4.1. Solution of the Stochastic Problem
Using a linear transformation involving multiplication by the terms
,
, as well as subtraction to remove the combination of derivatives on the left-hand side, we obtained a more standard form of the stochastic differential equation
where the respective dissipative terms (
) are counterbalanced by the auxiliary random forces
,
. There are no more mixed derivatives of variables in one equation, which results in a qualitative change from white to colored noise. The properties of
,
can be represented by the linear relations
For computational purposes, the system of Equation (
3) is converted to the Fourier domain in a standard way. Fourier images (coefficients), which we begin to denote by a tilde become functions of the angular frequency
. Despite the fact that
or
symbols implying frequency dependence may be redundant in the case of white noise, it emphasizes the dependence on
for general reasons in other situations. It is also worth noting that the complex conjugate’s asterisk label appears after the transition to the Fourier representation. Assuming that cross-correlations of Fourier images
,
vanish as a result of Equations (
2) and (
4), for the relations of the first and the second-order moments, we have
Of course, the properties of the above averages are sufficient to determine multivariate Gaussian random force statistics. More precisely, the consequences of the Gaussian process from the postulates for towards the statements for can be easily justified.
According to the Equation (
3), the respective coefficients
,
,
,
are present in
where
Note that here
is the label of newly introduced
-dependent matrix. The linearity of the problem implies that the solution
can be expressed in the terms of the inverse matrix
. The following elements of the matrix represent a solution of the linear response type
We see that the formulas contain
=
, in the form
=
with the auxiliary real-valued components
and
. We also state that
Next, we will use
often reflected in the results.
4.2. Statistical Averages, Responses to Random Perturbations
This section is about the change to mean values, which are important for the measurement process, interpretation, and data processing. Only the statistics of the sum
, not isolated
,
is observable in the experiment and allows comparison with the model. Thus, for many aspects of the study, only the behavior of
needs to be used to determine experimentally relevant correlations. To understand the statistics of
x, we focus on the Fourier spectrum of autocorrelation function
It is of course convenient to divide it into four independent terms. In the following, these are treated independently by means of Equation (
8). Partial results (so far without an emphasis on the
dependence) are
We can achieve a clearer relationship by including the correlations between the initially imposed random forces
and
from Equation (
5). The pairwise correlations take the form
with
;
, which define the following eight coefficients
For the relations above, we use a notation that also includes the auxiliary symbols
and
. They are interrelated to the combinations
of the prior
terms
Obviously, the emphasis on the symmetry
will help us to handle the complex numbers. Furthermore, we recognize that the identical pairs of indices provide that
. The advantage of the auxiliary notation by means of
is that we obtain
from Equation (
12) in the compact form
We continue the calculation to reveal the terms introduced by Equation (
16)
Note that we used
and
from Equation (
10) to express the result. After substituting these elements into Equation (
17), we come to the relation
which is important to derive measurable results. We will apply a similar procedure later to determine the pairwise correlations to prove the validity of the equipartition theorem.
4.3. Towards Fusing of Theory
and Experiment
Suppose that there are two finite formal limits relevant for the obtaining of the power spectral density in the form
The formula is understood as a postulate, which introduces the duration of the measurement time
[
19] into a part of the procedure at the formal level. The correlations in the Fourier domain can be formally taken as infinite for frequency
. The formal nature of the limits given by Equation (
21) makes it evident that considerations are not fully compatible with the Fourier framework because the measurement is dependent on assumptions about the large time (
) of the measurement.
The occurrence of
and
later in Equation (
23) can be interpreted as the contribution of the power spectral densities of two random force variants given by the
Wiener–Khinchin theorem
considered for
alternatives (see Equation (
2), where the correlation function is defined and integrated). If we extend the application of the formal limit by dividing Equation (
20) with
, we obtain the power spectrum density in the form
In this way, the physical meaning of the coefficients
is revealed. They can also be represented in an independent way by expressing their relation to the absolute temperature
However, this construct also provides information about the dissipative mechanisms. This is built with the idea that fluctuations from random forces are dissipated by the mechanisms represented by the parameters
,
. As provided below in
Section 4.5, the mean potential energy for the respective degrees of freedom can be compared to determine the equilibrium level of energy flow controlled by
and
. The Equation (
24) given above is essentially the case of the general
fluctuation–dissipation theorem introducing the natural heat unit
.
The fluctuation–dissipation theorem is a statistical thermodynamics statement that explains how fluctuations in a detailed balanced system determine its response to applied disturbances. According to this theorem, two opposing mechanisms are responsible for creating a detailed equilibrium in mechanical systems. On the one hand, there are the consequences of the dynamics of a microsphere attached to a nanowire that is damped by the surrounding fluid. Contributions from the internal damping mechanisms of the nanowire also fall into the same category. Even with this damping combination, the mechanical energy is converted into heat. On the other hand, the presence of damping is necessarily accompanied by fluctuations born in the viscous environment. In the case of the surrounding liquid, these fluctuations result in typical random Brownian collisions of liquid molecules with the microbead. In a standard way, the process is interpreted so that on microscopic scales, heat can be converted back into the mechanical energy of the microbead. The internal damping inside the nanowire acts likewise. Summarizing the above statements, we arrive at a specific form of the fluctuation-dissipation theorem, which states that a constant dissipation flux keeps the mean mechanical energy input invariant, while ensuring the production of new fluctuations.
As a result, let us emphasize an important point: Equation (
23) can be modified to account for the temperature effect. With the intention of linking theory with experiment, we attain the expression
The asymptotic, high-frequency consequence of this general result is
At this stage, we benefit from the choice of auxiliary parameters
. Returning to material details is possible using transformations
These auxiliary parameters are positive for a given model specification that operates exclusively with positive
. However, there is also another, more sophisticated level of interpretation. It is interesting and also productive to assume that the result can be written as a sum of two weighted Lorentzian functions
Here
play the role of free parameters, which incorporate information coming from previously introduced
,
. The change to
should be considered as an intermediate step along with other consequences. The key consequence is double Lorentzian form
It is based on the assumption that there exist some relations between
and the corner frequencies
. When Equations (
28) and (
29) are combined, we obtained
In the above solution, we use the consensus that the plus sign corresponds to
. The constraints that allow for such a solution are as follows:
If we consider the transformation to physical parameters in the sense of Equation (
27) to analyze the satisfaction of the above constraints, we obtain
Using Equation (
27), we confirm that
. Moreover, the trivial
implies
≥
. Therefore, there is no obvious contradiction with the fact that
,
correspond to
,
. It is also notable that the inverse transformations
→
→
become
The result is intriguing in terms of revealing the central tendency in and as representatives of the pair , .
4.4. Autocorrelation Function
The findings presented above can be augmented by using direct time representation. According to the well-known
Wiener–Khinchin relation, we have the consequence for the autocorrelation function in the form
As a result, for Equation (
29) as a specific version of
, we obtain a two-exponential autocorrelation function
where
It is worth noting that corner frequency parameters have a significant impact on autocorrelation decrease over time. At a first glance, we can see the essential property here where a pair of frequencies in the Lorentz form corresponds to a pair of damping terms with the typical decay times proportional to
and
. It should also be noted that, assuming that the physical parameters are constant, the temperature is directly manifested only in the amplitude
. The calculations above were performed with the help of a well-known auxiliary relation
with some auxiliary parameter
.
4.5. Sharing of Elastic Energy;
Rationale for Choosing ,
According to the principle of energy equipartition, average energy is evenly distributed among the various degrees of freedom of ergodic systems. As shown here, the implications of this principle are valuable tools for calculating the amplitudes of a pair of random forces. The equipartition principle can be applied to the mean elastic energies. We start by writing the energy for the Fourier modes corresponding to . It is worth noting that since the inertial term is considered negligible, the zero limit of the kinetic energy has no effect on the equipartition issues.
Using the integration techniques already discussed, we continue to utilize the formal limit approach (
) for the integration of the spectrum and averaging over the respective potential energy fluctuations as follows
The formulas below can be applied to complete the integration
These four coefficients include two spectral integrals
Going back to a spectral decomposition using a pair of Lorentzian forms (see Equation (
29)) in combination with Equations (
30) and (
34) gives the following result
As a consequence, the following relationship
=
can be used in the mean potential energies listed below
Finally, in accordance with Equation (
24), we have the confirmation of the equipartition in the form
4.6. The Spectrum Moments
In this subsection, we discuss the usefulness of introducing power spectral density integrals in cases where the frequency domain over which we integrate is divided into non-intersecting intervals. Frequency integration is motivated by the fact that providing excessive detail for spectrum characterization may be unnecessary in certain contexts. The second reason is that aggregation of data helps to suppress statistical errors. The third reason is the possibility of comparing only a few moments with the moments estimated by direct data processing.
Naturally, the analytical form of the model moments simplifies further processing. In our case, the specificity of the moments corresponding to Lorentzian and related spectral forms supports the overall validation process. Let the
regression-related (rr) moments obtained by analytical integration be referred to as
. This notation is used to mean that integration has occurred within the range between the lowest
and the highest
frequencies. Then
Because the interval length may diverge, we decided to use non-normalized moments. Recall that
and
are the two respective amplitudes of the exponentials corresponding to the autocorrelation function (see Equation (
37)). On this basis, using
,
as natural boundaries, we can define the system of three specific
regression-related spectral moments
with the total sum
. Other suitable boundary options are, of course, possible, such as those that do not depend on regression results but instead emerge entirely from generalized averaging procedures of the experimental spectrum.
4.7. Experimental Results and Their Regression
After the successful implementation of the experiment, we obtained data representing the observed dynamics
, which we have then transformed into corresponding Fourier images. The aim was to obtain an experimental power spectrum density
for the system of
frequencies
(see
Figure 3). Some of the evaluations have been performed according to the work of [
19]. Preprocessing with grouping of the adjacent experimental spectral points is a necessary methodological peculiarity. Frequency and spectrum groupings with eight points over the frequency decade were introduced. The effectiveness with which the representative grouping frequencies were allocated was evaluated. Naturally, the grouping process affects not only the locations of representative frequencies, but also the statistics of spectral points, potentially increasing the regression’s feasibility. The optimization of parametric combinations is made possible by data knowledge. Let us formally encapsulate the unknown model parameters in a single symbol
, resulting in the parameterized form of the double Lorentzian model
(see Equation (
29)). In addition to identifying the optimum, we will focus on estimating errors for various components of
.
The problem-specific emphasis is on the asymptotic behavior of the spectrum. Despite the fact that the density of the power spectrum decreases as
, the high frequency domain must be properly included in the regression due to its physical significance. Hence, a weighted regression of the squares of
deviations has been implemented. The preference can be defined as the minimization of the objective function
Here, the parameters and their combinations appear to be formally merged into the vector
This is subject to optimization. We used the standard global function optimizer, which was built on the concept of the [
26] work with the implementation (scypy.optimize.curve_fit(…)) to the SciPy library [
27]. The regression corresponding to
provides the corner frequencies
Finally, there is also fixed corresponding parametric combination
which represents the constant factor in
as defined by Equation (
29). The regression outcomes are depicted in
Figure 3.
Now there is a standard way to find out the autocorrelation function (see Equation (
37)) via the respective parameters
A posteriori evaluation methodology following the regression results provides an implication for the values of the
regression-related spectral moments
We see that the sum of the moments equals , as predicted.
Adjusting the integration boundaries can be important for the design of some alternative test moments. The premise of the adjustment is that these variants should be more closely linked to the measurement process, conditioned by the need to avoid spectral distortions known as “aliasing” and “motion blur”. The effects occur due to too superficial and insufficient sampling of the signal
captured by the camera. This means that the calculations must focus on bands with frequencies less than
, which in our case was set to around a quarter of the Nyquist frequency. The lower limit value
prevents the use of extremely low frequencies. Respecting the lower limit suppresses distortions caused by the apparatus background noise. Numerically, the boundaries we introduce are
and
. The following three moments
were created to express the properties of the experimental data set, which was achieved by partly reducing the impact of the regression results. Here we see that Simpson’s integration quadrature based on uniform data sampling (without grouping) also provides us with variants of spectral moments. However, even when using numerical integration, we must be careful if we subsequently perform comparisons and interpretations. The reason is that certain integrals approximated by a suitable summation can become dependent on the previous regression only by their integration boundaries when these are linked to regression parameters (
and
). Independence from regression can be achieved using descriptive spectrum characteristics (analogous to descriptive statistics). This means using characteristic frequencies in the role of integration boundaries. Then the results of the calculation are generalized spectral averages. The fact that we do not present more moment variants here is mainly related to the focus of this work.
When comparing the Equations (
53) and (
54), we see that only the central moments for the
,
band are close enough to each other, which means that
and
are not sufficient approximations of
and
, respectively. Results show that in the case of regression-related moments, there is only a slight and negligible rise in moments compared to the use of Simpson’s rule