---
title: Generalized Copy-Hopping Operators
url: https://www.emergentmind.com/topics/generalized-copy-hopping-operators
type: topic
---

# Generalized Copy-Hopping Operators

Searching arXiv for the cited paper and closely related terminology to ground the article in current arXiv records.
Generalized copy-hopping operators are memory-cognizant extensions of Simon’s 1955 random-copying neutral model in which the uniform copying rule over all previous sequence entries is replaced by an arbitrary age-dependent kernel \(k(a)\). In the generalized formulation, copying at time \(t+1\) depends on the age \(a=(t+1)-i\) of a candidate source entry born at time \(i\), so the copy mechanism is represented by an operator \(K_k\) that weights the past non-uniformly. This preserves the innovation–copy competition of the classical model while permitting analytical study and numeric simulation of finite-memory, exponential, bounded, hyperbolic, and heavy-tailed memory effects within a single framework [2102.11705].

## 1. Classical Simon dynamics and the original copy operator

Simon’s classical random-copying model evolves on a sequence of discrete time steps \(t=1,2,\ldots\). At each step \(t+1\), one of two events occurs: with probability \(\mu\), a brand new variant appears at position \(t+1\); with probability \(1-\mu\), the new entry \(s_{t+1}\) is copied from one of the previous entries \(s_i\), \(i=1,\ldots,t\). In the original model, this copy choice is uniform over all \(t\) existing entries [2102.11705].

Using the notation \(s_i\in\{1,2,3,\ldots\}\) for the variant label of the \(i\)-th entry and
\[
n_m(t)=|\{i\le t:s_i=m\}|
\]
for the abundance of variant \(m\) at time \(t\), the uniform copy kernel is
\[
P(i\mid t)=\frac{1}{t},\qquad i=1,\ldots,t.
\]
Equivalently, for any test function \(f\) on the existing entries,
\[
(K_0 f)(t+1):=(1-\mu)\,\frac{1}{t}\sum_{i=1}^t f(i),
\]
plus the innovation contribution.

The classical model is analytically tractable through a probability generating function (PGF). If \(H_m(\tau,t;x)=\sum_{n\ge 1}\Pr[\text{variant }m\text{ seeded at time }\tau\text{ has total count }n\text{ at time }t]\cdot x^n\), then, as stated with reference to O’Brien and Gleeson, one obtains the exact difference equation
\[
H_m(\tau,t)=(1-\mu)\,\frac{1}{t}\bigl[H_m(\tau,t+1)\bigr]^2+\bigl[1-(1-\mu)\tfrac{1}{t}\bigr]H_m(\tau,t+1),
\]
whose continuum limit is the Riccati equation
\[
-\partial_\tau H=\frac{1-\mu}{\tau}(H^2-H).
\]
With boundary condition \(H(\tau=t;x)=x\), the solution is
\[
H(\tau,t;x)=\frac{x}{\,x[1-(\tau/t)^{-(1-\mu)}]+(\tau/t)^{-(1-\mu)}}.
\]
From this, one recovers a geometric distribution in \(n\), and after averaging over \(\tau\in[1,t]\), the model yields the familiar power-law tail
\[
\Pr(n)\sim n^{-\frac{2-\mu}{1-\mu}}\qquad (t\gg 1).
\]

The significance of the classical operator is that its simplicity is tied directly to the assumption of uniform copying. Generalized copy-hopping operators retain the same discrete-time architecture but relax precisely that assumption.

## 2. Age-dependent kernels and the generalized copy-hopping operator

The memory-dependent generalization replaces the uniform pick \(1/t\) by an arbitrary age-kernel \(k(a)\ge 0\), \(a=1,2,\ldots,t\), normalized by
\[
\sum_{a=1}^t k(a)=1.
\]
At time \(t+1\), the age of a candidate copy born at time \(i\) is
\[
a=(t+1)-i.
\]

The generalized copy-hopping operator \(K_k\) acts on functions \(f\) on past indices as
\[
(K_k f)(t+1)=(1-\mu)\sum_{i=1}^t k\bigl((t+1)-i\bigr)f(i).
\]
Accordingly, the probability that the new entry \(s_{t+1}\) equals the variant of the \(i\)-th entry is
\[
P(i\to t+1)=(1-\mu)\,k\bigl((t+1)-i\bigr).
\]

This reformulation makes memory explicit at the operator level. Simon’s uniform rule is recovered as the special case \(k(a)=1/t\), but the generalized framework permits arbitrarily decaying or truncated memory profiles. The paper’s stated motivation is to move away from the uniform assumption, incorporate memory effects via an arbitrary age-dependent kernel, determine further information regarding the structure of sequences from the classical model, and show that previously proposed memory-dependent models can be studied as specific cases of the same framework [2102.11705].

A plausible implication is that “copy-hopping” is best understood not as a single stochastic rule but as a class of linear operators indexed by the age profile \(k(a)\). The resulting heterogeneity of memory is what generates distinct abundance regimes.

## 3. Master equation, abundance dynamics, and continuum approximation

For a fixed variant \(m\) that first appeared at time \(\tau\), let
\[
q_n(\tau,t)=\Pr[n_m(t)=n\mid \text{first occurrence at }\tau].
\]
During one step \(t\to t+1\), three mutually exclusive events occur:

1. The copy event selects one of the \(n\) existing members of \(m\), with probability \((1-\mu)\,H_m(t)\), and the abundance jumps \(n\to n+1\).
2. The copy event selects some other variant, with probability \((1-\mu)[1-H_m(t)]\), and the abundance stays at \(n\).
3. Innovation occurs, with probability \(\mu\), and the abundance stays at \(n\).

The corresponding exact discrete master equation is
\[
q_n(\tau,t+1)=(1-\mu)\,H_m(t)\,q_{n-1}(\tau,t)+\bigl[1-(1-\mu)\,H_m(t)\bigr]q_n(\tau,t),
\]
where the framework writes
\[
H_m(t)=\sum_{i=\tau}^t k(t+1-i).
\]
In a mean-field continuum approximation \(t\to\infty\), one writes
\[
n_m(t+dt)-n_m(t)\approx (1-\mu)\sum_{a\ge 1}k(a)\,n_m(t-a)\,dt,
\]
which leads to an integrodifferential equation for the mean abundance \(m(t)=E[n_m(t)]\):
\[
m'(t)=(1-\mu)\int_0^t k(a)\,m(t-a)\,da,\qquad m(0)=1.
\]

These equations show how the age-kernel enters the model at two levels. At the discrete level it modulates the one-step transition probabilities. At the continuum level it appears as a memory convolution. This suggests that generalization is not merely a perturbation of Simon’s tail exponent; it changes the effective temporal architecture of replication.

## 4. Asymptotic regimes and abundance-tail exponents

The generalized framework distinguishes sharply between classical, finite-mean, and heavy-tailed memory regimes [2102.11705].

For the classical kernel \(k(a)=1/t\), the full sequence-averaged distribution has tail exponent
\[
\alpha_{\mathrm{classic}}=\frac{2-\mu}{1-\mu}=1+\frac{1}{1-\mu}.
\]

For any memory kernel with finite mean
\[
T=\sum_a a\,k(a),
\]
the infinite-time PGF obeys
\[
H_\infty(x)=x\exp\{(1-\mu)[H_\infty(x)-1]\},
\]
and the single-variant abundance distribution at \(t\to\infty\) has the asymptotic form
\[
q_n(\infty)\sim A\,n^{-3/2}\,\exp[-n/\kappa],
\qquad
\kappa=\frac{2(1-\mu)^2}{\mu^2}.
\]
Thus, whenever \(\int a\,k(a)\,da<\infty\), the pure power-law exponent becomes \(3/2\) universally, but with an exponential cutoff.

For heavy-tailed memory
\[
k(a)\sim C\,a^{-1-\gamma},\qquad 0<\gamma<1,
\]
the memory has infinite mean, and one obtains a pure power law for the single-variant abundance:
\[
q_n(\infty)\sim n^{-\alpha},
\qquad
\alpha=1+\frac{\gamma}{1-\mu}.
\]
In this regime, long-range memory kernels slow the decay of the abundance-tail.

The main asymptotic distinction is therefore between finite-mean memory and infinite-mean memory. Finite-mean kernels erase the classical Simon exponent and replace it with a universal \(3/2\) law plus cutoff, whereas heavy-tailed kernels preserve pure power-law behavior but with an exponent continuously controlled by \(\gamma\) and \(\mu\). This gives the generalized operator framework a clear taxonomic role: the tail class is encoded by the decay properties of \(k(a)\).

## 5. Canonical kernels and embedded special cases

Several previously proposed memory-dependent models appear as explicit special cases of the generalized operator formalism [2102.11705].

The exponential memory kernel,
\[
k(a)=\beta^{-1}e^{-a/\beta},
\]
has finite mean \(T=\beta\). In this case the tail behaves as
\[
\sim n^{-3/2}e^{-n/(2\beta(1-\mu)/\mu)},
\]
and the mean abundance admits the closed form
\[
m(t)=\frac{1}{\mu}\bigl[1-(1-\mu)e^{-(\mu/\beta)\,t}\bigr].
\]

The bounded uniform memory kernel, identified with the BMPG model of Schaigorodsky et al.,
\[
k(a)=\frac{1}{\kappa}\quad \text{for }a=1,\ldots,\kappa;\qquad 0\text{ otherwise},
\]
leads in Laplace space to
\[
\widehat m(s)=\frac{1}{s-[((1-\mu)/\kappa)](1-e^{-\kappa s})}.
\]
For the regime
\[
\kappa\ll t\ll \kappa^2/(2\mu),
\]
one finds
\[
m(t)\approx \frac{2}{(1-\mu)\kappa}\,t,
\]
and the large-\(t\) tail again behaves as
\[
\sim n^{-3/2}e^{-n/\kappa^*}.
\]

The hyperbolic memory kernel of Cattuto et al.,
\[
k(a)\propto (a+\gamma)^{-1},
\]
has marginally divergent mean. In that case one observes a crossover from \(\alpha=3/2\) to \(\alpha=1+\gamma/(1-\mu)\) depending on \(t\) relative to a cutoff.

These examples clarify the scope of the operator formalism. Exponential and bounded kernels represent finite-memory or effectively finite-mean memory and therefore fall into the \(3/2\)-with-cutoff class. Hyperbolic and other slowly decaying kernels interpolate toward the infinite-mean regime. The generalized operator thus functions as a unifying language for models that otherwise appear disparate.

## 6. Operator-theoretic interpretation and terminological scope

The framework also admits a compact operator summary. Defining the linear copy operator \(K_k\) on sequence-space \(\ell^1(\{1,2,\ldots,t\})\) by
\[
(K_k\phi)_t(j):=(1-\mu)\sum_{i=1}^{t}k(t-i)\,\phi_i
\qquad (j=t+1),
\]
and an injection operator \(I_\mu\) that with probability \(\mu\) injects a new basis-vector \(e_{t+1}\), the full time-evolution operator on probability vectors is
\[
E_{t+1}=I_\mu E_t+K_k E_t.
\]
Within this perspective, the spectral radius and Perron–Frobenius theory of \(K_k\) determine whether the system admits a heavy-tailed stationary distribution. When \(\|K_k\|<1\), equivalently \(\mu>0\), one finds a unique steady PGF \(H_\infty\) solving
\[
H=x\exp[(1-\mu)(H-1)].
\]
The singularity of \(H_\infty(x)\) nearest \(x=1\) controls the tail exponent, and its location depends on \(\int a\,k(a)\,da\), specifically on whether that mean is finite or infinite [2102.11705].

This operator formulation organizes the model’s main outputs—mean-field growth rates, stationary PGFs, and abundance exponents—around spectral properties rather than around a single closed-form counting argument. A plausible implication is that the age-kernel is most naturally viewed as a structural input to the evolution operator, not merely as a phenomenological parameterization of recency.

The phrase “copy-hopping” also appears in a separate usage in scattering-amplitude theory. There, “copy-hopping” from adjoint to symmetric structure constants refers to replacing antisymmetric \(f^{abc}=\Tr[T^a,T^b]T^c\) by symmetric \(d^{abc}=\Tr\{T^a,T^b\}T^c\) in color–kinematics duality and double-copy constructions. That setting emphasizes manifest locality and the spanning of effective photon and graviton operators, rather than age-dependent copying in neutral sequence growth [2211.04431]. The two usages share the language of “copy,” but they refer to distinct operator structures and distinct research programs.

Source: https://www.emergentmind.com/topics/generalized-copy-hopping-operators