---
title: Nonreciprocal Random Networks
url: https://www.emergentmind.com/topics/nonreciprocal-random-networks
type: topic
---

# Nonreciprocal Random Networks

Nonreciprocal random networks are systems in which directed interactions are disordered or heterogeneous and the influence of node \(i\) on node \(j\) is not constrained to match the reverse influence of \(j\) on \(i\). In graph-theoretic form this is the condition \(P_{i\to j}\neq P_{j\to i}\) for some ordered pairs [2509.05253]; in dynamical models it appears as a nonsymmetric random coupling matrix \({\mathbf M}^T\ne \mathbf M\) [2306.07609], antisymmetric cross-species couplings \(K_{AB}=-K_{BA}=K\) [2509.04164], or quenched directed bonds with \(J_{ij}=-J_{ji}\) on a fraction of lattice edges [2606.21582]. The topic spans directed random graph ensembles, percolation, oscillator populations, disordered spin systems, random wave transport, and nonreciprocal electrical networks. Across these settings, the central issue is not merely the presence of asymmetry, but how asymmetry is distributed across pairs, embedded in structure, and coupled to disorder.

## 1. Definitions, reciprocity, and representation dependence

A basic random-network formulation assigns a directed link probability \(P_{i\to j}\) to each ordered pair \(i\neq j\), with nonreciprocity defined by \(P_{i\to j}\neq P_{j\to i}\) [2509.05253]. For adjacency matrices \(A\), one standard reciprocity measure is
\[
\rho[A] = \frac{\sum_{i\ne j}(A_{ij}-\bar A)(A_{ji}-\bar A)} {\sum_{i\ne j}(A_{ij}-\bar A)^2}
= \frac{r-\bar A}{1-\bar A},
\]
where \(\bar A\) is the average adjacency entry and \(r=L^\leftrightarrow/L\) is the fraction of links that are reciprocated. In the two-probability nonreciprocal ensemble, the mean reciprocity is approximately
\[
\langle \rho\rangle \approx -\frac{(p-q)^2}{(p+q)(2-p-q)},
\]
which is negative whenever \(p\neq q\); the ensemble is therefore anti-reciprocal rather than merely nonreciprocal [2509.05253].

A complementary benchmark comes from pairwise-symmetric random networks. If connection probabilities satisfy \(P_{ij}=P_{ji}\), the relative occurrence of reciprocal connections compared with an Erdős–Rényi graph of the same mean density,
\[
\varrho=\frac{\mathbb{E}(P_{ij}P_{ji})}{\mu^2},
\qquad
\mu=\mathbb{E}(P_{ij}),
\]
reduces to
\[
\varrho=1+\frac{\mathrm{Var}(P_{ij})}{\mu^2}.
\]
Hence any non-degenerate heterogeneity in symmetric pairwise probabilities forces \(\varrho>1\), while only the completely homogeneous case gives \(\varrho=1\) [1609.04245]. This establishes a sharp distinction between asymmetric random networks, which can be anti-reciprocal, and symmetric-but-heterogeneous networks, where reciprocal overrepresentation is mathematically generic.

Randomness is also representation-dependent. In projected bipartite models, a completely random bipartite graph with independent Bernoulli edges can yield a projected monopartite network whose single-link probability
\[
p=1-(1-\beta^2)^N
\]
and large-\(N\) degree distribution are Erdős–Rényi-like, yet whose clustering and higher-order statistics remain systematically non-random because projection induces correlations through shared latent memberships [2304.12258]. This suggests that in nonreciprocal settings as well, observed pairwise statistics need not determine the effective higher-order structure.

## 2. Directed random graph ensembles and percolation structure

A canonical graph ensemble for nonreciprocal random networks assigns exactly two possible link probabilities, \(p\) and \(q\), to opposite directions of each node pair, so that if \(P_{i\to j}=p\), then \(P_{j\to i}=q\), with self-loops excluded [2509.05253]. The ensemble has two limiting organizations. In the unstructured case, the assignment of which direction receives \(p\) or \(q\) is random for each unordered pair, so nodes are statistically indistinguishable. In the structured fully transitive case,
\[
P_{i\to j}= \begin{cases} p, & i<j,\\ q, & i>j, \end{cases}
\]
and node position in the ordering becomes a genuine structural variable.

In the unstructured sparse regime,
\[
p=\frac{\lambda_p}{N-1},\qquad q=\frac{\lambda_q}{N-1},
\]
both in-degree and out-degree converge to Poisson distributions with mean
\[
\frac{\lambda_p+\lambda_q}{2}.
\]
At finite \(N\), however, the joint degree statistics are not trivial. The exact generating function
\[
G(x,y)= \left[(1-p+px)(1-q+qy)-\frac{p-q}{2}(x-y)\right]^{N-1}
\]
implies
\[
\mathrm{Cov}(K_{\mathrm{out}},K_{\mathrm{in}})
=-(N-1)\left[\left(\frac{p+q}{2}\right)^2-pq\right]\le 0.
\]
Thus finite-size in- and out-degrees are anti-correlated even when the marginals look Erdős–Rényi-like.

Percolation in the unstructured model collapses to the directed Erdős–Rényi result with effective mean degree \((\lambda_p+\lambda_q)/2\). The subcritical reachability equation yields the threshold
\[
\lambda_p+\lambda_q=2.
\]
Above threshold, the giant out-component and giant in-component satisfy
\[
O=I= 1+\frac{2}{\lambda_p+\lambda_q} W\!\left( -\frac{\lambda_p+\lambda_q}{2} e^{-(\lambda_p+\lambda_q)/2} \right),
\]
and the giant strongly connected component obeys
\[
S=O^2.
\]

The structured transitive ensemble behaves differently at every level. Writing \(\alpha=(i-1)/(N-1)\in[0,1]\), node \(\alpha\) has Poisson out-degree and in-degree distributions with means
\[
\langle K_{\mathrm{out},\alpha}\rangle =(1-\alpha)\lambda_p+\alpha\lambda_q,\qquad
\langle K_{\mathrm{in},\alpha}\rangle =\alpha\lambda_p+(1-\alpha)\lambda_q.
\]
There is therefore no single degree distribution for an actual node. If one averages over \(\alpha\), the result is a mixed Poisson–uniform law rather than an ordinary Poisson distribution, and the averaged “typical node” acquires
\[
\overline{\mathrm{Cov}(K_{\mathrm{out}},K_{\mathrm{in}})}
= -\frac{(\lambda_p-\lambda_q)^2}{12},
\]
even though the node-wise covariance averaged over the network is zero. This noncommutation of node- and ensemble-averaging is one of the paper’s main methodological warnings.

To analyze connectivity in both structured and unstructured cases, the percolation problem can be formulated through self-consistent reachability probabilities. In the locally tree-like subcritical regime, if \(P_{i\to\triangle j}\) is the probability that node \(i\) can reach node \(j\) by some directed path, then
\[
P_{i\to\triangle j} \approx P_{i\to j} + \sum_{k\ne i,j} P_{i\to k}P_{k\to\triangle j}.
\]
In the continuum limit this becomes an integral equation for \(f_{\alpha\to\triangle\beta}\), and above threshold it leads to node-dependent cumulative giant-component probabilities \(F_S(\alpha)\), \(F_O(\alpha)\), and \(F_I(\alpha)\) [2509.05253].

For the structured transitive network, the percolation threshold is no longer \(\lambda_p+\lambda_q=2\) but the nonlinear critical curve
\[
\lambda_q=\lambda_p e^{\lambda_q-\lambda_p},
\]
equivalently,
\[
\frac{\lambda_q-\lambda_p}{\ln(\lambda_q/\lambda_p)} < 1 \Rightarrow \text{unpercolated},\qquad
\frac{\lambda_q-\lambda_p}{\ln(\lambda_q/\lambda_p)} > 1 \Rightarrow \text{percolated}.
\]
The giant out- and in-components obey implicit equations
\[
e^{\lambda_q O}-1 = \left(e^{\lambda_p O}-1\right)e^{\lambda_q-\lambda_p},\qquad
e^{\lambda_q I}-1 = \left(e^{\lambda_p I}-1\right)e^{\lambda_q-\lambda_p},
\]
so \(O=I\), but in general
\[
S\neq OI.
\]
The transitive structure therefore enlarges the unpercolated region and invalidates naive one-parameter reductions based on a single averaged degree distribution.

## 3. Random nonreciprocal couplings in oscillator populations

A different but closely related class of nonreciprocal random networks is realized by populations of heterogeneous phase oscillators with quenched random couplings,
\[
\dot\theta_j=\omega_j+\frac{J}{N}\sum_{k=1}^N M_{jk}\sin(\theta_k-\theta_j),
\]
where the intrinsic frequencies \(\omega_j\) are drawn from a symmetric unimodal distribution \(g(\omega)\) centered at zero, \(J>0\) is a global coupling strength, and the random couplings \(M_{jk}\) can be synchronizing or anti-synchronizing [2306.07609]. The relevant order parameter is not the usual global synchronization amplitude, which remains small through the transition, but the local field
\[
P_j(t)\equiv r_j e^{i\psi_j}=\frac1N\sum_{k=1}^N M_{jk}e^{i\theta_k}.
\]
The “volcano transition” refers to a change in the distribution of these local fields from being peaked at the origin to having a ring-shaped maximum away from zero, associated with a quasi-glassy state.

Nonreciprocity is introduced by decomposing the random coupling matrix into symmetric and antisymmetric parts,
\[
\mathbf M=\frac12\left[(1+\eta)\mathbf S+(1-\eta)\mathbf A\right],
\]
with \(\eta\in[-1,1]\). Here \(\eta=1\) is fully symmetric, \(\eta=0\) corresponds to uncorrelated mirror couplings, and \(\eta=-1\) is fully antisymmetric. The mirror-entry correlation is
\[
\mathrm{corr}(M_{jk},M_{kj})= \frac{\langle M_{jk}M_{kj}\rangle}{\langle M_{jk}^2\rangle} =\frac{2\eta}{1+\eta^2},
\]
with variance
\[
\langle M_{jk}^2\rangle=L(1+\eta^2).
\]

Two analytically tractable low-rank random matrix models expose how the detailed organization of nonreciprocity affects collective dynamics. In Model 1, each oscillator carries two \(L\)-dimensional random binary vectors \(\mathbf u_j,\mathbf v_j\in\{\pm1\}^L\), and
\[
S_{jk}=\mathbf u_j\cdot\mathbf u_k-\mathbf v_j\cdot\mathbf v_k,\qquad
A_{jk}=\mathbf u_j\cdot\mathbf v_k-\mathbf v_j\cdot\mathbf u_k.
\]
The matrix \(\mathbf M\) has rank \(2L\), except at \(\eta=0\) where the rank drops to \(L\). In Model 2, the antisymmetric part is built from independent random vectors,
\[
S_{jk}=\mathbf u_j\cdot\mathbf u_k-\mathbf v_j\cdot\mathbf v_k,\qquad
A_{jk}=\tilde{\mathbf u}_j\cdot\tilde{\mathbf v}_k-\tilde{\mathbf v}_j\cdot\tilde{\mathbf u}_k,
\]
so \(\mathbf S\) and \(\mathbf A\) are statistically independent, and \(\mathbf M\) has rank \(4L\), except at \(\eta=\pm1\) where it reduces to \(2L\).

Because only the first harmonic appears in the continuity equation for the phase density, the Ott–Antonsen ansatz closes the dynamics. The resulting linear stability problem shows that the volcano transition survives only when reciprocity is sufficiently strong, but the threshold is model-dependent. In Model 1,
\[
J_v^{(1)}=\frac{2}{\pi g(0)\sqrt{\eta}}
\]
and the transition exists only for \(\eta>0\); if \(\eta<0\), the relevant eigenvalues are purely imaginary and there is no volcano transition. In Model 2,
\[
J_v^{(2)}=\frac{4}{1+\eta}\frac{1}{\pi g(0)},
\]
so the transition exists for the wider range \(\eta>-1\), including some negatively correlated mirror couplings. The central inference is explicit in the paper: the same mirror-correlation coefficient can have different dynamical consequences depending on how the random matrix is constructed.

Full-rank Gaussian random couplings exhibit the same qualitative suppression of the volcano transition by nonreciprocity, but not the same quantitative thresholds. In the reciprocal case \(\tau=1\), numerical simulations place the transition below the low-rank extrapolation \(\tilde J_v=2/[\pi g(0)]\); for \(\tau<1\), the transition persists, the critical coupling increases as \(\tau\) decreases, the volcano width shrinks as reciprocity weakens, and the numerical divergence occurs near \(\tau\approx0.19\). Low-rank solvable models are therefore analytically revealing but not quantitatively universal.

## 4. Quenched disorder, nonreciprocal bonds, and nonequilibrium criticality

Nonreciprocal random networks also arise as spatially extended nonequilibrium spin systems. One formulation is the kinetic random-field nonreciprocal Ising model, a two-species system with spins \(\sigma_i^\alpha=\pm1\), same-species ferromagnetic coupling \(J\), antisymmetric cross-species coupling
\[
K_{AB}=-K_{BA}=K,
\]
and a site-dependent bimodal random field
\[
P(h_i)=\frac{1}{2}\big[\delta(h_i-h)+\delta(h_i+h)\big]
\]
[2509.04164]. The dynamics are single-spin-flip Glauber kinetics,
\[
\omega(\sigma_i^\alpha \to -\sigma_i^\alpha) = \frac{1}{2\tau}\left[1-\tanh\left(\frac{\Delta E_i^\alpha}{2k_B T}\right)\right].
\]
In mean-field form, the magnetizations obey
\[
\frac{d M_\alpha(t)}{dt} = -M_\alpha +\frac{1}{2} \Big[ \tanh(M_\alpha + K_{\alpha\beta} M_\beta + h) + \tanh(M_\alpha + K_{\alpha\beta} M_\beta - h) \Big].
\]

The relevant phase diagnostics are
\[
R=\sqrt{\frac{M_A^2+M_B^2}{2}},\qquad
S=M_A\frac{\partial M_B}{\partial t}-M_B\frac{\partial M_A}{\partial t}.
\]
Here \(R\) measures overall alignment amplitude, while \(S\) measures rotation in the \((M_A,M_B)\) plane. The characteristic nonequilibrium ordered state is the swap phase, a limit cycle in which the two species oscillate out of phase.

Its onset changes qualitatively with disorder strength. For \(\tilde h<\tilde h_c\), the disordered fixed point loses stability through a supercritical Hopf bifurcation and the limit cycle emerges continuously. For \(\tilde h>\tilde h_c\), the transition occurs through a saddle-node-of-limit-cycle (SNLC) bifurcation, with discontinuous jumps in \(R\) and \(S\), hysteresis, and a dip in the Binder cumulant. The change between these regimes is a nonequilibrium tricritical, or Bautin, point. The paper reports \(\tilde h_c\approx0.658\) in mean-field theory, \(1.35<\tilde h_c<1.40\) in effective-field theory, and a tricritical window \(2.0<\tilde h_c<2.2\) in 3D kinetic Monte Carlo. Finite-size scaling of the susceptibility yields \(\gamma/\nu \approx 1.968\pm 0.060\) at \(\tilde h=2.0\), and \(\gamma/\nu \approx 3.055\pm 0.069\) and \(3.076\pm 0.039\) at \(\tilde h=2.2\) and \(2.6\), respectively, consistent with continuous and first-order regimes. In the first-order regime the swap phase survives only above a threshold nonreciprocity \(\tilde K_c(\tilde h)\), which increases monotonically with disorder. At high disorder and subcritical nonreciprocity, the model also exhibits a droplet-induced swap phase cycling through eight metastable states.

A second spatially extended formulation places quenched nonreciprocity directly on a 2D square-lattice Ising ferromagnet. For each nearest-neighbor pair \(\langle i,j\rangle\),
\[
(J_{ij},J_{ji})= \begin{cases}
(+1,+1) & \text{with probability } 1-p,\\
(+1,-1) & \text{with probability } p/2,\\
(-1,+1) & \text{with probability } p/2,
\end{cases}
\]
so reciprocal ferromagnetic bonds are mixed with nonreciprocal bonds satisfying \(J_{ij}=-J_{ji}\) [2606.21582]. The Glauber flip rate is
\[
P_{\sigma_i\to -\sigma_i} = \frac{1-\sigma_i\tanh(h_i/T)}{2},\qquad
h_i=\sum_j J_{ij}\sigma_j.
\]

This model has a continuous nonequilibrium transition between an ordered ferromagnetic phase and a disordered paramagnetic phase, with the transition line extending to \(T=0\) at finite disorder density. Numerically the zero-temperature threshold is near \(p_c\simeq0.20\), while discrete mean-field theory gives
\[
p_c^{\mathrm{MF}}=\frac{1}{3}.
\]
A gauge-invariance argument shows that
\[
p_c(T)\le \frac{1}{2}
\]
for all \(T\ge0\): disorder-averaged long-range correlations can survive only if reciprocal bonds percolate, and the square-lattice percolation threshold is \(1/2\).

The most distinctive result is dynamical. Unlike equilibrium disordered Ising models, the zero-temperature disordered phase remains active rather than freezing. Because nonreciprocity destroys the relation between local selfish energy changes and a global energy function, deterministic \(T=0\) flips can persist in steady state. The paper identifies explicit \(2\times2\) nonreciprocal plaquette motifs with zero boundary field that execute an 8-step deterministic cycle, and more generally an infinite family of such motifs. Rare low-disorder regions locally order and reverse on broad timescales, producing a stretched-exponential autocorrelation tail
\[
C_t(t)\sim \exp\!\left[-(t/\tau^{(\mathrm{str})})^\alpha\right],\qquad \alpha<1.
\]
In the ordered phase, domain growth crosses over from curvature-driven
\[
L(t)\sim t^{1/2}
\]
to logarithmic activated coarsening
\[
L(t)\sim \ln t.
\]
Taken together, these results show that quenched nonreciprocity can both suppress static order and sustain athermal dynamics.

## 5. Random wave transport and localization in nonreciprocal media

In wave systems, nonreciprocal random networks appear as disordered propagation media whose transfer structure is direction-dependent. A representative case is one-dimensional random layered media composed of alternating layers, some magneto-optical, with random layer thicknesses and transfer-matrix dynamics [1110.6511]. The transmission coefficient for left incidence is
\[
T=\frac{1}{(\hat T)_{11}},\qquad \mathcal T=|T|^2,
\]
and the standard transmission decrement is
\[
\kappa = -\frac{\langle \ln \mathcal T\rangle}{2N}.
\]
In the localized regime,
\[
\ell_{\rm loc}^{-1} \approx \kappa.
\]

Under the short-wavelength condition
\[
\frac{\lambda_k}{2} \ll d,
\]
the layer phases are effectively random, and the localization length is governed primarily by interfaces:
\[
\kappa = \ell_{\rm loc}^{-1} = \ln\left|(\hat F^{ab})_{11}(\hat F^{ba})_{11}\right|.
\]
Nonreciprocity arises only when there are two uncoupled modes \(\sigma=\pm1\) such that time reversal maps \((\upsilon,\sigma)\mapsto(-\upsilon,-\sigma)\). Then reciprocal waves can have different transmittances,
\[
\mathcal T^\sigma \neq \mathcal T^{-\sigma}.
\]

In the Faraday geometry, with magnetization parallel to propagation, the eigenmodes are circularly polarized and the wavenumber becomes
\[
k = nk_0\sqrt{1+\sigma q} \simeq nk_0\left(1+\frac{\sigma q}{2}\right).
\]
The localization decrement acquires a first-order magneto-optical correction,
\[
\kappa^{(1)} = \frac{\sigma}{2}(q_a-q_b)\frac{n_a-n_b}{n_a+n_b},
\]
so the typical transmission difference is exponentially amplified as
\[
\mathcal T \sim e^{-2N\kappa},
\qquad
\exp[-2N\kappa^{(1)}].
\]
Numerically, the paper reports strong resonance splitting, with wavelength shift roughly
\[
\Delta \lambda_{\rm res}\sim -\sigma\,3.6\ \text{nm},
\]
large compared with the resonance width.

In the Voigt geometry, magnetization is transverse to propagation and the first-order magneto-optical effect enters mainly through phase rather than propagation constant. After phase averaging, the averaged localization decrement is reciprocal to \(O(q)\), and nonreciprocity survives mainly in individual TM-polarized transmission resonances. The estimated resonance shift is much smaller,
\[
\Delta \lambda_{\rm res}\sim -\sigma\,3\times 10^{-4}\ \text{nm}.
\]
Disorder therefore does not have a uniform effect: in one geometry it strongly amplifies broadband nonreciprocal transport, while in the other it leaves only weak resonance-level asymmetry.

## 6. Engineered nonreciprocal electrical networks: synthesis and quantization

Nonreciprocal networks are also studied as designed electrical systems rather than stochastic graphs. For lossless two-port microwave networks, coupling-matrix synthesis can be extended from reciprocal to nonreciprocal cases by allowing complex couplings [1408.2321]. The general lossless condition is
\[
[Y]^* + [Y]^t = 0,
\]
so the admittance matrix is skew-Hermitian symmetric rather than purely imaginary. For any lossless two-port, even when nonreciprocal,
\[
|S_{21}| = |S_{12}|,
\]
while the phases may differ,
\[
\angle S_{21} \neq \angle S_{12}.
\]

Minimum-order synthesis is obtained by choosing the polynomial ratio
\[
\frac{P_{21}(s)}{P_{12}(s)^\star}=c,\qquad |c|=1,
\]
often with \(c=\pm1\), so that
\[
P_{21}(s)=\pm P_{12}(s)^\star,\qquad
F_{11}(s)=\mp F_{22}(s)^\star.
\]
The crucial building block is the generalized lossless complex inverter with admittance matrix
\[
[Y] = j \begin{bmatrix} 0 & \widetilde{M}\\ \widetilde{M}^* & 0 \end{bmatrix},
\]
which reduces to a real inverter when \(\widetilde M\) is real and to a gyrator when \(\widetilde M\) is purely imaginary. Complex similarity transformations
\[
[\widetilde{M}]' = [R]\,[\widetilde{M}]\,([R]^*)^t
\]
convert the initial nonreciprocal transversal matrix into folded or star topologies, and phase rotations remove unnecessary complex phases. The final realization uses real inverters plus the minimum number of gyrators, equal to the nonreciprocity order of \(P_{21}(s)/P_{12}(s)\). In the second-order example treated explicitly in the paper, all but one coupling can be made reciprocal/real, leaving a single gyrator.

At the Hamiltonian level, exact quantization of nonreciprocal quasi-lumped electrical networks requires a geometric formulation that does not assume purely node-flux coordinates [2401.09120]. The branch state is represented by fluxes and charges,
\[
\phi^b,\qquad q^b,\qquad \dot{\phi}^b=v^b,\qquad \dot{q}^b=i^b,
\]
on an initial manifold
\[
\mathcal{M}_{2B}=\mathbb{R}^{2B-k}\times (S^1)^k.
\]
Kirchhoff laws together with transformer and gyrator constraints define a linear Pfaffian system
\[
\mathsf{F}\,\boldsymbol{\zeta}=0,\qquad \boldsymbol{\zeta}^T=(\boldsymbol{\phi}^T,\boldsymbol{q}^T).
\]
The fundamental object is the pre-symplectic two-form
\[
\omega_{2B},
\]
which, after immersion to the constrained manifold, yields a reduced form
\[
\omega=\imath^*\omega_{2B}=\tfrac12\,\omega_{\alpha\beta}\,dz^\alpha\wedge dz^\beta.
\]
The resulting coordinates \(z^\alpha\) are generally mixtures of flux- and charge-type variables.

When \(\det(\omega_{\alpha\beta})=0\), the paper applies the Faddeev–Jackiw method to classify zero modes as removable superfluous variables, gauge modes, or non-homogeneous-rank singularities. After reduction to a nondegenerate symplectic form, canonical quantization proceeds via
\[
[\hat\phi,\hat q]=i\hbar.
\]
The same formalism extends to transmission lines, multiport blackboxes, and frequency-dependent nonreciprocal linear systems. For example, nonreciprocal admittances of the form
\[
\mathsf Y(s)=\frac{\mathsf L^{-1}}{s}+\mathsf C\,s+\mathsf Y_g+\sum_k \frac{\mathsf D_k s+\mathsf E_k}{s^2+\Omega_k^2}
\]
and gyrator blocks
\[
\mathsf Y_g = G\, i\sigma_y
\]
can be incorporated into an exact enlarged Hamiltonian description. The framework also extends Caldeira–Leggett constructions to dissipative nonreciprocal multiports and yields a non-divergent input-output theory with intrinsic ultraviolet cutoff, explicitly relevant to superconducting circuits and chiral waveguide QED.

Taken together, these engineered-network results show that nonreciprocity is compatible with both exact synthesis and canonical quantization, provided the network is formulated in terms of the correct complex couplings or reduced symplectic manifold rather than by reciprocal circuit heuristics.

Source: https://www.emergentmind.com/topics/nonreciprocal-random-networks