--- title: Branching Diffusion with Interactions url: https://www.emergentmind.com/topics/branching-diffusion-with-interactions type: topic --- # Branching Diffusion with Interactions Searching arXiv for recent and foundational papers on interacting branching diffusions to ground the article. Branching diffusion with interactions denotes a class of stochastic population models in which individuals move in continuous space, typically by diffusion, while branching, death, or selection events depend on the current population configuration rather than on independent per-particle mechanisms alone. Recent formulations represent the population either by a finite atomic measure on an Ulam–Harris tree or by a weighted empirical measure, and place the interaction in rank, local occupancy, empirical-law dependence, pairwise catalytic terms, or nonlinear competition [2404.12964] [2601.11294] [2505.08563]. In this sense, the subject includes rank-based branching Brownian systems, McKean–Vlasov branching diffusions, mutually catalytic branching, spatially self-regulating diffusions, and multitype interacting branching processes, together with their hydrodynamic, mean-field, and genealogical limits. ## 1. State spaces, labeling schemes, and martingale formulations A common representation uses Ulam–Harris labels to encode genealogies. In the controlled interacting branching-diffusion framework, the label set is \[ \mathcal I_c=\{\varnothing\}\;\cup\;\bigcup_{n\ge1}\mathbb N^n, \] and the state space is \[ E\;=\;\Bigl\{\lambda=\sum_{i\in\mathcal V}\delta_{(i,x_i)}:\mathcal V\subset\mathcal I_c\text{ finite},\;x_i\in\mathbb R^d\Bigr\}, \] endowed with the weak-* topology. A controlled branching-diffusion is then a càdlàg, adapted \(E\)-valued process satisfying a martingale problem in which each labeled particle evolves by a diffusion, branches according to a configuration-dependent death rate, and produces \(k\) offspring with probability \(p_k(i,x,\lambda,a)\) [2601.11294]. An equivalent population-level notation appears in McKean–Vlasov branching diffusion. There one writes \[ Z_t=\sum_{k\in\mathcal K_t}\delta_{(k,X^k_t)}, \] with \(\mathcal K_t\) the set of alive particles and \(\mu_t=\mathrm{Law}(Z_t)\) the law of the whole branching process up to time \(t\). Each particle follows \[ dX^k_t = b\bigl(t,X^k_{\cdot\wedge t},\mu_t\bigr)\,dt + \sigma\bigl(t,X^k_{\cdot\wedge t},\mu_t\bigr)\,dW^k_t, \] dies at rate \(y(t,X^k_{\cdot\wedge t},\mu_t)\), and at death produces a random number of offspring with law \(p^\ell(t,X^k_{\cdot\wedge\tau},\mu_\tau)\) [2404.12964]. For large but finite populations, weighted empirical measures are standard. In the rank-based Brownian system one records \[ \mu^K_t = \frac1K\sum_{i\in V^K_t}\delta_{X^i_t}, \] where \(V_t^K\) is the set of alive labels and \(\langle\mu_t^K,1\rangle=N_t^K/K\) is the rescaled population size [2505.08563]. In mean-field interacting branching diffusions with logistic binary branching one likewise writes \[ \mu_t^K= \frac1K\sum_{n=1}^{N_t^K} \delta_{X_t^{n,K}}, \] so that total mass remains of order \(O(1)\) under the \(1/K\)-weighting [2101.04099]. Across these formulations, martingale problems provide the unifying analytic language. In the rank-based model, for each \(f\in C_b^{1,2}(\mathbb R_+\times\mathbb R)\), \[ \langle\mu^K_t,f_t\rangle =\langle\mu^K_0,f_0\rangle +\int_0^t\!\langle\mu^K_s,\partial_sf_s+L_{\mu^K_s}f_s+b(\cdot,\mu^K_s)f_s\rangle\,ds +M^{K,f}_t, \] with a local generator \(L_\nu\) and a martingale \(M^{K,f}\) whose quadratic variation is \(O(1/K)\) [2505.08563]. In the McKean–Vlasov setting, weak solutions are characterized by the martingale property of \[ M^\phi_t = \phi(Z_t)-\phi(Z_0)-\int_0^t \mathcal L^{\mu_s}\phi(Z_s)\,ds \] for smooth test functions \(\phi\), where \(\mathcal L^m\) is a nonlinear generator acting on the current configuration [2404.12964]. ## 2. Principal interaction mechanisms One major class is rank-based interaction. In the “Go or Grow” system, the rank of a particle in a configuration \(\{x_1,\dots,x_N\}\) is \[ \mathrm{rank}(x_i)=\#\{j:x_j\ge x_i\}. \] The \(K\) rightmost particles, with ranks \(1,\dots,K\), are in the Grow regime and branch at rate \(1\) each, while all others are in the Go regime, drift rightward with speed \(\chi>0\), and do not branch [2505.08563]. The motion between branching events is Brownian with piecewise-constant drift, \[ dX^i_s=\chi\,a(x,\nu)\,ds+\sqrt2\,dW^i_s, \] where \(a(x,\nu)\) indicates whether the particle has rank \(>K\) [2505.08563]. The model is explicitly tied to the Go or Grow hypothesis for cells in a capillary tube moving upwards a chemical gradient. A second class is mean-field interaction through smoothed empirical fields. In the quantitative mean-field model, each particle solves \[ dX_t^{n,K} = b\bigl(X_t^{n,K},H*\mu_t^K(X_t^{n,K})\bigr)\,dt +\sigma\bigl(X_t^{n,K},G*\mu_t^K(X_t^{n,K})\bigr)\,dB_t^n, \] with bounded Lipschitz kernels \(H,G\), birth rate \(r>0\), and death rate \((c/K)N_t^K\) [2101.04099]. The interaction therefore enters both the diffusion coefficients and the logistic mass regulation. A third class is nonlinear dependence on the law of the whole branching process. In McKean–Vlasov branching diffusion, the coefficients \(b,\sigma\), the death rate \(y\), and the offspring laws \(p^\ell\) depend on \(\mu_t=\mathrm{Law}(Z_t)\), so the interaction is measure-valued and path-dependent [2404.12964]. Under boundedness, measurability, and Lipschitz conditions in path and measure under \(W_1\), strong well-posedness is obtained [2404.12964]. Catalytic interaction produces a different mechanism. In finite-rate two-type mutually catalytic branching on a countable site space \(S\), \[ \begin{aligned} du_t(i)&= (Au_t)(i)\,dt+\sqrt{\gamma\,u_t(i)\,v_t(i)}\;dB^1_t(i),\\ dv_t(i)&= (Av_t)(i)\,dt+\sqrt{\gamma\,u_t(i)\,v_t(i)}\;dB^2_t(i), \end{aligned} \] where \(\mathrm{Cov}(B^1_t(i),B^2_t(i))=\rho\,t\) and \(\rho<0\) implements negative correlation [1109.6105]. In the discrete particle counterpart on \(\mathbb Z^d\), a type-1 particle at \(x\) branches at instantaneous rate \(y\,\eta_t(x)\), and symmetrically a type-2 particle branches at rate \(y\,\xi_t(x)\), where the catalyst is the local population of the opposite type [2310.17497]. Competition and self-regulation give another canonical structure. In the small-mass diffusion limit of a multitype locally self-regulating model, \[ dx^m_{\xi}(t) = \sum_{\eta\in G}\bar a(\xi,\eta)\bigl(x^m_\eta-x^m_\xi\bigr)\,dt + \gamma^m\,x^m_\xi(t)\,\Bigl(K^m-\sum_{n=1}^M\lambda_{m,n}x^n_\xi(t)\Bigr)\,dt + \sqrt{\gamma^m\,x^m_\xi(t)}\,dw^m_\xi(t), \] so migration, local carrying capacities, and cross-type competition are all explicit in the drift [1509.04023]. A related one-dimensional competition limit is \[ Z^x_t = x+\int_0^t f(Z^x_s)\,ds+2\int_0^t\sqrt{Z^x_s}\,dW_s, \] where \(f\) encodes net births minus deaths including competition [1304.0710]. Finally, interaction may act through the center of mass. In the dyadic branching diffusion with Ornstein–Uhlenbeck drift and mean-field attraction or repulsion, \[ dZ^i_t = dW^i_t-b\,Z^i_t\,dt+\frac{\gamma}{2^m}\sum_{j=1}^{2^m}(Z^j_t-Z^i_t)\,dt \quad\text{for }t\in[m,m+1), \] so the overall drift is the combination of Ornstein–Uhlenbeck confinement or expansion and pairwise mean-field attraction or repulsion [1610.02088]. ## 3. Large-population, hydrodynamic, and scaling limits The rank-based Go or Grow system admits a hydrodynamic limit as \(K\to\infty\). Writing \(u^K_t(dx)=\mu_t^K(dx)\), defining the cumulative tail \[ F(t,x)=\int_x^\infty u(t,y)\,dy, \] and weighting individuals by \(1/K\), the limit is the nonlinear integro-PDE \[ \partial_tu = \partial_{xx}u -\chi\,\partial_x\Bigl(u\,\mathbf 1_{\{F>1\}}\Bigr) + u\,\mathbf 1_{\{F\le1\}}, \] or, in terms of \(F\), \[ \partial_tF = \partial_{xx}F-\chi\,\partial_xA(F)+B(F), \qquad A(z)=(z-1)_+,\quad B(z)=\min\{z,1\}, \] with tightness and convergence established under finite third-moment initial data and vague convergence of the initial empirical measures [2505.08563]. Mean-field interacting branching diffusions with logistic binary branching admit both qualitative and quantitative limits. Under i.i.d. initial atoms, Lipschitz coefficients, and bounded Lipschitz kernels, one has \(\mu^K\Rightarrow\mu\) in \(D([0,T],\mathcal M^+(\mathbb R^d))\), where \(\mu\) is the unique continuous finite-measure solution of \[ \langle\mu_t,f(t,\cdot)\rangle = \langle\mu_0,f(0,\cdot)\rangle +\int_0^t\langle\mu_s,\partial_s f + L_{\mu_s}f + (r-c\langle\mu_s,1\rangle)f\rangle\,ds. \] Its total mass \(n_t=\langle\mu_t,1\rangle\) solves the logistic ODE \(dn_t=(r-c\,n_t)n_t\,dt\) [2101.04099]. With higher moments, the convergence rate is explicit: \[ \sup_{t\in[0,T]} E\bigl(\|\mu_t^K-\mu_t\|_{BL^*}\bigr) \le C_T\,[\,I_4(K)+R_{d,q}(K)^{1/2}\,]. \] The proof uses a coupling with an auxiliary process of McKean–Vlasov diffusions and optimal transport at branching times [2101.04099]. McKean–Vlasov branching diffusion itself is obtained as a large-population limit. For the \(N\)-particle interacting system in which coefficients depend on the empirical law, the law of the empirical measure is tight in \(\mathcal P(\mathcal P_1(D([0,T],E)))\), and any limit point is supported on solutions of the McKean–Vlasov martingale problem. If strong uniqueness holds, the entire sequence converges to the law of the unique McKean–Vlasov branching diffusion [2404.12964]. Several interacting branching models admit discrete-to-continuous scaling limits. In the multitype interacting branching framework, if \(Z^{(n)}\) is the discrete-state interacting multitype branching process and \(a_n,b_n\to\infty\) satisfy the scaling assumptions, then \[ \bigl(a_n^{-1}\,Z^{(n)}_{\,b_n\,t}\bigr)_{t\ge0} \Longrightarrow (Y_t)_{t\ge0} \quad\text{in }D([0,\infty),\mathbb R^d), \] where \(Y\) is the unique strong solution of the continuous-state interacting branching SDE with quadratic interaction terms [2203.09701]. In branching processes with competition, the renormalized population size converges to the nonlinear diffusion \(Z^x\), and the whole family \(\{Z^x_t:t\ge0,x\ge0\}\) is represented by local times of a reflected Brownian motion with local-time-dependent drift, yielding a generalized Ray–Knight theorem [1304.0710]. The finite-system scheme for infinite-rate mutually catalytic branching has a different time scale because the jumps lack finite second moments. For the mean-field site space \(S_N=\{1,\dots,N\}\), the correct scale for the total masses is \[ B_N = N/\log N, \] and the rescaled averages \(\widetilde Z_i^{(N)}(t)=Z_i^{(N)}(B_N t)\) converge to \[ dZ_i(t)=\sqrt{\gamma^*\,Z_1(t)\,Z_2(t)}\,dW_i(t), \qquad \gamma^*=8/\pi, \] with joint convergence of total masses and coordinate processes under the finite-system scheme [1510.01154]. ## 4. Traveling waves, invasion speeds, and genealogical front classification For the rank-based Go or Grow limit PDE, traveling waves are obtained from the ansatz \(u(t,x)=U(z)\), \(z=x-ct\), with cumulative tail \(G(z)\). The resulting ODE admits a minimal wave speed \[ c^*(\chi) =\begin{cases} \chi+\frac1\chi,&\chi>1,\\[0.5ex] 2,&\chi\le1, \end{cases} \] such that for each \(c\ge c^*(\chi)\) there is a monotone profile \(U_c(z)\ge0\) with \(U_c(-\infty)=1\) and \(U_c(+\infty)=0\), while no such wave exists for \(c1\), the front is pushed and \(c^*=\chi+1/\chi>2\) [2505.08563]. Microscopic finite-\(K\) simulations use the position of the \(K\)-th largest particle, \[ \xi_t^K=H_K(\mu_t^K), \] and the empirical speed estimator \[ \widehat v^K(T_0,T)=\frac{\xi_T^K-\xi_{T_0}^K}{T-T_0}. \] For each fixed \(K\), \(\xi_t^K\) grows roughly linearly in \(t\). As \(K\) grows, \(\widehat v^K\) concentrates around a deterministic limit, and numerically \[ \lim_{K\to\infty}\widehat v^K\approx c^*(\chi), \] with small finite-\(K\) corrections in the pulled regime \(\chi\le1\), reflecting the Bramson logarithmic shift [2505.08563]. The same model supports a genealogical criterion for pulled versus pushed fronts. If a particle sampled at time \(T\) is followed backward along its unique ancestral lineage, and the lineage is observed in the moving frame of the front, then Conjecture 5.5 states that for large \(K\) \[ d\hat Y_s=\beta(\hat Y_s)\,ds+\sqrt2\,dW_s, \qquad \beta(z)=c^*-\chi\,\mathbf 1_{\{z<0\}}+2\,\partial_z(\ln U_{c^*}(z)). \] In the pushed case \(\chi>1\), the adjoint operator admits a unique integrable invariant density \[ v_\infty(z) = \frac{\chi^2-1}{\chi^3}\exp\!\Bigl(\int_0^z\beta(y)\,dy\Bigr), \] and numerically the lineage histogram closely matches \(v_\infty\). In the pulled case \(\chi\le1\), the formal stationary solution is not integrable, so the lineage law drifts to \(+\infty\); ancestors come only from the ever-advancing tip. Exactly at \(\chi=1\), a hybrid “pushmi-pullyu” behavior is observed: the mode stays near \(0\), but the mean drifts out, in agreement with the Bramson correction exponent \(1/2\) [2505.08563]. A related rank-based branching-selection system also converges to a reaction-diffusion equation. In the \((\psi,r,N)\)-BBM family, the hydrodynamic limit is \[ U_t=\tfrac12 U_{xx}+r(t)\,G(U), \qquad G(U)=\int_{1-U}^{1}\psi(s)\,ds, \] and under additional assumptions the particle system has an asymptotic velocity \[ c_N = V_2-\frac{\pi^2}{2\,V_2[\log N]^2}+o\!\bigl((\log N)^{-2}\bigr), \qquad V_2=2\sqrt{G'(0)}=2. \] The minimal spreading speed of the PDE is exactly \(V_2\), and \(c_N\to V_2\), which is described there as a partial weak-selection principle [2605.04860]. ## 5. Long-time behavior, coexistence, extinction, and regulation In mutually catalytic branching with negative correlations, the long-time behavior is governed by the transience or recurrence of the migration chain. For each fixed \(\gamma\in(0,\infty]\) and \(\rho<0\), coexistence of the two types occurs if and only if the underlying migration chain is transient; equivalently, if \[ \sup_{i\in S}G(i,i)<\infty, \qquad G(i,j)=\int_0^\infty P^i\{X_s=j\}\,ds. \] Otherwise, in the recurrent case, \[ \lim_{t\to\infty}\langle u_t,1\rangle\,\langle v_t,1\rangle=0 \quad\text{a.s.} \] In particular, on \(S=\mathbb Z^d\), coexistence occurs precisely for \(d\ge3\) [1109.6105]. The discrete particle system with mutually catalytic interactions on \(\mathbb Z^d\) has the same phase boundary. For nearest-neighbor random walk, \(d\ge3\) implies coexistence possible, while \(d\le2\) implies coexistence impossible [2310.17497]. On finite tori, the empirical total masses converge after renormalization to the mutually catalytic diffusion \[ dX_t=\sqrt{y\,X_tY_t}\,dW_t^1, \qquad dY_t=\sqrt{y\,X_tY_t}\,dW_t^2 \] in dimensions \(d\ge3\), under the transience bound \(y<1/g(0)\) [2310.17497]. In branching diffusion with Ornstein–Uhlenbeck drift and pairwise attraction or repulsion, the center of mass decouples from the interaction parameter \(\gamma\). Writing \[ Z_t=\frac1{2^{[t]}}\sum_{i=1}^{2^{[t]}} Z_t^i, \] one obtains \[ dZ_t=2^{-[t]/2}\,dW_t-b\,Z_t\,dt. \] Hence, almost surely, \(Z_t\to0\) when \(b>0\), whereas \(e^{bt}Z_t\to N\) when \(b<0\), with \(N\) Gaussian [1610.02088]. For the relative system \(Y_t^i=Z_t^i-Z_t\), if \(b+\gamma>0\) then the empirical measure converges almost surely to \[ N\!\Bigl(0,\,(2(b+\gamma))^{-1}I_d\Bigr), \] while for \(b<0\) local extinction always occurs: every fixed bounded set is eventually empty [1610.02088]. Self-regulation and competition replace coexistence thresholds by carrying-capacity effects. In the multi-type spatial branching model for local self-regulation, the diffusion limit has local logistic drift \(\gamma^m x^m_\xi(K^m-\sum_n\lambda_{m,n}x^n_\xi)\), and the abstract states that the densities for large times do not depend on the initial distribution but mainly on the carrying capacities [1509.04023]. In the one-dimensional competition model, \[ Z_t^x = x+\int_0^t f(Z_s^x)\,ds+2\int_0^t\int_0^{Z_s^x}W(ds,du), \] subcriticality and extinction are characterized by \[ \Lambda(f) := \int_1^\infty \exp\!\Bigl(-\tfrac12\!\int_1^u\!\tfrac{f(r)}r\,dr\Bigr)\,du = \infty \quad\Longleftrightarrow\quad Z^x\text{ dies out a.s.} \] [1304.0710]. These results show that interacting branching diffusions exhibit several distinct long-time regimes: coexistence driven by transience, extinction driven by recurrence or outward drift, regulation by carrying capacities, and front propagation governed by minimal-wave-speed selection. ## 6. Duality, spine methods, control, and structural representations Duality is a principal tool when interaction destroys the classical branching property. For negatively correlated mutually catalytic branching, the exact mixed second-moment formula is \[ \mathbb E\bigl[u_t(i)\,v_t(j)\bigr] = \mathbb E^{i,j}\Bigl[u_0(X_t^1)\,v_0(X_t^2)\,e^{\,\rho\,\gamma\,L_t}\Bigr], \] where \(L_t\) is the collision local time of two independent migration chains [1109.6105]. Since \(\rho<0\), the factor \(e^{\rho\gamma L_t}\) decays on collisions, which is the key mechanism behind extinction in the recurrent case [1109.6105]. The same paper combines this with exit-time estimates for correlated Brownian motion in the first quadrant and uniform integrability bounds independent of the branching rate, allowing passage to the infinite-rate limit [1109.6105]. Exponential duality also appears in multitype self-regulating systems. In the exchangeable case, Greven–Sturm–Winter–Zähle construct a dual process \((\alpha,\kappa)\) and the duality function \[ H\bigl((\alpha,\kappa),x\bigr) = \exp\Bigl(-\sum_{\xi}\alpha_\xi\bar x_\xi\Bigr)\,x^\kappa, \] which yields uniqueness of the exchangeable diffusion martingale problem and supports equilibrium and coexistence analysis [1509.04023]. For mutually catalytic branching particle systems on finite tori, approximate self-duality based on \[ H\bigl((\xi,\eta),(u,v)\bigr) = \exp\!\bigl[-\langle\xi+\eta,u+v\rangle - i\,\langle\xi-\eta,u-v\rangle\bigr] \] is central in identifying the finite-system-scheme limit [2310.17497]. Spine constructions provide another structural method. For continuous type-space branching processes with interactions, Medous introduces a Girsanov-type change of measure that singles out a distinguished lineage and proves the many-to-one identity \[ \E_{\bar z}\Bigl[\sum_{u\in\mathcal G(t)}\psi(X_t^u,Z_t,t)\, H\bigl(u,(Z_s)_{0\le s\le t}\bigr)\Bigr] = \langle z,\psi(\cdot,z,0)\rangle\, \widehat\E_{\bar z}\Bigl[\xi_t\, H\bigl(E_t,(\widehat Z_s)_{s\le t}\bigr)\Bigr], \] with path-integral weight \(\xi_t\) determined by the nonlinear operator \(\mathcal G\) [2309.15449]. If \(\mathcal G\psi=\lambda\psi\), then \(\xi_t=e^{\lambda t}\), recovering a classical many-to-one formula. The same framework yields a generalized continuous-time Kesten–Stigum theorem with interactions and an alternative spine construction for exact simulation of stochastic size-dependent populations [2309.15449]. A complementary structural representation is the generalized Ray–Knight theorem for branching with competition. Under Hypothesis B, the reflected Brownian motion \[ H^z_s = B_s + \tfrac12 \int_0^s f'\bigl(z(H^z_r)+L^z_r(H^z_r)\bigr)\,dr + \tfrac12\,L^z_s(0) \] has local times whose inverse-local-time slices encode the diffusion \(Z^{x,z}\), and for \(z\equiv0\), \[ \{L_{S_x}(t):t\ge0\}_{x\ge0} \stackrel{(d)}= \{Z_t^x:t\ge0\}_{x\ge0}. \] This identifies the full branching diffusion field with the local-time profile of a single reflected Brownian motion whose drift depends on accumulated local time through \(f'/2\) [1304.0710]. Optimal control adds a PDE layer to interacting branching diffusions. In the controlled setting, the dynamic programming principle yields an infinite system of coupled Hamilton–Jacobi–Bellman equations on the disjoint union \(\bigsqcup_{\mathcal V}\mathbb R^{d|\mathcal V|}\), with value functions \[ v_{\mathcal V}(t,\mathbf x_{\mathcal V}) = v\Bigl(t,\sum_{i\in\mathcal V}\delta_{(i,x_i)}\Bigr). \] Under coercivity, continuity and growth bounds transfer to the value function, which is the unique continuous viscosity solution in the corresponding growth class; a comparison principle is proved by a doubling-of-variables argument with a penalization in \(|\mathcal V|\) [2601.11294]. In the permutation-invariant mean-field regime, optimal controls can be chosen symmetric, and the HJB system reduces to a single PDE on measure space, written in the paper as a master-equation involving non-local derivatives \(\partial_\mu U\) [2601.11294]. Taken together, these methods show that branching diffusion with interactions is not a single model but a framework. Its state spaces are measure-valued and genealogical, its nonlinearities may be rank-based, catalytic, mean-field, or competitive, and its analysis proceeds through martingale problems, hydrodynamic limits, scaling limits, traveling-wave theory, duality, spinal decompositions, and viscosity solutions. Source: https://www.emergentmind.com/topics/branching-diffusion-with-interactions