White Paper 01
Date: 2026-01-09
Status: Complete draft (self-contained)
| The Adaptive Resistance Principle (ARP) is a dynamical rule for networks in which edge conductances strengthen in proportion to carried activity and decay in time in the absence of activity. In its canonical form, ARP is an ordinary differential equation (ODE) on an edge conductance variable $G_{ij}(t)$ driven by the magnitude of current $ | I_{ij}(t) | $. |
The Adaptive Impedance Network (AIN) generalizes ARP to frequency-aware, passive electrical analogs by allowing not only conductance/resistance but also capacitance and inductance to adapt under local excitation. AIN supplies a unified template for (i) self-organizing routing and graph optimization, (ii) adaptive analog computation, and (iii) control-like “memory depth” via time constants.
This paper provides the mathematical definition, equilibrium behavior, discrete-time integration guidance, stability notes, and practical parameterization. It also describes a minimal simulation recipe and testable predictions.
Many systems that “learn paths” or “discover structure” can be modeled as networks with adaptive edges. The edges represent conductance, trust, affinity, capacity, or an effective coupling. The key design requirement is:
ARP is the simplest continuous-time rule satisfying these requirements.
AIN adds two additional edge-states (capacitance and inductance) so the network can represent not just static routing but frequency-dependent behavior (buffering, inertia, damping).
ARP/AIN will remind readers of several known ideas:
The distinct claim of this paper is not that these inspirations are new, but that a single, parameter-minimal ODE template (reinforce on activity, decay otherwise) can serve as a clean foundation across routing, analog computation, and adaptive impedance modeling.
Consider a graph with nodes $i,j$ and an edge state per ordered pair (or per undirected edge). Denote:
Let $I_{ij}(t)$ be the (possibly signed) current through the edge at time $t$ under a chosen driving condition (in circuit analogs, driven by sources/sinks; in routing analogs, driven by demand).
The canonical ARP update is:
\[\boxed{\ \frac{dG_{ij}}{dt} = \alpha_G\,|I_{ij}(t)| - \mu_G\,G_{ij}(t).\ }\]Parameters:
Interpretation:
| If an edge carries no current ($ | I_{ij} | =0$), conductance decays exponentially. |
If this model is instantiated as a physical circuit analog, then:
Therefore:
\[[\alpha_G] = \frac{\mathrm{S}}{\mathrm{A}\cdot\mathrm{s}},\quad [\mu_G] = \frac{1}{\mathrm{s}}.\]For non-physical analogs (e.g., routing weights), these “units” can be interpreted as consistent scaling dimensions.
In coupled networks, “every edge wins” unless there is competition. A practical, stable-by-default variant is to combine (i) bounded reinforcement and (ii) an explicit global (or local) resource constraint.
One canonical choice is:
\[\boxed{\ \frac{dG_{ij}}{dt} = \alpha_G\,\sigma(|I_{ij}|) - \mu_G\,G_{ij}\ }\]subject to a conductance budget, for example:
\[\boxed{\ \sum_{(i,j)\in E} G_{ij}(t) \le G_{\mathrm{total}}\ }\]Implementation note (discrete time): after the Euler update, if $\sum G_{ij} > G_{\mathrm{total}}$, rescale
\[G_{ij} \leftarrow G_{ij}\,\frac{G_{\mathrm{total}}}{\sum G_{ij}}.\]This simple normalization forces edges to compete and reliably produces sparse backbones.
A compact scalar that summarizes “how conductive the used network is” is:
\[\boxed{\ \mathrm{ANE}(t) = \frac{\sum_{ij} G_{ij}(t)\,|I_{ij}(t)|}{\sum_{ij} |I_{ij}(t)|}.\ }\]This is a weighted average of conductance over active edges.
Role clarification: in this paper, $\mathrm{ANE}(t)$ is a diagnostic (an instrument panel), not assumed to be an objective function and not claimed to be a Lyapunov function.
| Fix a single edge $(i,j)$ and assume $ | I_{ij}(t) | $ is constant in time at value $ | I | $ (a local quasi-static assumption often valid when the network dynamics are slower than instantaneous current solving). |
Then the ODE is linear:
\[\frac{dG}{dt} + \mu_G G = \alpha_G |I|.\]With initial condition $G(0)=G_0$:
\[G(t) = G_{\mathrm{eq}} + (G_0 - G_{\mathrm{eq}})e^{-\mu_G t},\]where
\[\boxed{\ G_{\mathrm{eq}} = \frac{\alpha_G}{\mu_G}\,|I|.\ }\]If the current itself depends on conductance (as it does in circuits), the equilibrium becomes a fixed point of a coupled system. In practice, this coupling is often what yields emergent path selection.
ARP treats an edge as purely resistive (conductive). AIN augments each edge with additional passive degrees of freedom:
AIN uses the same reinforce/decay structure, but the driving signals differ:
with $\alpha_C,\mu_C,\alpha_L,\mu_L$ as analogous gains and decay rates.
For angular frequency $\omega$ (and assuming linear components at that instant), define:
\[Z_{ij}(t,\omega) = R_{ij}(t) + j\omega L_{ij}(t) + \frac{1}{j\omega C_{ij}(t)}.\]Using $R_{ij}=1/G_{ij}$,
\[\boxed{\ Z_{ij}(t,\omega) = \frac{1}{G_{ij}(t)} + j\omega L_{ij}(t) - \frac{j}{\omega C_{ij}(t)}.\ }\]This creates a network whose effective coupling depends on frequency and whose parameters adapt to local excitation.
In most computational uses, ARP/AIN are integrated in discrete time with step $\Delta t$.
To preserve nonnegativity, clamp:
\[G_{ij}^{k+1} \leftarrow \max(G_{\min},\, G_{ij}^{k+1}).\]A small $G_{\min}>0$ avoids division-by-zero when converting to resistance.
Then clamp to $C_{\min},L_{\min}$ similarly.
A conservative condition for explicit Euler stability on the decay term is:
\[\Delta t \lesssim \frac{1}{\mu_G},\quad \Delta t \lesssim \frac{1}{\mu_C},\quad \Delta t \lesssim \frac{1}{\mu_L}.\]In practice, use $\Delta t$ one order of magnitude smaller than the fastest decay constant.
ARP/AIN do not by themselves define the currents and voltages; they define how edge parameters adapt given those signals. A typical simulation loop is:
In a purely resistive analog, with node potentials $\phi_i$, edge current is
\[I_{ij} = G_{ij}(\phi_i-\phi_j).\]Potentials are determined by boundary conditions (sources/sinks) and Kirchhoff constraints.
In many routing-like setups:
This is a positive feedback loop that concentrates flow into certain structures, while the decay term prevents all edges from accumulating.
Below is a schematic of the coupled loop. “Solve flows” can be a circuit solve, a routing demand solve, or any rule that maps edge states to currents/voltages.
(boundary conditions / demand)
|
v
+-------------------+
| Solve flows |
| I_ij, V_ij |
+-------------------+
|
v
+-------------------+
| Adapt edges |
| dG/dt ~ |I| - G |
| dC/dt ~ |V| - C |
| dL/dt ~ |dI/dt|-L|
+-------------------+
|
v
updated (G,C,L)
|
+----> repeat
| If $ | I(t) | $ is bounded, $ | I(t) | \le I_{\max}$, then |
This comparison implies $G(t)$ remains bounded by a quantity on the order of $(\alpha_G/\mu_G)I_{\max}$ plus transient, assuming the flow solver does not create unbounded currents.
| Continuous-time ODE preserves nonnegativity if $G(0)\ge 0$ and $ | I | \ge 0$. |
Discrete time requires clamping or a positivity-preserving integrator to avoid numerical undershoot.
For physical interpretability, maintain:
\[G_{ij}(t) > 0,\quad C_{ij}(t) > 0,\quad L_{ij}(t) > 0.\]This does not “prove passivity” in the full nonlinear adaptive system, but it prevents trivial nonphysical parameter values.
Sections 7.1–7.3 are local guarantees (single-edge behavior under bounded driving) and numerical hygiene. The full ARP/AIN system is typically coupled:
As a result, global behaviors can include rapid concentration, slow drift, and (depending on the flow solver and constraints) oscillations or limit cycles.
Practical takeaway:
ARP parameters are interpretable through equilibrium and timescale.
Pick a decay constant $\mu_G$ such that
Example: if you want unused edges to halve in strength every 10 seconds, set $\mu_G = \ln 2 / 10$.
If typical active edge current magnitude is $I_{\mathrm{typ}}$ and you want typical active edge equilibrium conductance $G_{\mathrm{typ}}$, set
\[\alpha_G = \mu_G\,\frac{G_{\mathrm{typ}}}{I_{\mathrm{typ}}}.\]AIN parameters follow the same logic:
\[C_{\mathrm{eq}} = \frac{\alpha_C}{\mu_C}|V|,\quad L_{\mathrm{eq}} = \frac{\alpha_L}{\mu_L}\left|\frac{dI}{dt}\right|.\]Real implementations often add saturation to prevent runaway reinforcement under extreme signals:
\[\frac{dG}{dt} = \alpha_G \sigma(|I|) - \mu_G G,\]| where $\sigma$ is a bounded function (e.g., $\sigma(x)=\tanh(x/I_0)$ or $\sigma(x)=x/(1+x/I_0)$). This is optional; the canonical ARP form is linear in $ | I | $. |
Below is a minimal architecture that works across routing-like tasks.
| If AIN is enabled, update $C$ and $L$ similarly using $ | V | $ and $ | dI/dt | $. |
Consider two parallel paths from source $s$ to sink $t$.
Assume a fixed applied potential difference $\Delta\phi$ between $s$ and $t$ and ignore internal node structure so currents satisfy:
\[I_A = G_A\,\Delta\phi,\quad I_B = G_B\,\Delta\phi.\]ARP updates become:
\[\frac{dG_A}{dt} = \alpha_G |I_A| - \mu_G G_A = \alpha_G \Delta\phi\,G_A - \mu_G G_A = (\alpha_G\Delta\phi-\mu_G)G_A.\]Similarly for $G_B$.
This toy shows:
This motivates either:
The full ARP approach is best understood as “reinforce used edges while the system continuously recomputes flows”; the network structure then concentrates activity.
A practical rule of thumb: if a deployment shows runaway amplification, add (i) saturation, (ii) a conductance budget/normalization, or (iii) a source model that limits current supply.
ARP/AIN is not just a metaphor; it implies measurable behaviors.
If sources are turned off, conductances decay with time constant $1/\mu_G$. When sources are turned on again, previously reinforced edges re-amplify faster if they retained higher $G$.
Under repeated routing tasks, a sparse subgraph should carry most of the current while other edges decay toward $G_{\min}$.
| If AIN is active, repeated high-frequency excitation tends to increase $L$ (via $ | dI/dt | $), while repeated voltage excursions increase $C$. The resulting $Z(t,\omega)$ can develop band-like preferences. |
| Current vs conductance scatter $ | I_{ij} | $ vs $G_{ij}$. |
These are deliberately sharp statements intended to be testable:
ARP
\[\frac{dG_{ij}}{dt} = \alpha_G |I_{ij}| - \mu_G G_{ij}\] \[G_{ij,\mathrm{eq}} = \frac{\alpha_G}{\mu_G}|I_{ij,\mathrm{eq}}|\]AIN
\[\frac{dC_{ij}}{dt} = \alpha_C |V_{ij}| - \mu_C C_{ij}\] \[\frac{dL_{ij}}{dt} = \alpha_L \left|\frac{dI_{ij}}{dt}\right| - \mu_L L_{ij}\] \[Z_{ij}(t,\omega) = \frac{1}{G_{ij}(t)} + j\omega L_{ij}(t) - \frac{j}{\omega C_{ij}(t)}\]ANE (one canonical metric)
\[\mathrm{ANE}(t) = \frac{\sum_{ij} G_{ij}(t)\,|I_{ij}(t)|}{\sum_{ij}|I_{ij}(t)|}\]