
Citation: Carlos Gutierrez-Merino, Dorinda Marques-da-Silva, Sofia Fortalezas, Alejandro K. Samhan-Arias. The critical role of lipid rafts nanodomains in the cross-talk between calcium and reactive oxygen and nitrogen species in cerebellar granule neurons apoptosis by extracellular potassium deprivation[J]. AIMS Molecular Science, 2016, 3(1): 12-29. doi: 10.3934/molsci.2016.1.12
[1] | Juan Manuel Pastor, Silvia Santamaría, Marcos Méndez, Javier Galeano . Effects of topology on robustness in ecological bipartite networks. Networks and Heterogeneous Media, 2012, 7(3): 429-440. doi: 10.3934/nhm.2012.7.429 |
[2] | Yunhua Liao, Mohamed Maama, M. A. Aziz-Alaoui . Consensus dynamics and coherence in hierarchical small-world networks. Networks and Heterogeneous Media, 2025, 20(2): 482-499. doi: 10.3934/nhm.2025022 |
[3] | A. Marigo . Robustness of square networks. Networks and Heterogeneous Media, 2009, 4(3): 537-575. doi: 10.3934/nhm.2009.4.537 |
[4] | Yacine Chitour, Guilherme Mazanti, Mario Sigalotti . Stability of non-autonomous difference equations with applications to transport and wave propagation on networks. Networks and Heterogeneous Media, 2016, 11(4): 563-601. doi: 10.3934/nhm.2016010 |
[5] | Giuseppe Buttazzo, Filippo Santambrogio . Asymptotical compliance optimization for connected networks. Networks and Heterogeneous Media, 2007, 2(4): 761-777. doi: 10.3934/nhm.2007.2.761 |
[6] | Yao-Li Chuang, Tom Chou, Maria R. D'Orsogna . A network model of immigration: Enclave formation vs. cultural integration. Networks and Heterogeneous Media, 2019, 14(1): 53-77. doi: 10.3934/nhm.2019004 |
[7] | Juan Manuel Pastor, Javier García-Algarra, José M. Iriondo, José J. Ramasco, Javier Galeano . Dragging in mutualistic networks. Networks and Heterogeneous Media, 2015, 10(1): 37-52. doi: 10.3934/nhm.2015.10.37 |
[8] | Rosa M. Benito, Regino Criado, Juan C. Losada, Miguel Romance . Preface: "New trends, models and applications in complex and multiplex networks". Networks and Heterogeneous Media, 2015, 10(1): i-iii. doi: 10.3934/nhm.2015.10.1i |
[9] | Hyeontae Jo, Hwijae Son, Hyung Ju Hwang, Eun Heui Kim . Deep neural network approach to forward-inverse problems. Networks and Heterogeneous Media, 2020, 15(2): 247-259. doi: 10.3934/nhm.2020011 |
[10] | Michael Herty, Veronika Sachers . Adjoint calculus for optimization of gas networks. Networks and Heterogeneous Media, 2007, 2(4): 733-750. doi: 10.3934/nhm.2007.2.733 |
In this paper we focus on the discrete graph model introduced by Hu and Cai [14] and further studied in [2,6,9,15] which describes optimal transportation networks in a primarily biological context. Typical examples are leaf venation in plants, mammalian circulatory systems that convey nutrients to the body through blood circulation, or neural networks that transport electric charge. Understanding the properties of optimal transportation networks and mechanisms of their development, function, and adaptation has been the subject of an active field of research, see, e.g., [4,7,24,30].
Mathematical modeling of transportation networks is traditionally based on the frameworks of mathematical graph theory and discrete energy optimization. The model studied in [14] falls into this category, being formulated as an energy functional depending on edge conductivities of a given undirected discrete graph, constrained by a local mass conservation law. The mass conservation law is imposed in terms of a linear system of equations for the material pressures known as Kirchhoff law. The energy consists of a pumping term, describing the kinetic energy of the material flow through the network, and a metabolic term, which reflects the biological motivation of the model. The metabolic term is a function of the edge conductivities and is assumed to be of power-law form, with exponent $ \gamma > 0 $. For biological applications, one usually assumes that $ \gamma\in [1/2, 1] $. We note that for $ \gamma < 1 $ the energy functional is nonconvex, while for $ \gamma = 1 $ it is convex, and for $ \gamma > 1 $ it is strictly convex [14].
The first goal of this paper is to provide two new results regarding the (local) minimizers of the energy. In particular, it has been shown [6] that for $ \gamma < 1 $, each minimizer of the energy is a loop-free graph, i.e., an undirected graph where each pair of nodes is connected by at most one path. Here we show the complementary claim, namely, that if $ \gamma < 1 $, every admissible loop-free graph represents a local minimizer of the energy. The term "admissible" refers to a graph for which the Kirchhoff law is solvable when all edges with zero conductivity are treated as void (nonexistent). Moreover, we show that for $ \gamma = 1 $ every (local) minimizer is a set containing a loop-free graph.
The second goal of this paper is to incorporate the concept of robustness into the model [14]. In particular, the fact that, for biological applications where $ \gamma\leq 1 $, the energy minimizers are loop-free, does not correspond to the transportation networks observed in most real organisms, where typically (many) loops are present. For a striking example of a highly redundant pattern in leaf venation, we refer to [17, Figure 1]. There, the authors assert that the high redundancy of paths from the leaf base to any point on the leaf surface might be very advantageous with regard to local damages, i.e., they promote robustness of the transportation network against the removal of edges. We therefore propose to incorporate the robustness aspect into the discrete model [14] by extending the energy functional by a term that measures the connectivity of the network.
The concept of network robustness (or resilience) is usually defined as the ability of a network to maintain its function when subjected to disturbances or attacks [1,27]. In our paper, we understand the term robustness as the ability of the network to withstand disturbance of its connectivity through the removal of edges. A generally accepted measure of connectivity of a discrete graph is the Cheeger constant [19,21], also called the isoperimetric number or minimum cut. Loosely speaking, the Cheeger constant is a numerical measure of whether or not a graph has a "bottleneck", i.e., a small subset of edges whose removal destroys the connectivity of the graph. The Cheeger constant is strictly positive if and only if the underlying (undirected) graph is connected. Graphs with a small (but positive) Cheeger constant have a "bottleneck", in the sense that there are two "large" sets of vertices with "few" links (edges) between them. On the other hand, the Cheeger constant is "large" if any possible division of the vertex set into two subsets has "many" links between those two subsets. The fundamental problem is that the calculation of the isoperimetric number of graphs with multiple edges is NP hard [19]. To circumvent this issue and to obtain a practically tractable optimization problem, we propose to replace it by a related quantity, the algebraic connectivity, also called the Fiedler number [8]. This is defined as the second smallest eigenvalue of the matrix Laplacian of the connectivity matrix of the graph; for a connected graph, this is the smallest nonzero eigenvalue. The Fiedler number is another generally accepted measure of network robustness [18,21], and is related to the Cheeger constant by the well-known Cheeger inequalities [19]. These state that the isoperimetric number is bounded from below by one half of the Fiedler number. Consequently, knowing that a given graph is "well-connected" due to a sufficiently large value of its Fiedler number guarantees robustness also in terms of its Cheeger constant.
In this paper, we take into account the connectivity of the graph in terms of the Fiedler number calculated from the weighted adjacency matrix, with weights equal to the edge conductivities. This is motivated by the fact that, if a certain transport path is severed due to the removal of some edges, then the "alternative" (backup) path should have enough transport capacity, i.e., high enough conductivity. The Fiedler number is then a Lipschitz continuous function of the edge conductivities. Moreover, as long as it represents a simple eigenvalue (i.e., of multiplicity one) of the matrix Laplacian, it is differentiable with respect to the conductivities.
We thus formulate a new energy functional, where a multiple of the Fiedler number is subtracted from the pumping (kinetic) and metabolic energy terms. We then present two toy examples demonstrating how the modified energy functional enforces connectivity of the optimal transportation network. Moreover, since the Fiedler number is a concave function of the edge conductivities, the problem remains convex as long as the metabolic exponent $ \gamma\geq 1 $. The convexity and Lipschitz continuity of the modified energy functional facilitate the application of the projected subgradient method for its numerical minimization. We then present results of two numerical experiments, further documenting how the connectivity of the graph and its energy expenditure grow with increasing relative weight of the Fiedler number in the modified energy functional.
This paper is organized as follows. In Section 2, we describe the discrete network formation model of Hu and Cai [14] and establish some of its fundamental properties (convexity and boundedness of the kinetic energy). We also introduce the matrix Laplacian and Fiedler number of the graph. In Section 3, we provide results about the structure of optimal transportation networks in the cases $ \gamma < 1 $ and $ \gamma = 1 $. In Section 4, we extend the energy functional such that it also accounts for robustness of the transportation network in terms of the Fiedler number of the weighted graph. We then establish basic mathematical properties of the modified functional with $ \gamma = 1 $, namely, boundedness from below, coercivity, and convexity. We also provide insights into the new model by studying two toy examples. Finally, in Section 5, we implement the projected subgradient method to find global minimizers of the modified functional numerically. We present two examples where we demonstrate how the structure of the optimal graphs depends on the required robustness of the network.
The discrete network formation model introduced by Hu and Cai [14] is posed on a prescribed undirected connected graph $ G = (V, E) $, consisting of a finite set of vertices $ V $, also called nodes, and a finite set of edges $ E $. The number of vertices shall be denoted by $ | V| $ in the sequel. Any pair of vertices is connected by at most one edge and no vertex is not connected to itself (i.e., the graph does not contain "self-loops"). The edge between vertices $ i\in V $ and $ j\in V $ is denoted by $ (i, j)\in E $. Since the graph is undirected, we refer by $ (i, j) $ and $ (j, i) $ to the same edge. For each edge $ (i, j)\in E $ of the graph $ G $, we prescribe its length $ L_{ij} = L_{ji} > 0 $.
Throughout the paper, we assume that the graph $ G $ is connected, i.e., that every pair of vertices in $ V $ is connected by an undirected path (sequence of edges) in $ E $.
We treat the graph as a transportation structure, and in each vertex $ i\in V $ we have the pressure $ P_i\in{\mathbb{R}} $ of the medium flowing through the network. The oriented flux (flow rate) from vertex $ i\in V $ to $ j\in V $ is denoted by $ Q_{ij} $; obviously, we have the antisymmetry $ Q_{ij} = -Q_{ji} $. For biological networks, the Reynolds number of the flow is typically small and the flow is predominantly in the laminar (Poiseuille) regime. Then the flow rate between vertices $ i\in V $ and $ j\in V $ along the edge $ (i, j)\in E $ is proportional to the conductance $ C_{ij}\geq 0 $ of the edge and the pressure drop $ P_i-P_j $,
$ Qij:=CijPi−PjLijfor all (i,j)∈E. $ | (2.1) |
The local mass conservation in each vertex is expressed in terms of the Kirchhoff law
$ ∑j∈N(i)CijPi−PjLij=Sifor all i∈V. $ | (2.2) |
Here $ N(i) $ denotes the set of vertices connected to $ i\in V $ through an edge, i.e., $ N(i) : = \{ j\in V; \, (i, j)\in E \}. $ Moreover, $ S = (S_i)_{i\in V} $ is the strength of the flow source ($ S_i > 0 $) or sink ($ S_i < 0 $) at vertex $ i $, which has to be prescribed as a datum to the model. Given the matrix of nonnegative conductivities $ C = (C_{ij})_{(i, j)\in E} $ and the prescribed positive lengths $ L = (L_{ij})_{(i, j)\in E} $, the Kirchhoff law (2.2) is a linear system of equations for the vector of pressures $ P = (P_i)_{i\in V} $. Clearly, a necessary condition for the solvability of (2.2) is the global mass conservation $ S\in{\mathbb{R}}^{|V|}_0 $, where here and in the sequel we denote
$ R|V|0:={v∈R|V|;|V|∑i=1vi=0}. $ | (2.3) |
Remark 1 If $ C $ is the adjacency matrix of a weighted connected graph, then the solution $ P $ of Eq (2.2) always exists and is unique up to an additive constant. Although $ E $ itself is connected, $ C $ might be supported on a proper subset of $ E $ and may therefore correspond to a disconnected graph; we then shortly say that $ C $ is disconnected. In this case, the solution of (2.2) may not exist or, if it exists, it is not unique (up to an additive constant), as demonstrated in Example 2 in Appendix A. In the case when $ C $ is disconnected and (2.2) is solvable, we denote the connected components of $ C $ by $ \Gamma_1, \dots, \Gamma_n $ and observe that $ \sum_{j\in \Gamma_k}S_j = 0 $ for every $ 1\le k\le n $. Then $ P $ is uniquely defined up to $ n $ additive constants, one for every component $ \Gamma_k $. Therefore, the matrix of fluxes $ Q $ is still uniquely defined by Eq (2.1). On the other hand, if Eq (2.2) is not solvable, then $ C $ is disconnected.
The conductivities $ C_{ij} = C_{ji} \geq 0 $ of the edges are subject to an energy optimization process. Hu and Cai [14] proposed an energy cost functional consisting of a pumping power term and a metabolic cost term. According to Joule's law, the power (kinetic energy) needed to pump material through an edge $ (i, j)\in E $ is proportional to the pressure drop $ P_i-P_j $ and the flow rate $ Q_{ij} $ along the edge, i.e.,
$ (Pi−Pj)Qij=Q2ijCijLij. $ |
The metabolic cost of maintaining the edge is assumed to be proportional to its length $ L_{ij} $ and the power of its conductivity $ C_{ij}^{\gamma} $, with an exponent $ \gamma > 0 $. For instance, in blood vessels, the metabolic cost is proportional to the cross-section area of the vessel [20]. Modeling the blood flow by Hagen-Poiseuille's law, the conductivity is proportional to the square of the cross-section area, implying $ \gamma = 1/2 $ for blood vessel systems. For models of leaf venation, the material cost is proportional to the number of small tubes, which is proportional to $ C_{ij} $, and the metabolic cost is due to the effective loss of the photosynthetic power at the area of the venation cells, which is proportional to $ C_{ij}^{1/2} $. Consequently, the effective value of $ \gamma $ typically used in models of leaf venation lies between $ 1/2 $ and $ 1 $, see, e.g., [15]. The energy cost functional is thus given by
$ E[C]:=∑(i,j)∈E(Qij[C]2Cij+νγCγij)Lij, $ | (2.4) |
where $ Q_{ij} = Q_{ij}[C] $ is given by Eq (2.1) with pressures calculated from the Kirchhoff law (2.2), and $ \nu > 0 $ is the so-called metabolic coefficient. Note that every edge of the graph $ G $ is counted exactly once in the above sum, i.e., we identify each edge $ (i, j) $ with $ (j, i) $ and the energy can also be written as
$ {{\mathcal{E}}}[C] : = \frac12 \sum\limits_{i\in V} \sum\limits_{j\in V} \left( \frac{Q_{ij}[C]^2}{C_{ij}}+\frac{\nu}{\gamma} C_{ij}^{\gamma}\right) L_{ij}, $ |
where we set $ L_{ij} : = 0 $ whenever $ (i, j)\notin E $. If $ C $ is such that the linear system (2.2) is not solvable, we formally set $ {\mathcal{E}}[C]: = +\infty $. In the sequel, we shall sometimes address the kinetic (pumping) and metabolic parts of the energy separately. For this purpose, we denote
$ {{\mathcal{E}}_{\mathrm{kin}}}[C] : = \sum\limits_{(i,j)\in E} \frac{Q_{ij}[C]^2}{C_{ij}} L_{ij}, \qquad {{\mathcal{E}}_{\mathrm{met}}}[C] : = \frac{\nu}{\gamma} \sum\limits_{(i,j)\in E} C_{ij}^\gamma L_{ij}. $ | (2.5) |
In order to find the optimal transportation structure for a given vector of sources and sinks $ S\in{\mathbb{R}}^{|V|}_0 $, one needs to find the (global) minimum of the energy functional (2.4), coupled to the Kirchhoff law (2.1) and (2.2). The energy functional is to be minimized over the convex set $ \mathcal{C} $ of symmetric matrices with nonnegative elements,
$ \mathcal{C} : = \left\{ C\in{\mathbb{R}}^{|V|\times |V|}; \, C_{ij} = C_{ji} \geq 0 \text{ for all } i, j\in V, \text{ with } C_{ij} = C_{ij} = 0 \text{ if } (i,j)\notin E \right\}. $ | (2.6) |
If the optimal solution has $ C_{ij} = C_{ij} = 0 $ for any $ (i, j)\in E $, then the corresponding edge $ (i, j) $ is considered nonexistent. We collected some basic mathematical properties of the functional $ {\mathcal{E}} = {\mathcal{E}}[C] $ in Appendix A. An important property in the context of the optimization task (2.1)–(2.4) is the convexity of the problem.
Proposition 1. The energy functional (2.4), constrained by the Kirchhoff law (2.1) and (2.2), is strictly convex for $ \gamma > 1 $ and convex for $ \gamma = 1 $.
This result is not new, see for instance [10,11] for an analogous result for the continuum version of the problem. For the sake of the reader, we offer its proof for the discrete model (2.1)–(2.4) in the Appendix (Lemma 8). Convexity of the functional shall be instrumental in employing the projected subgradient method in Section 5.
Definition 1. For any $ C\in \mathcal{C} $, we define the matrix Laplacian $ {\mathcal{L}}[C] $ as
$ {\mathcal{L}}[C] : = D - C, $ |
where $ D $ is the diagonal matrix of row/column sums of $ C $ (recall that $ C\in \mathcal{C} $ is symmetric). For simplicity, we shall often write just $ {\mathcal{L}} $ instead of $ {\mathcal{L}}[C] $.
With this definition, we can express the Kirchhoff law (2.2) in the form
$ {\mathcal{L}}[\widetilde C] P = S, $ | (2.7) |
where we introduce the matrix $ \widetilde C_{ij} : = C_{ij}/L_{ij} $, with $ \widetilde C_{ij} : = 0 $ if $ L_{ij} = 0 $ (recall that $ L_{ij} = 0 $ means that $ (i, j)\notin E $ and, consequently, $ C_{ij} = 0 $). Moreover, we denoted $ P = (P_i)_{i\in V} $ the vector of pressures and $ S = (S_i)_{i\in V} $ the vector of sources and sinks.
We also note that if $ P $ is a solution of (2.7), then an easy calculation gives the following expression for the kinetic energy (2.5),
$ {{\mathcal{E}}_{\mathrm{kin}}}[C] = P^T {\mathcal{L}}[\widetilde C] P. $ | (2.8) |
Definition 2. The Fiedler number $ {\mathfrak{f}}[C] $ of a matrix $ C\in \mathcal{C} $ is the second smallest eigenvalue of the matrix Laplacian $ {\mathcal{L}}[C] $, where multiple eigenvalues are counted separately [8].
Lemma 1. Let $ 0 \neq S \in R_0^{|V|} $. For any $ C\in \mathcal{C} $, we have
$ {{\mathcal{E}}_{\mathrm{kin}}}[C] \leq \frac{\left\| {S} \right\|^2}{{\mathfrak{f}}[\widetilde C]}, $ | (2.9) |
where $ {{\mathcal{E}}_{\mathrm{kin}}}[C] $ is defined in Eq (2.5) and $ \widetilde C_{ij} : = C_{ij}/L_{ij} $, with $ \widetilde C_{ij} : = 0 $ if $ L_{ij} = 0 $. Here and in the sequel, $ \left\| {S} \right\| $ denotes the $ \ell^2 $-norm of the vector $ S $.
Proof. For the case when $ C\in \mathcal{C} $ is such that the Kirchhoff law (2.7) is not solvable, we have set $ {{\mathcal{E}}_{\mathrm{kin}}}[C] = +\infty $. But then the graph composed of edges $ (i, j) $ with $ C_{ij} > 0 $ is not connected, see Remark 1, thus $ {\mathfrak{f}}[C] = 0 $ and (2.9) holds.
Let us now fix $ C\in \mathcal{C} $ such that the Kirchhoff law (2.7) admits a solution $ P\in{\mathbb{R}}^{|V|} $. Since $ P $ can be shifted by an additive constant, we choose $ P\in{\mathbb{R}}^{|V|}_0 $. Therefore, denoting $ \mathbf{1} : = (1, \ldots, 1)\in{\mathbb{R}}^{|V|} $, we have $ P^T \mathbf{1} = 0 $. The Courant-Fisher (minimax) theorem [22] gives
$ {\mathfrak{f}}[\widetilde C] = \min \left\{ \frac{v^T {\mathcal{L}}[\widetilde C]v}{\left\| {v} \right\|^2}; \, v\in{\mathbb{R}}^{|V|}, \, v^T \mathbf{1} = 0 \right\}. $ | (2.10) |
Consequently,
$ {\mathfrak{f}}[\widetilde C] \left\| {P} \right\|^2 \leq P^T {\mathcal{L}}[\widetilde C] P. $ | (2.11) |
Taking a scalar product of Eq (2.7) with $ P $ and applying the Cauchy-Schwartz and Young inequalities yields
$ P^T {\mathcal{L}}[\widetilde C] P = P^T S \leq \left\| {S} \right\| \left\| {P} \right\| \leq \frac{\left\| {S} \right\|^2}{2 {\mathfrak{f}}[\widetilde C]} + \frac{\left\| {P} \right\|^2}{2} {\mathfrak{f}}[\widetilde C]. $ |
Combining this with Eq (2.11), we obtain
$ P^T {\mathcal{L}}[\widetilde C] P \leq \frac{\left\| {S} \right\|^2}{{\mathfrak{f}}[\widetilde C]} $ |
and we conclude by using Eq (2.8).
In this section, we shall provide results about the structure of optimal transportation networks for $ \gamma < 1 $ and $ \gamma = 1 $. Let us recall that we assume the graph $ G = (V, E) $ to be connected.
In [6, Theorem 2.1], the authors proved that every global minimizer of the energy functional (2.4) with $ \gamma < 1 $, constrained by the Kirchhoff law (2.1) and (2.2), is loop-free (again, we remind that edges with zero conductivities are treated as nonexistent). By a very slight modification of the proof, one can extend the claim to local minimizers. Here we prove the opposite claim, namely, that for $ \gamma < 1 $ and every spanning tree of the connected graph $ (V, E) $, we can find a set of conductivities $ C\in \mathcal{C} $ supported on the spanning tree which is a local minimizer of the constrained energy.
Theorem 1. Let $ \gamma < 1 $ and let $ (V, \tilde E) $, $ \tilde E \subseteq E $, be a spanning tree of the connected graph $ (V, E) $. Then there exists a set of conductivities $ \tilde C\in \mathcal{C} $, supported on $ \tilde E $, which is a local minimizer of the energy (2.4) constrained by the Kirchhoff law (2.1) and (2.2), on the set $ \mathcal{C} $.
Proof. We observe that since $ (V, \tilde E) $ is a tree, the fluxes $ \tilde Q_{ij} $ for $ (i, j)\in\tilde E $ are uniquely determined by $ S $. Indeed, for any edge $ (i, j)\in\tilde E $, there is a unique split of the vertices into two disjoint sets $ V^{(i)} $, $ V^{(j)} $ such that nodes from $ V^{(i)} $ are connected to $ i $ by paths in $ \tilde E \setminus \{(i, j)\} $, and analogously for nodes in $ V^{(j)} $. By the local mass conservation, the flux through $ (i, j) $ is then given by
$ \tilde Q_{ij} = \sum\limits_{k\in V^{(i)}} S_k - \sum\limits_{k\in V^{(j)}} S_k. $ | (3.1) |
Once we identified the fluxes by the above prescription, we construct the conductivities $ \tilde C\in \mathcal{C} $ as follows,
$ ˜Cij:=(˜Q2ij/ν)1/(γ+1) for all (i,j)∈˜E,:=0 for all (i,j)∈E∖˜E. $ | (3.2) |
Obviously, $ \tilde C $ is supported on $ \tilde E $. We now claim that $ \tilde C $ represents a local minimizer of the energy (2.4) constrained by the Kirchhoff law (2.1) and (2.2) on the set $ \mathcal{C} $.
Let us first observe that since $ (V, \tilde E) $ is connected, the Kirchhoff law (2.2) admits a solution $ \tilde P\in{\mathbb{R}}^{|V|} $. Clearly, the fluxes $ \tilde Q_{ij} $ constructed in (3.1) verify the relation (2.1). Moreover, we recall the formula [9, Lemma 2.1] for the derivative
$ ∂E[˜C]∂˜Cij=−(˜Q2ij˜C2ij−ν˜Cγ−1ij)Lij. $ | (3.3) |
Then, by Eq (3.2), we have
$ \frac{\partial {\mathcal E}[\tilde C]}{\partial \tilde C_{ij}} = 0 \qquad \mbox{for all } (i,j)\in\tilde E \mbox{ with } \tilde C_{ij} > 0. $ |
Moreover, using Eq (2.1), we recast Eq (3.3) as
$ \frac{\partial {\mathcal{E}}[\tilde C]}{\tilde C_{ij}} = -\frac{(\tilde P_j-\tilde P_i)^2}{L_{ij}}+\nu \tilde C_{ij}^{\gamma-1}L_{ij}, $ |
which gives
$ \frac{\partial {\mathcal E}[\tilde C]}{\partial \tilde C_{ij}} = +\infty \qquad \mbox{for all } (i,j)\in E \mbox{ with } \tilde C_{ij} = 0, $ |
where the partial derivative is taken from the right, i.e., as $ \tilde C_{ij}\to 0+ $. Since $ \tilde C\in \mathcal{C} $ is a unique element of $ \mathcal{C} $ with these properties, we conclude that it is a local minimizer of $ {\mathcal{E}} = {\mathcal{E}}[C] $ on $ \mathcal{C} $.
Remark 2. Combining the claims of Theorem 1 and [6, Theorem 2.1] implies that the set of local minimizers of the energy functional (2.4) with $ \gamma < 1 $ is the set of spanning trees of the graph $ (V, E) $. Consequently, a possible way to find the global minimizer is to search the set of spanning trees and identify the one with the smallest value of the energy $ {\mathcal{E}} = {\mathcal{E}}[C] $. The value of the energy for any given tree is calculated using the procedure described in the proof of Theorem 1. Clearly, this approach quickly turns computationally infeasible with the growing size of the graph. For instance, the number of spanning trees for a complete graph with $ | V| $ nodes is $ | V|^{| V|-2} $ by Cayley's formula. For planar graphs, the number of possible spanning trees grows exponentially, see, e.g., [5].
In this section, we consider the energy functional $ {\mathcal{E}} = {\mathcal{E}}[C] $ with $ \gamma = 1 $. By Proposition 1, $ {\mathcal{E}} $ is a convex function on $ \mathcal{C} $ and, therefore, the set $ M $ of its minimizers,
$ M:={˜C∈C;E(˜C)=minC∈CE(C)} $ | (3.4) |
is a closed convex subset of $ \mathcal{C} $. Moreover, due to the coercivity
$ {\mathcal{E}}[C] \geq\nu \sum\limits_{(i,j)\in E} C_{ij} L_{ij}, $ |
it is nonempty. Its structure is characterized by the following theorem.
Theorem 2. Let $ \gamma = 1 $ and let $ M $ be defined by Eq (3.4). Then the extremal points of $ M $ represent loop-free graphs, and vice-versa, the loop-free elements of $ M $ are extremal points of $ M $.
Proof. We first prove that the extremal points of $ M $ represent loop-free graphs. Let $ \tilde C\in M $. Since $ {\mathcal{E}}[\tilde C] < +\infty $, the Kirchhoff law (2.2) admits a solution $ \tilde P\in{\mathbb{R}}^{|V|} $. Formula (A.11) gives
$ \frac{\partial {\mathcal{E}}[\tilde C]}{\partial\tilde C_{ij}} = -\frac{(\tilde P_j-\tilde P_i)^2}{L_{ij}}+\nu L_{ij}. $ |
If $ \tilde C_{ij} > 0 $ for some $ (i, j)\in E $, we have $ \frac{\partial {\mathcal{E}}[\tilde C]}{\partial\tilde C_{ij}} = 0 $ and, consequently,
$ (˜Pj−˜Pi)2=νL2ij>0. $ | (3.5) |
For contradiction, let us now assume that $ \tilde C $ contains a loop, i.e., that there is a chain of edges
$ T={(i0,i1),…,(iK−2,iK−1),(iK−1,i0)}⊂E, $ | (3.6) |
such that $ \tilde C_{i_j, i_{j+1}} > 0 $ for $ j = 0, \dots, K-1 $ (we count the indices modulo $ K $, so $ i_{K}\equiv i_0 $). We also assume that $ {\mathcal T} $ is the shortest loop, i.e., that all $ i_0, \dots i_{K-1} $ are mutually different. By Eq (3.5), we know that $ \tilde P_{i_j}\neq \tilde P_{i_{j+1}} $ and, with (2.1), the corresponding fluxes $ \tilde Q_{i_j, i_{j+1}} $ are all nonvanishing.
We rewrite Eq (3.5) as
$ ˜Pij−˜Pij+1=sign(˜Pij−˜Pij+1)√νLij,ij+1 $ | (3.7) |
and sum these identities over $ j = 0, \dots, K-1 $ to obtain
$ 0=K−1∑j=0sign(˜Pij−˜Pij+1)Lij,ij+1. $ | (3.8) |
We now perturb $ \tilde Q $ by adding a circular flow along $ {\mathcal T}. $ For $ (i, j)\in E\setminus {\mathcal T} $ we set
$ Q_{ij}: = \tilde Q_{ij}, \qquad C_{ij}: = \tilde C_{ij}, $ |
while for $ l = 0, \dots, K-1 $,
$ Q_{i_l,i_{l+1}}: = \tilde Q_{i_l,i_{l+1}}+\varepsilon,\qquad C_{i_l,i_{l+1}}: = \tilde C_{i_l,i_{l+1}}+\frac{\varepsilon}{\sqrt{\nu}}\cdot{\rm{sign}}(\tilde P_{i_l}-\tilde P_{i_{l+1}}), $ | (3.9) |
with some $ \varepsilon\neq 0 $. Obviously, if $ |\varepsilon| $ is small enough, we have $ C\in \mathcal{C}. $ Also the local mass conservation $ \sum_{j\in N(i)} Q_{ij}=S_i $ is verified for all $ i\in V. $ Moreover, if $ (i, j)\not\in {\mathcal T} $, then $ Q_{ij}=C_{ij} \frac{\tilde P_i-\tilde P_j}{L_{ij}} $ remains true. On the other hand, for $ (i_l, i_{l+1})\in{\mathcal T} $ we use Eqs (3.7) and (3.9) to get
$ Cil,il+1⋅˜Pil−˜Pil+1Lil,il+1=[˜Cil,il+1+ε√ν⋅sign(˜Pil−˜Pil+1)]⋅˜Pil−˜Pil+1Lil,il+1=˜Qil,il+1+ε√ν⋅sign(˜Pil−˜Pil+1)⋅sign(˜Pil−˜Pil+1)√ν=˜Qil,il+1+ε=Qil,il+1. $ |
Finally, using Eqs (3.5), (3.9), and (3.8) successively, we have
$ K−1∑l=0[Cil,il+1(˜Pil−˜Pil+1)2L2il,il+1+νCil,il+1]Lil,il+1=2νK−1∑l=0Cil,il+1Lil,il+1=2νK−1∑l=0[˜Cil,il+1+ε√ν⋅sign(˜Pil−˜Pil+1)]Lil,il+1=2νK−1∑l=0˜Cil,il+1Lil,il+1+2ε√νK−1∑l=0sign(˜Pil−˜Pil+1)Lil,il+1=2νK−1∑l=0˜Cil,il+1Lil,il+1, $ |
so that one more application of (3.5) gives
$ \sum\limits_{l = 0}^{K-1}\left[C_{i_l,i_{l+1}}\frac{(\tilde P_{i_l}- \tilde P_{i_{l+1}})^2}{L_{i_l,i_{l+1}}^2}+\nu C_{i_l,i_{l+1}}\right]L_{i_l,i_{l+1}} = \sum\limits_{l = 0}^{K-1}\left[\tilde C_{i_l,i_{l+1}}\frac{(\tilde P_{i_l}-\tilde P_{i_{l+1}})^2}{L_{i_l,i_{l+1}}^2}+\nu \tilde C_{i_l,i_{l+1}}\right] L_{i_l,i_{l+1}}. $ |
This immediately implies that $ {\mathcal E}[C] = {\mathcal E}[\tilde C] $. It follows that if $ \tilde C $ contains a loop, then it can be written as a convex combination of $ C^{(+)} $ and, respectively, $ C^{(-)}\in \mathcal{C} $, which are obtained from $ \tilde C $ by taking (small) positive and, respectively, negative $ {\varepsilon} $ in Eq (3.9). Therefore, $ \tilde C $ is not an extremal point of $ M $. This concludes the proof that all extremal points of $ M $ are loop-free.
Let us now prove the converse claim, namely, that every loop-free element of $ M $ is an extremal point. Let $ C\in M $ be loop-free and assume, for contradiction, that $ C $ can be written as $ C = \alpha C^{(1)}+(1-\alpha)C^{(2)} $ for some $ 0 < \alpha < 1 $ and $ C^{(1)}, C^{(2)}\in M $. But then
$ \bigl\{ (i,j)\in E:C_{ij} > 0 \bigr\} = \bigl\{(i,j)\in E:C^{(1)}_{ij} > 0 \bigr\}\cup \bigl\{(i,j)\in E:C^{(2)}_{ij} > 0 \bigr\}, $ |
which means that $ C^{(1)} $ and $ C^{(2)} $ are also loop-free. As $ {\mathcal{E}}(C) = {\mathcal{E}}(C^{(1)}) = {\mathcal{E}}(C^{(2)}) < +\infty $, the Kirchhoff law (2.2) is solvable with $ C $, $ C^{(1)} $ and, respectively, $ C^{(2)} $, and (2.1) gives the corresponding fluxes $ Q $, $ Q^{(1)} $ and, respectively, $ Q^{(2)} $. As for loop-free networks the fluxes are uniquely determined by the sources/sinks $ S $, we have $ Q = Q^{(1)} = Q^{(2)} $ and, in turn, $ C = C^{(1)} = C^{(2)} $. Hence, $ C $ is an extremal point.
The Cheeger constant (also called the isoperimetric number) of a graph is a numerical measure of whether or not a graph has a "bottleneck". For $ G = (V, E) $, its Cheeger constant $ {\mathfrak{h}}(G) $ is defined as
$ {\mathfrak{h}}(G) = \min \left\{ \frac{|\partial W|}{|W|}; \; \emptyset \neq W \subset V, |W| \leq \frac{|V|}{2} \right\}, $ |
where $ \partial W \subset E $ denotes the set of edges having one end in $ W $ and the other end in $ V\setminus W $.
The Cheeger constant is strictly positive if and only if $ G $ is a connected graph. Intuitively, if the Cheeger constant is small but positive, then there exists a "bottleneck", in the sense that there are two "large" sets of vertices with "few" links (edges) between them. The Cheeger constant is large if any possible division of the vertex set into two subsets has "many" links between those two subsets. In other words, it is a measure of resilience of the network against disturbance of the connectivity through the removal of edges.
In general, the calculation of the isoperimetric number of graphs with multiple edges is NP hard [19]. Consequently, we consider a related quantity, the algebraic connectivity, also called the Fiedler number, defined as the second smallest eigenvalue of the matrix Laplacian $ {\mathcal{L}} = {\mathcal{L}}[A] $ of the adjacency matrix $ A $. For a connected graph, which we assume throughout this paper, it is the smallest nonzero eigenvalue of $ {\mathcal{L}}[A] $. The Fiedler number $ {\mathfrak{f}}(G) $ is another classical measure of connectivity of the graph, or robustness with respect to edge removal. It is related to the isoperimetric number by the well-known Cheeger inequalities [19],
$ \frac{{\mathfrak{h}}(G)^2}{2\Delta(G)} \leq {\mathfrak{f}}(G) \leq 2 {\mathfrak{h}}(G), $ |
where $ \Delta(G) $ is the maximum degree for the nodes in $ G $.
In our paper, we shall work with the generalized Fiedler number $ {\mathfrak{f}}[C] $, calculated as the second smallest eigenvalue of the weighted matrix Laplacian $ {\mathcal{L}}[C] $ with weights given by the edge conductivities $ C\in \mathcal{C} $. This reflects the modeling assumption that edges with higher conductivities contribute more to overall network robustness (for instance, because they may be more resilient against severing). Our idea is to modify the energy functional (2.4) to take robustness into account in terms of the generalized Fiedler number $ {\mathfrak{f}}[C] $. Consequently, for $ \mu > 0 $ we introduce the modified energy functional $ {\mathcal{F}} = {\mathcal{F}}[C] $,
$ {\mathcal{F}}[C] : = {\mathcal{E}}[C] - \mu\,{\ell} \, \frac{|V|-1}{2} \, {\mathfrak{f}}[C], $ | (4.1) |
where $ {\mathcal{E}} = {\mathcal{E}}[C] $ is defined in (2.4) and
$ {\ell} : = \min\limits_{(i,j)\in E} L_{ij} > 0. $ | (4.2) |
The reason for scaling the second term in (4.1) by $ {\ell} $ is that the energy (2.4) is homogeneous with respect to the edge lengths, i.e., a multiplication of $ L=(L_{ij})_{(i, j)\in E} $ by a positive factor leads to a rescaling of the value of the energy by the same factor. Consequently, the modified energy (4.1) has the same scaling property. The motivation for introducing the factor $ \frac{|V|-1}{2} $ is clarified in Lemma 2 below.
Recalling that the model (2.1)–(2.4) bears relevance in biological applications as long as $ \gamma\leq 1 $, we consider the modified energy functional (4.1) exclusively with $ \gamma = 1 $ in the sequel. The reason for excluding the values $ \gamma < 1 $ is that $ {\mathcal{F}} = {\mathcal{F}}[C] $ with $ \mu > 0 $ is then, in general, not bounded from below.
Example 1. Let us consider a complete graph with all edges of the same length, i.e., $ L_{ij} = {\ell} $ for all $ (i, j)\in E $. Moroever, for some $ c > 0 $, let us take $ C_{ij} = c $ for all $ (i, j)\in E $. Then we have $ {\mathfrak{f}}[C] = c |V| $ and
$ F[C]=Ekin[C]+ν∑(i,j)∈ECγijLij−μℓ|V|−12f[C]=Ekin[C]+ℓ|V||V|−12(νcγ−μc). $ |
Lemma 1 gives
$ {{\mathcal{E}}_{\mathrm{kin}}}[C] \leq \frac{\left\| {S} \right\|^2}{{\mathfrak{f}}[\widetilde C]} = \frac{\left\| {S} \right\|^2 {\ell}}{c |V|}, $ |
with $ \widetilde C_{ij} : = C_{ij}/L_{ij} = c/{\ell} $. Consequently, we readily have $ \lim_{c\to+\infty} {\mathcal{F}}[C] = -\infty $ whenever $ \gamma < 1 $ and $ \mu > 0 $.
On the other hand, for $ \gamma = 1 $ we have the following Lemma establishing boundedness of $ {\mathcal{F}} = {\mathcal{F}}[C] $ from below if $ \mu\leq\nu $ and coercivity if $ \mu < \nu $. It is a direct consequence of [8], Claim 3.5], which gives an upper bound on the value of the Fiedler number. We provide the complete proof here for the sake of the reader.
Lemma 2. Let $ \gamma = 1 $ and $ \mu \leq \nu $. Then the modified energy functional $ {\mathcal{F}} = {\mathcal{F}}[C] $, defined in Eq (4.1), satisfies
$ {\mathcal{F}}[C] \geq {{\mathcal{E}}_{\mathrm{kin}}}[C] \geq 0 \qquad for\ all\ C\in \mathcal{C}. $ | (4.3) |
Moreover, if $ \mu < \nu $, then $ {\mathcal{F}} = {\mathcal{F}}[C] $ is coercive in the sense that there exists $ \alpha > 0 $ such that
$ {\mathcal{F}}[C] \geq \alpha \sum\limits_{i = 1}^{|V|} \sum\limits_{j = 1}^{|V|} C_{ij} \qquad for\ all\ C\in \mathcal{C}. $ | (4.4) |
Proof. Let $ {\mathcal{L}} $ be the Laplacian matrix corresponding to $ C\in \mathcal{C} $ and let $ {\mathfrak{f}}[C] $ be the Fiedler number of $ C $. We use the Courant-Fisher representation (2.10) of $ {\mathfrak{f}}[C] $ in the form
$ {\mathfrak{f}}[C] = \min \left\{v^T {\mathcal{L}} v;\, v\in{\mathbb{R}}^{|V|}, \, \left\| {v} \right\| = 1,\, v^T\mathbf{1} = 0 \right\}. $ | (4.5) |
We claim that the matrix
$ \widetilde {\mathcal{L}} : = {\mathcal{L}} - \left( I - \frac{\mathbf{1}\otimes\mathbf{1}}{|V|} \right) {\mathfrak{f}}[C] $ |
is positive semidefinite. Let us choose any vector $ y\in{\mathbb{R}}^{|V|} $, then we have the decomposition $ y = \alpha \mathbf{1} + \beta v $, where $ v\perp\mathbf{1} $ and $ \left\| {v} \right\| = 1 $. Since, trivially, $ \widetilde {\mathcal{L}} \mathbf{1} = \mathcal{O} $, where $ \mathcal{O} $ denotes the zero vector in $ {\mathbb{R}}^{|V|} $, we have
$ y^T \widetilde {\mathcal{L}} y = \beta^2 v^T \widetilde {\mathcal{L}} v = \beta^2 \left(v^T {\mathcal{L}} v - {\mathfrak{f}}[C] \right) \geq 0, $ |
where the nonnegativity follows from Eq (4.5). Indeed, $ \widetilde {\mathcal{L}} $ is positive semidefinite, and, consequently, all its diagonal elements are nonnegative. In particular,
$ \min\limits_{i\in V} \widetilde {\mathcal{L}}_{ii} = \min\limits_{i\in V} {\mathcal{L}}_{ii} - \left(1 - |V|^{-1} \right) {\mathfrak{f}}[C] \geq 0, $ |
so that
$ {\mathfrak{f}}[C] \leq \frac{|V|}{|V| - 1} \min\limits_{i\in V} {\mathcal{L}}_{ii} = \frac{|V|}{|V| - 1} \min\limits_{i\in V} \sum\limits_{j = 1}^{|V|} C_{ij} \leq \frac{1}{|V| - 1} \sum\limits_{i = 1}^{|V|} \sum\limits_{j = 1}^{|V|} C_{ij}. $ |
Therefore, using the convention $ L_{ij} = C_{ij} = 0 $ if $ (i, j)\notin E $, we finally conclude that
$ F[C]=Ekin[C]+ν2|V|∑i=1|V|∑j=1CijLij−μℓ|V|−12f[C]≥Ekin[C]+(ν2min(i,j)∈ELij−μℓ2)|V|∑i=1|V|∑j=1Cij=Ekin[C]+(ν−μ)ℓ2|V|∑i=1|V|∑j=1Cij, $ |
with $ {\ell} = \min_{(i, j)\in E} L_{ij} > 0 $ as defined in Eq (4.2). For $ \mu\leq\nu $ this immediately gives (4.3). Moreover, if $ \mu < \nu $, then Eq (4.4) holds with $ \alpha : = (\nu-\mu) \frac{{\ell}}{2} > 0 $.
Remark 3. The statement of Lemma 2 is optimal in the following sense: Considering again the setting from Example 1, we have
$ F[C]=Ekin[C]+ν∑(i,j)∈ECijLij−μℓ|V|−12f[C]≤‖S‖2ℓc|V|+ℓ|V||V|−12(ν−μ)c. $ |
Consequently, $ \lim_{c\to+\infty} {\mathcal{F}}[C] = 0 $ if $ \mu = \nu $, and $ \lim_{c\to+\infty} {\mathcal{F}}[C] = -\infty $ if $ \mu > \nu $.
Let us note that the idea of promoting robustness of the transportation network by means of the Fiedler number $ {\mathfrak{f}}[C] $ can also be realized by minimizing (2.4) subject to (2.1) and (2.2) on the subset of $ \mathcal{C} $ where $ {\mathfrak{f}}[C]\geq \mu $, for a given $ \mu > 0 $. With this setting, the values of $ \gamma < 1 $ are perfectly admissible. We plan to explore this direction in a future work.
A fundamental property of the functional (4.1) with $ \gamma = 1 $ is its convexity.
Lemma 3. The functional (4.1) with $ \gamma = 1 $ is convex on the set $ \mathcal{C} $.
Proof. We recast (4.5) as
$ {\mathfrak{f}}[C] = \min \left\{\sum\limits_{(i,j)\in E} C_{ij}(v_i - v_j)^2;\, v\in{\mathbb{R}}^{|V|}, \, \left\| {v} \right\| = 1,\, v^T\mathbf{1} = 0\right\}. $ |
Therefore, $ {\mathfrak{f}}[C] $ is a minimum of linear functionals in $ C $ and thus concave. The claim follows directly from the convexity of $ {\mathcal{E}} = {\mathcal{E}}[C] $ with $ \gamma = 1 $ established in Proposition 1.
Finally, we have the following trivial observation about monotonicity of the Fiedler number of the minimizers of (4.1) with respect to the values of $ \mu\geq 0 $.
Lemma 4. Fix $ 0 \leq \mu_1 \leq \mu_2 $ and assume that $ C^1 $, and, respectively, $ C^2\in \mathcal{C} $ are global minimizers of (4.1) on $ \mathcal{C} $ with $ \mu_1 $ and, respectively, $ \mu_2 $. Then $ {\mathfrak{f}}[C^1] \leq {\mathfrak{f}}[C^2] $.
Proof. For brevity, let us denote $ \bar\mu_i : = \mu_i\, {\ell} \, \frac{|V|-1}{2} $ for $ i = 1, 2 $. According to the assumptions, we have
$ E[C1]−ˉμ1f[C1]≤E[C2]−ˉμ1f[C2],E[C2]−ˉμ2f[C2]≤E[C1]−ˉμ2f[C1]. $ |
Adding these two inequalities gives
$ (\bar\mu_1 - \bar\mu_2) \left({\mathfrak{f}}[C^1] - {\mathfrak{f}}[C^2] \right) \geq 0, $ |
and since we have $ \bar\mu_1 \leq \bar\mu_2 $, the claim directly follows.
Let us now present two toy models where, in a very simple setting, we are able to provide explicit results for the minimizers of (4.1) with $ \gamma = 1 $ and their dependence on the value of $ \mu > 0 $.
Let us consider a network consisting of three nodes $ V = \{0, 1, 2\} $ and three (unoriented) edges $ E = \{(0, 1), (0, 2), (1, 2)\} $ of unit length. The sources and sinks are given by
$ S_0 = 1,\qquad S_1 = -1,\qquad S_2 = 0. $ |
Taking into account the symmetry of the problem, we denote $ C_0 $ the conductivity of the edge (0, 1), and $ C_1 $ the conductivity of the other two edges, (0, 2) and (1, 2). Moreover, let us denote $ Q_0 $ the flux from vertex 0 to vertex 1 through the edge (0, 1). Obviously, then the flux from vertex 0 to vertex 2 through the edge (0, 2) is $ 1-Q_0 $ and the flux from vertex 1 to vertex 2 through the edge (1, 2) is $ Q_0-1 $. The Kirchhoff law yields, by a simple calculation,
$ Q_0 = \frac{2C_0}{2C_0 + C_1}, \qquad 1-Q_0 = \frac{C_1}{2C_0 + C_1}. $ |
Then the energy functional (2.4) with $ \gamma = 1 $ reads
$ \mathcal{E}[C] = \frac{Q_0^2}{C_0} + 2\frac{(1-Q_0)^2}{C_1} + \nu\left( C_0 + 2 C_1 \right) $ | (4.6) |
$ = \frac{2}{2C_0 + C_1} + \nu\left( C_0 + 2 C_1 \right), $ | (4.7) |
with metabolic coefficient $ \nu > 0 $. The global minimizer $ \bar C $ of $ \mathcal{E}[C] $ is $ \bar C_0 = 1/\sqrt\nu $, $ \bar C_1 = 0 $, with $ \mathcal{E}[\bar C] = 2\sqrt\nu $. Obviously, with $ \bar C_1 = 0 $, only the edge $ (0, 1) $ is present and, consequently, the graph is disconnected. However, since $ S_2 = 0 $, the Kirchhoff law (2.2) is solvable and the network still fulfills its task of transporting unit mass from node $ 0 $ to node 1.
The eigenvalues of the matrix Laplacian of the corresponding weighted adjacency matrix are
$ \lambda_0 = 0, \qquad \lambda_1 = 3C_1, \qquad \lambda_2 = 2C_0 + C_1. $ |
Consequently, the Fiedler number is $ {\mathfrak{f}}[C] = \min\{ 2C_0 + C_1, 3C_1 \} $ and the modified energy (4.1) reads
$ {\mathcal{F}}[C] = \frac{2}{2C_0 + C_1} + \nu\left( C_0 + 2 C_1 \right) - \mu \min\{ 2C_0 + C_1, 3C_1 \}, $ | (4.8) |
with $ \mu \geq 0 $. We note that due to the concavity of $ {\mathfrak{f}}[C] = \min\{ 2C_0 + C_1, 3C_1 \} $ and strict (but non-uniform) convexity of the kinetic energy term $ \frac{2}{2C_0 + C_1} $, the functional $ {\mathcal{F}}[C] $ is non-uniformly strictly convex on the positive quadrant $ {\mathbb{R}}^2_+ $. Lemma 2 states that $ {\mathcal{F}}[C] $ is bounded from below if $ \mu\leq\nu $ and coercive if $ \mu < \nu $. By inspecting the case $ C_0 = C_1 $, we see that this is also a necessary condition. In fact, for $ \mu = \nu $, choosing $ C_0 = C_1 = c $ with $ c > 0 $, we have $ {\mathcal{F}}[C] = \frac{2}{3c} $ and, consequently, $ {\mathcal{F}} = {\mathcal{F}}[C] $ does not have a minimizer. We thus study the minimizers of (4.8) for $ 0 < \mu < \nu $. The result of this simple exercise, where we chose $ \nu: = 1 $, is displayed in Figure 1. We observe that for $ \mu < 1/2 $, the optimal conductivities $ \bar C $ are $ \bar C_0 = 1/\sqrt{\nu} $, $ \bar C_1 = 0 $, i.e., the same as for $ \mu = 0 $. Therefore, choosing $ \mu < 1/2 $ does not enforce any improvement of the robustness of the network. On the other hand, for $ \mu > 1/2 $, the optimal conductivities are
$ \bar C_0 = \bar C_1 = \frac13 \sqrt{\frac{2}{\nu-\mu}}. $ |
In the right panel of Figure 1, we observe that the modified optimal energy $ {\mathcal{F}}[\bar C] $ is continuous, constant for $ 0 < \mu < 1/2 $, and decaying to zero as $ \mu\to 1- $. The optimal kinetic-metabolic energy $ {\mathcal{E}}[\bar C] $ is also constant for $ 0 < \mu < 1/2 $, but has a discontinuity at $ \mu = 1/2 $ and increases to $ +\infty $ as $ \mu\to 1- $. This increasing branch of $ {\mathcal{E}}[\bar C] $ can be interpreted as the energetic cost of enforcing the robustness of the network. Finally, we observe that the Fiedler number $ {\mathfrak{f}}[\bar C] $ of the minimizer is a monotone function of $ \mu $, as claimed by Lemma 4. However, in general it is not strictly monotone.
We consider a network consisting of three nodes $ V = \{0, 1, 2\} $ and three (unoriented) edges $ E = \{(0, 1), (0, 2), (1, 2)\} $ of unit length and conductivities $ C_{01} $, $ C_{02} $ and, respectively, $ C_{12} $. The sources and sinks are given by
$ S_0 = 1,\qquad S_1 = -1/3,\qquad S_2 = -2/3. $ |
Let us denote by $ Q = Q[C] $ the flux from vertex $ 0 $ to vertex $ 1 $. Then the Kirchhoff law (2.2) gives
$ Q = \frac{C_{01}(C_{02}+3C_{12})}{3(C_{01}C_{02}+C_{01}C_{12}+C_{02}C_{12})}. $ | (4.9) |
Moreover, due to the mass conservation in vertex $ 1 $, the flux from $ 0 $ to $ 2 $ equals $ 1-Q $, and the flux from vertex $ 1 $ to $ 2 $ is $ Q-1/3 $.
The energy functional (2.4) with $ \gamma = 1 $ reads
$ {\mathcal{E}}[C] = \frac{Q^2}{C_{01}} + \frac{(1-Q)^2}{C_{02}} + \frac{(Q-1/3)^2}{C_{12}} + \nu \left( C_{01} + C_{02} + C_{12} \right). $ | (4.10) |
A lengthy but simple calculation reveals that the energy $ {\mathcal{E}}[C] $ is globally minimized for $ \bar C_{01} = \frac{1}{3\sqrt\nu} $, $ \bar C_{02} = \frac{2}{3\sqrt\nu} $, and $ \bar C_{12} = 0 $, with $ Q = 1/3 $.
The eigenvalues of the matrix Laplacian of the corresponding weighted adjacency matrix are $ \lambda_0 = 0 $ and
$ λ1=C01+C02+C12−√C201+C202+C212−C01C02−C01C12−C02C12,λ2=C01+C02+C12+√C201+C202+C212−C01C02−C01C12−C02C12. $ |
Consequently, the Fiedler number is $ {\mathfrak{f}}[C] = \lambda_1 $. Since $ {\ell} = 1 $ and $ |V|-1 = 2 $, the functional (4.1) takes the form
$ {\mathcal{F}}[C] = {\mathcal{E}}[C] - \mu \lambda_1, $ | (4.11) |
with $ {\mathcal{E}}[C] $ given by Eq (4.10). We minimize $ {\mathcal{F}}[C] $ with respect to the variables $ C_{01} $, $ C_{02} $, $ C_{12}\geq 0 $, with $ Q = Q[C] $ given by Eq (4.9). The result of this optimization problem with $ \nu: = 1 $, obtained by an application of the MATLAB function ${\tt fminsearch}$, is displayed in Figure 2. In the left panel, we observe that for $ \mu < 1/2 $ we have positive optimal conductivities $ \bar C_{01} \neq \bar C_{02} $, while $ \bar C_{12} = 0 $. In this regime, the optimal $ Q $ remains equal to $ 1/3 $. Consequently, similarly as in the example in Section 4.1, for $ \mu < 1/2 $ the connectivity of the network is not improved by "activation" of the edge $ (1, 2) $. However, the Fiedler number $ {\mathfrak{f}}[\bar C] $ is slowly increasing for $ 0\leq\mu\leq 1/2 $, from $ {\mathfrak{f}}[\bar C] = 1 - \sqrt{1/3} \approx 0.423 $ for $ \mu = 0 $, to $ {\mathfrak{f}}[\bar C] \approx 0.527 $ for $ \mu = 1/2 $. Let us also observe that $ \lambda_1 < \lambda_2 $, i.e., the Fiedler number is a simple eigenvalue of the matrix Laplacian.
On the other hand, for $ \mu > 1/2 $ we have the minimizer $ \bar C_{01} = \bar C_{02} = \bar C_{12} = :c > 0 $, so that the modified energy functional reads
$ {\mathcal{F}}[c] = \frac{Q^2 + (1-Q)^2 + (Q-1/3)^2}{c} + 3 (\nu-\mu) c, $ |
and has the global minimizer $ c = \frac19 \sqrt{ \frac{14}{\nu-\mu} } $, $ Q = 4/9 $. Consequently, for $ \mu > 1/2 $ the network robustness is enforced by "activation" of the edge $ (1, 2) $. It is interesting to observe that, despite the "nonsymmetry" of the values of $ S_i $, all three edges have the same optimal conductivity. Moreover, we obviously have $ \lambda_1 = \lambda_2 $, i.e., the Fiedler number is a double eigenvalue of the matrix Laplacian.
In this section, we present a projected subgradient method [25] for minimization of the modified energy functional $ {\mathcal{F}} = {\mathcal{F}}[C] $ given by (4.1) with $ \gamma = 1 $, constrained by the Kirchhoff law (2.1) and (2.2). The reason for choosing the subgradient method is that the functional is convex (by Lemma 3), and Lipschitz continuous — this follows from Lemma 7 of the Appendix, combined with the classical Hoffman-Wielandt inequality [13] for the Lipschitz continuity of the eigenvalues of a normal matrix. Then, by the Rademacher theorem, the functional is almost everywhere differentiable. In fact, the Fiedler number $ {\mathfrak{f}} = {\mathfrak{f}}[C] $ is differentiable (with respect to elements $ C_{ij} $ of $ C $) in points $ C $ where it is a simple eigenvalue of the matrix Laplacian $ {\mathcal{L}}[C] $, see, e.g., [16,23,28].
On the other hand, if $ {\mathfrak{f}}[C] $ is a multiple eigenvalue of $ {\mathcal{L}}[C] $, then it does not admit a derivative in classical sense. However, it is well-known that the Clarke subdifferential of the function which maps a symmetric matrix to its $ m $-th smallest eigenvalue can be explicitly calculated [26]. The subdifferential is identical for all choices of the index $ m $ corresponding to equal eigenvalues, and, moreover, it coincides with the and Michel-Penot subdifferential [12]. Assuming that $ C\in \mathcal{C} $ represents a connected graph, the Fiedler number $ {\mathfrak{f}}[C] $ is the second smallest eigenvalue of the matrix Laplacian $ {\mathcal{L}}[C] $, and we are interested in calculating (any element of) its subdifferential with respect to the symmetric matrix $ C $. As the mapping $ C\mapsto {\mathcal{L}}[C] $ is analytic, we make use of the result provided in [29].
Lemma 5. Let $ C\in \mathcal{C} $ represent a connected graph and let $ {\mathfrak{f}}[C] $ be its Fiedler number of multiplicity $ r\geq 1 $, i.e., the $ r $-fold eigenvalue of the matrix Laplacian $ {\mathcal{L}}[C] $. Let $ \Theta\subset{\mathbb{R}}^{|V|\times|V|} $ be the (Clarke) subdifferential of $ {\mathfrak{f}}[C] $ at $ C $ and let the unit vector $ v\in{\mathbb{R}}^{|V|} $ be any element of the eigenspace of $ {\mathcal{L}}[C] $ corresponding to $ {\mathfrak{f}}[C] $. Then, denoting $ \mathcal{V}[v]_{ij} : = (v_i-v_j)^2 $ for all $ i, j\in V $, we have
$ \mathcal{V}[v] \in \Theta. $ | (5.1) |
Proof. We apply [29, Theorem 3.7] — in particular, formula (3.13), adapted to our notation, reads
$ \Theta = \mbox{co} \left\{ A\in{\mathbb{R}}^{|V|\times |V|};\, A_{ij} = u^T W^T \frac{\partial {{\mathcal{L}}[C]}}{\partial {C_{ij}}} W u\, \mbox{ for all } i,j\in V,\, u\in{\mathbb{R}}^r, |u| = 1 \right\}. $ |
Here $ W\in{\mathbb{R}}^{|V|\times r} $ denotes the matrix of column orthonormal basis vectors of the eigenspace of $ {\mathcal{L}}[C] $ corresponding to $ {\mathfrak{f}}[C] $. Without loss of generality, we choose $ v $ to be its first column. Moreover, in the partial derivative $ \frac{\partial {{\mathcal{L}}[C]}}{\partial {C_{ij}}}\in{\mathbb{R}}^{|V|\times|V|} $, the symmetry of $ C $ is taken into account, i.e., for $ i\neq j $ it quantifies the sensitivity of $ {\mathcal{L}}[C] $ with respect to changes in both $ C_{ij} $ and $ C_{ji} $. A trivial calculation gives then, for $ i\neq j $ and $ k\neq m $,
$ \frac{\partial {{\mathcal{L}}_{km}}}{\partial {C_{ij}}} = -\delta_{k,i}\delta_{m,j} -\delta_{k,j}\delta_{m,i}, \qquad \frac{\partial {{\mathcal{L}}_{kk}}}{\partial {C_{ij}}} = \delta_{k,i} + \delta_{k,j}, $ |
where we use the shorthand notation $ {\mathcal{L}}_{km} $ for the $ (k, m) $-element of $ {\mathcal{L}}[C] $. As $ {\mathcal{L}}[C] $ does not depend on the diagonal elements of $ C $, we have $ \frac{\partial {{\mathcal{L}}[C]}}{\partial {C_{ii}}} = 0 $ for all $ i\in V $. We then easily calculate, for all $ i, j\in V $,
$ A_{ij} = \sum\limits_{s = 1}^r \sum\limits_{\sigma = 1}^r \Bigl( W_{is} W_{i\sigma} + W_{js}W_{j\sigma} - W_{is}W_{j\sigma} - W_{js}W_{i\sigma} \Bigr) u_s u_\sigma. $ |
Finally, choosing $ u: = (1, 0, \dots, 0)^T $, we get $ A_{ij} = v_i^2 + v_j^2 - 2v_iv_j $ and we conclude that $ \mathcal{V}[v] \in \Theta. $
Obviously, Lemma 5 does apply also to the case when $ {\mathfrak{f}}[C] $ is a simple eigenvalue of $ {\mathcal{L}}[C] $. The Fiedler number is then differentiable in the classical sense and we have
$ \frac{\partial {{\mathfrak{f}}[C]}}{\partial {C_{ij}}} = (v_i - v_j)^2 \qquad\mbox{for all } i,j\in V, $ |
where $ v\in{\mathbb{R}}^{|V|} $ is the corresponding normalized eigenvector.
We now collected all of the ingredients needed for establishing the projected subgradient method for minimization of the functional (4.1). We initialize the method by choosing $ C^{(0)}\in \mathcal{C} $ such that the Kirchhoff law (2.2) with $ C^{(0)} $ admits a solution. In our practical realization, we draw the values for the elements $ C^{(0)}_{ij} $, $ (i, j)\in E $, randomly from the uniform distribution on the interval $ (0, 1) $. The elements $ C^{(0)}_{ij} $ for $ (i, j)\notin E $ are all set to zero. Then, we fix some $ K\in{\mathbb{N}} $ and for $ k = 0, \dots, K $, we perform the subgradient step
$ C^{(k+1/2)}_{ij} = C_{ij}^{(k)} + \tau_k \left(\frac{\left(P^{(k)}_i - P^{(k)}_j\right)^2}{L_{ij}} - \nu L_{ij} + \mu\,{\ell} \, \frac{|V|-1}{2} \, \left(v^{(k)}_i - v^{(k)}_j\right)^2 \right) $ | (5.2) |
for all $ (i, j)\in E $. Here we used the formula (A.11) for the derivative of the kinetic term of the energy, and $ P^{(k)} $ denotes any solution of the Kirchhoff law (2.1)–(2.2) with $ C: = C^{(k)} $. Moreover, $ v^{(k)} $ is any normalized eigenvector of the matrix Laplacian $ {\mathcal{L}}[C^{(k)}] $. We use the diminishing step size $ \tau_k > 0 $, in particular, $ \tau_k : = \tau_0 / \sqrt{k} $ for some $ \tau_0 > 0 $. The subgradient step is followed by the projection step
$ C^{(k+1)} = \mathbb{P}_{ \mathcal{C}}\left[ C^{(k+1/2)} \right]. $ | (5.3) |
Here $ \mathbb{P}_{ \mathcal{C}}: {\mathbb{R}}^{|V|\times |V|}_{\mathrm{sym}} \to \mathcal{C} $ denotes the projection onto the set $ \mathcal{C} $ and is realized by simply trimming the negative elements to zero,
$ \mathbb{P}_{ \mathcal{C}}[C]_{ij} : = \max\{C_{ij},0\} \qquad\mbox{for all } i,j\in V. $ |
After performing the projection step, we check whether $ C^{(k+1)} $ remained in the domain of $ {\mathcal{F}} = {\mathcal{F}}[C] $, which is equivalent to the solvability of the Kirchhoff law (2.1) and (2.2) with $ C: = C^{(k+1)} $. Obviously, due to continuity, $ {\mathcal{F}}[C^{(k)}] < +\infty $ implies the same for $ C^{(k+1)} $ for small enough step size $ \tau_k > 0 $. In practical numerical realization of the method, we interrupt the calculation if we detect that the linear system (2.1) and (2.2) becomes badly conditioned, and restart with a reduced $ \tau_0 > 0 $.
The subgradient method is not a descent method, i.e., it is not guaranteed that $ {\mathcal{F}}[C^{(k+1)}] \leq {\mathcal{F}}[C^{(k)}] $. Therefore, we set
$ \widetilde C^{(K)} : = \mbox{argmin}_{k = 0,\dots,K} {\mathcal{F}}[C^{(k)}]. $ |
Then, due to the convexity and Lipschitz continuity of the functional $ {\mathcal{F}} = {\mathcal{F}}[C] $ on its domain, the standard theory [25] provides convergence of the sequence $ (\widetilde C^{(K)})_{K > 0} $ to a global minimizer. Its existence follows from the coercivity of $ {\mathcal{F}} = {\mathcal{F}}[C] $, Lemma 2, as long as $ \mu < \nu $.
We apply the projected subgradient method (5.2) and (5.3) to search for optimal transportation networks in two cases: a complete graph with $ 7 $ nodes, and a leaf-shaped graph with 122 nodes and 323 edges. We study how the resulting optimal networks depend on the value of the parameter $ \mu\geq 0 $. In particular, we are interested in the number of active edges (i.e., the number of nonzero elements $ C_{ij} $), the multiplicity of the Fiedler number, and the values of the kinetic, metabolic, and modified energies.
We applied the projected subgradient method (5.2) and (5.3) for minimization of the functional (4.1) with $ \gamma: = 1 $ and $ \nu: = 1 $, on a complete graph $ G = (V, E) $ with $ 7 $ nodes located in $ {\mathbb{R}}^2 $,
$ V = \{1,2,\ldots, 7\}, \qquad E = \{ (i,j) = (j,i);\; i, j\in V \}. $ |
The node locations and the source/sink intensities are specified in Table 1.
node | $ x $ | $ y $ | $ S $ |
1 | 0.100 | 0.000 | 0.164 |
2 | 0.874 | –0.104 | 0.794 |
3 | 0.581 | 0.790 | –0.128 |
4 | –0.043 | 1.0547 | 0.936 |
5 | –0.945 | 0.371 | –0.299 |
6 | –0.818 | –0.161 | 0.750 |
7 | –0.206 | –1.023 | –2.217 |
We chose $ \tau_0: = 10^{-1} $ and $ K: = 10^6 $. To check for consistency of the method, we ran the calculation multiple times for every fixed value of $ \mu\geq 0 $, each time with a different initial $ C^{(0)}\in \mathcal{C} $, with elements drawn from the uniform random distribution on $ (0, 1) $. In all of these runs (with a fixed value of $ \mu\geq 0 $), the method converged to the same minimizer, up to a relative error of the order $ 10^{-12} $.
The graphs corresponding to the minimizers for $ \mu\in\{0, 0.2, 0.4, 0.6, 0.8, 1.0\} $ are plotted in Figure 3, where the thickness of the line segments is proportional to the square root of the conductivity $ C_{ij}\geq 0 $ of the corresponding edge. Edges with $ C_{ij} = 0 $ are excluded from the plot. We observe that for $ \mu = 0 $, the optimal transportation structure is loop-free, which corresponds to the result of Theorem 2. For $ \mu = 0.2 $, one loop is present, consisting of nodes $ \{1, 4, 7\} $. With an increasing value of $ \mu $, the graph successively becomes denser.
Statistical properties of the optimal graphs in dependence on the value of $ \mu\in[0, 1] $ are plotted in Figure 4. In the top left panel, we plot the values of the energy $ {\mathcal{E}} = {\mathcal{E}}[C] $ given by (2.4) and the modified energy $ {\mathcal{F}} = {\mathcal{F}}[C] $ given by (4.1). We observe that the value of $ {\mathcal{E}}[C] $ increases with increasing $ \mu\geq 0 $. This can be seen as the extra energy expenditure for securing the robustness of the network. In the top right panel, we plot the kinetic (star-shaped markers) and metabolic (circular markers) energies of the optimal networks, defined in (2.5). The kinetic energy decreases with increasing $ \mu $, which is due to the fact that less pumping power is necessary if the network consists of more edges, or edges with higher conductivities. However, this means that the "material expense" is higher, i.e., the metabolic energy increases. In the bottom left panel, we plot the smallest three nonzero eigenvalues (i.e., the second, third and fourth smallest) of the matrix Laplacian $ {\mathcal{L}}[C] $. This is to understand the multiplicity of the Fiedler number $ {\mathfrak{f}} = {\mathfrak{f}}[C] $. We observe that the Fiedler number is simple for $ \mu\lesssim 0.6 $, and turns to double for $ \mu\gtrsim 0.6 $. Finally, in the bottom left panel of Figure 4, the number of positive $ C_{ij} $ (i.e., the number of edges present in the graph) is plotted. We again notice the monotonicity, i.e., increasing the value of $ \mu $ indeed leads to the graph becoming denser.
Inspired by the possible application of the model to simulate leaf venation patterns, we generated a planar graph in the form of a triangulation of leaf-shaped domain with $ |V| = 122 $ nodes and $ 323 $ edges, see Figure 5. We prescribed a single source, $ S_i = 1 $, for the left-most node ("stem" of the leaf), and $ S_j = -(|V|-1)^{-1} $ for all other nodes.
We again applied the projected subgradient method (5.2) and (5.3) for minimization of the functional (4.1) with $ \gamma: = 1 $ and $ \nu: = 1 $. We chose $ \tau_0: = 10^{-1} $ and $ K: = 10^7 $, and $ \mu\in [0, 5] $. Although Lemma 2 guarantees boundedness from below of $ {\mathcal{F}} = {\mathcal{F}}[C] $ only for $ \mu\leq 1 $, this condition is sufficient, but not necessary. As $ {\mathcal{F}} = {\mathcal{F}}[C] $ is obviously bounded from below on every compact subset of $ \mathcal{C} $, the projected subgradient method would indicate a possible unboundedness from below by divergence of the sequence of iterates. This, however, did not happen for $ \mu\in [0, 5] $.
The optimal graphs found for $ \mu\in\{0, 1, 2, 3, 4, 5\} $ are plotted in the correspondingly labeled panels of Figure 6. Again, the thickness of the line segments is proportional to the square root of the conductivity of the corresponding edge. Edges with $ C_{ij} = 0 $ are excluded from the plot.
Statistical properties of the graphs in dependence on the value of $ \mu\in [0, 5] $ are plotted in Figure 7. Here, in the bottom left panel, it is interesting to observe that the Fiedler number seems to be a double eigenvalue for $ \mu\in\{2.5, 2.75\} $, and for $ \mu\gtrsim 0.4 $. In the bottom right panel, we see that the number of active edges is not monotonically increasing with $ \mu $. Indeed, around the value of $ \mu\simeq 2.5 $, the number of nonzero elements of $ \widetilde C $ drops slightly. We hypothesize that this "anomaly" may be related to the multiplicity of the Fiedler number observed in the left bottom panel around the same value of $ \mu $. However, we are not able to provide a rigorous explanation.
Both authors contributed equally to this work.
The authors declare they have not used Artificial Intelligence (AI) tools in the creation of this article.
The authors declare there is no conflict of interest.
Here we prove some mathematical properties of the functional $ {\mathcal{E}} = {\mathcal{E}}[C] $ and the set $ {\mathcal C} $, defined by Eqs (2.4) and (2.6), respectively.
Lemma 6. For any fixed $ S\in{\mathbb{R}}^{|V|}_0 $, the set
$ {C∈C:E[C]<+∞} $ | (A.1) |
is an open convex subset of $ \mathcal{C} $.
Proof. We start by showing that (A.1) is convex. For that sake, we take $ C^1, C^2\in \mathcal{C} $ with $ {\mathcal{E}}[C^1] < +\infty $, and $ {\mathcal{E}}[C^2] < +\infty $. Let $ 0 < \alpha < 1 $ and put $ C: = \alpha C^1+(1-\alpha)C^2. $ We want to show that also $ {\mathcal{E}}[C] < +\infty $, i.e., that (2.2) is solvable with $ C $.
Let $ x\in{\mathbb{R}}^{|V|} $ be fixed. Then
$ xTL[C]x=αxTL[C1]x+(1−α)xTL[C2]x. $ | (A.2) |
Since all of the three Laplacian matrices appearing in the above identity are positive semidefinite, we observe that the left-hand side of (A.2) is zero if, and only if, both terms on the right-hand side of (A.2) vanish. Therefore, the null space of $ {\mathcal{L}}[C] $ is the intersection of the null spaces of $ {\mathcal{L}}[C^1] $ and $ {\mathcal{L}}[C^2] $ and the range space of $ {\mathcal{L}}[C] $ is the sum of the range spaces of $ {\mathcal{L}}[C^1] $ and $ {\mathcal{L}}[C^2]. $ We conclude that $ S $ is in the range space of $ {\mathcal{L}}[C] $ and (2.2) is solvable for $ C $.
Next, we show that (A.1) is open in $ \mathcal{C} $. Let $ C\in \mathcal{C} $ with $ {\mathcal{E}}[C] < +\infty $ and let $ \Gamma_1, \dots, \Gamma_n\subset V $ be the connected components of $ C $. Then (2.2) is solvable if, and only if, $ \sum_{j\in\Gamma_k}S_j = 0 $ for every $ 1\le k\le n. $ Let us define
$ \mathcal{U}[C] : = \left\{ C'\in \mathcal{C}; \, C_{ij} > 0 \implies C'_{ij} > 0 \mbox{ for all } (i,j)\in E \right\}. $ |
Obviously, any $ C'\in \mathcal{U}[C] $ has either the same connected components as $ C $, or fewer but larger components, stemming from establishing new connections among $ \Gamma_1, \dots, \Gamma_n\subset V $. In both cases, (2.2) is solvable with $ C' $ and, therefore, $ {\mathcal{E}}[C'] $ is also finite. Consequently, $ \mathcal{U}[C] $ is an open neighborhood of $ C $ in the set $ \{C\in \mathcal{C}:{\mathcal{E}}[C] < +\infty\} $.
In the sequel, we shall denote, for any matrix $ A\in{\mathbb{R}}^{|V|\times |V|} $ and any vector $ x\in{\mathbb{R}}^{|V|} $,
$ \left\| {A} \right\|_\infty : = \max\limits_{i, j \in V} |A_{ij}|, \qquad |x|_\infty : = \max\limits_{i \in V} |x_{i}|. $ |
Moreover, we recall that $ {\mathbb{R}}^{|V|}_0 $ was defined in Eq (2.3) as the set of vectors $ x\in{\mathbb{R}}^{|V|} $ with a vanishing sum.
Next, we study the continuity of the pressures $ P = P[C] $, the fluxes $ Q = Q[C] $, and the energy $ {\mathcal{E}} = {\mathcal{E}}[C] $ as functions of $ C $. We restrict ourselves to the case when the Kirchhoff law (2.2) has a unique solution $ P = P[C]\in{\mathbb{R}}^{|V|}_0 $. This happens if, and only if, the matrix Laplacian $ {\mathcal{L}}[C] $ of $ C $ is nonsingular on $ {\mathbb{R}}_0^{|V|} $, which in turn holds if, and only if, $ C $ represents a weighted connected graph.
Example 2. If $ C $ represents an unconnected graph, then (2.2) might admit no solution at all or the solution might not be unique in $ {\mathbb{R}}^{|V|}_0 $. In both cases, we might interpret $ P = P[C] $ as a set-valued function of $ C $, cf. [3]. Unfortunately, the following example shows that $ P $ does not need to be continuous.
Let $ G = (V, E) $ be a complete graph on four vertices $ V = \{1, 2, 3, 4\} $. Let $ L_{ij} = 1 $ for $ i\not = j $, and let $ S = (1, -1, 1, -1) $. Furthermore, set
$ C^0: = \left(0100100000010010\right). $ |
Then the graph represented by $ C^0 $ is disconnected with two components $ U_1 = \{1, 2\} $ and $ U_2 = \{3, 4\} $. Since $ S_1+S_2 = S_3+S_4 = 0 $, the equation $ {\mathcal{L}}[C^0]P = S $ has non-unique solutions, namely every $ P\in\{(a, a-1, b, b-1), a, b\in{\mathbb{R}}\} $. Even when we restrict $ P $ to be an element of $ {\mathbb{R}}_0^4 $, we still have the non-unique solutions $ P = (a, a-1, 1-a, -a) $ with $ a\in{\mathbb{R}} $.
For $ t > 0 $, we define $ C^t $ by adding to $ C^0 $ the edge (2, 3) with weight $ t $. Namely, we define $ C^t_{2, 3} = C^t_{3, 2} = t $ and $ C^t_{ij} = C^0_{ij} $ for $ \{i, j\}\not = \{2, 3\}. $ Then $ {\mathcal{L}}[C^t]P = S $ for every $ P = (c+1, c, c, c-1) $, $ c\in{\mathbb{R}} $ and in $ {\mathbb{R}}_0^4 $, the solution is unique, namely $ P = (1, 0, 0, -1) $.
On the other hand, if $ t < 0 $, we define $ C^t $ by adding an edge (1, 4) with a weight $ -t > 0 $ to $ C^0 $. Then $ {\mathcal{L}}[C^t]P = S $ for every $ P = (c, c-1, c+1, c) $, $ c\in{\mathbb{R}} $. Again, in $ {\mathbb{R}}_0^4 $, the solution is unique, namely $ P = (0, -1, 1, 0) $. Using the notions of set-valued analysis [3], we observe that the mapping $ t\to \{P\in{\mathbb{R}}_0^4:{\mathcal{L}}[C^t]P = 0\} $ is upper semi-continuous but not lower semi-continuous in $ t = 0 $.
We formulate the next result as a general statement about matrix Laplacian $ {\mathcal{L}}[K] $ of a symmetric matrix $ K $ with non-negative entries. Later on, we apply this result to the matrix $ K_{ij} = C_{ij}/L_{ij} $, $ (i, j)\in E $.
Lemma 7. Let $ S\in {\mathbb{R}}^{|V|}_0 $, let $ K^\ast\in {\mathbb{R}}_+^{|V|\times|V|} $ be a symmetric matrix with non-negative entries representing a connected weighted undirected graph, and let $ P^\ast\in {\mathbb{R}}^{|V|}_0 $ be the unique solution of the linear system
$ {\mathcal{L}}(K^\ast) P^\ast = S $ | (A.3) |
Then there exists a small neighborhood of $ K^\ast $ (relative to $ {\mathbb{R}}_+^{|V|\times|V|} $) on which (A.3) is uniquely solvable and defines $ P = P(K) $ as a differentiable, Lipschitz-continuous function of $ K $. Moreover, the fluxes $ Q_{ij} = K_{ij} (P_j - P_i) $ are Lipschitz-continuous functions of $ K $, as well.
Remark 4. To be more specific, Lemma 7 ensures that to a given $ K^\ast\in {\mathbb{R}}_+^{|V|\times|V|} $, there exists $ \rho > 0 $, which in general depends on $ K^\ast $, such that if $ K\in {\mathbb{R}}_+^{|V|\times|V|} $ verifies $ \left\| {K - K^\ast} \right\|_\infty \leq \rho, $ then there exists a unique solution $ P\in {\mathbb{R}}^{|V|}_0 $ of the linear system $ {\mathcal{L}}(K) P = S. $ Moreover, there exists a constant $ c $, independent of $ \rho $, such that the fluxes
$ Q_{ij} : = K_{ij} (P_j - P_i), \qquad Q_{ij}^\ast : = K_{ij}^\ast (P_j^\ast - P_i^\ast) $ |
satisfy
$ \left\| {Q-Q^\ast} \right\|_\infty \leq c \left\| {K - K^\ast} \right\|_\infty. $ | (A.4) |
Proof. Since $ {\mathcal{L}}(K^\ast) $ is a nonsingular operator on the space $ {\mathbb{R}}^{|V|}_0 $, a classical perturbation result implies unique solvability of
$ {\mathcal{L}}(K) P = S $ | (A.5) |
on $ {\mathbb{R}}^{|V|}_0 $ for $ K\in{\mathbb{R}}_+^{|V|\times|V|} $ with
$ \left\| {K - K^\ast} \right\|_\infty \leq \rho $ | (A.6) |
and $ \rho $ small enough. Let us denote $ \delta K : = K-K^\ast $ and $ \delta {\mathcal{L}} : = {\mathcal{L}}(K)-{\mathcal{L}}(K^\ast) $. Note that $ \delta{\mathcal{L}} $ depends linearly on $ \delta K $, $ \left\| {\delta {\mathcal{L}}} \right\|_\infty \leq (|V|-1)\left\| {\delta K} \right\|_\infty $, and that $ \delta{\mathcal{L}} $ maps $ {\mathbb{R}}_0^{|V|} $ into itself.
In particular, denoting $ \delta P : = P-P^\ast $, we have
$ ({\mathcal{L}}(K^\ast) + \delta {\mathcal{L}})(P^\ast + \delta P) = S. $ |
Using (A.3), applying $ {\mathcal{L}}(K^\ast)^{-1} $ to both sides of this equation and adding $ P^\ast $ gives
$ P^\ast+({\mathcal{L}}(K^\ast))^{-1} \delta {\mathcal{L}} P^\ast + \delta P+({\mathcal{L}}(K^\ast))^{-1} \delta {\mathcal{L}}\delta P = P^\ast, $ |
which can be reformulated as
$ \delta P = \left[ \left( I + ({\mathcal{L}}(K^\ast))^{-1} \delta {\mathcal{L}} \right)^{-1} - I \right] P^\ast. $ | (A.7) |
Expanding the expression $ \left(I + ({\mathcal{L}}(K^\ast))^{-1} \delta {\mathcal{L}} \right)^{-1} $ into the von Neumann series, we have
$ δP=∞∑k=1[−(L(K∗))−1δL]kP∗. $ | (A.8) |
As $ \delta{\mathcal{L}} $ depends linearly on $ \delta K $, it follows that $ P = P(K) $ is differentiable at $ K^\ast. $
In the next step, we denote by $ \left\| {\cdot} \right\| $ the operator norm induced by the vector norm $ |\cdot|_\infty $ on the space $ {\mathbb{R}}^{|V|}_0 $. By norm equivalence, there exists a constant $ c > 0 $ such that $ \left\| {\cdot} \right\| \leq c \left\| {\cdot} \right\|_\infty $ and we recall that $ \left\| {\delta {\mathcal{L}}} \right\|_\infty \leq (|V|-1)\left\| {\delta K} \right\|_\infty $. Combining this with (A.8), we obtain
$ |δP|∞≤∞∑k=1‖(L(K∗))−1δL‖k|P∗|∞≤|P∗|∞⋅∞∑k=1‖(L(K∗))−1‖k⋅‖δL‖k≤|P∗|∞⋅∞∑k=1[c(|V|−1)‖(L(K∗))−1‖]k‖δK‖k∞=|P∗|∞⋅Λ‖δK‖∞1−Λ‖δK‖∞, $ | (A.9) |
where we denoted $ \Lambda : = c \left\| {({\mathcal{L}}(K^\ast))^{-1}} \right\| (N-1) $ and assumed that $ \left\| {\delta K} \right\|_\infty < \Lambda^{-1} $.
Now, denoting $ (\Delta P)_{ij} : = P_j - P_i $ and analogously for $ (\Delta P^\ast)_{ij} $, we have for every $ i, j \in V $,
$ |Qij−Q∗ij|=|Kij(ΔP)ij−K∗ij(ΔP∗)ij|≤|(Kij−K∗ij)((ΔP)ij−(ΔP∗)ij)|+|(Kij−K∗ij)(ΔP∗)ij)|+|K∗ij((ΔP)ij−(ΔP∗)ij)|≤2‖δK‖∞|δP|∞+‖δK‖∞‖ΔP∗‖∞+2‖K∗‖∞|δP|∞,$ |
where we used the estimate $ | (\Delta P)_{ij} - (\Delta P^\ast)_{ij} | \leq 2 |\delta P|_\infty $. Therefore, using (A.6) and (A.9), we estimate
$ \bigl| Q_{ij} - Q_{ij}^\ast \bigr| \leq 2 \left(\left\| {\delta K} \right\|_\infty + \left\| {K^\ast} \right\|_\infty \right) \frac{\Lambda \left\| {\delta K} \right\|_\infty}{1 - \Lambda \left\| {\delta K} \right\|_\infty} + \left\| {\delta K} \right\|_\infty \left\| {\Delta P^\ast} \right\|_\infty. $ |
Finally, we choose $ \rho : = 1/(2\Lambda) $, so that for $ \left\| {\delta K} \right\|_\infty \leq \rho $, we have
$ \frac{1}{1 - \Lambda \left\| {\delta K} \right\|_\infty} \leq 2, \qquad \Lambda \left\| {\delta K} \right\|_\infty^2 \leq \frac12 \left\| {\delta K} \right\|_\infty, $ |
which gives
$ \bigl| Q_{ij} - Q_{ij}^\ast \bigr| \leq \left( 2 + 4 \Lambda \left\| {K^\ast} \right\|_\infty + \left\| {\Delta P^\ast} \right\|_\infty \right) \left\| {\delta K} \right\|_\infty $ |
and an obvious choice of the constant $ c $ concludes (A.4).
Corollary 1. The functional $ {\mathcal{E}} = {\mathcal{E}}[C] $ is continuous on $ \mathcal{C} $ and totally differentiable on the set $ \{C\in \mathcal{C}: C\ \mathit{\text{corresponds to a connected graph}}\}\subset \mathcal{C} $.
Proof. By Lemma 7, $ {\mathcal{E}} $ is continuous in $ C $ whenever $ {\mathcal{E}}[C] < +\infty $. It is therefore enough to show that $ {\mathcal{E}}(C) $ is large on a small neighborhood of a given $ C^\ast $ with $ {\mathcal{E}}(C^\ast) = +\infty. $ In that case, (2.2) is not solvable and, therefore, $ C^\ast $ represents a disconnected graph. Let us assume for simplicity, that it has only two connected components $ \Gamma_1, \Gamma_2 \subset V $ and define $ A: = \sum_{j\in \Gamma_1}S_j = -\sum_{j\in \Gamma_2}S_j > 0 $. If (2.2) is solvable for $ C\in \mathcal{C} $ with $ \|C-C^\ast\|_\infty < {\varepsilon} $, then there must be some new edges between $ \Gamma_1 $ and $ \Gamma_2 $, with weights bounded by $ {\varepsilon} $, that transfer the mass $ A $ from $ \Gamma_1 $ to $ \Gamma_2 $. Therefore, $ {{\mathcal{E}}_{\mathrm{kin}}}(C)\ge \min_{(i, j)\in E} A^2 L_{i, j}/({\varepsilon} |V|^2) $, which tends to infinity as $ {\varepsilon}\to 0 $. We thus conclude that $ {\mathcal{E}} = {\mathcal{E}}[C] $ is continuous on $ \mathcal{C} $.
The statement about total differentiability of $ {\mathcal{E}} = {\mathcal{E}}[C] $ follows directly from Lemma 7.
Finally, we give a proof of the convexity of the energy functional $ {\mathcal{E}} = {\mathcal{E}}[C] $ for $ \gamma\geq 1 $. It follows directly from the convexity of the pumping (kinetic) part of the energy $ {{\mathcal{E}}_{\mathrm{kin}}}[C] $, coupled to the Kirchhoff law (2.1) and (2.2). If (2.2) is not solvable, we again set $ {{\mathcal{E}}_{\mathrm{kin}}}[C]: = +\infty. $
Lemma 8. The pumping energy $ {{\mathcal{E}}_{\mathrm{kin}}}[C] $ defined in (2.5), constrained by the Kirchhoff law (2.1) and (2.2), is a convex functional on the set $ \mathcal{C} $.
Proof. By Lemma 6, the set $ \{C\in \mathcal{C}:{\mathcal{E}}(C) < +\infty\} $ is an open, convex subset of $ \mathcal{C} $ and the same obviously holds true with $ {\mathcal{E}} $ replaced by $ {{\mathcal{E}}_{\mathrm{kin}}} $. To show the convexity of $ {{\mathcal{E}}_{\mathrm{kin}}} $, it is therefore enough to prove that the Hessian matrix of $ {{\mathcal{E}}_{\mathrm{kin}}}[C] $ is positive semidefinite for every $ C\in \mathcal{C} $ with $ {{\mathcal{E}}_{\mathrm{kin}}}[C] < +\infty. $
We use (2.1) to express the kinetic energy as
$ {{\mathcal{E}}_{\mathrm{kin}}}[C] = \sum\limits_{(i,j)\in E} C_{ij} \frac{(P_j-P_i)^2}{L_{ij}}. $ | (A.10) |
We note that, by Lemma 7, $ P $ restricted by the condition $ \sum_{j\in V} P_j = 0 $ is a differentiable function of $ C $. The first-order derivative of $ {{\mathcal{E}}_{\mathrm{kin}}}[C] $ with respect to the element $ C_{km} $ reads, see [9, Lemma 2.1],
$ \frac{\partial {}}{\partial {C_{km}}} {{\mathcal{E}}_{\mathrm{kin}}}[C] = - \frac{(P_k - P_m)^2}{L_{km}}. $ | (A.11) |
Consequently, the second-order derivative with respect to the elements $ C_{km} $ and $ C_{\alpha\beta} $ reads
$ \frac{\partial^2 {}}{\partial {C_{km}} \partial {C_{\alpha\beta}}} {{\mathcal{E}}_{\mathrm{kin}}}[C] = - \frac{1}{L_{km}} \frac{\partial {}}{\partial {C_{\alpha\beta}}} (P_k - P_m)^2. $ |
We fix a vector $ \varphi\in{\mathbb{R}}^{|V|} $, multiply the Kirchhoff law (2.2) by $ \varphi_i $, and sum over $ i\in V $, using the standard symmetrization trick on the left-hand side,
$ \frac12 \sum\limits_{i\in V} \sum\limits_{j\in V} \frac{ C_{ij}}{L_{ij}} (P_i-P_j) (\varphi_i-\varphi_j) = \sum\limits_{i\in V} S_i \varphi_i. $ |
We take a derivative of the above identity with respect to $ C_{km} $,
$ \frac{1}{L_{km}} (P_k-P_m)(\varphi_k-\varphi_m) + \frac12 \sum\limits_{i\in V} \sum\limits_{j\in V} \frac{ C_{ij}}{L_{ij}} (\varphi_i-\varphi_j) \frac{\partial {}}{\partial {C_{km}}} (P_i-P_j) = 0, $ |
where we took into account the symmetry $ C_{km} = C_{m k} $. We now choose $ \varphi : = \frac{\partial {}}{\partial {C_{\alpha\beta}}} P $, which gives
$ \frac{1}{L_{km}} (P_k-P_m) \frac{\partial {}}{\partial {C_{\alpha\beta}}} (P_k-P_m) + \frac12 \sum\limits_{i\in V} \sum\limits_{j\in V} \frac{ C_{ij}}{L_{ij}} \frac{\partial {}}{\partial {C_{\alpha\beta}}} (P_i-P_j) \frac{\partial {}}{\partial {C_{km}}} (P_i-P_j) = 0. $ |
Consequently,
$ \frac{1}{L_{km}} \frac{\partial {}}{\partial {C_{\alpha\beta}}} (P_k-P_m)^2 = - \sum\limits_{i\in V} \sum\limits_{j\in V} \frac{ C_{ij}}{L_{ij}} \frac{\partial {}}{\partial {C_{\alpha\beta}}} (P_i-P_j) \frac{\partial {}}{\partial {C_{km}}} (P_i-P_j) $ |
and
$ \frac{\partial^2 {}}{\partial {C_{km}} \partial {C_{\alpha\beta}}} {{\mathcal{E}}_{\mathrm{kin}}}[C] = \sum\limits_{i\in V} \sum\limits_{j\in V} \frac{ C_{ij}}{L_{ij}} \frac{\partial {}}{\partial {C_{\alpha\beta}}} (P_i-P_j) \frac{\partial {}}{\partial {C_{km}}} (P_i-P_j). $ |
Now, fix any $ \xi\in {\mathbb{R}}^{|V|\times|V|} $ and denote
$ \Xi_{ij} : = \sum\limits_{k\in V} \sum\limits_{m\in V} \xi_{km} \frac{\partial {}}{\partial {C_{km}}} (P_i-P_j). $ |
Then we have
$ \sum\limits_{k, m} \sum\limits_{\alpha,\beta} \left( \frac{\partial^2 {{{\mathcal{E}}_{\mathrm{kin}}}[C] }}{\partial {C_{km}} \partial {C_{\alpha\beta}}} \right) \xi_{km} \xi_{\alpha\beta} = \sum\limits_{i\in V} \sum\limits_{j\in V} \frac{ C_{ij}}{L_{ij}} \Xi_{ij}^2 \geq 0, $ |
where the nonnegativity follows from $ C_{ij}\geq 0 $. We conclude that the Hessian matrix of $ {{\mathcal{E}}_{\mathrm{kin}}}[C] $ is positive semidefinite for every $ C\in \mathcal{C} $ with $ {{\mathcal{E}}_{\mathrm{kin}}}[C] < +\infty $, and, therefore, $ {{\mathcal{E}}_{\mathrm{kin}}} $ is convex in $ \mathcal{C} $.
[1] |
Contestabile A (2002) Cerebellar granule cells as a model to study mechanisms of neuronal apoptosis or survival in vivo and in vitro. The Cerebellum 1: 41-55. doi: 10.1080/147342202753203087
![]() |
[2] | Gallo V, Kingsbury A, Balazs R, et al. (1987) The role of depolarization in the survival and differentiation of cerebellar granule cells in culture. J Neurosci 7: 2203-2213. |
[3] |
Balazs R, Gallo V, Kingsbury A (1988) Effect of depolarization on the maturation of cerebellar granule cells in culture. Devel Brain Res 40: 269-276. doi: 10.1016/0165-3806(88)90139-3
![]() |
[4] |
D’Mello SR, Galli C, Ciotti T, et al. (1993) Induction of apoptosis in cerebellar granule neurons by low potassium: inhibition of death by insulin-like growth factor I and cAMP. Proc Natl Acad Sci USA 90: 10989-10993. doi: 10.1073/pnas.90.23.10989
![]() |
[5] | Copani A, Bruno VMG, Barresi V, et al. (1994) Activation of metabotropic glutamate receptors prevents neuronal apoptosis in culture. J Neurochem 64: 101-108. |
[6] |
Ciani E, Rizzi S, Paulsen RE, et al. (1997) Chronic pre-explant blockade of the NMDA receptor affects survival of cerebellar granule cells explanted in vitro. Devel Brain Res 99: 112-117. doi: 10.1016/S0165-3806(96)00187-3
![]() |
[7] | Sparapani M, Virgili M, Bardi G (1998) Ornithine decarboxylase activity during development of cerebellar granule neurons. J Neurochem 71: 1898-1904. |
[8] | Martin-Romero FJ, Garcia-Martin E, Gutierrez-Merino C (1996) Involvement of free radicals in signaling of low-potassium induced apoptosis in cultured cerebellar granule cells. Int J Dev Biol Suppl.1: 197S-198S. |
[9] |
Martin-Romero FJ, Garcia-Martin E, Gutierrez-Merino C (2002) Inhibition of the oxidative stress produced by plasma membrane NADH oxidase delays low-potassium induced apoptosis of cerebellar granule cells. J Neurochem 82: 705-715. doi: 10.1046/j.1471-4159.2002.01023.x
![]() |
[10] | Nardi N, Avidan G, Daily D, et al. (1997) Biochemical and temporal analysis of events associated with apoptosis induced by lowering the extracellular potassium concentration in mouse cerebellar granule neurons. J Neurochem 68: 750-759. |
[11] | Schulz JB, Beinroth S, Weller M, et al. (1998) Endonucleolytic DNA fragmentation is not required for apoptosis of cultured rat cerebellar granule neurons. Neurosci Lett 27: 9-12. |
[12] | Marks N, Berg MJ, Guidotti A, et al. (1998) Activation of caspase-3 and apoptosis in cerebellar granule cells. J Neurosci Res 52: 334-341. |
[13] |
Allsopp TE, McLuckie J, Kerr LE, et al. (2000) Caspase 6 activity initiates caspase 3 activation in cerebellar granule cell apoptosis. Cell Death Differ 7: 984-993. doi: 10.1038/sj.cdd.4400733
![]() |
[14] | Eldadah BA, Ren RF, Faden AI (2000) Ribozyme-mediated inhibition of caspase-3 protects cerebellar granule cells from apoptosis induced by serum-potassium deprivation. J Neurosci 20: 179-186. |
[15] |
Cowling V, Downward J (2002) Caspase-6 is the direct activator of caspase-8 in the cytochrome c-induced apoptosis pathway: absolute requirement for removal of caspase-6 prodomain. Cell Death Differ 9: 1046-1056. doi: 10.1038/sj.cdd.4401065
![]() |
[16] |
Valencia A, Morán J (2001) Role of oxidative stress in the apoptotic cell death of cultured cerebellar granule neurons. J Neurosci Res 64: 284-297. doi: 10.1002/jnr.1077
![]() |
[17] | Simons M, Beinroth S, Gleichmann M, et al. (1999) Adenovirus-mediated gene transfer of inhibitors of apoptosis protein delays apoptosis in cerebellar granule neurons. J Neurochem 72: 292-301. |
[18] | Wigdal SS, Kirkland RA, Franklin JL, et al. (2002) Cytochrome c release precedes mitochondrial membrane potential loss in cerebellar granule neurons apoptosis: lack of mitochondrial swelling. J Neurochem 82: 1029-1038. |
[19] |
Samhan-Arias AK, Marques-da-Silva D, Yanamala N, et al. (2012) Stimulation and clustering of cytochrome b5 reductase in caveolin-rich lipid microdomains is an early event in oxidative stress-mediated apoptosis of cerebellar granule neurons. J Proteomics 75: 2934-2949. doi: 10.1016/j.jprot.2011.12.007
![]() |
[20] |
Bobba A, Atlante A, Giannattasio S, et al. (1999) Early release and subsequent caspase-mediated degradation of cytochrome c in apoptotic cerebellar granule neurons. FEBS Lett 457: 126-130. doi: 10.1016/S0014-5793(99)01018-2
![]() |
[21] | McGinnis KM, Gnegy ME, Wang KK (1999) Endogenous bax translocation in SH-SY5Y human neuroblastoma cells and cerebellar granule neurons undergoing apoptosis. J Neurochem 72: 1899-1906. |
[22] |
Miller TM, Moulder KL, Knudson CM, et al. (1997) Bax deletion further orders the cell death pathway in cerebellar granule cells and suggests a caspase-independent pathway to cell death. J Cell Biol 139: 205-217. doi: 10.1083/jcb.139.1.205
![]() |
[23] | Cregan SP, MacLaurin JG, Craig CG, et al. (1999) Bax-dependent caspase-3 activation is a key determinant in p53-induced apoptosis in neurons. J Neurosci 19: 7860-7869. |
[24] |
Tanabe H, Eguchi Y, Kamada S, et al. (1997) Susceptibility of cerebellar granule neurons derived from Bcl-2- deficient and transgenic mice to cell death. Eur J Neurosci 9: 848-856. doi: 10.1111/j.1460-9568.1997.tb01434.x
![]() |
[25] |
Gleichmann M, Beinroth S, Reed JC, et al. (1998) Potassium deprivation-induced apoptosis of cerebellar granule neurons: cytochrome c release in the absence of altered expression of Bcl-2 family proteins. Cell Physiol Biochem 8: 194-201. doi: 10.1159/000016282
![]() |
[26] | Galli C, Meucci O, Scorziello A, et al. (1995) Apoptosis in cerebellar granule cells is blocked by high KCl, forskolin and IGF-I through distinct mechanisms of action: the involvement of intracellular calcium and RNA synthesis. J Neurosci 15: 1172-1179. |
[27] |
Kubo T, Nonomura T, Enokido Y, et al. (1995) Brain derived neurotrophic factor (BDNF) can prevent apoptosis of rat cerebellar granule neurons in culture. Devel Brain Res 85: 249-258. doi: 10.1016/0165-3806(94)00220-T
![]() |
[28] |
Chang JY, Korolev VV, Wang JZ (1996) Cyclic AMP and pituitary adenylate cyclase-activating polypeptide (PACAP) prevent programmed cell death of cultured cerebellar granule cells. Neurosci Lett 206: 181-184. doi: 10.1016/S0304-3940(96)12468-X
![]() |
[29] |
Campard PK, Crochemore C, Rene F, et al. (1997) PACAP type I receptor activation promotes cerebellar neuron survival through the cAMP/PKA signaling pathway. DNA Cell Biol 16: 323-333. doi: 10.1089/dna.1997.16.323
![]() |
[30] | Ikeuchi T, Shimoke K, Kubo T, et al. (1998) Apoptosis- inducing and -preventing signal transduction pathways in cultured cerebellar granule neurons. Hum Cell 11: 125-140. |
[31] |
Koulich E, Nguyen T, Johnson K, et al. (2001) NFkappaB is involved in the survival of cerebellar granule neurons: association of NF-kappabeta phosphorylation with cell survival. J Neurochem 76: 1188-1198. doi: 10.1046/j.1471-4159.2001.00134.x
![]() |
[32] | D’Mello SR, Borodezt K, Soltoff SP (1997) Insulin-like growth factor and potassium depolarization maintain neuronal survival by distinct pathways: possible involvement of PI 3-kinase in IGF-I signaling. J Neurosci 17: 1548-1560. |
[33] |
Dudek H, Datta SR, Franke TF, et al. (1997) Regulation of neuronal survival by the serine-threonine protein kinase Akt. Science 275: 661-665. doi: 10.1126/science.275.5300.661
![]() |
[34] |
Shimoke K, Kubo T, Numakawa T, et al. (1997) Involvement of phosphatidylinositol- 3 kinase in prevention of low K+-induced apoptosis of cerebellar granule neurons. Devel Brain Res 101: 197-206. doi: 10.1016/S0165-3806(97)00065-5
![]() |
[35] | Bhave SV, Ghoda L, Hoffman PL (1999) Brain-derived neurotrophic factor mediates the anti-apoptotic effect of NMDA in cerebellar granule neurons: signal transduction cascade and site of ethanol action. J Neurosci 19: 3277-3286. |
[36] |
Shimoke K, Yamagishi S, Yamada M, et al. (1999) Inhibition of phosphatidylinositol 3-kinase activity elevates c-Jun N-terminal kinase activity in apoptosis of cultured cerebellar granule neurons. Devel Brain Res 112: 245-253. doi: 10.1016/S0165-3806(98)00172-2
![]() |
[37] |
Le-Niculescu H, Bonfoco E, Kasuya Y, et al. (1999) Withdrawal of survival factors results in activation of the JNK pathway in neuronal cells leading to Fas ligand induction and cell death. Mol Cell Biol 19: 751-763. doi: 10.1128/MCB.19.1.751
![]() |
[38] | Vaudry D, Gonzalez BJ, Basille M, et al. (2000) PACAP acts as a neurotrophic factor during histogenesis of the rat cerebellar cortex. Ann N Y Acad Sci 921: 293-299. |
[39] | Cavallaro S, Copani A, D’Agata V, et al. (1996) Pituitary adenylate cyclase activating polypeptide prevents apoptosis in cultured cerebellar granule neurons. Mol Pharmacol 50: 60-66. |
[40] | Villalba M, Bockaert J, Journot L (1997) Pituitary adenylate cyclase-activating polypeptide (PACAP-38) protects cerebellar granule neurons from apoptosis by activating the mitogen-activated protein kinase (MAP kinase) pathway. J Neurosci 17: 83-90. |
[41] |
Journot L, Villalba M, Bockaert J (1998) PACAP-38 protects cerebellar granule cells from apoptosis. Ann N Y Acad Sci 865: 100-110. doi: 10.1111/j.1749-6632.1998.tb11168.x
![]() |
[42] |
Franklin JL, Johnson Jr EM (1992) Suppression of programmed neuronal death by sustained elevation of cytoplasmic calcium. Trends Neurosci 15: 501-508. doi: 10.1016/0166-2236(92)90103-F
![]() |
[43] |
Franklin JL, Johnson Jr EM (1994) Block of neuronal apoptosis by a sustained increase of steady-state free Ca2+ concentration. Philos Trans R Soc Lond B Biol Sci 345: 251-256. doi: 10.1098/rstb.1994.0102
![]() |
[44] |
Gutierrez-Martin Y, Martin-Romero FJ, Henao F, et al. (2005) Alteration of cytosolic free calcium homeostasis by SIN-1: high sensitivity of L-type Ca2+ channels to extracellular oxidative/nitrosative stress in cerebellar granule cells. J Neurochem 92: 973-989. doi: 10.1111/j.1471-4159.2004.02964.x
![]() |
[45] |
Copani A, Casabona V, Bruno A, et al. (1998) The metabotropic glutamate receptor mGlu5 controls the onset of developmental apoptosis in cultured cerebellar neurons. Eur J Neurosci 10: 2173-2184. doi: 10.1046/j.1460-9568.1998.00230.x
![]() |
[46] | Borodetz K, D’Mello SRD (1998) Decreased expression of the metabotropic glutamate receptor-4 gene is associated with neuronal apoptosis. J Neurosci Res 53: 531-541. |
[47] |
Garcia-Bereguiain MA, Samhan-Arias AK, Martin-Romero FJ, et al. (2008) Hydrogen sulfide raises cytosolic calcium in neurons through activation of L-type Ca2+ channels. Antioxid Redox Signal 10: 31-42. doi: 10.1089/ars.2007.1656
![]() |
[48] |
Balazs R, Jorgensen OS, Hack N (1988) N-methyl-D-aspartate promotes the survival of cerebellar granule cells in culture. Neuroscience 27: 437-451. doi: 10.1016/0306-4522(88)90279-5
![]() |
[49] |
Balazs A, Hack N, Jorgensen OS (1990) Selective stimulation of excitatory amino acid receptor subtypes and the survival of cerebellar granule cells in culture: Effect of kainic acid. Neuroscience 37: 251-258. doi: 10.1016/0306-4522(90)90211-L
![]() |
[50] |
Balazs R, Hack N, Jorgensen OS (1990) Interactive effects involving different classes of excitatory amino acid receptors and the survival of cerebellar granule cells in culture. Int J Devel Neurosci 8: 347-359. doi: 10.1016/0736-5748(90)90068-D
![]() |
[51] | Yan GM, Lin SZ, Irwin RP, et al. (1995) Activation of muscarinic cholinergic receptor blocks apoptosis of cultured cerebellar granule neurons. Mol Pharmacol 47: 248-257. |
[52] | Mattson MP (1996) Calcium and free radicals: mediators of neurotrophic factor and excitatory transmitter-regulated developmental plasticity and cell death. Perspect Dev Neurobiol 3: 79-91. |
[53] |
Altman J (1982) Morphological development of the rat cerebellum and some of its mechanisms. Exp Brain Res Suppl 6: 8-49. doi: 10.1007/978-3-642-68560-6_2
![]() |
[54] |
Burgoyne RD, Graham ME, Cambray-Deakin M (1993) Neurotrophic effects of NMDA receptor activation on developing cerebellar granule cells. J Neurocytol 22: 689-695. doi: 10.1007/BF01181314
![]() |
[55] |
Monti B, Contestabile A (2000) Blockade of the NMDA receptor increases developmental apoptotic elimination of granule neurons and activates caspases in the rat cerebellum. Eur J Neurosci 12: 3117-3123. doi: 10.1046/j.1460-9568.2000.00189.x
![]() |
[56] |
Hack N, Hidaka H, Wakefield MJ, et al. (1993) Promotion of granule cell survival by high K+ or excitatory amino acid treatment and Ca2+/calmodulin-dependent protein kinase activity. Neuroscience 57: 9-20. doi: 10.1016/0306-4522(93)90108-R
![]() |
[57] |
See V, Boutillier AR, Bito H, et al. (2001) Calcium/calmodulin-dependent protein kinase IV (CaMKIV) inhibits apoptosis induced by potassium deprivation in cerebellar granule neurons. FASEB J 15: 134-144. doi: 10.1096/fj.00-0106com
![]() |
[58] |
Samhan-Arias AK, Martin-Romero FJ, Gutierrez-Merino C (2004) Kaempferol blocks oxidative stress in cerebellar granule cells and reveals a key role for the plasma membrane NADH oxidase activity in the commitment to apoptosis. Free Radic Biol Med 37: 48-61. doi: 10.1016/j.freeradbiomed.2004.04.002
![]() |
[59] | Schulz JB, Weller M, Klockgether T (1996) Potassium deprivation-induced apoptosis of cerebellar granule neurons: a sequential requirement for new mRNA and protein synthesis, ICE-like protease activity, and reactive oxygen species. J Neurosci 16: 4696-4706. |
[60] |
Atlante A, Gagliardi S, Marra E, et al. (1998) Neuronal apoptosis in rats is accompanied by rapid impairment of cellular respiration and is prevented by scavengers of reactive oxygen species. Neurosci Lett 245: 127-130. doi: 10.1016/S0304-3940(98)00195-5
![]() |
[61] |
Hockenbery DM, Oltvai ZN, Yin X-M, et al. (1993) Bcl-2 functions in an antioxidant pathway to prevent apoptosis. Cell 75: 241-251. doi: 10.1016/0092-8674(93)80066-N
![]() |
[62] |
Kane DJ, Sarafian TA, Anton R, et al. (1993) Bcl-2 inhibition of neuronal death: decreased generation of reactive oxygen species. Science 262: 1274-1277. doi: 10.1126/science.8235659
![]() |
[63] |
Mao GD, Poznansky MJ (1992) Electron spin resonance study on the permeability of superoxide radicals in lipid bilayers and biological membranes. FEBS Lett 305: 233-236. doi: 10.1016/0014-5793(92)80675-7
![]() |
[64] |
Samhan-Arias AK, Garcia-Bereguiain MA, Martin-Romero FJ, et al. (2009) Clustering of plasma membrane-bound cytochrome b5 reductase within ‘lipid rafts’ microdomains of the neuronal plasma membrane. Mol Cell Neurosci 40: 14-26. doi: 10.1016/j.mcn.2008.08.013
![]() |
[65] | Borgese N, Meldolesi J (1980) Localization and biosynthesis of NADH-cytochrome b5 reductase, an integral membrane protein, in rat liver cells. I. Distribution of the enzyme activity in microsomes, mitochondria, and Golgi complex. J Cell Biol 85: 501-515. |
[66] |
Chatenay-Rivauday C, Cakar ZP, Jenö P, et al. (2004) Caveolae: biochemical analysis. Mol Biol Rep 31: 67-84. doi: 10.1023/B:MOLE.0000031352.51910.e9
![]() |
[67] | May JM (1999) Is ascorbic acid an antioxidant for the plasma membrane? FASEB J 13: 995-1006. |
[68] |
Martin-Romero FJ, Gutierrez-Martin Y, Henao F, et al. (2002) The NADH oxidase activity of the plasma membrane of synaptosomes is a major source of superoxide anion and is inhibited by peroxynitrite. J Neurochem 82: 604-614. doi: 10.1046/j.1471-4159.2002.00983.x
![]() |
[69] |
Samhan-Arias AK, Duarte RO, Martin-Romero FJ, et al. (2008) Reduction of ascorbate free radical by the plasma membrane of synaptic terminals from rat brain. Arch Biochem Biophys 469: 243-254. doi: 10.1016/j.abb.2007.10.004
![]() |
[70] | Samhan-Arias AK, Gutierrez-Merino C (2014) Cytochrome b5 as a pleiotropic metabolic modulator in mammalian cells, In: Thom R. Editor, Cytochromes b and c: Biochemical properties, biological functions and electrochemical analysis, 1 Ed., New York (USA): Hauppauge, Chapter 2: 39-80. |
[71] | Samhan-Arias AK, López-Sánchez C, Marques-da-Silva D, et al. (2015) High expression of cytochrome b5 reductase isoform 3/cytochrome b5 system in the cerebellum and pyramidal neurons of adult rat brain. Brain Struct Funct 1-16. |
[72] | Percy MJ, Lappin TR (2008) Recessive congenital methaemoglobinaemia: cytochrome b5 reductase deficiency. Br J Haematol 141: 298-308. |
[73] |
Ewenczyk C, Leroux A, Roubergue A, et al. (2008) Recessive hereditary methaemoglobinaemia, type II: delineation of the clinical spectrum. Brain 131: 760-761. doi: 10.1093/brain/awm337
![]() |
[74] |
Huang YH, Tai CL, Lu YH, et al. (2012) Recessive congenital methemoglobinemia caused by a rare mechanism: Maternal uniparental heterodisomy with segmental isodisomy of a chromosome 22. Blood Cells Mol Dis 49: 114-117. doi: 10.1016/j.bcmd.2012.05.005
![]() |
[75] |
Leroux A, Junien C, Kaplan J, et al. (1975) Generalised deficiency of cytochrome b5 reductase in congenital methaemoglobinaemia with mental retardation. Nature 258: 619-620. doi: 10.1038/258619a0
![]() |
[76] |
Toelle SP, Boltshauser E, Mössner E, et al. (2004) Severe neurological impairment in hereditary methaemoglobinaemia type 2. Eur J Pediatr 163: 207-209. doi: 10.1007/s00431-004-1409-x
![]() |
[77] | Aalfs CM, Salieb-Beugelaar GB, Wanders RJA, et al. (2000) A case of methemoglobinemia type II due to NADH-cytochrome b5 reductase deficiency: determination of the molecular basis. Hum Mutat 16: 18-22 |
[78] | Marques-da-Silva D, Gutierrez-Merino C (2014) Caveolin-rich lipid rafts of the plasma membrane of mature cerebellar granule neurons are microcompartments for calcium/reactive oxygen and nitrogen species cross-talk signaling. Cell Calcium 56: 108-123. |
[79] | Gutierrez-Merino C, Marques-da-Silva D, Fortalezas S, et al. (2014) Cytosolic calcium homeostasis in neurons: Control systems, modulation by reactive oxygen and nitrogen species, and space and time fluctuations, In: Heinbockel T. Editor, Neurochemistry, 1 Ed., Rijeka (Craotia): InTech, Chapter 3: 59-110. |
[80] |
Pike LJ (2006) Rafts defined: a report on the keystone symposium on lipid rafts and cell function. J Lipid Res 47: 1597-1598. doi: 10.1194/jlr.E600002-JLR200
![]() |
[81] |
O’Connell KMM, Martens JR, Tamkun MM (2004) Localization of ion channels to lipid raft domains within the cardiovascular system. Trends Cardiovasc Med 14: 37-42. doi: 10.1016/j.tcm.2003.10.002
![]() |
[82] |
Head BP, Insel PA (2007) Do caveolins regulate cells by actions outside of caveolae? Trends Cell Biol 17: 51-57. doi: 10.1016/j.tcb.2006.11.008
![]() |
[83] | Wu G, Lu ZH, Nakamura K, et al. (1996) Trophic effect of cholera toxin B subunit in cultured cerebellar granule neurons: modulation of intracellular calcium by GM1 ganglioside. J Neurosci Res 44: 243-254. |
[84] |
Wu G, Xie X, Lu ZH, et al. (2001) Cerebellar neurons lacking complex gangliosides degenerate in the presence of depolarizing levels of potassium. Proc Natl Acad Sci USA 98: 307-312. doi: 10.1073/pnas.98.1.307
![]() |
[85] |
Marques-da-Silva D, Samhan-Arias AK, Tiago T, et al. (2010) L-type calcium channels and cytochrome b5 reductase are components of protein complexes tightly associated with lipid rafts microdomains of the neuronal plasma membrane. J Proteomics 73: 1502-1510. doi: 10.1016/j.jprot.2010.02.014
![]() |
[86] |
Davare MA, Dong F, Rubin CS, et al. (1999) The A-kinase anchor protein MAP2B and cAMP-dependent protein kinase are associated with class C L-type calcium channels in neurons. J Biol Chem 274: 30280-30287. doi: 10.1074/jbc.274.42.30280
![]() |
[87] |
Razani B, Rubin CS, Lisanti MP (1999) Regulation of cAMP-mediated Signal Transduction via Interaction of Caveolins with the Catalytic Subunit of Protein Kinase A. J Biol Chem 274: 26353-26360. doi: 10.1074/jbc.274.37.26353
![]() |
[88] | Suzuki T, Du F, Tian Q-B, et al. (2008) Ca2+/calmodulin-dependent protein kinase IIα clusters are associated with stable lipid rafts and their formation traps PSD-95. J Neurochem 104: 596-610. |
[89] | Pinard CR, Mascagni F, McDonald AJ (2005) Neuronal localization of Cav1.2 L-type calcium channels in the rat basolateral amygdala. Brain Res 1064: 52 - 55. |
[90] | Samhan-Arias AK, García-Bereguiaín MA, Gutierrez-Merino C (2007) Plasma membrane-bound cytochrome b5 reductase forms a large network of redox centres that co-localizes with cholera toxin B binding sites in cerebellar granule neurons in culture, In: Society for Free Radical Research (SFRR) Editor, Proceedings of the European Meeting of the SFFR, Bologna (Italy): Medimond, 147-150. |
[91] | Samhan-Arias AK, Gutiérrez-Merino C (2008) Plasma membrane-bound cytochrome b5 reductase is associated with lipid rafts in cerebellar granule neurons in culture, In: Grune T. Editor, Proceedings of the European Meeting of the Society for Free Radical Research, 1 Ed., Bologna (Italy): Medimond, 75-78. |
[92] | Silva DM, Samhan-Arias AK, Garcia-Bereguiain MA, et al. (2009) Major plasma membrane-associated redox centres co-localize with L-type calcium channels in neuronal lipid rafts microdomains, In: Caporosi D., Pigozzi F., Sabatini S. Editors, Free Radicals, Health and Lifestyle, 3 Eds., Bologna (Italy): Medimond, 127-130. |
[93] |
Marques-da-Silva D, Gutierrez-Merino C (2012) L-type voltage-operated calcium channels, N-methyl-D-aspartate receptors and neuronal nitric-oxide synthase form a calcium/redox nano-transducer within lipid rafts. Biochem Biophys Res Commun 420: 257-262. doi: 10.1016/j.bbrc.2012.02.145
![]() |
[94] |
Sato Y, Sagami I, Shimizu T (2004) Identification of Caveolin-1-interacting Sites in Neuronal Nitric-oxide Synthase. J Biol Chem 279: 8827-8836. doi: 10.1074/jbc.M310327200
![]() |
[95] |
Samhan-Arias AK, Garcia-Bereguiain MA, Martin-Romero FJ, et al. (2006) Regionalization of plasma membrane-bound flavoproteins of cerebellar granule neurons in culture by fluorescence energy transfer imaging. J Fluorescence 16: 393-401. doi: 10.1007/s10895-005-0065-5
![]() |
[96] | Gutierrez-Merino C (2008) Redox modulation of neuronal calcium homeostasis and its deregulation by reactive oxygen species, In: Gutierrez-Merino C. and Leeuwenburgh C. Editors, Free Radicals in Biology and Medicine, 2 Eds., Kerala (India): Research Signpost, 67-101. |
[97] |
Marchetti C, Usai C (1996) High affinity block by nimodipine of the internal calcium elevation in chronically depolarized rat cerebellar granule neurons. Neurosci Lett 207: 77-80. doi: 10.1016/0304-3940(96)12492-7
![]() |
[98] |
Maric D, Maric I, Barker JL (2000) Developmental changes in cell calcium homeostasis during neurogenesis of the embryonic rat cerebral cortex. Cereb Cortex 10: 561-573. doi: 10.1093/cercor/10.6.561
![]() |
[99] |
Arakawa Y, Nishijima C, Shimizu N, et al. (2002) Survival-promoting activity of nimodipine and nifedipine in rat motoneurons: implications of an intrinsic calcium toxicity in motoneurons. J Neurochem 83: 150-156. doi: 10.1046/j.1471-4159.2002.01126.x
![]() |
[100] |
Samhan-Arias AK, Gutierrez-Merino C (2014) Purified NADH-Cytochrome b5 Reductase Is a Novel Superoxide Anion Source Inhibited by Apocynin: Sensitivity to nitric oxide and peroxynitrite. Free Radic Biol Med 73: 174-189. doi: 10.1016/j.freeradbiomed.2014.04.033
![]() |
[101] |
Hidalgo C, Donoso P (2008) Crosstalk between calcium and redox signalling: from molecular mechanisms to health implications. Antioxid Redox Signal 10: 1275-1312. doi: 10.1089/ars.2007.1886
![]() |
[102] |
Szabó C, Ischiropoulos H, Radi R (2007) Peroxynitrite: biochemistry, pathophysiology and development of therapeutics. Nat Rev Drug Discov 6: 662-680. doi: 10.1038/nrd2222
![]() |
[103] |
Bredt DS, Snyder SH (1994) Nitric oxide: a physiologic messenger molecule. Annu Rev Biochem 63: 175-195. doi: 10.1146/annurev.bi.63.070194.001135
![]() |
[104] |
Parekh AB (2008) Ca2+ microdomains near plasma membrane Ca2+ channels: impact on cell function. J Physiol 586: 3043-3054. doi: 10.1113/jphysiol.2008.153460
![]() |
[105] |
Neher E (1998) Vesicle pools and Ca2+ microdomains: new tools for understanding their roles in neurotransmitter release. Neuron 20: 389-399. doi: 10.1016/S0896-6273(00)80983-6
![]() |
[106] | Neher E (1998) Usefulness and limitations of linear approximations to the understanding of Ca2+ signals. Cell Calcium 24: 345-357. |
[107] |
Willmott NJ, Wong K, Strong AJ (2000) Intercellular Ca2+ waves in rat hippocampal slice and dissociated glial-neuron cultures mediated by nitric oxide. FEBS Lett 487: 239-247. doi: 10.1016/S0014-5793(00)02359-0
![]() |
[108] | Marques-da-Silva D (2012) Estudio de los microdominios de sistemas redox y de transporte de calcio en la membrana plasmática de neuronas. PhD Thesis, University of Extremadura. |
[109] |
Contestabile A, Ciani E (2004) Role of nitric oxide in the regulation of neuronal proliferation, survival and differentiation. Neurochem Int 45: 903-914. doi: 10.1016/j.neuint.2004.03.021
![]() |
[110] |
Contestabile A (2008) Regulation of transcription factors by nitric oxide in neurons and in neural-derived tumor cells. Prog Neurobiol 84: 317-328. doi: 10.1016/j.pneurobio.2008.01.002
![]() |
[111] |
Grueter CE, Abiria SA, Wu Y, et al. (2008) Differential regulated interactions of calcium/calmodulin-dependent protein kinase II with isoforms of voltage-gated calcium channel beta subunits. Biochemistry 47: 1760-1767. doi: 10.1021/bi701755q
![]() |
[112] |
Paratcha G, Ibáñez CF (2002) Lipid rafts and the control of neurotrophic factor signaling in the nervous system: variations on a theme. Curr Opin Neurobiol 12: 542-549. doi: 10.1016/S0959-4388(02)00363-X ![]() |
[113] | Inoue H, Miyaji M, Kosugi A, et al. (2002) Lipid rafts as the signaling scaffold for NK cell activation: tyrosine phosphorylation and association of LAT with phosphatidylinositol 3-kinase and phospholipase C-gamma following CD2 stimulation. Eur J Immunol 32: 2188-2198. |
[114] |
Zheng F, Soellner D, Nunez J, et al. (2008) The basal level of intracellular calcium gates the activation of phosphoinositide 3-kinase - Akt signaling by brain-derived neurotrophic factor in cortical neurons. J Neurochem 106: 1259-1274. doi: 10.1111/j.1471-4159.2008.05478.x
![]() |
[115] |
Hudmon A, Schulman H, Kim J, et al. (2005) CaMKII tethers to L-type Ca2+ channels, establishing a local and dedicated integrator of Ca2+ signals for facilitation. J Cell Biol 171: 537-547. doi: 10.1083/jcb.200505155
![]() |
[116] | Lee TS, Karl R, Moosmang S, et al. (2006) Calmodulin kinase II is involved in voltage-dependent facilitation of the L-type Cav1.2 calcium channel: Identification of the phosphorylation sites. J Biol Chem 281: 25560-25567. |
[117] |
Coultrap SJ, Bayer KU (2014) Nitric Oxide Induces Ca2+-independent Activity of the Ca2+/Calmodulin-dependent Protein Kinase II (CaMKII). J Biol Chem 289: 19458-19465. doi: 10.1074/jbc.M114.558254
![]() |
[118] |
Coultrap SJ, Zaegel V, Bayer KU (2014) CaMKII isoforms differ in their specific requirements for regulation by nitric oxide. FEBS Lett 588: 4672-4676. doi: 10.1016/j.febslet.2014.10.039
![]() |
[119] | Müller U, Hildebrandt H (2002) Nitric Oxide/cGMP-Mediated Protein Kinase A Activation in the Antennal Lobes Plays an Important Role in Appetitive Reflex Habituation in the Honeybee. J Neurosci 22:8739-8747. |
[120] |
De Jongh KS, Murphy BJ, Colvin AA, et al. (1996) Specific phosphorylation of a site in the full-length form of the alpha 1 subunit of the cardiac L-type calcium channel by adenosine 3',5'-cyclic monophosphate-dependent protein kinase. Biochemistry 35: 10392-10340. doi: 10.1021/bi953023c
![]() |
[121] |
Mitterdorfer J, Froschmayr M, Grabner M, et al. (1996) Identification of PK-A phosphorylation sites in the carboxyl terminus of L-type calcium channel alpha 1 subunits. Biochemistry 35: 9400-9406. doi: 10.1021/bi960683o
![]() |
[122] |
Gao T, Yatani A, Dell’Acqua ML, et al. (1997) cAMP-dependent regulation of cardiac L-type Ca2+ channels requires membrane targeting of PKA and phosphorylation of channel subunits. Neuron 19: 185-196. doi: 10.1016/S0896-6273(00)80358-X
![]() |
[123] |
Puri TS, Gerhardstein BL, Zhao XL, et al. (1997) Differential effects of subunit interactions on protein kinase A- and C-mediated phosphorylation of L-type calcium channels. Biochemistry 36: 9605-9615. doi: 10.1021/bi970500d
![]() |
node | $ x $ | $ y $ | $ S $ |
1 | 0.100 | 0.000 | 0.164 |
2 | 0.874 | –0.104 | 0.794 |
3 | 0.581 | 0.790 | –0.128 |
4 | –0.043 | 1.0547 | 0.936 |
5 | –0.945 | 0.371 | –0.299 |
6 | –0.818 | –0.161 | 0.750 |
7 | –0.206 | –1.023 | –2.217 |