 1. Introduction
 2. Calculation of the correlation functions
 3. Analysis of twotime correlation functions using different time coordinate systems
 4. 2TCFs for stationary and nonstationary systems
 5. Examples
 6. Discussion
 7. Alternative representation of the twotime correlation function
 8. Summary
 A1. Trajectories in a twotime correlation function
 References
 1. Introduction
 2. Calculation of the correlation functions
 3. Analysis of twotime correlation functions using different time coordinate systems
 4. 2TCFs for stationary and nonstationary systems
 5. Examples
 6. Discussion
 7. Alternative representation of the twotime correlation function
 8. Summary
 A1. Trajectories in a twotime correlation function
 References
research papers
On the use of twotime correlation functions for Xray photon correlation spectroscopy data analysis
^{a}Department of Physics, University of Warwick, Gibbet Hill Road, Coventry CV4 7AL, UK, and ^{b}XMaS, The UK–CRG Beamline, ESRF – The European Synchrotron, CS40220, F38043 Grenoble Cedex 09, France
^{*}Correspondence email: oier.bikondoa@esrf.fr
Multitime correlation functions are especially well suited to study nonequilibrium processes. In particular, twotime correlation functions are widely used in Xray photon correlation experiments on systems out of equilibrium. Onetime correlations are often extracted from twotime correlation functions at different sample ages. However, this way of analysing twotime correlation functions is not unique. Here, two methods to analyse twotime correlation functions are scrutinized, and three illustrative examples are used to discuss the implications for the evaluation of the correlation times and functional shape of the correlations.
1. Introduction
Xray photon correlation spectroscopy (XPCS), the equivalent of dynamic et al., 2008; Sutton, 2008; Gutt & Sprung, 2015; Madsen et al., 2015; Bikondoa, 2016). XPCS allows one to probe the dynamics of fluctuations on short length scales (100 nm) and long time scales ( s) (Malik et al., 1998; Madsen et al., 2010). Information about the dynamics is obtained by studying the time correlation of the intensity scattered by a system in a dynamic regime when illuminated with coherent light. Under coherent illumination, the farfield pattern of light scattered by a sample shows a grainy intensity distribution called speckle (Sutton et al., 1991). The intermediate scattering function of the sample, , is obtained from the normalized intensity autocorrelation of the speckles, , through the Siegert relation:^{1}
using Xrays instead of visible light, is a powerful technique to study the dynamics of soft and hard condensed matter (Grübelwhere I_{t} and are the intensities at times t and and at momentum transfer . τ is a delay time. The superscript (2) marks that the intensity autocorrelation is a secondorder correlation on the electric field. The bar indicates an ensemble average over wavevectors with equivalent momentum transfer value and for which it is expected that the correlations are statistically equivalent. The brackets denote a time average.^{2} is the static The optical contrast is a factor that is used to account for the degree of spatial coherence of the incident radiation and is given by the variance of the intensity () divided by its mean value (Madsen et al., 2010). The calculation of in equation (1) assumes that a time average can be performed over the entire measurement (Goodman, 1985). This assumption is valid for systems in equilibrium because for such systems the autocorrelation g^{(2)} depends only on and the time difference or delay time τ between measurements. That is, g^{(2)} is timeshift invariant and does not depend on the specific time when the measurement was made (observation time). is a onetime correlation function (1TCF).
For nonequilibrium systems (i.e. for systems with average properties changing with time) the time average in equation (1) is not suitable because the dynamics are evolving and may strongly depend on the observation time. For those systems the evolution of the correlations can still be captured by using a more general expression than equation (1), namely a twotime correlation function (2TCF) (Brown et al., 1997):
Corr is the autocovariance of the intensity normalized by its standard deviation. Different correlation functions are also used (Sutton et al., 2003):
or
where
For random Gaussian fluctuations the standard deviation equals the average intensity (Brown et al., 1997; Loudon, 1983). Therefore, and the different correlation functions [equations (2), (3) and (4)] are equivalent.
The use of 2TCFs for XPCS was introduced, to our knowledge, by Brown et al. (1997), who studied the time correlations in the intensity scattered by a phase ordering system using numerical simulations. The 2TCF is generally represented as a twodimensional graph of the value of , or for a fixed , with axes t_{1} and t_{2} (see §2.2). Brown et al. (1997) introduced an alternative coordinate system, which has subsequently been widely used in the XPCS literature (Malik et al., 1998; Brown et al., 1999; Livet et al., 2001; Sutton et al., 2003; Fluerasu et al., 2005; Ludwig et al., 2005; Fluerasu et al., 2007; Müller et al., 2011; Orsi et al., 2010, 2012; Chushkin et al., 2012; Livet & Sutton, 2012; Bikondoa et al., 2012; Ruta et al., 2012). Using this alternative coordinate system, approximated 1TCFs are often extracted from the 2TCF at different sample ages or observation times. We show below that in some cases employing such a coordinate system to extract approximate 1TCFs may pose interpretation problems. For such cases, we put forward another coordinate system to extract the 1TCFs and propose a clearer graphical representation of the 2TCFs.
This article is organized as follows: in §2 we describe the calculation of the autocorrelation (§2.1) and the twotime correlation function (§2.2) in discrete form. The extraction of 1TCFs from the 2TCF using different coordinate systems is examined in §3. Some properties of the 2TCF for stationary and nonstationary systems, such as the time reversal symmetry, the functional shape and decay times, are analysed in §4. Model examples of 2TCFs that reflect the differences between the analysis done using one coordinate system or another are presented in §5. A discussion about the coordinate system that should be used for different cases follows in §6. In §7 we propose, in our view, a clearer graphical representation of the 2TCFs. A summary (§8) and an appendix that introduces a geometric interpretation of the multitime correlation functions in terms of metric spaces (Appendix A) close the article.
2. Calculation of the correlation functions
2.1. Autocorrelation function
We start by constructing the onetime correlation function (autocorrelation) for a generic set of data. The time autocorrelation function of a process u(t) is defined by (Goodman, 1985)
Let us consider that in a experiment we measure intensity fluctuations in time and at points in reciprocal space.^{3} We can express a set of measured intensity fluctuations at different times as an ntuple:
where the terms I_{j} are the intensity fluctuations measured at times and are real numbers. A generic sample function is displayed in Fig. 1. The autocorrelation function is defined, in discrete form, by
where . The terms of equation (8) corresponding to the different delay times (τ) are shown in Table 1. There are terms for a given delay time. For , (N + 1) terms are averaged, for , N terms and so on.

2.2. Twotime correlation function
If the process is not stationary, the statistical properties of the fluctuations will evolve over time. Thus, the summation and averaging that is done over the measurement time in equation (8) is not appropriate. A more general expression, namely a twotime correlation function (2TCF), is obtained if the average in equation (8) is not performed. The 2TCF is very useful to analyse the dynamics of nonequilibrium systems (Sutton et al., 2003). The temporal fluctuations and the variance of the 2TCF are also used to investigate dynamical heterogeneities in glassy systems through the analysis of higherorder correlations and multipoint dynamic susceptibilities [see Orsi et al. (2012) and Conrad et al. (2015) for recent XPCS work and references therein for details on the use of higherorder correlations to study dynamical heterogeneities].
The 2TCF is obtained by calculating the Cartesian product of [equation (7)] with itself (see also Appendix A for the calculation of the 2TCF using the terminology of metric spaces) and ensemble averaging over equivalent momentum transfer vectors or pixels, when using a twodimensional detector (Lumma et al., 2000):
The 2TCF is a symmetric matrix by construction.^{4} That is, the 2TCF is symmetric upon index swapping, i.e. : I_{i}^{j} = I_{j}^{i}, or, equivalently, C(t_{1},t_{2}) = C(t_{2},t_{1}) = C(t_{1},t_{2})^{T}, where T denotes the transpose operation. The time difference between two elements of the 2TCF matrix is obtained using an L^{1}metric (also known as Manhattan, cityblock or taxicab metric; Deza & Deza, 2014); the temporal distance (in units of scaled time) between two points I_{i1}^{j1} and I_{i2}^{j2} is obtained from the sum of the absolute value of the differences between their row and column indexes:
The elements with equal row and column indexes (terms of the form I_{i}^{i}) are `equaltime' terms. For a generic equaltime term I_{i}^{i}, if the start of the experiment is taken as t = 0 for i = 0, the time elapsed from the start of the experiment is t_{obs} = i. We shall call this elapsed time the observation time t_{obs}. The temporal distances [equation (10)] can be converted into absolute time differences by multiplying them by the time step . The autocorrelation function equation (8) is obtained by averaging the terms along lines parallel to the t_{1} = t_{2} diagonal.
3. Analysis of twotime correlation functions using different time coordinate systems
The evolution of the correlation functions is often quantified by selecting slices of the 2TCFs at different observation times. These slices can be taken in different ways, using different coordinate systems. We discuss here the two most common procedures in the literature. Before proceeding, we should note, however, that other time variables such as t_{1}, t_{2} could also be used to define an `observation time'.
3.1. Conventional coordinate system
Starting from an equaltime term I_{i}^{i} in the 2TCF matrix [equation (9)], 1TCFs at different observation times can be extracted by taking the delay time along rows or columns in equation (9), i.e. lines with t_{1} = constant or t_{2} = constant. The terms in these 1TCFs have the form and the delay time τ is given by . This way of extracting 1TCFs arises directly from equation (8), removing the summation over i and the normalization factor that takes into account the number of terms summed for each delay time. We shall call this coordinate system the `conventional coordinate system' (CCS), although we note that this coordinate system is rarely used in the XPCS literature to analyse 2TCFs. The terms at different delay times are shown in Table 2. The autocorrelation [equation (8)] is obtained from the 2TCF by averaging the terms at equal delay times τ of all the CCS1TCFs extracted at different observation times.

3.2. Alternative coordinate system
An `alternative coordinate system' (ACS) was introduced by Brown et al. (1997) and has become the customary coordinate system in XPCS to analyse the 2TCFs to extract 1TCFs from them at different sample ages. In this ACS, the sample age is taken along the t_{1} = t_{2} diagonal and defined as t_{age}: = (t_{2}+t_{1}) / 2. The delay time magnitude is the same as in the CCS system (i.e. ), but starting from an equal term I_{i}^{i}, the delay time direction is taken along lines perpendicular to the t_{1} = t_{2} diagonal (see Fig. 2b). Cuts of the 2TCF along these perpendicular lines are 1TCFs and are defined as `constant sample age' cuts. These 1TCFs are symmetric by construction around the value. The terms at different delay times that are obtained for a given observation time t_{obs} = i, following the definition of Brown et al. (1997), are schematically shown in equation (11) (bold elements) and displayed in Table 2. Using the ACS, the autocorrelation function [equation (8)] is also obtained by averaging the terms at different delay times.
3.3. Differences between the CCS and ACS onetime correlation functions
There are essential differences between the 1TCFs that are extracted using the conventional or the alternative coordinate systems. The CCS1TCF that passes through an equaltime term I_{i}^{i} has terms of the form , while the ACS1TCF comprises terms of the form (τ even) or (τ odd) (see Table 2). Thus, for a 1TCF extracted for an observation time^{5} t_{obs} = A we can note two differences:
(1) The terms of the CCS1TCF are of the form and thus are directly related to the intensity measured at time t_{obs} = A. On the other hand, in a `constant sample age' cut of the 2TCF using the ACS, the terms have the form : terms obtained from the multiplication of intensities measured at times before and after t_{obs} = A are mixed (see terms in Table 2). The ACS1TCF correlates terms that are equidistant (in time) from the observation time t_{obs} = A, but these terms are not directly related to the intensity at the observation time t_{obs} = A, except for the terms .
(2) The number of terms of the 1TCFs extracted for t_{obs} = A using the CCS or the ACS are different. For the CCS, the longest 1TCF that can be extracted is at t_{obs} = 0 (beginning of the experiment). In contrast, using the ACS, the longest delay times accessible are for observation times , while the delay times accessible close to the start or end of the experiment are much shorter (see Fig. 3).
The first difference arises from the difficulty of defining precisely a `constant sample age' in the case of time correlation functions. Time correlation functions are constructed by multiplying terms measured at different times. What happens at a certain time is related to another event at another time. `Constant sample age' is then ambiguous and may be interpreted or defined in different ways. On one hand, a `constant' age may be considered what happens to the state of the sample at time t_{age} = A when it is related to its state at other times. This interpretation would be in line with the analysis done using the CCS. Or it may be interpreted as what happens when events that occur before and after t_{age} = A and at an equidistant time delay are compared, which would correspond to the ACS analysis.
The second difference is just a consequence of the choice of the coordinate system. However, the direction in which the delay time is taken it is extremely important when performing quantitative analysis, because, for nonequilibrium systems, the relaxation times obtained from CCS or ACS1TCFs will be different (see §5.3).
4. 2TCFs for stationary and nonstationary systems
In stationary systems (strictly speaking, for widesense stationary systems; see Goodman, 1985), the 1TCFs depend only on the time difference, not on the observation time. Therefore, the 2TCF of a widesense stationary system is a Toeplitz matrix, i.e. the following relationship between the terms of the 2TCF holds: . In addition, as the 2TCF is symmetric around the t_{1} = t_{2} diagonal, then . The 2TCF [equation (9)] of a widesense stationary process thus has the form
For widesense stationary systems, the CCS and ACS 1TCFs are therefore equivalent, except for the number of terms for each 1TCF. The time symmetry is also maintained for stationary processes. The autocorrelation function of a real (i.e. not complex) stationary process has the following property (Goodman, 1985):
This property is fulfilled for the CCS and ACS1TCFs of stationary processes because (CCS) and (ACS). However, in nonstationary processes equation (13) does not necessarily hold (i.e. the time symmetry is broken). Besides, the 1TCFs will generally depend on the observation time t_{obs} and the delay time τ. The breakdown of the time symmetry is well reflected in the CCS coordinate system: nonstationary processes yield asymmetric CCS1TCFs. However, even for nonstationary processes, equation (13) is fulfilled for the ACS1TCFs because they are symmetric by construction.
5. Examples
It is illustrative to compare the CCS1TCF and ACS1TCF for some model, extreme cases. Three examples are presented below: the first two examples are based on simple mathematical functions and the third is based on the integration of a partial differential equation that has been proposed to describe the evolution of a semiconductor surface upon ion beam sputtering (Castro et al., 2005). These examples have been chosen not for their physical relevance but because they reflect well some of the issues that arise when using different coordinate systems to extract 1TCFs from 2TCFs. The first example (§5.1: intensity following a step function) manifests that the ACS convention breaks the causality by mixing terms before and after an event has happened. In the second example (§5.2: sinusoidal intensity variation), the ACS1TCFs give skewed correlation functions and the skewness depends on the observation time. The third example (§5.3: 2TCF of selforganized nanostructure formation dynamics on a surface due to sputtering) shows that, for an ageing system, the choice of the delay time direction has a direct effect on the correlation times and can also affect the functional shape of the correlation function.
In all the examples, we assume that the functions used in the calculations are representative of the dynamics of the system, i.e. that proper corrections, normalization and ensemble averaging of the raw data have been performed, as would indeed be required in a real DLS or XPCS experiment (for details, see e.g. Chu, 2007; Wong & Wiltzius, 1993; Madsen et al., 2010; Madsen et al., 2015).
5.1. Correlation function of a step intensity function
We consider a dynamical system yielding intensity fluctuations in the scattered signal that can be described by a step function:
The signal profile is plotted in Fig. 2(a): the signal jumps from 1 to −1 at T_{1} = 650 and goes back to 1 at T_{2} = 800. The corresponding 2TCF is shown in Fig. 2(b). CCS and ACS 1TCFs extracted at observation times are displayed in Fig. 3.
We observe that the CCS1TCFs are correlated from the observation time t_{obs} and until the end of the period () and that in the following period they are anticorrelated (i.e. C = −1). However, the ACS1TCFs show correlation from the observation time until delay times that are twice those of the CCS1TCFs. This is due to the different delay time directions and the Manhattan geometry of the 2TCF; the CCS1TCF follows a line while the ACS1TCFs follow a staircase trajectory [see equation (11)]. For this reason, the ACS1TCFs change sign at different delay times than the CCS1TCFs. The ACS1TCFs are correlated for delay times that are longer than the difference between the observation time and the switching of the intensity which, physically, is inconsistent.
5.2. Correlation function of a periodically oscillating intensity
Let us consider a system where, because of its dynamics, the scattered intensity fluctuates around a constant mean value in a sinusoidal way with angular frequency ω. The intensity fluctuation can be represented as , where is the phase at time t = 0. For simplicity, we take in the following. For such a signal, comparing the signal at time t with itself for different delay times τ, it is expected that the correlation of the signal should vary periodically from positive to negative. The autocorrelation, as calculated using equation (6), is
The 2TCF is shown in Fig. 4. The terms of the 2TCF have the form . The 1TCFs that are extracted from the 2TCF following the CCS or the ACS convention have the following form:
The 1TCFs extracted at t_{obs} = 284, 610, 895 with are shown in Fig. 5. Using the CCS, the (a priori) expected behaviour is reflected in the 1TCFs, namely, the correlation oscillates periodically from positive values to negative ones and vice versa. The amplitude of the oscillations of a 1TCF at time t_{obs} is determined by the value of . The amplitudes of the CCS1TCFs are symmetric around zero.
However, the 1TCFs obtained with the ACS have a different behaviour: they may sometimes be always positive or negative. The behaviour of the 1TCF extracted at t_{obs} following the ACS convention can be determined more easily by rewriting equation (16) as
where we have used trigonometric identities to rewrite the expression. The two terms in equation (17) are positive and the amplitude of the ACS1TCFs will oscillate between . Thus, if (), will always be positive (negative). This can be observed in the top panel of Fig. 5 (t_{obs} = 284): the correlation is always positive. For other t_{obs} values, the amplitude variation of the correlations is not symmetrical and will be skewed to positive or negative values unless .
5.3. Surface evolution under ion beam sputtering
Ion beam sputtered surfaces are nonequilibrium systems that show ageing (Bikondoa et al., 2013). One theoretical approach to describe the temporal evolution and dynamics of such systems is the continuum theory, which uses partial differential equations to describe the evolution of the surface height (MuñozGarcía et al., 2009). Fig. 6 displays the 2TCF obtained from numerical simulations integrating an equation that describes the evolution of semiconductor surfaces under ion bombardment [for more details on such systems and the calculation of the 2TCF, see Bikondoa et al. (2012), and references therein]. In Fig. 7, we have extracted CCS and ACS 1TCFs from Fig. 6, for t_{obs} = 100. In the case of the ACS1TCF (open circles), only delay values up to are accessible. For the CCS1TCF (crosses), a delay time up to can be extracted. The two 1TCFs have been fitted using a stretched exponential , where is the correlation time and α is the Kohlrausch–Williams–Watts exponent (Pecora, 2008). The value of the exponent α depends on the microscopic nature of the dynamics (Madsen et al., 2010). Only the values in the range have been used for the fit. This example shows (see values in Fig. 7) that, for a nonequilibrium system, there may be important differences in both the correlation times and the α exponents that are obtained using one convention or the other. Such differences may be extremely important when interpreting results from 2TCFs and modelling the underlying dynamics (see the Discussion).
6. Discussion
Which coordinate system should be used when analysing 2TCFs and extracting 1TCFs? A priori, either of the two coordinate systems can be used, provided, obviously, the comparison with theoretical models is done accordingly. This is the procedure followed in the pioneering work of Brown et al. (1997), in which computer simulations are used to study the statistical properties of speckles arising from the scattering of coherent radiation by a phaseordering system. Theoretical models for such systems predict twopoint, twotime correlation functions of the (i.e. the scalar field that describes the inhomogeneity of the system). The which is obtained by averaging the scattered intensity over an ensemble of initial conditions, is related to the modulus square of the Fourier transform of the Brown et al. (1997) analysed the intensity 2TCFs using the ACS1TCF reference system, and the comparison with theoretical models and the scaling functions that they predict was done taking into account the ACS modified coordinates. The same procedure has been used in subsequent theoretical and experimental work on related or similar systems (Brown et al., 1997; Livet et al., 2001; Fluerasu et al., 2005). However, for most cases the interpretation of the CCS1TCFs is more straightforward because its calculation is in line with the usual way of calculating time correlation functions in statistical mechanics: a function of the state of the system at an initial time is multiplied by the value of the function at another, later time t; the autocorrelation function is defined as the ensemble average of that product (Zwanzig, 1965). CCS1TCFs are also in accordance with the use of dynamic correlations and response functions to analyse how a function of the system responds to a perturbation applied at a certain time t_{p} [for an account of the relationships between response and correlation functions, see Chaikin & Lubensky (1995) or Cugliandolo et al. (1994)]. The time symmetry is broken by applying an external field or force at time t_{p}. The response function will be nonzero only for . To account for this, a step function dependence on the time is often included in the definition of the response function. As shown in the example of §5.1, causality between terms of the correlation function is not retained for the ACS1TCFs and events that happen before and after the perturbation has occurred (i.e. t_{p}) are then mixed. Thus, for the analysis of such systems, the use of CCS1TCFs seems to be better suited. The same applies for quenched systems: ACS1TCFs would mix events prior and subsequent to the quenching. This could be avoided if the ACS analysis is restricted to areas in the TCF that are not crossed by any of the `events'. That would entail remaining inside a single square area (either red or blue, in Fig. 2) without crossing the boundary to another area.
Extracting CCS1TCFs from the 2TCFs is an equivalent procedure to that employed to analyse the contact dynamics on granular piles subjected to weak vibrations using multispeckle diffusive wave spectroscopy (MDWS) (Kabla & Debrégeas, 2004). A waiting time is used to account for the number of vibrations the system has suffered before the measurement starts and a delay time for the number of vibrations after the waiting time. The waiting time is equivalent to the `observation time' (t_{obs}) that has been defined above. The slow dynamics in glasses studied with dynamic have also been analysed in a similar manner, using a waiting time or sample age (Cipelletti et al., 2000). In these two studies, the 2TCF is not explicitly employed. We note here that theories of nonequilibrium phenomena are generally expressed in terms of correlations that follow the CCS formulation (see e.g. Van Vliet, 2008; Berthier et al., 2011).
Ageing phenomena in glasses and other outofequilibrium systems have been extensively studied with XPCS using 2TCFs and ACS1TCFs (Madsen et al., 2015; Bikondoa, 2016). Thus, to compare quantitative values extracted from ACS1TCFs with values obtained using other experimental techniques (e.g. MDWS or DLS) or theoretical predictions, it may be necessary to perform a coordinate change to analyse the results appropriately. Unfortunately, this point is not always clear in the literature. Instances can be found in which the width of the diagonal contour is taken as being proportional to the (Ruta et al., 2013; Bikondoa et al., 2013) – i.e. the ACS1TCF convention is used – and where quantitative values of the and the stretching parameter at different sample ages are reported. However, it would have been more natural to report quantitative values obtained following the CCS1TCF convention, as this is the one habitually used in glassy systems theory (Wolynes & Lubchenko, 2012). But because the ageing is so slow in the systems studied by Ruta et al. (2013) and Bikondoa et al. (2013), the ACS and CCS1TCFs are essentially equivalent. In other work (Müller et al., 2011), it is unclear if the 1TCFs extracted from a 2TCF that has sharpcut division due to avalanche dynamics follow the CCS or the ACS convention. The example of the step function presented here in §5.1, suggests that the CCS1TCFs would be more suitable to analyse avalanchetype dynamics, and this may have been the procedure followed by Müller et al. (2011). But the reference provided by Müller et al. (2011) to explain how the 1TCF has been calculated corresponds to work where the ACS1TCF was used (Malik et al., 1998). Which reference system has been used by Shinohara et al. (2015) to extract 1TCFs from 2TCFs is not clear either. As there are different possible ways to extract 1TCFs from 2TCFs, it is important to explain precisely how the analysis has been carried out.
Nonequilibrium systems are arguably the most interesting cases to use the 2TCF. Equilibrium systems are time translation invariant (Forster, 1995) but time symmetry is not retained in nonequilibrium systems. This symmetry break is reflected in the CCS1TCFs, but ACS1TCFs keep the symmetry for both equilibrium and nonequilibrium processes. On this basis, the CCS convention seems more convenient for the analysis of dynamic processes on nonequilibrium systems. But this does not preclude using the ACS convention if the theoretical analysis justifies it, as was done by Brown et al. (1997) and in subsequent work on the nonequilibrium dynamics of ordering systems and firstorder transitions (see the references in the Introduction). Notwithstanding, we remark that for equilibrium systems both coordinate systems lead to the same result and that for systems in quasiequilibrium the quantitative differences may be minor. The 2TCFs could certainly be analysed using other slicing methods if the dynamics under study and their physical interpretation require it. A generic procedure to extract onetime correlation functions from multitime correlation functions is presented in Appendix A.
7. Alternative representation of the twotime correlation function
We propose an alternative way to display graphically the 2TCF in a way that the CCS coordinate system is more apparent. The 2TCF elements are plotted according to the following matrix:
For a generic matrix term I_{i}^{j} in equation (18), the observation time is t_{obs} = i and the delay time . Graphically representing equation (18), the observation and delay times are along the vertical and horizontal axes, respectively (see Fig. 8). Negative/positive delay times correspond to going backward/forward in time. One advantage of this representation is that the 1TCFs at different observation times are visualized more easily as horizontal lines. The autocorrelation function is obtained by averaging the rows instead of having to average diagonals. It also shows that with increasing sample age there are fewer terms for each of the 1TCFs. In this graphical representation, the skewness and kurtosis of the peak at could be used to quantify the degree of departure from equilibrium and the correlation times. This assertion should still be cautioned: further theoretical developments are needed to verify if indeed the skewness and kurtoisis can meaningfully be related to the deviation from equilibrium, but the idea looks attractive.
8. Summary
We have compared two coordinate systems that are used to analyse twotime correlation functions and extract onetime correlation functions from them. We have shown that taking onetime correlation functions along rows or columns (CCS1TCFs) is more compatible with the way autocorrelation functions are generally calculated and theoretical results reported. In certain cases, these CCS1TCFs are more consistent physically and do not present causality problems. Importantly, the CCS1TCFs are not necessarily symmetric by construction and thus a lack of time symmetry indicates that the system is not stationary. For nonequilibrium systems, the correlation and delay times that are obtained with this coordinate system differ from the ones that are obtained using the convention introduced by Brown et al. (1997) (ACS1TCFs). A new graphical representation of the 2TCFs has been introduced, where the observation time is represented along the vertical axis and the delay time along the horizontal.
APPENDIX A
Geometric description of multitime correlation functions
We show here that multitime (equivalently, multipoint) correlation functions can be conveniently expressed in terms of the formalism of metric spaces. Correlation functions of lower order are obtained using an adequate metric and defining a geometric trajectory in the multidimensional space. We describe how to construct generic αtime correlation functions from operations between Ntuples and how onetime correlation functions can be extracted from them. We pay special attention to the case and the physical interpretation of the possible trajectories. We restrict ourselves to the correlation between only one variable. The generalization to correlations between different variables (cross correlations) is straightforward. A comprehensive discussion of arbitraryorder correlation functions using a tensor formalism, with special emphasis on coherence properties, is given by Mandel & Wolf (1995).
Let the tuple be a set of measurements of the variable X(t) made at times . Thus, the tuple indexes are related to the time the measurement was done. The time difference (or temporal distance) between measurements is . Using the we build an αdimensional array
where α is the order of the correlation function we want to obtain. Each element of the array is a tuple with α elements.
For each term in the array , we define the function = = = , where . is the product of the elements of each αtuple and yields the correlation function of order α. From , to extract an thorder correlation function we need to select a dimensional subset of . Here, we sketch a method to obtain onetime correlation functions from an thorder correlation function.
First, we need to use a metric that defines the distance between the elements in the set. The set and the metric define a `metric space' (Reed & Barry, 1980). To extract a onetime correlation function from we define a trajectory on , . Starting from a point , the trajectory is chosen such that it joins points that are at consecutively larger distances in the grid. The distance depends on the metric used.
The trajectories starting from a point can be generically described as a set of points at successive r distances from P:
where d(y,x) is the metric. As explained above, for a generic element , the indexes indicate which element of the tuple of measurements X(t) are being multiplied when the function is calculated, and are related to the time when the elements were measured. The Manhattan metric gives the distance between two elements in as the sum of the absolute differences between their indexes:
where and are two generic points in . The Manhattan distance corresponds to the case p = 1 of the L_{p} norm (Deza & Deza, 2014):
The Manhattan distance is the equivalent of the delay time. In an grid, there are many different ways (`trajectories') to join points at monotonically increasing Manhattan distances. In general, the onetime correlation functions along different trajectories starting at a point P will not be equivalent. We analyse in the following the trajectories on .
A1. Trajectories in a twotime correlation function
A 2TCF can be represented by a twodimensional grid or matrix (see §2.2). The Moore neighbourhood of a point in a twodimensional grid is the set of points surrounding it (Deza & Deza, 2014). If we denote the surrounding points using four cardinal (N, E, S, W) and four intercardinal points (NE, SE, SW, NW), the equaltime diagonal (i.e. terms of the form X_{i,i}) goes from the SW corner to the NE one (see Fig. 9). The allowed trajectories following oneunit step sizes of the Manhattan distance have individual steps going only along any of the four cardinal directions. With the Manhattan metric, trajectories along the intercardinal directions are obtained as staircase paths. X^{(2)} = X×X is symmetric by definition upon index swapping (i.e. X_{i,j} = X_{j,i}), so we restrict ourselves to trajectories that remain in only one part of X^{(2)}, under the equaltime diagonal. Under there requisites, the most relevant trajectories, or at least those with a clear physical interpretation, are the trajectories starting at an equaltime point P = (p,p) and which go only eastwards (E), southwards (S) or southeastwards (SE):
E. The eastwards trajectory mixes the event (measurement) at point P with measurements done at later times. This trajectory is equivalent to the usual autocorrelation function [equation (8)] except that there is no average between the trajectories that start at every point of the equaltime diagonal. The pair terms in the trajectory are of the form , where τ is given by the Manhattan distance. Averaging all the E trajectories for every point P on the equaltime diagonal, one obtains the usual autocorrelation function.
S. The southward trajectory relates the event (measurement) at point P with measurements done at earlier times. Thus, it can be interpreted as a correlation function where the delay time goes backwards in time. The terms in the trajectory are of the form . The S trajectory of any point P is the same as the W trajectory. Averaging all the S trajectories for every point P on the equaltime diagonal, one obtains the usual autocorrelation function.
SE. Using the Manhattan distance, the SE trajectory can only be obtained following a staircaselike trajectory. Depending on the choice of the term at a Manhattan distance equal to 1, the starting point P will be at the bottom or the top of the stair riser. In the SE trajectory, the event at time P only appears in the term at Manhattan distances 0 and 1. Terms at relate events that happen before and after the event at P. The terms are of the form .
E, S or purely SE trajectories can be obtained using a Chebyshev metric instead of the Manhattan one as the selecting rule for the terms along a trajectory. The Moore neighbourhood of a point P is the set of points that are at a Chebyshev distance equal to 1. The Chebyshev metric corresponds to the case of the L_{p} metric [equation (22)] and is defined as follows:
Two points in a grid at distance d_{Ch} = 1 can be joined by a unit displacement along any of the cardinal or intercardinal directions, i.e. by the movement of the king on a chessboard (see Fig. 10). The Chebyshev distance is also called the `chessboard' or `kingmove' metric (Deza & Deza, 2014). A 1TCF extracted from an SE trajectory starting at a point P of the equaltime diagonal and joining points at increasing Chebyshev distances is composed of terms arising from the multiplication of two events that happen at delay times  d_{Ch} and + d_{Ch}, respectively (see Fig. 10).
There is an important difference between the 1TCFs obtained using a Manhattan or a Chebyshev metric. The 1TCFs obtained with a Manhattan metric always relate events that are at a unit delay time, whatever the direction of the steps is. However, using the Chebyshev metric, the delay time between the events that are related depends on the direction chosen. For 1TCFs along only E or S (or staircase trajectories), the delay time is always 1. Along diagonals, the delay time between the related events is 2. That is, for a step with a Chebyshev distance equal to 1, the time step can in fact be 1 or 2.
It is clear that, depending on the metric used and the trajectories chosen, many different 1TCFs can be constructed, which, in general, will not be equivalent. Other common metrics (for example, the Euclidean, which corresponds to the L_{p} with p = 2 case, and coincides with the Manhattan one if ) can yield completely different 1TCFs from the Manhattan or Chebyshev metrics. In this particular case of timecorrelation functions, the Manhattan norm yields a clear physical picture for any dimensions, because d_{Man} = 1 always relates events that are separated by the same delay time, independently of the direction chosen in the αmultidimensional space. For point (position) correlation functions obtained from measurements on a plane, Euclidean metrics would be better suited. The physics of the problem treated will determine which metric should be used.
Footnotes
^{1}The Siegert relation is valid if the coherence volume contains a large number of independent scatterers. In that case the central limit theorem conditions are met (Goodman, 1985). The total scattered field, which is the sum of many independent scatterers, is a random variable with a Gaussian probability distribution and the time correlation function of the intensity can be factorized to obtain the Siegert relation (Pusey, 2002).
^{2}The order in which the time and ensemble averages are performed can be very important. For example, information about the ergodicity of the system may be lost if the time average is done before the ensemble average (Pusey & Van Megen, 1989), but when using area detectors the time average should be done first to preserve the speckle visibility and to be able to extract measurement errors directly from azimuthal variations of (Lumma et al., 2000). However, multispeckle dynamic (DLS)/PCS analysis is often done by performing the ensemble average first (Cipelletti & Weitz, 1999; Chushkin et al., 2012). Thus, the choice of the top or bottom formula between the braces in equation (1) depends on the case under study. This aspect is beyond the scope of this manuscript. For more details, see the aforementioned references.
^{3}For the purposes of this paper, we will ignore the dependence of the intensity fluctuations on the scattering vector and, to simplify the notation, we will not consider the normalization terms. See for example Chushkin et al. (2012), Madsen et al. (2015), Möller et al. (2016) and references therein for technical details about the measurement of intensity fluctuations and data processing.
^{4}In the usual matrix representation, the lowest row and column term is at the top left corner. Here, we represent the matrix setting the term with the lowest row and column index at the bottom left corner as the 2TCF in XPCS is generally represented graphically in this manner.
^{5}In the following, we consider that using the CCS the 1TCF is extracted along the rows. If it were extracted along the columns, the result would be equivalent owing to the symmetry of the 2TCF matrix along the I^{i}_{i} diagonal.
Acknowledgements
OB gratefully acknowledges the financial support of the UK Engineering and Physical Sciences Research Council (EPSRC). The author would also like to thank Gerardina Carbone and Didier Wermeille for invaluable discussions.
References
Berthier, L., Biroli, G., Bouchaud, J.P., Cipelletti, L. & Van Saarloos, W. (2011). Dynamical Heterogeneities in Glasses, Colloids and Granular Media. Oxford University Press. Google Scholar
Bikondoa, O. (2016). Xray and Neutron Techniques for Nanomaterials Characterization, edited by C. S. S. R. Kumar. Berlin: SpringerVerlag. Google Scholar
Bikondoa, O., Carbone, D., Chamard, V. & Metzger, T. H. (2012). J. Phys. Condens. Matter, 24, 445006. Web of Science CrossRef PubMed Google Scholar
Bikondoa, O., Carbone, D., Chamard, V. & Metzger, T. H. (2013). Sci. Rep. 3, 1850. Web of Science CrossRef PubMed Google Scholar
Brown, G., Rikvold, P. A., Sutton, M. & Grant, M. (1997). Phys. Rev. E, 56, 6601–6612. Web of Science CrossRef CAS Google Scholar
Brown, G., Rikvold, P. A., Sutton, M. & Grant, M. (1999). Phys. Rev. E, 60, 5151–5162. Web of Science CrossRef CAS Google Scholar
Castro, M., Cuerno, R., Vázquez, L. & Gago, R. (2005). Phys. Rev. Lett. 94, 016102. Web of Science CrossRef PubMed Google Scholar
Chaikin, P. M. & Lubensky, T. C. (1995). Principles of Condensed Matter Physics. Cambridge University Press. Google Scholar
Chu, B. (2007). Laser Light Scattering: Basic Principles and Practice. Mineola: Dover Publications. Google Scholar
Chushkin, Y., Caronna, C. & Madsen, A. (2012). J. Appl. Cryst. 45, 807–813. Web of Science CrossRef CAS IUCr Journals Google Scholar
Cipelletti, L., Manley, S., Ball, R. C. & Weitz, D. A. (2000). Phys. Rev. Lett. 84, 2275–2278. Web of Science CrossRef PubMed CAS Google Scholar
Cipelletti, L. & Weitz, D. A. (1999). Rev. Sci. Instrum. 70, 3214–3221. Web of Science CrossRef CAS Google Scholar
Conrad, H., Lehmkühler, F., Fischer, B., Westermeier, F., Schroer, M. A., Chushkin, Y., Gutt, C., Sprung, M. & Grübel, G. (2015). Phys. Rev. E, 91, 042309. Web of Science CrossRef Google Scholar
Cugliandolo, L. F., Kurchan, J. & Parisi, G. (1994). J. Phys. I, 4, 1641–1656. Google Scholar
Deza, M. & Deza, E. (2014). Encyclopedia of Distances, 3rd ed. Heidelberg: Springer. Google Scholar
Fluerasu, A., Moussaïd, A., Madsen, A. & Schofield, A. (2007). Phys. Rev. E, 76, 010401. Web of Science CrossRef Google Scholar
Fluerasu, A., Sutton, M. & Dufresne, E. M. (2005). Phys. Rev. Lett. 94, 055501. Web of Science CrossRef PubMed Google Scholar
Forster, D. (1995). Hydrodynamic Fluctuations, Broken Symmetry and Correlation Functions. Boulder: Westview Press. Google Scholar
Goodman, J. W. (1985). Statistical Optics. New York: John Willey and Sons. Google Scholar
Grübel, G., Madsen, A. & Robert, A. (2008). Soft Matter Characterization, edited by R. Borsali & R. Pecora, pp. 953–995. Berlin: Springer. Google Scholar
Gutt, C. & Sprung, M. (2015). Xray Diffraction: Modern Experimental Techniques, edited by O. H. Seeck & B. M. Murphy, pp. 385–419. Singapore: Pan Stanford Publishing. Google Scholar
Kabla, A. & Debrégeas, G. (2004). Phys. Rev. Lett. 92, 035501. Web of Science CrossRef PubMed Google Scholar
Livet, F., Bley, F., Caudron, R., Geissler, E., Abernathy, D., Detlefs, C., Grübel, G. & Sutton, M. (2001). Phys. Rev. E, 63, 036108. Web of Science CrossRef Google Scholar
Livet, F. & Sutton, M. (2012). C. R. Phys. 13, 227–236. Web of Science CrossRef CAS Google Scholar
Loudon, R. (1983). The Quantum Theory of Light, 2nd ed. New York: Oxford University Press. Google Scholar
Ludwig, K., Livet, F., Bley, F., Simon, J.P., Caudron, R., Le Bolloc'h, D. & Moussaid, A. (2005). Phys. Rev. B, 72, 144201. Web of Science CrossRef Google Scholar
Lumma, D., Lurio, L. B., Mochrie, S. G. J. & Sutton, M. (2000). Rev. Sci. Instrum. 71, 3274–3289. Web of Science CrossRef CAS Google Scholar
Madsen, A., Fluerasu, A. & Ruta, B. (2015). Synchrotron Radiation and FreeElectron Lasers, edited by E. Jaeschke, S. Khan, J. R. Schneider & J. B. Hastings. Berlin: Springer. Google Scholar
Madsen, A., Leheny, R. L., Guo, H., Sprung, M. & Czakkel, O. (2010). New J. Phys. 12, 055001. Web of Science CrossRef Google Scholar
Malik, A., Sandy, A. R., Lurio, L. B., Stephenson, G. B., Mochrie, S. G. J., McNulty, I. & Sutton, M. (1998). Phys. Rev. Lett. 81, 5832–5835. Web of Science CrossRef CAS Google Scholar
Mandel, L. & Wolf, E. (1995). Optical Coherence and Quantum Optics. Cambridge University Press. Google Scholar
Möller, J., Chushkin, Y., Prevost, S. & Narayanan, T. (2016). J. Synchrotron Rad. 23, 929–936. Web of Science CrossRef IUCr Journals Google Scholar
Müller, L., Waldorf, M., Gutt, C., Grübel, G., Madsen, A., Finlayson, T. R. & Klemradt, U. (2011). Phys. Rev. Lett. 107, 105701. Web of Science PubMed Google Scholar
MuñozGarcía, J., Vázquez, L., Cuerno, R., SánchezGarcía, J. A., Castro, M. & Gago, R. (2009). Toward Functional Nanomaterials, edited by Z. M. Wang, pp. 323–398. New York: Springer. Google Scholar
Orsi, D., Cristofolini, L., Baldi, G. & Madsen, A. (2012). Phys. Rev. Lett. 108, 105701. Web of Science CrossRef PubMed Google Scholar
Orsi, D., Cristofolini, L., Fontana, M. P., Pontecorvo, E., Caronna, C., Fluerasu, A., Zontone, F. & Madsen, A. (2010). Phys. Rev. E, 82, 031804. Web of Science CrossRef Google Scholar
Pecora, R. (2008). Soft Matter Characterization, edited by R. Borsali & R. Pecora, pp. 953–995. Berlin: Springer. Google Scholar
Pusey, P. N. (2002). Neutrons Xrays and Light: Scattering Methods Applied to Soft Condensed Matter, edited by P. Lindner & Th. Zemb, pp. 203–220. Amsterdam: Elsevier North Holland. Google Scholar
Pusey, P. N. & Van Megen, W. (1989). Phys. A Stat. Mech. Appl. 157, 705–741. CrossRef CAS Web of Science Google Scholar
Reed, M. & Barry, S. (1980). Methods of Modern Mathematical Physics. I. Functional Analysis. New York: Academic Press. Google Scholar
Ruta, B., Baldi, G., Monaco, G. & Chushkin, Y. (2013). J. Chem. Phys. 138, 054508. Web of Science CrossRef PubMed Google Scholar
Ruta, B., Chushkin, Y., Monaco, G., Cipelletti, L., Pineda, E., Bruna, P., Giordano, V. M. & GonzalezSilveira, M. (2012). Phys. Rev. Lett. 109, 165701. Web of Science CrossRef PubMed Google Scholar
Shinohara, Y., Yamamoto, N., Kishimoto, H. & Amemiya, Y. (2015). J. Synchrotron Rad. 22, 119–123. Web of Science CrossRef CAS IUCr Journals Google Scholar
Sutton, M. (2008). C. R. Phys. 9, 657–667. Web of Science CrossRef CAS Google Scholar
Sutton, M., Laaziri, K., Livet, F. & Bley, F. (2003). Opt. Express, 11, 2268–2277. Web of Science CrossRef PubMed CAS Google Scholar
Sutton, M., Mochrie, S. G. J., Greytak, T., Nagler, S. E., Berman, L. E., Held, G. A. & Stephenson, G. B. (1991). Nature, 352, 608–610. CrossRef Web of Science Google Scholar
Van Vliet, C. M. (2008). Equilibrium and Nonequilibrium Statistical Mechanics. Singapore: World Scientific. Google Scholar
Wolynes, P. G. & Lubchenko, V. (2012). Editors. Structural Glasses and Supercooled Liquids: Theory, Experiment and Applications. Hoboken: Wiley. Google Scholar
Wong, A. P. Y. & Wiltzius, P. (1993). Rev. Sci. Instrum. 64, 2547–2549. CrossRef CAS Web of Science Google Scholar
Zwanzig, R. (1965). Annu. Rev. Phys. Chem. 16, 67–102. CrossRef CAS Web of Science Google Scholar
This is an openaccess article distributed under the terms of the Creative Commons Attribution (CCBY) Licence, which permits unrestricted use, distribution, and reproduction in any medium, provided the original authors and source are cited.