Variational Procedure of Deriving Diffusion Equation for Spreading in Porous Media

We proposed the mathematical model and concrete example of how to use the notion of functional derivatives in order to arrive at a macroscopic equation for dispersion in disordered media. In the sake of simplicity, we considered the case of random process being a Gaussian process.


Introduction
The problem of deriving the governing equations of spreading matters in porous media, derived under different propositions was considered in [1][2][3].Thus in [1], we thoroughly used the fact that centered Gaussian processes are completely determined by the (two points) correlation function.A multipoint correlation appears when a linear fixed-point equation is averaged.That each of these correlations splits into a finite (but increasing with the number of points) number of products of correlation functions was quite important for us.It is possible to use the diagrammatic method with more general processes [3].Then we have much more terms in the expansions.Even if we replace the coefficients of the fixed-point equation for u by functions of a centered Gaussian process, the method is not of a simple use.It is possible to obtain similar results via a variational method, which we will explain for the example of problem, already considered in [1,2].Not surprisingly, we then will retrieve the already obtained macroscopic equation.Then, we will use the variational method for a variant of equation including functions of a centered Gaussian process.

Statement of the Method
The idea goes back to so-called variational (functional) derivatives [4]. Let  is named a variational (or functional) derivative.It is obvious that this derivative is also a functional that depends on function     and point t (as a parameter).In this way we can define the second functional derivative of This is again a functional with respect to     that depends on the couple of parameters   , t t    and so on.In the case when functional  (this is the most interesting case for us) the functional derivative satisfies chain rule: And with Then ratio Also notice that the functional derivative of functional and this is very convenient for dif-ferentiation procedure.
For following purposes we take the functional  .Then we will obtain, according to above formulas:   over function η(τ) about a point η ≈ 0 looks like [4]: where functional shift operator understood in the sense of expansion over infinite integration limits.

Application of Variational Method
In [1][2][3] we used the special summation procedure of diagrams of a certain type and received the diffusion equation with fractional derivatives.In this section, we will exploit another method for to receive the self-contained systems of governing equations for diffusion problems that we have used also in [5].We again consider the spreading of matter in a porous medium such (1) rules the particles transfer on the small scale and we repeat this equation here once more: Averaging with respect to realizations   x  of the random porosity yields This equation contains the unknown u , and also u  and . However, the Furutsu-Novikov formula connects the new unknowns to functional derivatives of u itself, since [4,[6][7][8] the concentration is a functional of . Indeed, we have: , and so on, are functional derivatives of increasing order k or of functional , with respect to By substituting all above into (1), we derive: We see here that together with averaged concentration u we have also averaged values of functional deriva- tives for of different orders: By taking the functional derivative of (1) with respect to  at point 1 x , we obtain Here in turn is a random functional   x  , with two free coordinates which are This equation contains the unknown function and also new unknown functions, which are . If we apply Furutsu-Novikov formula to functionals 1 f and 1 f x substituting the results into (3), we find: (see below ( 4)) We should have written here "and so on", since after that we should take the functional derivative over   , , 2 f x x x .Iteratively using the procedure would lead us to an infinite system of equations for the sequence of functions . The structure of the thus obtained equations is such that the k-th equation in the hierarchy, with unknown k f on the left hand-side has f  , and the concentration u in its right hand-side.And we are led to set the natural question "When can we break this chain of equations?".It turns out that under quite reasonable assumptions, this problem can be solved.
Here we consider random media, where cumulant functions entering Equation (4) are of the even order where 0 a denotes the amplitude of the telegraph or normal random field  which also is to the two-point correlation function of process  .That is why in (4) and as well as in subsequent equations, which are not given here, we can leave only the addends containing the second order correlation functions.In this case, instead of above we get, respectively: We mentioned that if porosity represents itself a normal random field, then ( 5) is exact because all cumulant except one are zero.As we consider small porosity fluctuations, then in the right-hand side of (5) the third and the fourth addend in the right-hand part, containing , , f x x x , will be neglected because they are proportional to .Since also the functional derivative of we arrive at the following system of equations, which turns out to be closed and has the following form: where is the averaged concentration, while is the first functional derivative of the concentration, with respect to    6)-( 7) will be solved in the unbounded domain x     , starting from the following initial condition: This leads us immediately to the following initial condition for Indeed, we have In order to give the integrals on the right hand-side of (6) a concrete meaning, we now specialize the Gaussian process   x  by assigning a definite expression to its correlation function.

Basic Equation Evolution
Let us assume that the random field   x  is homogeneous and isotropic.Then, without any concrete definition of the correlation function 2 , let us note that the latter depends only on the modulus of its arguments' difference . Our ultimate aim is to derive an equation related only to the mean concentration u .The procedure of getting such as equation seems evident.It is clearly seen from ( 7) that However, such a procedure is rather intricate in   , x t - representation.In ( 6) and (7) it is more convenient to turn to   , q k -representation, i.e. to use Laplace transform with respect to time (q-parameter) and Fourier transform over with respect to space (k-parameter).After all the necessary calculations instead of ( 6) and (7), we get, re-spectively: In the last relations the argument of q   0 , u q k and   1 0 , , f q k k is omitted for the sake of simplicity.We also have Also note, that if we choose 0 in (11) instead of argument , then we get the following equation for function . Substituting in (10) the thus obtained expression, we get an algebraic equation with respect to with A and B being defined by Thus, we have explicit solutions to ( 6) and ( 7) in

 
, q k -representation.Therefore, Equation (12) in representation is equivalent to basic equation [1], which we reproduce here: (3 ) Hence successive approximation method and functional derivative method lead us to similar results for problems such that both methods are available.

Conclusion
So, in this paper, we presented an example of how to use the notion of a functional derivative in order to arrive at a macroscopic equation for dispersion in disordered media.In fact, we also used the present method for equation, which had been derived for a different type of disordered medium, made of inter-twisted tubes, such that a onedimensional approach has physical meaning.Hence, the example can only serve formally for two reasons.Indeed, the sample paths of Gaussian processes can take negative values, which are not good when the existence of solutions is needed.The drawback can be removed by considering ε being replaced by the exponential of a Gaussian process.We will not do it, but the presented method works fairly well for this case and gives the results already obtained via Feynman diagrams in our work [1].

2 x
and then average the obtained equations over the realizations of  and apply Furutsu-Novikov's formula again.As a result, we would obtain an equation for we should assume that in a porous media model, and due to the definition of the porosity   x  we should take for granted that .Considering the fluctuations to be weak, we 0 1 a  neglect cumulant of the order higher than two.Moreover, each time   x  is a centered Gaussian process, all the cumulant are equal to zero, except 2 the concentration that appears in the right-hand part of this equation.The function get the final integro-differential equation for u .