq.
ber of parameters, and yi , fi , and ␴i are the data, the 共16兲 is accurate only near ␶ ⫽ 0, not very useful if one
fit, and the uncertainty in the data, respectively, at a wishes to fit the whole data set to obtain as much
given delay time ␶i . information from it as possible. In contrast, one de-
The results for the second model function, as given rives the expansion in terms of moments about the
in Eq. 共18兲, are shown in Table 2. This function is mean, Eq. 共22兲, by making an expansion about ⌫៮ .
difficult to fit at small delay times because B is not Equation 共22兲 should be most accurate near ⌫៮ ; this
specified, and it is difficult to fit at large delay times seems a more appropriate point of expansion when
because the positive term in the exponential that is trying to determine the distribution function.
increasing as ␶2 makes the function unstable for large
␶. In the region in which a fit is obtained the pa-
rameters vary as more data points are fitted so that Table 4. Comparison of the Robustness of the Fits of Model Function
2 关Eq. 共18兲兴 and Model Function 3 关Eq. 共23兲兴 to DLS Data for POPC
Vesicles Extruded through 200-nm Poresa
Table 2. Fit of Eq. 共18兲 to DLS Data for POPC Vesicles Extruded
through 200-nm Poresa Number of Iterations
Model Required for Initial ⌫ Initial ␮2
␶max ⌫ ␮2 Function Convergence 共1兾ms兲 共1兾ms2兲
共ms兲 ␹2 B ␤ 共1兾ms兲 共1兾ms2兲
2 3 1 0.1
1.64 0.227 0.9962 0.3792 0.9033 0.122 2 20 2 0.1
3.28 0.213 0.99843 0.37704 0.9062 0.1015 2 16 1 0.2
6.55 0.284 0.99961 0.37542 0.9051 0.0824 2 8 1 0.01
13.11 0.308 0.99981 0.37511 0.9001 0.0763 3 3 1 0.1
26.2 0.486 1.00011 0.37452 0.9001 0.063 3 6 2 0.1
52.8 Unstable Unstable Unstable Unstable Unstable 3 3 1 0.2
a
3 3 1 0.01
The subscripts refer to the error in the final digit of the param-
eter. The fit was made to data ranging from 6.4 ␮s to a maximum a
Initial values of B ⫽ 1.0 and ␤ ⫽ 0.4 were used in all cases.
delay time ␶max. The fits were made to data ranging from 6.4 ␮s to 6.55 ms.

4090 APPLIED OPTICS 兾 Vol. 40, No. 24 兾 20 August 2001

 In principle, a fit that uses Eq. 共23兲 could include The author gratefully acknowledges helpful discus-
␮3. For the vesicle data, including ␮3 sometimes sions with Art Bailey.
decreases ␹2 but sometimes increases it; this param-
eter does not seem to be significant for these data. References and Note
Furthermore, use of ␮3 does not change the values of 1. D. E. Koppel, “Analysis of macromolecular polydispersity in
the other parameters significantly in these fits. intensity correlation spectroscopy: the method of cumulants,”
J. Chem. Phys. 57, 4814 – 4820 共1972兲.
4. Conclusions 2. B. J. Berne and R. Pecora, Dynamic Light Scattering 共Krieger,
Data from polydisperse samples as measured in DLS Malabar, Fla., 1990兲.
3. P. Stepanek, “Data analysis in dynamic light scattering,” in
experiments can be analyzed in terms of the moments
Dynamic Light Scattering, W. Brown, ed. 共Oxford University,
about the mean of the distribution function that de- Oxford, UK, 1993兲, pp. 177–240.
scribes the polydispersity of the sample. The tradi- 4. R. Pecora, Dynamic Light Scattering 共Plenum, New York, 1985兲.
tional fitting function, as derived from the cumulants 5. B. Chu, Laser Light Scattering: Basic Principles and Practice
of the distribution, has several problems associated 共Academic, Boston, 1991兲.
with it: It results in parameter values that depend 6. K. Schätzel, “Single-photon correlation techniques,” in Dynamic
on the number of data points fitted, and it does not Light Scattering, W. Brown, ed. 共Oxford University, Oxford, UK,
permit an independent fit of the baseline B. This 1993兲, pp. 76 –148.
paper has demonstrated how a more robust fitting 7. A. Stuart and J. K. Ord, Kendall’s Advanced Theory of Statistics
function can be obtained by the avoidance of the log- 共Wiley, New York, 1994兲, Chap. 4.
arithm of g共1兲共␶兲 and the direct expansion of g共1兲共␶兲 in 8. P. N. Pusey, D. E. Koppel, D. W. Schaefer, R. D. Camerini-
Otero, and S. H. Koenig, “Intensity fluctuation spectroscopy of
terms of the moments about the mean. The function laser light scattered by solutions of spherical viruses,” Biochem-
can be fitted to the entire data set, gives consistent istry 13, 952–960 共1974兲.
results for fitting parameters when different num- 9. For details of extrusion methods for vesicle production as used
bers of points are fitted, is more robust to bad guesses in these experiments, please see B. J. Frisken, C. Asman, and
of the initial parameters, and permits an indepen- P. J. Patty, “Studies of vesicle extrusion,” Langmuir 16, 928 –
dent fit of the baseline B. 933 共2000兲, and references therein.

20 August 2001 兾 Vol. 40, No. 24 兾 APPLIED OPTICS 4091