betaFeedback

WorksPhysics and Astronomy

Network analysis of memristive device circuits: dynamics, stability and correlations

Abstract

Networks with memristive devices are a potential basis for the next generation of computing devices. They are also an important model system for basic science, from modeling nanoscale conductivity to providing insight into the information-processing of neurons. The resistance in a memristive device depends on the history of the applied bias and thus displays a type of memory. The interplay of this memory with the dynamic properties of the network can give rise to new behavior, offering many fascinating theoretical challenges. But methods to analyze general memristive circuits are not well described in the literature. In this paper we develop a general circuit analysis for networks that combine memristive devices alongside resistors, capacitors and inductors and under various types of control. We derive equations of motion for the memory parameters of these circuits and describe the conditions for which a network should display properties characteristic of a resonator system. For the case of a purely memresistive network, we derive Lyapunov functions, which can be used to study the stability of the network dynamics. Surprisingly, analysis of the Lyapunov functions show that these circuits do not always have a stable equilibrium in the case of nonlinear resistance and window functions. The Lyapunov function allows us to study circuit invariances, wherein different circuits give rise to similar equations of motion, which manifest through a gauge freedom and node permutations. Finally, we identify the relation between the graph Laplacian and the operators governing the dynamics of memristor networks operators, and we use these tools to study the correlations between distant memristive devices through the effective resistance.

Introduction

Once regarded as somewhat niche, memristance is now recognized as a ubiquitous feature of nature, especially at the nanoscale1. As a result, memristive devices have been the subject of intense research over the past decade. In addition to the wide applicability of these components, ranging from scalable memory devices to machine learning and reservoir computing2–7, there is also interest in their fundamental dynamical properties8–12. For example, it has long been known that the introduction of memristive devices in a circuit can lead to chaotic dynamics and there are numerous papers studying the properties of the Chua attractors13,14.

However, less is known for generic circuits with memory, and even fewer exact results have been obtained in this regard. While little theoretical progress has been made for the most general cases of electronic circuits with memory, purely memristive circuits are amenable to an analytical approach: the equations for networks of these devices can often be obtained analytically, unveiling the most bizarre and unexpected properties of these dynamical circuits15,16. For instance, it has been shown that the asymptotic states of certain memristive-resistive circuits can be well described by the Curie-Weiß model and its ferromagnetic-paramagnetic transition. Further, it has been argued that random circuits of memristive devices (analyzed through the lens of their Lyapunov function) exhibit a ferromagnetic-glass transition17,18.

Obtaining a full dynamic picture of this behavior is difficult because the nonlinearity of the memristive devices makes the equations of motion impossible to solve analytically. This paper provides tools to understand the properties of generic, first-order memristive devices with window functions in networks. In particular, we derive dynamical equations for generic circuits with memristive devices, and we derive Lyapunov functions for the infinite family of first-order memristive circuits with sharp boundaries (e.g., discontinuous window functions).

We also generalize techniques used for network circuit analysis to the case of memristive device networks. Using these, we demonstrate that these circuits frequently give rise to analytically nontrivial or intractable dynamical equations. In doing so, we demonstrate mathematical techniques, including correlation analysis and Lyapunov functions, which can provide information on the dynamics, evolution, and stability of a network of memristive devices. Further, we discuss how different circuits give rise to similar dynamics through gauge freedom and node permutation in the governing equations.

This work builds upon previous works to study the dynamics of memristor networks most of which use the graph Laplacian or projector operators11,19–21. We identify the relation between the graph Laplacian and the projection operators. Finally we use this correspondence to study the effective resistance in memristor networks and construct effective circuits which allow us to study the correlations between distant memristors.

Graph theory for circuit dynamics

In network analysis of electrical circuits, an electrical circuit is mapped to a graph G=(V,E)\mathcal{G} = (V,E). A graph G\mathcal{G} consists of a set of nn vertices/nodes VV and a set of mm edges EE. An undirected edge may be denoted {k,l}\{k,l\}, but in circuit analysis, each edge is assigned an arbitrary direction from kk to ll denoted (k,l)(k,l). The graph allows no self-edges and no duplicate edges. As a simple example, let us work with the triangle graph defined by,

V={1,2,3},E={(1,2),(2,3),(1,3)}(1)V = \{1,2,3\}, \quad E = \{(1,2),(2,3),(1,3)\} \tag*{(1)}

A representation of this graph is shown in Figure 1. Note that the edge between 1 and 3 is not cyclical. A graph representation

A simple graph with three edges connecting three nodes without an oriented cycle

Figure 1. A simple graph with three edges connecting three nodes without an oriented cycle.

of a circuit generally consists of edges corresponding to circuit elements and nodes, i.e., junctions in the circuit. Currents and voltages are assigned in the graph to satisfy Kirchhoff’s current law (KCL) and Kirchhoff’s voltage law (KVL). A circuit is solved when it satisfies these constraints given a circuit topology; thus, network analysis of a circuit consists of solving for valid current and voltage configurations.

A current configuration i={ie,e∈E}\mathbf{i} = \{i_e, e \in E\} is a set of numbers associated with the edges which satisfy KCL. KCL could be enforced on each vertex by building an n×mn \times m incidence matrix BB, where each row represents a node and each column an edge; the matrix entries take values of ±1\pm1 corresponding to a directed edge oriented towards or away from each node, and 0 when an edge is not incident on a node. For our triangle graph, this is,

B=n1n2n3(101−1100−1−1)e12e23e13(2)B = \begin{array}{c} n_1\\ n_2\\ n_3 \end{array} \left( \begin{array}{ccc} 1 & 0 & 1\\ -1 & 1 & 0\\ 0 & -1 & -1 \end{array} \right) \begin{array}{c} e_{12}\quad e_{23}\quad e_{13} \end{array} \tag*{(2)}

eije_{ij} are edges linking nodes ii and jj (edges are indexed with superscripts, eke^k), and nin_i are indexed nodes. For each vertex kk and edge ee, let bk,e=1b_{k,e}=1 if kk is the start, −1-1 if kk is the finish, and 0 if the edge does not include kk. KCL are enforced such that a current vector i⃗\vec{i} satisfies,

∑ebk,eie=0.(3)\sum_e b_{k,e}i_e = 0. \tag*{(3)}

If i⃗\vec{i} is an allowed current configuration then

Bi⃗=0.(4)B\vec{i}=0. \tag*{(4)}

The operator BB may be interpreted as a type of divergence on the graph, and the relationship Bi⃗=0B\vec{i}=0 is equivalent to saying that the current configurations must be divergence-less. In this sense, they are similar to magnetic fields.

Similarly, a voltage configuration v={ve,e∈E}\mathbf{v} = \{v_e, e \in E\} is a set of numbers associated with the edges that satisfy KVL. That is, we define a cycle, ll on the graph as a sequence of nodes (k1,…kj)(k_1,\ldots k_j) such that {ki,ki+1}∈E\{k_i,k_{i+1}\} \in E (in either direction) and {kj,k1}∈E\{k_j,k_1\} \in E. We can find the oriented edge set of the cycle ElE_l, wherein the orientation matches the direction traversed in ll (these edges do not necessarily lie in EE). We construct a cycle matrix AA; each row represents a unique cycle in the graph, and each column an edge in EE. The matrix entries take values of ±1\pm1 when the edge is part of the cycle and 0 otherwise; the sign indicates whether the orientation of the edge in EE aligns with the direction of the cycle (+1+1) or is opposite to it (−1-1). The triangle graph has two possible cycles, 1→2→3→11 \rightarrow2 \rightarrow3 \rightarrow1 or 1→3→2→11 \rightarrow3 \rightarrow2 \rightarrow1. We can include both for now,

A=cycle1cycle2(11−1−1−11)e12e23e13(5)A = \begin{array}{c} \mathrm{cycle}_1\\ \mathrm{cycle}_2 \end{array} \left( \begin{array}{ccc} 1 & 1 & -1\\ -1 & -1 & 1 \end{array} \right) \begin{array}{c} e_{12}\quad e_{23}\quad e_{13} \end{array} \tag*{(5)}

These are not independent, and we must eliminate one cycle to solve for a voltage configuration, generating a reduced cycle matrix. KVL are enforced on a voltage vector v⃗\vec{v} for each cycle in the graph; let al,ea_{l,e} be 1 if the orientation matches that in the graph, −1-1 if it is opposite, and 0 if e∉Ele \notin E_l. In this case, we can write,

∑eal,eve=0.(6)\sum_e a_{l,e}v_e = 0. \tag*{(6)}

To solve for a voltage configuration we will have to do some extra work later to determine which cycles we include. If v⃗\vec{v} is a voltage configuration then

Av⃗=0.(7)A\vec{v} = 0. \tag*{(7)}

AA may be interpreted as a type of curl on the graph that gives the circulation of a vector. In this sense, the relation Av⃗=0A\vec{v} = 0 is analogous to saying that the voltage drops are curl-less and thus analogous to electric fields.

The current and voltage configurations belong to a set of all current and voltage configurations. In the supplementary material, we prove that these current and voltage configurations are dual. Thus, we have two valid representations of the electrical properties of the circuit. Here, we define two ways of approaching the solution of a circuit: the node method and the loop method.

In the node method, the voltage configuration is calculated by finding voltage potentials of the nodes in the circuit. One potential value is set as the ground, and the n−1n-1 remaining potential values are considered as the unknowns. For the n−1n-1 non-root nodes, we write down KCL and solve for the voltage potential in succession via Gaussian elimination.

In the loop method, we must have a way of choosing cycles. To do this, we introduce a spanning tree TT. A spanning tree is the minimum number of edges to connect all nodes. The remaining cycle edges can be added to the spanning tree; when an individual cycle edge is added to the spanning tree it gives a unique cycle whose orientation we assign by that of the included edge. This is called the fundamental cycle of the edge, also called a fundamental loop. From the reduced cycle matrix, we write down KVL. A current i⃗\vec{i} may then be expressed by the fundamental loop currents jej_e as a weighted sum of the fundamental cycles. Writing a reduced cycle matrix A~\tilde{A} in terms of the fundamental loops, we have

i⃗=A~tj⃗.(8)\vec{i} = \tilde{A}^{t}\vec{j}. \tag*{(8)}

Given a spanning tree TT, we can find unique paths from any arbitrary initial node, the root, and any other node kk. We can sum the voltages at the nodes along this path; if an edge is oriented along the walk from root to node, the voltage value of that edge is added; if the walk and edge are unaligned, the voltage is subtracted. We define this walk pkp_k, and the voltage configuration can be written in terms of this fundamental walk:

v⃗=Btp⃗,(9)\vec{v} = B^{t}\vec{p}, \tag*{(9)}

which reproduces the relation between the voltage and potential above. These identities will be useful to derive the results below.

Resistors and memristive devices

A memristive device can be treated as a resistor with variable resistance; its resistance can take continuous values between a low and high resistance. The state of the resistance between two limiting values can be parameterized by a variable xx, which is constrained by some window function such that 0≤x≤10 \le x \le1. This can be thought of as the first-order term in a polynomial expansion for the resistance in terms of an dimensional parameter. Thus, the resistance varies linearly with xx; later, we will generalize this to include higher-order terms of xx to describe nonlinear memristance. In typical conditions, the resistance of a memristive device is bounded between two fixed high and low resistance states, corresponding to the ’off’ and ’on’ states. The resistance of these states is given by the values Roff≫Ron>0R_{\mathrm{off}} \gg R_{\mathrm{on}} > 0. As the resistance state depends on the history of applied voltage or bias, there is a memory stored in the resistance value, and we will refer to xx as the internal memory parameter. For the case of widely used metal oxide memristive devices, a simple toy model for the evolution of the resistance is the following1:

R(x)=Ronx+Roff(1−x), the direct parametrization,(10)R(x) = R_{\mathrm{on}}x + R_{\mathrm{off}}(1-x), \text{ the direct parametrization,} \tag*{(10)}
ddtx(t)=Ronβi(t)−αx(t),(11)\frac{d}{dt}x(t) = \frac{R_{\mathrm{on}}}{\beta}i(t) - \alpha x(t), \tag*{(11)}

In the direct parametrization, R(1)=RonR(1) = R_{\mathrm{on}} and R(0)=RoffR(0) = R_{\mathrm{off}}. In equation (11), i(t)i(t) is the current flowing in the device at time tt, α\alpha is a dampening parameter with units of inverse time, and β\beta can be thought of as the inverse learning rate with units of time per voltage. These constants control the decay and reinforcement timescales. We define a scaling variable ξ=Roff−RonRon\xi= \frac{R_{\mathrm{off}}-R_{\mathrm{on}}}{R_{\mathrm{on}}}. α\alpha, β\beta, ξ\xi can be measured experimentally. Alternatively, a similar formulation of the resistance dynamical equation is given,

R(x)=Ron(1−x)+Roffx, the flipped parametrization,(12)R(x) = R_{\mathrm{on}}(1-x) + R_{\mathrm{off}}x,\ \text{the flipped parametrization,} \tag*{(12)}
ddtx(t)=αx(t)−Ronβi(t).(13)\frac{d}{dt}x(t) = \alpha x(t) - \frac{R_{\mathrm{on}}}{\beta}i(t). \tag*{(13)}

In the flipped parametrization R(1)=RoffR(1) = R_{\mathrm{off}} and R(0)=RonR(0) = R_{\mathrm{on}}. Here we define the scaling factor χ=Roff−RonRoff\chi= \frac{R_{\mathrm{off}}-R_{\mathrm{on}}}{R_{\mathrm{off}}} which is bounded χ<1\chi< 1. These two parameterizations are inequivalent models of the memristive device dynamics. From the point of view of the dynamics of the single memristive device, the two parametrizations are related by the following change of variables: α↔−α\alpha\leftrightarrow-\alpha, β↔−β\beta\leftrightarrow-\beta and ξ↔−χ\xi\leftrightarrow-\chi, i.e., ξ(−RoffRon)=χ\xi\left(\frac{-R_{\mathrm{off}}}{R_{\mathrm{on}}}\right) = \chi.

Resistance characterization

The model given above, which has been previously studied11,22, has a simple time dependency.

x(t)=e±α(t−t0)x0∓1β′∫t0te∓α(s−t)i(s) ds(14)x(t) = e^{\pm\alpha(t-t_0)}x_0 \mp\frac{1}{\beta'}\int_{t_0}^{t} e^{\mp\alpha(s-t)}i(s)\,\mathrm{d}s \tag*{(14)}

References

  1. [1]Pershin, Y. V., & Ventra, M. D. (2011). Memory effects in complex materials and nanoscale systems. Advances in Physics, 60, 145–227.
  2. [2]Traversa, F. L., & al., . et . (2014). Dynamic computing random access memory. Nanotechnology, 25, 285201.
  3. [3]3J. P. Carbajal, J. Dambre, M. Hermans, and B. Schrauwen, “Memristor models for machine learning,” Neural Computation 27 (2015), 10.1162/NECO_a_00694, arXiv:1406.2210.DOI
  4. [4]Du, C., & al., . et . (2017). Reservoir computing using dynamic memristors for temporal information processing. Nature Communications, 8. https://doi.org/10.1038/s41467-017-02337-y
  5. [5]Milano, G., Pedretti, G., Montano, K., & al., . et . (2022). In materia reservoir computing with a fully memristive architecture based on self-organizing nanowire networks. 195–202.
  6. [6]Fu, K., Zhu, R., Loeffler, A., & al., . et . (2020). Reservoir computing with neuromemristive nanowire networks. Proceedings of the International Joint Conference on Neural Networks (IJCNN).
  7. [7]7W. Xu, J. Wang, and X. Yan, “Advances in memristor-based neural networks,” Frontiers in Nanotechnology 3 (2021), 10.3389/fnano.2021.645995.DOI
  8. [8]Pershin, Y. V., Slipko, V. A., & Ventra, M. D. (2013). Complex dynamics and scale invariance of one-dimensional memristive networks. Physical Review E, 87, 022116.
  9. [9]9A. Diaz-Alvarez, R. Higuchi, P. Sanz-Leon, I. Marcus, Y. Shingaya, A. Z. Stieg, J. K. Gimzewski, Z. Kuncic, and T. Nakayama, “Emergent dynamics of neuromorphic nanowire networks,” Scientific Reports 9, 14920 (2019).DOI
  10. [10]10R. Zhu, J. Hochstetter, A. Loeffler, A. Diaz-Alvarez, T. Nakayama, J. T. Lizier, and Z. Kuncic, “Information dynamics in neuromorphic nanowire networks,” Scientific Reports 11, 13047 (2021).DOI
  11. [11]Caravelli, F., Traversa, F. L., & Ventra, M. D. (2017). Complex dynamics of memristive circuits: Analytical results and universal slow relaxation. Physical Review E, 95, 022410.
  12. [12]12R. Riaza, “Dynamical properties of electrical circuits with fully nonlinear memristors,” Nonlinear Analysis: Real World Applications 12, 3674–3686 (2011).arxiv.org/abs/1008.2528
  13. [13]13T. Matsumoto, “A chaotic attractor from chua’s circuit,” IEEE Transactions on Circuits and Systems 31, 1055–1058 (1984).DOI
  14. [14]Nadan, R. N. (1993). Chua’s Circuit: A Paradigm for Chaos. In World Scientific. <a href="https://doi.org/https://www.worldscientific.com/doi/pdf/10.1142/1997" rel="noopener noreferrer" target="_blank">https://doi.org/https://www.worldscientific.com/doi/pdf/10.1142/1997</a>
  15. [15]15A. Zegarac and F. Caravelli, “Memristive networks: From graph theory to statistical physics,” EPL (Europhysics Letters) 125, 10001 (2019).DOI
  16. [16]Tellini, B., Bologna, M., Chandia, K. J., & Maccucci, M. (2021). Revisiting the memristor concept within basic circuit theory. International Journal of Circuit Theory and Applications, 49, 3488–3506. <a href="https://doi.org/https://onlinelibrary.wiley.com/doi/pdf/10.1002/cta.3111" rel="noopener noreferrer" target="_blank">https://doi.org/https://onlinelibrary.wiley.com/doi/pdf/10.1002/cta.3111</a>
  17. [17]17F. Caravelli and F. C. Sheldon, “Phases of memristive circuits via an interacting disorder approach,” https://arxiv.org/abs/2009.00114 (2020).
  18. [18]18F. Caravelli and P. Barucca, “A mean-field model of memristive circuit interaction,” EPL (Europhysics Letters) 122, 40008 (2018).DOI
  19. [19]19M. Stern, A. J. Liu, and V. Balasubramanian, “Physical effects of learning,” Physical Review E 109 (2024), 10.1103/physreve.109.024311.DOI
  20. [20]Jiang, J. R., Ren, J. M., & Tan, S. (2022). Stability analysis for memristor-based stochastic multi-layer neural networks with coupling disturbance. Chaos, Solitons & Fractals, 165, 12771.
  21. [21]21P. B. Dhivakaran, M. Gowrisankar, and A. Vinodkumar, “Bipartite synchronization of fractional order multiple memristor coupled delayed neural networks with event triggered pinning control,” Neural Processing Letters 56, 50 (2024).DOI
  22. [22]22F. Caravelli, “Asymptotic behavior of memristive circuits,” Entropy 21, 789 (2019).DOI
  23. [23]23F. C. Sheldon, A. Kolchinsky, and F. Caravelli, “The computational capacity of Irc, memristive and hybrid reservoirs,” Phys. Rev. E 106 (2022).DOI
  24. [24]24V. Baccetti, R. Zhu, Z. Kuncic, and F. Caravelli, “Ergodicity, lack thereof, and the performance of reservoir computing with memristive networks,” (2023), arXiv:2310.09530 [cond-mat.dis-nn].DOI
  25. [25]25O. L. White, D. D. Lee, and H. Sompolinsky, “Short-term memory in orthogonal neural networks,” Physical review letters 92, 148102 (2004).pubmed.ncbi.nlm.nih.gov/15089576
  26. [26]26F. Caravelli, F. Sheldon, and F. L. Traversa, “Global minimization via classical tunneling assisted by collective force field formation,” Science Advances 7, 1542 (2021), https://www.science.org/doi/pdf/10.1126/sciadv.abh1542.DOI
  27. [27]27Y. N. Joglekar and S. J. Wolf, “The elusive memristor: properties of basic electrical circuits,” Eur. J. of Phys. 30, 661–675 (2009).DOI
  28. [28]28A. Ascoli, R. Tetzlaff, L. O. Chua, J. P. Strachan, and R. S. Williams, “History erase effect in a non-volatile memristor,” IEEE Transactions on Circuits and Systems I: Regular Papers 63, 389–400 (2016).DOI
  29. [29]29F. Caravelli, “Cycle equivalence classes, orthogonal weingarten calculus, and the mean field theory of memristive systems,” arXiv:2304.14890 (2023).arXiv
  30. [30]30Note that this is the result is the same as the bottom corner of a block-wise matrix inversion (Schur inverse),
  31. [31]31N. Biggs, Algebraic Graph Theory, 2nd ed., Cambridge Mathematical Library (Cambridge University Press, 1974).DOI
  32. [32]32P. G. Doyle and J. L. Snell, “Random walks and electric networks,” (2000), arXiv:math/0001057 [math.PR].DOI

Paper details

vlnc:W_y4uq2q27
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents