In this paper
AbstractIntroductionFrom Cramér to Energy distancesLearning from Cramér distanceThe decay in Energy distanceThe fast diffusion caseConclusionsAcknowledgementReferencesWorksMathematics
On the domain of attraction of source–type solutions of the porous medium equations
Abstract
We study the rate of convergence to Barenblatt profiles of the solutions to porous medium equations in , by means of a certain class of Energy distances, a class of discrepancies between probability measures which have been recently widely used in problems related to AI, in particular for tasks related to machine learning. If , it is shown both that the rate of decay in Energy distance is improved with respect to known results which are based on the decay of relative entropy, and that, with respect to this distance, any initial value of bounded mass belongs to the domain of attraction of the source–type solution. Similar results are shown to hold for the fast diffusion case.
Introduction
The study of the asymptotic behavior of the solutions of the nonlinear diffusion equation
where , , is a classical problem which, starting with the porous medium case, characterized by , has been originally studied in the 70’s of the last century in a pioneering paper by Kamin [21] (cf. [34, 36] for an exhaustive presentation of the results obtained on this topic up to the recent years). In a physical sense, the solution to (1) represents the time-density of the diffusion process.
If we consider the Cauchy Problem for equation (1) posed in with integrable and nonnegative initial data
the source-type solutions of equation (1) provide the asymptotic representation of a large class of solutions starting from satisfying (2) [34].
The source-type or Barenblatt solutions of equation (1) are given by [5]
where is a positive free constant, and , the profile, has the form
In (4), the function denotes as usual the positive part of .
In this range of the parameter , for a given initial value of given mass , the constant can be chosen to satisfy the rule of mass conservation
Further, in presence of additional conditions on the initial density, such as boundedness of the relative (to the Barenblatt solution ) Ralston-type entropy of order [10], or the boundedness of the relative Renyi entropy of order [9, 26], it is possible to reckon an explicit rate of convergence to the right source-type solution in . It is interesting to note that the boundedness of energy allows to establish a direct connection between the asymptotic stabilization property of the porous medium flow and the stabilization to the Gaussian profile embodied in the classical central limit theorem [33]. Moreover, the boundedness of energy is essential to characterize in a precise way the best matching Barenblatt profile [14, 15]
Diffusion processes described by (1) arise in various fields, such as plasma physics, kinetic theory of gases, solid state physics, filtration models and transport in porous media. In some cases, including many metals and ceramic materials the diffusion coefficient can, over a wide range of temperatures, be approximated as for some exponent such that [25], which leads to an exponent in (1). Also, the special case emerges in the central limit approximation to Carleman’s model of the Boltzmann equation, according to [23, 24]. It is interesting to remark that, while less relevant from the point of view of applications, in mathematics the fast diffusion case is better known than the porous medium case for its properties in connection with functional analysis.
In what follows, we approach the problem of identifying an explicit rate of convergence of the solution to equation (1) to the source-type solution in a new way, which requires only that the initial datum has a finite mass. This broadened range of attraction of the Barenblatt solution can be obtained only if the distance in which the explicit rate of convergence is proven is weaker than the classical -distance; moreover, while as usual the rate of convergence is closely dependent on the space dimension , the proposed approach works in this generality only in dimension .
To this end, we make use of arguments similar to the ones recently introduced in [4], which are based on applying Cramér distance [11] to one-dimensional Fokker–Planck equations with linear drift and nonlinear diffusion. As extensively discussed in [4], the Cramér distance—also known as the continuous ranked probability score—has gained renewed attention in AI-related problems, particularly in machine learning [7], where it is used for tasks such as distribution estimation, regression, and image generation. As first remarked in [7], the main motivation for adopting the Cramér distance, a metric for comparing one-dimensional probability distributions, lies in its advantageous properties, most notably its unbiased sample gradients, which make it especially amenable to optimization via stochastic gradient descent.
As we shall detail in this paper, the results proven in [4] for one-dimensional Fokker–Planck type equations follow by resorting to properties typical of Cramér’s distance, closely linked to some of the various forms in which this distance can be expressed. As a matter of fact, an interesting related distance, which, as shown in [1], shares most properties of Cramér’s distance, while designed to also compare multivariate probability distributions, is the Energy distance introduced by Székely [30, 31, 32] by resorting to one of the alternative expressions of Cramér’s distance.
The classical Energy distance, of wide use in Statistics, has been recently generalized in [1], where the relationships with other well-known metrics—such as the Wasserstein distance [37] and the Fourier-based metrics [9, 18]—were thoroughly analyzed, in reason of their applicability in the field of artificial intelligence [2, 3, 12, 27], as well as in the context of partial differential equations [13, 16, 17].
In this paper, we show that a certain class of the new Energy–type distances considered in [1] are very-well adapted to study the rate of convergence to the source–type solutions to non linear diffusion equations. These metrics are directly connected with the Cramér distance in the one-dimensional case. However, at variance with the one-dimensional Cramér distance, this class of -dimensional Energy distances do not work well on drift operators for , and cannot be directly applied to study the large-time behavior of multi-dimensional Fokker–Planck equations.
In more details, in § 2 we briefly report the definitions and the main properties of both Cramér and Energy distances, as well as their connections. The equivalence between these two distances in the one-dimensional case relies on arguments first applied to obtain an alternative definition of Gini index [38]. Then, in § 3 we will show how a particular class of these distances can be fruitfully employed to study convergence to source–type solutions to the porous medium equation (1), departing from any initial value of bounded mass. § 4 will be devoted to proving that an explicit rate of convergence follows as soon as . In § 5, we will show that a similar argument can be used to study the convergence rate to source–type solutions in the fast diffusion case if belongs to a certain range and if . In Section 6, we will outline the conclusions of the paper.
From Cramér to Energy distances
Let be a probability density over , . We will associate with a random variable , such that for a subset of the reals , we have
Given , we denote by the class of all probability densities over such that
If , the quantity defines the (absolute) mean value of , say .
Given a probability density in , , its Fourier transform is defined by
If , the (cumulative) distribution function of the one-dimensional random variable is given by
Given two one-dimensional random variables and and their cumulative distribution functions and , the Cramér distance between and is expressed by [11]
Note that, as written, Cramér distance is not a metric, while its square root is.
One of the interesting properties of Cramér’s distance is that it can be equivalently written in different ways. It is immediate to show that it can be written in terms of Fourier transforms. Indeed, owing to Parseval’s theorem, Cramér distance (7) can be equivalently written as
Also, assuming that the one-dimensional random variables and are mutually independent, and that and are independent copies of and , Cramér’s distance can be expressed as
Note that (9) implies that the Cramér distance of and is bounded as soon as . This third way of expressing Cramér’s distance has been shown in [32] (cf. [4] for a detailed proof based on the Gini index [19, 20]).
Expressing the Cramér distance as in (9) allows us to compute a weak metric between two probability distributions in closed form on every , with . In a series of papers [30, 31, 32], Székely introduced the definition of Energy distance between two multidimensional random variables and with densities , and, respectively, , with , . The Energy distance in , , of order , with is defined by
As for the Cramér distance, in dimension the Energy distance of order can be equivalently expressed in Fourier transform. In this case it holds [1, 32]
where the constant , determined as in [29] takes the value
where is the Gamma function.
The passage to Fourier transform is the easiest way to verify that right-hand side in (10) is nonnegative, so that twice the mixed integral dominates the sum of the last two. Further, expression (18) allows one to relate in a precise way the energy distance of order to homogeneous Sobolev spaces of fractional order.
Let us note that in one dimension, and for , the Energy distance is nothing but twice the Cramér distance, as defined by (9), so that
Learning from Cramér distance
In what follow, we will show how Cramér distance works in connection with the porous medium equation (1). We assume that the initial value for the Cauchy problem for (1) satisfies conditions (2), and all solutions satisfy well-known regularity properties that are enough to justify all the computation that follows [35].
A key property of identity (13) is that for any given probability density it holds
namely that the function
coincides with the delta function centered in .
This implies that, if and are the two probability density functions solving the one-dimensional porous medium equation (1) departing from (respectively ) satisfying (2) with mass equal to one, and such that is bounded, integrating twice by parts one has
Hence, the Cramér distance between any two solutions to the porous medium equation is decreasing in time.
The key point in deriving this result is that the kernel in the definition of Cramér distance satisfies (14).
This result can be easily generalized to any dimension by resorting to the concept of fundamental solution of the Laplacian in , namely a function that satisfies in the equation in the sense of distributions [28]. The fundamental solutions read
where denotes the surface area of the unit sphere in , given by
Owing to (16) we introduce, for , a (normalized) variant of the Energy distance between two -dimensional probability densities in the form
The Energy distance of negative order , with , defined in (17) is a particular case of the energy distances with negative exponent introduced and studied in [1]. As shown in in [1], the distance is, up to a multiplicative constant given in (12), equal to a Fourier based metric, which in this case gives
so that is the square of the distance in the Sobolev space .
The decay in Energy distance
Let us now consider two probability density functions and , solutions of the -dimensional porous medium equation (1) departing from (respectively ) and such that is bounded. By means of Green’s third identity, we can repeat the computations in (13) for any dimension , obtaining the decay
where
Similar computations have been done in [4] in the (simpler) linear case for the classical Energy distance corresponding to .
In the next step, we obtain a lower bound on the functional in terms of the energy distance .
First of all, since , inequality
implies
Further, since , thanks to the classical interpolation inequality
and to the fact that and are two probability densities on , so that
we get the inequality
which shows that, as , the functional controls the -th power of the square of the -distance between and .
Now, consider that, for , we can bound from above the Energy distance between two probability densities by their -distance resorting to a Nash-type argument. For any given positive constant we have
where
Optimizing with respect to we obtain the bound
where the constant is explicit
Owing to (24) we finally have
Substituting into (19) we obtain that the time derivative of the Energy distance satisfies the inequality
where the explicit constant is given by
The differential inequality (30) can be integrated to give the time-decay
In (31)
Note that, for any given the value of is decreasing with , and
being the lower value reached as .
Hence we proved
Theorem 1. Let , and let and be the solutions of the -dimensional porous medium equation (1) departing from the initial values (respectively ) of equal mass such that the Energy distance is bounded. Then, decays to zero as at the explicit rate , where is defined in (32).
We now apply Theorem 1 to the case in which the initial configuration is the Dirac delta measure centered in zero. Since the solution to the porous medium equation departing from is nothing but the Barenblatt solution of mass one defined in (3), we get both the convergence to zero and the rate of decay to the source-type solution in the Energy distance , provided the Energy distance is bounded at the initial time . Indeed, when this condition is met, belongs to the domain of attraction of the source-type solution with respect to the Energy distance .
Applying definition (17) to the pair yields
Since the Energy distance is non-negative, and , is bounded as soon as
is bounded.
Owing to the Hardy–Littlewood–Sobolev inequality [22], stating that
for all , with , and , an explicit sufficient condition ensuring the boundedness of the Energy distance in (34) is to assume that with .
Let the initial density , and let its Fourier transform. Also, let denote the Gaussian density
Owing to expression (18), and using the bound , we obtain
Indeed, since both and , the convolution of and , belong to , they also belong to when . In particular, their Energy distances from the Dirac delta located in zero, according to the Hardy–Littlewood–Sobolev inequality, are bounded.
Therefore we have
Theorem 2. Let , and let be the solution of the -dimensional porous medium equation (1) departing from a probability density . Then, the Energy distance , with defined as in (3), decays to zero as at the explicit rate , where is defined in (32).
Theorem 2 shows that, if , the domain of attraction of the source–type solution defined as in (3) with respect to the Energy distance contains all initial values in . Hence, the domain of attraction of the Barenblatt solution is larger than the one found in [8] by resorting to entropy methods, since there convergence in the stronger -distance requires the initial density to have the second moment bounded. Moreover, we stress that in [8] the convergence of the solution to the porous medium equation towards the source-type solution in was shown with rate , where
so that, for , we have . However, if , the two rates coincide, while if , the rate of convergence in Energy distance is better than the rate of convergence in , and does not collapse to zero if .
The fast diffusion case
Let us now consider the fast diffusion case. To avoid confusion with the previous Sections, let us denote with the characteristic exponent of the nonlinear diffusion (1) in the fast diffusion regime. Moreover, since we want to study the large time behavior of the solution to equation (1) in which the Barenblatt profile has bounded mass, we assume that . It is immediate to realize that all the computations of Section 3 leading to equation (19), with defined in (20) continue to hold. Hence, as in Section 4 we still obtain that the Energy distance satisfies
where now, for ,
In this case, however, we can not proceed as in Section 4, since inequality (21) is true if and only if . On the other hand, for any given , inequality (21) can be rewritten in the form
namely, since
that implies
We now provide a lower bound on the right-hand side of (40).
Let and two probability density functions and . Thanks to the pointwise inequality
for any given positive constant , the Holder inequality allows us to conclude that
with and such that . We then look for , , and such that
Solving the system, we find the values
Note that , while .
Since , and are probability densities, owing to the classical inequality
valid for any given and , we obtain
Substituting into (41), and taking into account (40) we finally get, for any
where and are given in (42), and
Hence controls the -distance of the probability densities and .
Since in (43) satisfies , the Beckner form of the Hausdorff–Young inequality [6]
which holds true for and , allows us to bound the -distance of the Fourier transforms of the probability densities and .
Since , the Hausdorff–Young inequality implies
Therefore, inequality (44) allows us to conclude with the following
where, clearly, the exponent , while
Thanks to (47), we can proceed as in § 4 to obtain
where is a given value and is the constant in (26).
By the Holder inequality, we can estimate the right hand side of the latter inequality as it follows
where and are such that . If , then , so that, making use of (47) we obtain
Notice that is integrable on the set if and only if
or, equivalently,
It is immediate to verify that, to solve (53) on , must be greater than
Notice that if , then . Moreover, if , then . For such values of , we have
where we denoted
and
All in all, we get
By optimizing the right hand side of (57) with respect to we get
where the constant , that can be explicitly computed in terms of , , and , as given by (26), (48), (55) and (56), takes the value
What is important to remark is that, apart from the different value of the constant , the previous computations show that in the fast diffusion case the time decay of the energy distance follows the law (30) relative to the porous medium equation.
If we now go back to the notation , where now , we conclude that the decay is given by (31), where now is well defined only when the fast diffusion coefficient satisfies the bound
Note that in this case, the measure of the interval in which the coefficient is well defined converges to zero as .
All in all, we proved the following.
Theorem 3. Let , and let and be the solutions of the -dimensional fast diffusion equation (1) departing from the initial values (respectively ) of equal mass such that the Energy distance is bounded. Then, if satisfies condition (60), decays to zero as at the explicit rate , where is defined in (32).
Likewise, since the reasoning leading to Theorem 2, and in particular the estimate (36), do not depend on the exponent of the diffusion equation (1), but only on the properties of the initial value, we have the following.
Theorem 4. Let , and let be the solution of the -dimensional fast diffusion equation (1) departing from a probability density . Then, if satisfies condition (60), the Energy distance decays to zero as at the explicit rate , where is defined in (32).
Conclusions
In this paper, we studied the rate of convergence of the solution to the porous medium equation to the Barenblatt source–type solution with the same mass by means of a particular class of Energy distances, which are equivalent to the square of the distance in the Sobolev space . It is shown that, as , the rate of decay is improved with respect to known results which are based on the decay of relative entropy, and that the domain of attraction of the Barenblatt solution includes all initial value of finite mass.
With few modifications, it has been further shown that analogous results hold in the fast diffusion case, with some limitations in the diffusion exponent that are mainly due to guarantee that the source–type solution has a bounded mass.
The present results suggest that both Cramér and Energy distances could be of high interest in the study of both theoretical and numerical problems related to nonlinear diffusion equations.
Acknowledgement
This work has been written within the activities of GNFM (Gruppo Nazionale per la Fisica Matematica) of INdAM (Istituto Nazionale di Alta Matematica), Italy. G.T. acknowledges support of the Institute for Applied Mathematics and Information Technologies (IMATI), Pavia, Italy. G.T. wishes to thank the Editors for inviting him to contribute to this volume dedicated to Giuseppe Savaré, a friend and colleague at the Department of Mathematics at the University of Pavia for many years.
References
- [1]Auricchio, G., Brigati, G., Giudici, P., and Toscani, G., From kinetic theory to AI: a rediscovery of high-dimensional divergences and their properties. Mathematical Models and Methods in Applied Sciences, 36 (2026) 1–49.arxiv.org/abs/2507.11387
- [2]Auricchio, G., Codegoni, A., Gualandi, S., Toscani, G., and Veneroni, M., On the equivalence between Fourier-based and Wasserstein metrics, Rendiconti Lincei. Matematica e Applicazioni, 31 (2020) 627–649.arxiv.org/abs/2005.06530
- [3]Auricchio, G., Codegoni, A., Gualandi, S., and Zambon, L., The Fourier discrepancy function, Communications in Mathematical Sciences, 21 (2023) 627–639.
- [4]Auricchio, G., and Toscani, G., Energy distance and evolution problems: a promising tool for kinetic equations. Ricerche di Matematica 75 (2026) 849–863.arxiv.org/abs/2510.09123
- [5]Barenblatt, G.I., On some unsteady motions of a liquid and gas in a porous medium. Akad. Nauk SSSR Prikl. Mat. Meh. 16 (1952) 67–78.
- [6]Beckner, W., Inequalities in Fourier analysis. Annals of Mathematics, 102 (1) (1975) 159–182DOI
- [7]Bellemare, M. G., Danihelka, I., Dabney, W., Mohamed, S., Lakshminarayanan, B., and Hoyer, S., and Munos, R., The Cramér distance as a solution to biased Wasserstein gradients, arXiv preprint arXiv:1705.10743 (2017).arxiv.org/abs/1705.10743
- [8]Carrillo, J.A. and Toscani, G. Asymptotic L¹-decay of solutions of the porous medium equation to self-similarity. Indiana University Mathematics Journal, (2000) Apr 1:113–42.DOI
- [9]Carrillo, J.A. and Toscani, G., Contractive probability metrics and asymptotic behavior of dissipative kinetic equations, Rivista di Matematica dell’Università di Parma, 6 (2007) 75–198.
- [10]Carrillo, J.A. and Toscani, G., Renyi entropy and improved equilibration rates to self-similarity for nonlinear diffusion equations, Nonlinearity, 27 (2014) 3159–3177.DOI
- [11]Cramér, H., On the composition of elementary errors, Skandinavisk Aktuarietidskrift, 11, (1928) 141–180.
- [12]Cuturi, M. and Doucet, A., Fast computation of Wasserstein barycenters, International conference on machine learning, (2014) 685–693.arxiv.org/abs/1310.4375
- [13]Di Francesco, M., Fornasier, M., Hütter, J.C. and Matthes, D., Asymptotic behavior of gradient flows driven by nonlocal power repulsion and attraction potentials in one dimension. SIAM Journal on Mathematical Analysis, 46 (6), (2014) 3814–3837.
- [14]Dolbeault, J., and Toscani, G., Best matching Barenblatt profiles are delayed, J. Phys. A: Math. Theor., 48 (2015) 065206.DOI
- [15]Dolbeault, J., and Toscani, G., Nonlinear diffusions: extremal properties of Barenblatt profiles, best matching and delays. Nonlinear Anal. Series A, 138 (2016) 31–43.arxiv.org/abs/1501.03646
- [16]Fornasier, M., Haskovec, J. and Steidl, G., Consistency of variational continuous-domain quantization via kinetic theory. Applicable Analysis, 92 (6), (2013) 1283–1298.arxiv.org/abs/1112.1403
- [17]Fornasier, M. and Hütter, J.C., Consistency of probability measure quantization by means of power repulsion–attraction potentials. Journal of Fourier Analysis and Applications, 22 (3) (2016) 694–749.arxiv.org/abs/1310.1120
- [18]Gabetta, E., Toscani, G., and Wennberg, B., Metrics for probability measures and the trend to equilibrium for solutions of the Boltzmann equation, Journal of Statistical Physics, 81 (1995) 901–934.
- [19]Gini, C., Sulla misura della concentrazione e della variabilità dei caratteri, Atti del Reale Istituto Veneto di Scienze, Lettere ed Arti, 73 (1914) 1203–1248. English translation in Metron, 3–38, (2005).
- [20]Gini, C., Measurement of inequality of incomes, The Economic Journal, 31 (1921) 124–126.DOI
- [21]Kamenomostskaya (Kamin), S., The asymptotic behaviour of the solution of the filtration equation, Israel J. Math. 14 (1973) 76–87.DOI
- [22]Lieb, E.H., Sharp constants in the Hardy-Littlewood-Sobolev and related inequalities, Ann. Math. 118 (1983) 349–374.
- [23]Lions, P.L. and Toscani, G., Diffusive limit for finite velocity Boltzmann kinetic models. Revista Matematica Iberoamericana 13 (3) (1997) 473–513.DOI
- [24]McKean, H.P., The central limit theorem for Carleman’s equation. Israel Journal of Mathematics 21 (1) (1975) 54–92.DOI
- [25]Rosenau, P., Fast and superfast diffusion processes. Physical review letters, 74 (7)(1995) 1056–1059pubmed.ncbi.nlm.nih.gov/10058923
- [26]Savaré G., and Toscani, G., The concavity of Rényi entropy power. IEEE Transactions on Information theory, 60 (5) (2014) 2687–93.DOI
- [27]Scagliotti, A. and Farinelli, S., Normalizing flows as approximations of optimal transport maps via linear-control neural ODEs, Nonlinear Analysis, 257 (2025), 113811.arxiv.org/abs/2311.01404
- [28]Smoller, J., Shock Waves and Reaction–Diffusion Equations (2 ed.). Springer New York 1974.
- [29]Stein, E.M, and Weiss, G., Introduction to Fourier analysis on Euclidean spaces, Princeton University Press, Princeton, NJ (1970)
- [30]Székely, G.J., Potential and kinetic energy in statistics. Lecture Notes. Budapest Institute of Technology (Technical University) (1989).
- [31]Székely, G.J., E-statistics: Energy of statistical samples. Bowling Green State University, Department of Mathematics and Statistics Technical Report No. 03-05, 2003 (also technical reports by same title, from 2000–2003 and NSA Grant # MDA 904-02-1-0091 (2000–2002).)DOI
- [32]Székely, G.J. and Rizzo, M.I., Energy statistics: A class of statistics based on distances. Journal of Statistical Planning and Inference, 143 (2013) 1249–1272DOI
- [33]Toscani, G., A central limit theorem for solutions of the porous medium equation, J. Evol. Equ. 5 (2005) 185–203DOI
- [34]Vázquez, J.L, Asymptotic behaviour for the porous medium equation posed in the whole space. Nonlinear Evolution Equations and Related Topics: Dedicated to Philippe Bénilan. Basel: Birkhäuser Basel, 2003. 67–118.DOI
- [35]Vázquez, J.L., The porous medium equation: mathematical theory. Clarendon Press, 2006.
- [36]Vázquez, J.L., The mathematical theories of diffusion: nonlinear and fractional diffusion. In: Nonlocal and Nonlinear Diffusions and Interactions: New Methods and Directions: Cetraro, Italy 2016. Cham: Springer International Publishing, 2017. 205–278.arxiv.org/abs/1706.08241
- [37]Villani, C., Optimal transport: old and new, Vol. 338, (Springer-Verlag, Basel, 2009).
- [38]Xu, K., How has the literature on Gini’s index evolved in the past 80 years? Dalhousie University, Economics Working Paper (2003).DOI