An Energy-Based Steepest Descent Evolution Principle: Applications in Continuum and Graph Theory

1. Introduction

The principle here presented belongs to a long tradition of variational formulations in mathematics and physics, which share the idea that dynamical laws can be derived from extremal properties of functionals. Classical examples include the Principle of Least Action in mechanics [1], the Maximum Entropy Principle in statistical physics [2], and variational formulations of free energy in thermodynamics.

We introduce the Steepest Descent Evolution Principle (SDEP), an energy-based abstract evolutionary principle designed to describe the evolution of system states across different contexts. The principle is broad and generative: it may illuminate existing laws in physics, suggest new applications, and potentially inspire developments in engineering and applied sciences. Our aim is not to claim universality, but to provide a framework that can be tested, extended, or even challenged in diverse situations.

In formal terms, the principle applies when a system with states u is endowed with an energy functional $Q\left(u\right)$. In such cases, we propose that the evolution of $u\left(t\right)$ is not arbitrary, but rather follows an optimal path that, at each instant, yields the steepest decrease of $Q\left(u\right)$. Equivalently, $u\left(t\right)$ evolves in the direction of the steepest descent of $Q\left(u\right)$.

We present this formulation in abstract terms, offering a general language that can be applied to the modeling of different problems. In particular, we will see in the Applications Sections that when the proposed formulation is applied to Dirichlet energies in continuum and in graph theory, it leads naturally to the classical Heat and Diffusion equations and to the respective forms of the Laplace operator in two very different scenarios.

Readers are invited to explore the potential of the proposed framework with other choices of $Q\left(u\right)$, whose form will depend on the particular application. Possible examples include entropy functionals, potential energies, total variation functionals, loss functionals in data science, consensus functionals in networks, and cost functionals in engineering or economics. Readers may discover new contexts where the principle applies, or situations where it does not. We hope this openness encourages exploration of both its reach and its boundaries.

2. Energy-based Steepest Descent Evolution Principle (SDEP)

2.1. Statement of the Principle

Many dynamical systems in nature are endowed with an underlying energy functional $Q\left(u\right)$ characterizing the states u of the system. In such cases, the observed behavior suggests that the system does not evolve arbitrarily, but instead follows a principle of maximal energy decrease. We formalize this observation in the following principle:

Steepest Descent Evolution Principle (SDEP): In many systems that can be fully characterized by an energy functional $Q\left(u\right)$ of their states u, the time evolution of u appears to be governed by a principle of steepest descent, in the sense that, at each instant, the system evolves not in any arbitrary direction, but specifically in the direction of maximal decrease of $Q\left(u\right)$. Whenever this behavior occurs in a system $\mathcal{S}$, we will say that $\mathcal{S}$ satisfies the Steepest Descent Evolution Principle. We will use the acronym SDEP to refer to this principle.

Throughout the paper, for brevity and ease of reading, we will often use the term evolution principle or just principle to refer to the above-defined Steepest Descent Evolution Principle (SDEP).

2.2. SDEP: Mathematical Formulation in Normed Vector Spaces

The general abstract formulation presented in this paper involves several theoretical aspects of Functional Analysis, including normed vector spaces, Banach spaces, Hilbert spaces, directional derivatives, and the Riesz representation theorem. These concepts form the foundation for many of the results discussed here. Readers who are not familiar with these topics may consult standard references such as Rudin [3], Kreyszig [4], and Reed & Simon [5] for detailed expositions.

We now translate the narrative statement of the Steepest Descent Evolution Principle into mathematical form. Assume we have a system $\mathcal{S}$. Let X denote the normed vector space of possible states u of $\mathcal{S}$, and let $u\left(t\right)\in X$ represent the system state at time t. The energy of the system is given by a functional $Q:X\to \mathbb{R}$, so $Q\left(u\right)$ represents the energy of the system at state u.

To describe the micro-evolution of the state, we write

$$u(t+\delta t)=u\left(t\right)+\delta tv\left(t\right)+o\left(\delta t\right),$$

where $v\left(t\right)\in X$ denotes the velocity at time t. Passing to the limit $\delta t\to 0$ yields the continuous-time evolution law

$${\displaystyle \frac{du}{dt}}\left(t\right)=v\left(t\right),$$

(1)

so the velocity $v\left(t\right)$ is precisely the time derivative of the state. We further decompose the velocity into direction and magnitude as

$$v\left(t\right)=\alpha \left(t\right)v^{*}\left(t\right),$$

where $v^{*}\left(t\right)$ is a unit vector in X representing the direction of motion at time t, and $\alpha \left(t\right)\ge 0$ is a scalar coefficient representing the magnitude of the velocity at such time.

Now, assume $\mathcal{S}$ satisfies the SDEP. With the above-introduced notation at hand, the principle asserts that, at each instant, the system state u evolves in the direction that produces the steepest decrease of Q. To determine such direction, we need first to be able to determine the directional variation of $Q\left(u\right)$ for all possible directions. We therefore define the directional derivative of Q in a given direction as follows:

Definition 1

(Directional Derivative). Let $Q:X\to \mathbb{R}$ be a functional defined on a normed space X. The directional derivative of Q at $u\in X$ in the direction $w\in X$ is defined as

$$D_{w}Q\left(u\right)=\underset{\delta t\to 0^{+}}{lim}{\displaystyle \frac{Q(u+\delta tw)-Q\left(u\right)}{\delta t}},$$

provided the limit exists.

With this definition, all directional derivatives can be determined. Among them, the one that yields the maximal decrease of Q defines the evolution direction $v^{*}$. In other words,

$$v^{*}\left(t\right)=arg\underset{\parallel w\parallel =1}{min}{D}_{w}Q\left(u\left(t\right)\right),$$

(2)

so $v^{*}$ is the direction of steepest descent of Q at state u.1 Equation (2) makes explicit that the system evolves in the direction of maximum energy decrease. Note that the magnitude $\alpha \left(t\right)$ of the velocity remains free and determined by modeling choices or physical considerations.

The mathematical formulation above already provides a precise statement of the principle. In the following subsections, we will progressively enrich the structure of the state space X. This enrichment is not required for the validity of the principle itself, but it equips us with additional analytical tools to more easily determine the evolution direction of $v^{*}$ in practical cases.

2.3. SDEP: Mathematical Formulation in Banach Spaces

In the previous subsection, the Steepest Descent Evolution Principle was described in terms of the directional derivative, which provides a way to evaluate the variation of the energy functional Q along a given direction. In this subsection, we further specialize the state space X to a Banach space in order to refine the mathematical formulation of the principle. Proceeding in analogy with Taylor expansions, we seek a formulation that captures the local behavior of Q in a single operator, rather than through separate directional derivatives. While directional derivatives describe the variation of Q along individual directions w, the Fréchet derivative provides a unified operator that simultaneously encodes all directional derivatives in one linear functional form. For this purpose, we enrich the state space X from a normed vector space to a Banach space, ensuring completeness and enabling a rigorous definition of differentiability.

Definition 2

(Fréchet Derivative). Let X be a Banach space and $Q:X\to \mathbb{R}$ a functional. We say that Q is Fréchet differentiable at $u\in X$ if there exists a bounded linear operator $DQ\left(u\right):X\to \mathbb{R}$ such that

$$\underset{\begin{array}{c}\parallel w\parallel \to 0\\ w\in X\end{array}}{lim}{\displaystyle \frac{\left|Q\right(u+w)-Q(u)-DQ(u\left)\right[w\left]\right|}{\parallel w\parallel }}=0.$$

The operator $DQ\left(u\right)$ is called the Fréchet derivative of Q at u.

This definition expresses that the local expansion of Q near u takes the form

$$Q(u+w)=Q\left(u\right)+DQ\left(u\right)\left[w\right]+o(\parallel w\parallel ),$$

(3)

so that $DQ\left(u\right)$ is the linear functional capturing the first-order variation of Q. Equivalently, for every $w\in X$,

$$DQ\left(u\right)\left[w\right]=D_{w}Q\left(u\right),$$

showing that the Fréchet derivative unifies all directional derivatives into a single operator.

With this operator in hand, the evolution principle given in Equation (2) can be rewritten in Fréchet form as

$$v^{*}\left(t\right)=arg\underset{\parallel w\parallel =1}{min}DQ\left(u\left(t\right)\right)\left[w\right],$$

(4)

which makes explicit that the system evolves in the direction $v^{*}$ along which the Fréchet derivative of the energy functional attains its minimum value.

For later use in the Application Sections, we point out that the Taylor-series–like expansion of Q around u, given in Equation (3), can alternatively be written as:

$$Q(u+\epsilon h)=Q\left(u\right)+\epsilon DQ\left(u\right)\left[h\right]+o\left(\epsilon \right),$$

(5)

where we have simply made the substitution $w=\epsilon h$, with $h\in X$ and $\epsilon$ a scalar.

2.4. SDEP: Mathematical Formulation in Hilbert Spaces, Gradient Flows

In finite-dimensional Euclidean spaces ${\mathbb{R}}^n$, the gradient of a function $f:{\mathbb{R}}^n\to \mathbb{R}$ is characterized by the first-order Taylor expansion:

$$f(x+w)=f\left(x\right)+\langle {\nabla }_{x}f\left(x\right),w\rangle +o(\parallel w\parallel ),w\in {\mathbb{R}}^n.$$

(6)

Here the directional derivative of f at x in the direction w is given by

$$Df\left(x\right)\left[w\right]=\langle {\nabla }_{x}f\left(x\right),w\rangle ,$$

(7)

so the gradient ${\nabla }_{x}f\left(x\right)$ is precisely the element that represents the linear functional $Df\left(x\right)$ via the Euclidean inner product. For the Steepest Descent Evolution Principle, we extend these ideas to general ambient state spaces X, as described below.

We saw in the previous subsection that when X is a Banach space we can define directional derivatives $DQ\left(u\right)\left[w\right]$. Now, when the ambient space X is specialized to a Hilbert space H, the same notion of a gradient, in the sense of Equations (6) and (7), appears in the functional setting. In this case, the Fréchet derivative $DQ\left(u\right):H\to \mathbb{R}$ is a bounded linear functional, and by the Riesz Representation Theorem it can be represented by a unique element of the Hilbert space H, denoted ${\nabla }_{H}Q\left(u\right)$, satisfying

$$DQ\left(u\right)\left[w\right]={\langle {\nabla }_{H}Q\left(u\right),w\rangle }_H,\forall w\in H.$$

(8)

We call ${\nabla }_{H}Q\left(u\right)$ the functional gradient of Q at u, since it generalizes the classical notion of gradient given in Equations (6) and (7).2

With the equivalence defined in Equation (8), the mathematical expression of the evolution principle given in Equation (4) takes the form:

$$v^{*}\left(t\right)=arg\underset{\parallel w\parallel =1}{min}{\langle {\nabla }_{H}Q\left(u\left(t\right)\right),w\rangle }_H,$$

(9)

It is immediate that the minimizing direction occurs when w is chosen to be

$$w=-{\displaystyle \frac{{\nabla }_{H}Q\left(u\right)}{\parallel {\nabla }_H{Q\left(u\right)\parallel }_H}},$$

that is, the normalized negative gradient. Therefore, the descent direction associated with the evolution principle is

$$v^{*}\left(t\right)=-{\displaystyle \frac{{\nabla }_{H}Q\left(u\left(t\right)\right)}{\parallel {\nabla }_H{Q\left(u\left(t\right)\right)\parallel }_H}}.$$

Then, using this expression, the evolution equation (1) associated with the evolution principle becomes

$${\displaystyle \frac{du}{dt}}\left(t\right)=-\alpha \left(t\right){\displaystyle \frac{{\nabla }_{H}Q\left(u\left(t\right)\right)}{\parallel {\nabla }_H{Q\left(u\left(t\right)\right)\parallel }_H}}.$$

Since $\alpha$ can be freely adjusted by modeling choices, we define $\overline{\alpha }={\displaystyle \frac{\alpha }{\parallel {\nabla }_H{Q\left(u\right)\parallel }_H}}$. Thus the above equation can be written simply as:

$${\displaystyle \frac{du}{dt}}\left(t\right)=-\overline{\alpha }\left(t\right){\nabla }_{H}Q\left(u\left(t\right)\right),$$

(10)

This shows that the SDEP naturally leads to the dynamics of gradient flows. In what follows, to ease the notation we will drop the time stamp $\left(t\right)$, so the dynamic equation of the SDEP can be written simply as:

$${\displaystyle \frac{du}{dt}}=-\overline{\alpha }{\nabla }_{H}Q\left(u\right),$$

(11)

The above equations will be referred to as the SDEP dynamic equations.

SDEP: Operational Perspective

In operational terms, the development presented here can be understood as a progressive simplification of the calculus required to determine the evolution equations associated with the Steepest Descent Evolution Principle applied to an energy functional Q. We began with the explicit computation of directional derivatives, which can be cumbersome in practice. We then introduced the Fréchet derivative, which provides a unified linear functional valid for all directions simultaneously. Finally, in Hilbert spaces, the Riesz representation identifies this functional with a gradient vector, allowing for direct and elegant computation of the evolution equations. Thus, each step introduces a more powerful notion that progressively streamlines the operational determination of the dynamics.

3. Application I: The Steepest Descent Evolution Principle in Continuum

In this section, we apply the Steepest Descent Evolution Principle (SDEP) to the continuum setting. The Dirichlet energy plays a fundamental role in modeling diffusion and heat propagation processes. [6,7]. We therefore select this functional as the energy model $Q\left(u\right)$ for the evolution principle. With this choice, and assuming that the system is governed by the SDEP, we derive the associated dynamics using the mathematical formulation presented in Section 2. We show that the resulting SDEP dynamical equation coincides with the classical heat (diffusion) equation. Furthermore, we demonstrate that in the continuum the Laplace operator generates the direction of steepest descent of the system’s energy. To proceed with this application, we begin by recalling the necessary analytical definitions in the continuum framework.

3.1. Function Space and Energy Functional

In this continuum application, consider a bounded domain $\Omega \subset {\mathbb{R}}^d$, where a given system has states $u\left(x\right)$ throughout the domain. For example, u may represent the temperature or any other diffusion-related property. In this case, the ambient state space X of the states u is selected to be the Hilbert space $X=H^1(\Omega )$, which denotes the Sobolev space of square-integrable functions with square-integrable weak derivatives. For any $u\in H$, we define the continuum Dirichlet energy functional:

$$Q_{\Omega }\left(u\right):={\displaystyle \frac{1}{2}}{\int }_{\Omega }{\left|{\nabla }_{x}u\left(x\right)\right|}^{2}dx$$

(12)

This functional quantifies the smoothness of u over $\Omega$, penalizing large spatial gradients. We assume that the physical states u are governed by the Steepest Descent Evolution Principle (SDEP), with the Dirichlet energy $Q_{\Omega }\left(u\right)$ serving as the associated energy functional $Q\left(u\right)$ (see statement in Section 2.1). We will apply the mathematical formalism described in Section 2 to this particular case. As the section progresses, we will clarify the physical meaning and importance of the evolution principle.

3.2. Fréchet Derivative

To apply the mathematical formulation of the SDEP in the continuum to the Dirichlet energy $Q_{\Omega }\left(u\right)$ defined above, we need to determine the corresponding derivatives of this energy functional. This step is essential because the principle requires the identification of directional derivatives, which are obtained through the Fréchet derivative (see Section 2.3).

To determine the Fréchet derivative, we mimic the general Taylor-type expansion introduced in Equation (5) in the particular case of the energy functional $Q_{\Omega }\left(u\right)$. Consider a variation around u of the form $u+\epsilon h$, where $h\in H$ is a test function and $\epsilon \in \mathbb{R}$ a scalar parameter. In this case, the particular instance of Equation (5) is:

$$Q_{\Omega }(u+\epsilon h)=Q_{\Omega }\left(u\right)+\epsilon D{Q}_{\Omega }\left(u\right)\left[h\right]+o\left(\epsilon \right)\mathrm{as}\epsilon \to 0.$$

(13)

where $D{Q}_{\Omega }\left(u\right)\left[h\right]$ denotes the Fréchet derivative.

To obtain the explicit form of the Fréchet derivative, we evaluate $Q_{\Omega }(u+\epsilon h)$ in the functional form given in Equation (12). Expanding the integrand, we obtain:

$$|{\nabla }_x{(u+\epsilon h)|}^2=|{\nabla }_{x}u+\epsilon {\nabla }_x{h|}^2=|{\nabla }_x{u|}^2+2\epsilon {\nabla }_{x}u·{\nabla }_{x}h+{\epsilon }^2{\left|{\nabla }_{x}h\right|}^2.$$

Substituting this into Equation (12) and comparing terms with Equation (13) yields the expression of the Fréchet derivative:

$$D{Q}_{\Omega }\left(u\right)\left[h\right]={\int }_{\Omega }{\nabla }_{x}u\left(x\right)·{\nabla }_{x}h\left(x\right)dx.$$

(14)

3.3. Functional Gradient and State Dynamics

Since in the present case the ambient state space X of the states u is a Hilbert space ($X=H^1(\Omega )$), we can formulate the evolution principle using the mathematical framework presented in Section 2.4. According to Equation (11) in that section, if the SDEP holds, the states u must satisfy the following dynamical equation:

$${\displaystyle \frac{du}{dt}}=-\overline{\alpha }{\nabla }_H{Q}_{\Omega }\left(u\right),$$

(15)

where ${\nabla }_H{Q}_{\Omega }\left(u\right)$ is the functional gradient of $Q_{\Omega }\left(u\right)$. To obtain the explicit dynamics defined by Equation (15), we need to determine the gradient ${\nabla }_H{Q}_{\Omega }\left(u\right)\in H$ via the Riesz representation theorem, as indicated in Section 2.4. Then, according to Equation (8) in that section, under the $L^2(\Omega )$ inner product, we seek a function ${\nabla }_H{Q}_{\Omega }\left(u\right)\in H$ such that:

$$D{Q}_{\Omega }\left(u\right)\left[h\right]={\int }_{\Omega }{\nabla }_H{Q}_{\Omega }\left(u\right)\left(x\right)h\left(x\right)dx\mathrm{for}\mathrm{all}h\in H$$

(16)

Using integration by parts in the expression of $D{Q}_{\Omega }\left(u\right)\left[h\right]$ given in Equation (14), and assuming homogeneous Neumann or Dirichlet boundary conditions, we can rewrite it as:

$$D{Q}_{\Omega }\left(u\right)\left[h\right]=-{\int }_{\Omega }({\nabla }_x·{\nabla }_{x}u\left(x\right))h\left(x\right)dx=-{\int }_{\Omega }\left({\Delta }_{x}u\left(x\right)\right)h\left(x\right)dx$$

(17)

where ${\Delta }_x$ denotes the Laplace operator in the continuum. Thus, comparison of (17) with (16) implies that the functional gradient of the Dirichlet energy is:

$${\nabla }_H{Q}_{\Omega }\left(u\right)=-{\Delta }_{x}u$$

(18)

This result shows that the functional gradient of the Dirichlet energy in the continuum setting is precisely the Laplacian. Conversely, it demonstrates that the Laplace operator defines the direction of steepest descent of the system’s energy. Furthermore, substituting Equation (18) into Equation (15) leads to the main result: the SDEP dynamical equation for the states u is

$${\displaystyle \frac{du}{dt}}=\overline{\alpha }{\Delta }_{x}u,$$

(19)

which coincides with the classical heat (diffusion) equation. This establishes that the classical heat equation is precisely the dynamical equation induced by the SDEP.

4. Application II: The Steepest Descent Evolution Principle on Graphs

In this section, we apply the abstract variational framework of the Steepest Descent Evolution Principle (SDEP) presented in Section 2 as a general recipe for developing physical equations in the discrete setting of graphs. In this context, we select the graph Dirichlet energy functional $Q_{\mathcal{G}}\left(u\right)$ as the energy model for deriving the dynamical equations associated with the SDEP. This choice is motivated by the central role of the Dirichlet energy in the dynamics of graph-based systems, particularly in modeling diffusion processes [8,9]. Assuming the validity of the SDEP, we show that its dynamical equations specialize in this case to diffusion equations on graphs, with the well-known graph Laplacian as the induced operator. To proceed with this application, we begin by introducing the necessary theoretical definitions about graphs.

4.1. Graphs Fundamentals

A weighted graph $\mathcal{G}$ is a pair $\mathcal{G}=(V,W)$, where $V=\{{\mathrm{v}}_1,{\mathrm{v}}_2,\dots ,{\mathrm{v}}_n\}$ is a finite set of n vertices, and $W:V\times V\to {\mathbb{R}}_{\ge 0}$ is a weight function assigning a non-negative real number $W({\mathrm{v}}_i,{\mathrm{v}}_j)$ to each pair $\{{\mathrm{v}}_i,{\mathrm{v}}_j\}$ of vertices in V. We assume $\mathcal{G}$ is undirected, meaning $W({\mathrm{v}}_i,{\mathrm{v}}_j)=W({\mathrm{v}}_j,{\mathrm{v}}_i)$ for all ${\mathrm{v}}_i,{\mathrm{v}}_j\in V$, and that $\mathcal{G}$ has no self-loops, so $W({\mathrm{v}}_i,{\mathrm{v}}_i)=0$${\mathrm{v}}_i,{\mathrm{v}}_j\in V$. Using W, we define the set of undirected edges of $\mathcal{G}$ as

$$E=\{\{{\mathrm{v}}_i,{\mathrm{v}}_j\}:W({\mathrm{v}}_i,{\mathrm{v}}_j)>0\}$$

(20)

so that vertices ${\mathrm{v}}_i$ and ${\mathrm{v}}_j$ are connected by an edge if and only if $W({\mathrm{v}}_i,{\mathrm{v}}_j)>0$. We introduce the shorthand notation $w_{ij}:=W({\mathrm{v}}_i,{\mathrm{v}}_j)$, so that W may be identified with a symmetric matrix in ${\mathbb{R}}^{n\times n}$ having nonnegative entries and zeros on the diagonal.

We also introduce the notion of functions and functionals on graphs, which will serve as central objects in our analysis: a) A real-valued function on the graph is a map $f:V\to \mathbb{R}$, assigning a scalar value $f\left({\mathrm{v}}_i\right)\in \mathbb{R}$ to each vertex ${\mathrm{v}}_i\in V$. The space of all such functions is denoted ${\mathbb{R}}^V$. b) A functional on the graph is a mapping $F:{\mathbb{R}}^V\to \mathbb{R}$, which assigns a scalar value $F\left(f\right)$ to each function f.

4.2. System State and Dirichlet Energy Functional

Consider a undirected graph $\mathcal{G}=(V,W)$ as defined in the previous subsection. Assume a dynamical system ${\mathcal{S}}_{\mathcal{G}}$ is defined on the Graph. The state of the system is described by a function $u:V\to \mathbb{R}$, which defines the nodal states $u_i=u\left({\mathrm{v}}_i\right)$ at each vertex ${\mathrm{v}}_i\in V$. Then, assume that the energetic state of such system is defined by the so-called Dirichlet Energy functional $Q_{\mathcal{G}}\left(u\right)$ which is given by:

$$Q_{\mathcal{G}}\left(u\right):={\displaystyle \frac{1}{2}}\sum _{\{{\mathrm{v}}_i,{\mathrm{v}}_j\}\in E}{w}_{ij}{\left(u_i-u_j\right)}^2$$

(21)

This graph Dirichlet energy is a weighted measure of the total variation of u between connected vertices. In the equation above each edge $\{{\mathrm{v}}_i,{\mathrm{v}}_j\}$ is counted exactly once. For convenience, we reformulate it as a double sum over all vertex pairs, so $Q_{\mathcal{G}}\left(u\right)$ can equivalently be re-written as:

$$Q_{\mathcal{G}}\left(u\right)={\displaystyle \frac{1}{4}}\sum _{i=1}^n\sum _{j=1}^n{w}_{ij}{\left(u_i-u_j\right)}^2$$

(22)

Note: The compensating factor of ${\displaystyle \frac{1}{2}}$ was added in the equation to take into account the double-counting

4.3. SDEP on Graphs with Associated Dirichlet Energy

Since we are now equipped with an energy functional, let us assume that the system ${\mathcal{S}}_{\mathcal{G}}$ satisfies the SDEP. We are then interested in determining its associated evolution equations. To this end, note that the state space X of functions u belongs to ${\mathbb{R}}^V$. Equipped with the standard inner product, this space is a Hilbert space $H={\mathbb{R}}^V$, which coincides with ${\mathbb{R}}^n$. Consequently, we can apply the mathematical framework developed in Section 2.3 and Section 2.4. Following that framework, we compute the Fréchet derivative $D{Q}_{\mathcal{G}}\left(u\right)\left[h\right]$ and the functional gradient ${\nabla }_{{\mathbb{R}}^V}{Q}_{\mathcal{G}}\left(u\right)$ of the energy functional $Q_{\mathcal{G}}\left(u\right)$. The subindex ${\mathbb{R}}^V$ in the gradient notation emphasizes that the functional gradient is taken with respect to the Hilbert space structure of $H={\mathbb{R}}^V$.

4.4. Fréchet Derivative

To compute the Fréchet derivative of the Dirichlet energy functional $Q_{\mathcal{G}}$, we consider a perturbation of the form $u+\epsilon h$, where $h:V\to \mathbb{R}$ is a test function and $\epsilon \in \mathbb{R}$ is a scalar parameter. Using the definition of the Fréchet derivative given in Equation (5) of Section 2.3, we require that $Q_{\mathcal{G}}\left(u\right)$ satisfies

$$Q_{\mathcal{G}}(u+\epsilon h)=Q_{\mathcal{G}}\left(u\right)+\epsilon D{Q}_{\mathcal{G}}\left(u\right)\left[h\right]+o\left(\epsilon \right),$$

(23)

Evaluating Equation (22) at $u+\epsilon h$, we obtain

$$Q_{\mathcal{G}}(u+\epsilon h)={\displaystyle \frac{1}{4}}\sum _{i=1}^n\sum _{j=1}^n{W}_{ij}{\left((u_i+\epsilon h_i)-(u_j+\epsilon h_j)\right)}^2.$$

After expanding and grouping the terms proportional to $\epsilon$ in the equation above, and comparing with Equation (23), we find that the Fréchet derivative is

$$D{Q}_{\mathcal{G}}\left(u\right)\left[h\right]={\displaystyle \frac{1}{2}}\sum _{i=1}^n\sum _{j=1}^n{W}_{ij}(u_i-u_j)(h_i-h_j).$$

(24)

4.5. Functional Gradient

We now compute the gradient of the Dirichlet energy functional $Q_{\mathcal{G}}\left(u\right)$, based on the Fréchet derivative in Equation (24). Recall from Equation (8) that the gradient ${\nabla }_{{\mathbb{R}}^V}{Q}_{\mathcal{G}}\left(u\right)\in X={\mathbb{R}}^V$ is defined via the Riesz representation of the linear operator $D{Q}_{\mathcal{G}}\left(u\right)\left[h\right]$ under the Euclidean inner product, as:

$$D{Q}_{\mathcal{G}}\left(u\right)\left[h\right]={\langle {\nabla }_{{\mathbb{R}}^V}Q\left(u\right),h\rangle }_{{\mathbb{R}}^V}=\sum _{i=1}^n{\left({\nabla }_{{\mathbb{R}}^V}{Q}_{\mathcal{G}}\left(u\right)\right)}_i{h}_i$$

(25)

To match this form, we rewrite the expression for the Fréchet derivative in Equation (24). We need to isolate the $h_i$ as factors. Observe that each term $(u_i-u_j)(h_i-h_j)$ contributes to both $h_i$ and $h_j$, so:

$$D{Q}_{\mathcal{G}}\left(u\right)\left[h\right]={\displaystyle \frac{1}{2}}\sum _{i=1}^n\sum _{j=1}^n{W}_{ij}(u_i-u_j)h_i+{\displaystyle \frac{1}{2}}\sum _{i=1}^n\sum _{j=1}^n{W}_{ij}(u_j-u_i)h_j$$

Permuting indices $i\leftrightarrow j$ in the second sum, we obtain:

$$D{Q}_{\mathcal{G}}\left(u\right)\left[h\right]=\sum _{i=1}^n\left(\sum _{j=1}^n{W}_{ij}(u_i-u_j)\right)h_i$$

(26)

Comparison of the right-hand side of this equation with Equation (25) shows that the functional gradient ${\nabla }_{{\mathbb{R}}^V}{Q}_{\mathcal{G}}\left(u\right)$ of $Q_{\mathcal{G}}\left(u\right)$ is an operator in ${\mathbb{R}}^V$ (equivalently, a vector in ${\mathbb{R}}^n$) with components:

$${\left({\nabla }_{{\mathbb{R}}^V}{Q}_{\mathcal{G}}\left(u\right)\right)}_i=\sum _{j=1}^n{W}_{ij}(u_i-u_j)\mathrm{for}\mathrm{all}i=1,\dots ,n.$$

(27)

Notice that the right-hand side of Equation (27) is the well-known form of the graph Laplacian operator $L_{\mathcal{G}}\left(u\right)$ commonly used in graph theory and applications:

$$L_{\mathcal{G}}\left(u\right)=\sum _{j=1}^n{W}_{ij}(u_i-u_j)\mathrm{for}\mathrm{all}i=1,\dots ,n.$$

(28)

Thus, we can state that the functional gradient of the graph Dirichlet energy is the graph Laplacian operator:

$${\nabla }_{{\mathbb{R}}^V}{Q}_{\mathcal{G}}\left(u\right)=L_{\mathcal{G}}\left(u\right)$$

(29)

4.6. SDEP Evolution Equations

Finally, let us derive the explicit form of the dynamic equations associated to the SDEP for the current application. From the general form given in Equation (11), in the discrete graph setting we have that:

$${\displaystyle \frac{du}{dt}}=-\overline{\alpha }{\nabla }_{{\mathbb{R}}^V}{Q}_{\mathcal{G}}\left(u\right),$$

(30)

Finally, using Equation (29), the graph-setting SDEP dynamic equation becomes:

$${\displaystyle \frac{du}{dt}}=-\overline{\alpha }{L}_{\mathcal{G}}\left(u\right),$$

(31)

or in component form:

$${\displaystyle \frac{d{u}_i}{dt}}=-\overline{\alpha }\sum _{j=1}^n{W}_{ij}(u_i-u_j)\mathrm{for}i=1,\dots ,n.$$

(32)

Equations (31) and (32) are commonly used in applications related to diffusion on graphs. They can now be interpreted as a direct consequence of the Steepest Descent Evolution Principle.

5. Summary of Analogy

The following table summarizes the correspondence between the graph and continuum settings:

Concept	Graph Setting	Continuum Setting
Domain	Vertex set V	Region $\Omega \subset {\mathbb{R}}^d$
Function space	${\mathbb{R}}^V$	$H^1(\Omega )$
Dirichlet energy	$Q_{\mathcal{G}}\left(u\right)={\displaystyle \frac{1}{4}}\sum W_{ij}{(u_i-u_j)}^2$	$Q_{\Omega }\left(u\right)={\displaystyle \frac{1}{2}}\int {|\nabla u|}^{2}dx$
Gradient	$\nabla Q_{\mathcal{G}}\left(u\right)=L_{\mathcal{G}}\left(u\right)$	$\nabla Q_{\Omega }\left(u\right)=L_{\Omega }\left(u\right)$
Laplace Op.	$L_{\mathcal{G}i}\left(u\right)={\sum }_j{W}_{ij}(u_i-u_j)$	$L_{\Omega }\left(u\right)=-\Delta u$

Concept	Graph Setting	Continuum Setting
Domain	Vertex set V	Region $\Omega \subset {\mathbb{R}}^d$
Function space	${\mathbb{R}}^V$	$H^1(\Omega )$
Dirichlet energy	$Q_{\mathcal{G}}\left(u\right)={\displaystyle \frac{1}{4}}\sum W_{ij}{(u_i-u_j)}^2$	$Q_{\Omega }\left(u\right)={\displaystyle \frac{1}{2}}\int {|\nabla u|}^{2}dx$
Gradient	$\nabla Q_{\mathcal{G}}\left(u\right)=L_{\mathcal{G}}\left(u\right)$	$\nabla Q_{\Omega }\left(u\right)=L_{\Omega }\left(u\right)$
Laplace Op.	$L_{\mathcal{G}i}\left(u\right)={\sum }_j{W}_{ij}(u_i-u_j)$	$L_{\Omega }\left(u\right)=-\Delta u$

6. Conclusions

We have introduced the energy-based Steepest Descent Evolution Principle (SDEP) and developed its mathematical formalism in general variational terms, making it applicable across different spaces and contexts. As illustrative applications, we examined both continuum and graph settings. By adopting Dirichlet energies as test cases, we showed that the SDEP framework recovers classical diffusion dynamics: in the continuum, the heat equation with the Laplace operator as generator of steepest descent, and in graphs, the discrete diffusion equations governed by the graph Laplacian, likewise arising as the generator of steepest descent. These results establish the SDEP as a unified variational recipe for deriving dynamical laws in distinct mathematical scenarios.

The broader significance of the SDEP lies in its openness. Rather than claiming universality, we have presented a principle that can be tested, extended, or challenged in diverse situations. Its formulation in terms of energy functionals suggests natural avenues for future exploration, including entropy-based models, potential energies, total variation, consensus functionals, and loss functionals in data science. Each of these choices may reveal new dynamical equations or highlight the boundaries of the principle’s applicability.

In summary, the SDEP offers a conceptual bridge between abstract variational formulations and concrete physical or network-based processes. By grounding evolution in the steepest descent of energy, it provides a generative framework that connects classical laws with new possibilities, inviting further investigation into both its reach and its limitations.