---
title: Conditional Splitting Probabilities
url: https://www.emergentmind.com/topics/conditional-splitting-probabilities
type: topic
---

# Conditional Splitting Probabilities

“Conditional splitting probabilities” denotes a family of context-dependent constructions in which a probability law, a first-passage event, or a structured prediction problem is decomposed into conditionally defined components. In recent arXiv usage, the term appears in at least four technically distinct senses: as the normalized integer- and half-integer-supported subsequences of a Poisson trinomial distribution [2603.09019]; as autoregressive probabilities of selecting a split point inside a span in constituency and discourse parsing [2106.15760]; as first-passage probabilities conditioned on phase-space variables such as position, velocity, or orientation in active matter and related stochastic processes [2606.28709], [2603.13621], [2402.05005]; and as conditional level-to-level factors in rare-event simulation, multilevel splitting, and conditional samplers [1405.2800], [0711.2037], [1909.03566].

## 1. Scope of the term across research areas

The phrase does not name a single standardized object. Instead, it consistently refers to probabilities defined after a “split” of either support, state space, trajectory space, or structure space. The split may be algebraic, geometric, algorithmic, or combinatorial.

| Domain | Conditional object | Representative form |
|---|---|---|
| Poisson trinomial laws | pmf restricted to one lattice component | $\mathbb{P}(X=k \mid X\in\mathbb Z)$ |
| Parsing | split-point distribution inside a span | $P_\theta(k_t \mid (i_t,j_t), y_{<t}, x)$ |
| First-passage theory | hitting probability conditioned on initial state | $\pi_\lambda(z,w)$ or $p_R(x,\theta)$ |
| Rare-event simulation | level-to-level survival factor | $\mathbb P(F_i\mid F_{i-1})$ |

This terminological plurality has methodological consequences. In distribution theory, the split is usually a decomposition of a probability mass function into interleaved subsequences. In NLP, it is a factorization of a tree into a sequence of local split decisions. In first-passage settings, it is a competition between exits, targets, or absorbing sets. In rare-event computation, it is a product decomposition over nested subsets or levels. A common misconception is that the phrase always refers to classical gambler’s-ruin-type hitting probabilities; the recent literature shows that it also names conditional lattice laws, pointer-network outputs, and exact-sampling factorizations.

## 2. Lattice splitting in Poisson trinomial distributions

For independent variables \(X_1,\dots,X_n\) taking values in \(\{0,\tfrac12,1\}\), with \(\mathbb P(X_i=\tfrac12)=T_i\), \(\mathbb P(X_i=1)=W_i\), and \(\mathbb P(X_i=0)=L_i:=1-T_i-W_i\), the sum \(X=\sum_i X_i\) has what the paper calls a Poisson trinomial distribution [2603.09019]. Writing \(H=2X\) and \(G(w)=\prod_{i=1}^n(L_i+T_i w+W_i w^2)=\sum_{m=0}^{2n} a_m w^m\), the pmf splits into an even part and an odd part,
\[
G(w)=p(w^2)+w\,q(w^2),
\]
where \(p(z)=\sum_{k=0}^n a_{2k}z^k\) and \(q(z)=\sum_{k=0}^{n-1} a_{2k+1}z^k\). Equivalently, the support of \(X\) splits into integers and half-integers, controlled by the parity of the number \(S\) of \(\tfrac12\)-valued summands:
\[
X\in\mathbb Z \iff S \text{ even},\qquad X\in\mathbb Z+\tfrac12 \iff S \text{ odd}.
\]

In this setting, “conditional splitting probabilities” are the normalized subsequences
\[
\mathbb{P}(X=k\mid X\in\mathbb Z)=\frac{a_{2k}}{p(1)},\qquad
\mathbb{P}\!\left(X=k+\tfrac12 \mid X\in\mathbb Z+\tfrac12\right)=\frac{a_{2k+1}}{q(1)}.
\]
The paper’s central structural result is that each normalized subsequence is a Poisson binomial distribution. Via Hurwitz stability of \(G\), the Hermite–Biehler theorem, and real non-positive zeros of \(p\) and \(q\), both conditional laws inherit log-concavity, unimodality, and the usual “one or two adjacent modes” structure of Poisson binomials. The same analysis yields quantitative stability: if \(\mu=\mathbb E[X]\), \(\mu_{\mathrm{even}}=\mathbb E[X\mid X\in\mathbb Z]\), and \(\mu_{\mathrm{odd}}=\mathbb E[X\mid X\in\mathbb Z+\tfrac12]\), then
\[
|\mu_{\mathrm{even}}-\mu|\le \tfrac12,\qquad |\mu_{\mathrm{odd}}-\mu|\le \tfrac12,
\]
and any two modes of the two conditional laws are within distance \(5/2\). The resulting picture is that parity splitting can substantially reshape pointwise masses while leaving the center and the peak tightly constrained.

A useful correction to an intuitive but false expectation is that the two split components need not behave as unrelated subdistributions. In the nondegenerate case they are highly structured, each lies in the Poisson binomial family, and their means and modes remain close to the unconditional mean. When \(\mathbb P(X\in\mathbb Z)=0\) or \(\mathbb P(X\in\mathbb Z+\tfrac12)=0\), no genuine splitting occurs; then necessarily every \(T_i\in\{0,1\}\), and \(X\) reduces to a shifted Poisson binomial law.

## 3. Autoregressive split-point distributions in parsing

In constituency parsing, “conditional splitting probabilities” are distributions over candidate split positions inside a current span. In the boundary-based seq2seq framework of Shen et al., a span \((i_t,j_t)\) is split at step \(t\) by choosing a boundary index \(k_t\), and the model parameterizes
\[
P_\theta(k_t \mid (i_t,j_t), y_{<t}, x),
\]
where \(x\) is the input sentence and \(y_{<t}\) is the sequence of previous splitting decisions [2106.15760]. The full structure probability factors autoregressively as a product of these conditional splitting probabilities:
\[
P_\theta((T)\mid x)=\prod_{t=1}^{|(T)|}P_\theta\big((i_t,j_t)\to k_t \mid ((i,j)\to k)_{<t},x\big).
\]

The model instantiates this factorization with a 3-layer bidirectional LSTM encoder, fencepost boundary representations \(h_k=[\overrightarrow h_k,\overleftarrow h_{k+1}]\), span representations \(r_{i,j}=W_1h_i+W_2h_j\), a 3-layer unidirectional decoder LSTM, and a biaffine pointer that scores every boundary position. After softmax,
\[
a_{t,i}=P_\theta(k_t=i\mid (i_t,j_t), y_{<t}, x).
\]
For syntax, valid splits satisfy \(i_t<k_t<j_t\). For discourse, the framework relaxes this to \(i_t<k_t\le j_t\), where \(k_t=j_t\) means “stop splitting here,” so discourse segmentation becomes a special case of the same conditional splitting process rather than a separate preprocessing stage.

This interpretation of splitting probabilities is algorithmic rather than measure-theoretic. The probabilities are local decisions in a top-down depth-first decoder, but because they are conditioned on the decoder state they encode the entire splitting history. The paper emphasizes that this gives structurally consistent trees without chart-based global inference. With beam size treated as constant, decoding complexity is \(\mathcal O(n)\) on GPU and \(\mathcal O(n^2)\) on CPU. Empirically, the model reaches 93.77 F1 on PTB without pretraining, 95.7 F1 with BERT, and in end-to-end discourse parsing improves relation F1 by 1.3 points over the previous best joint model; syntactic parsing runs at 1,127 sentences/s on GPU, while end-to-end discourse parsing is roughly 44× faster than CODRA and 4–5× faster than the previous neural joint parser [2106.15760].

## 4. First-passage, active matter, and memory-dependent splitting

In stochastic transport, splitting probabilities are classical first-passage objects: the probability that one boundary or target is reached before another, conditioned on an initial state. Recent active-matter work makes that conditioning explicit in phase space. For confined run-and-tumble particles, active Brownian particles, and active Ornstein–Uhlenbeck particles in a slit \(z\in[0,\lambda]\), the conditional splitting probability
\[
\pi_\lambda(z,w)
\]
is the probability that a particle starting at position \(z\) with projected velocity \(w\) reaches the right wall before the left. The backward equation
\[
0=w\,\partial_z\pi_\lambda(z,w)+\hat L_w\pi_\lambda(z,w)
\]
is paired with the stationary forward equation for the bulk density, and the paper derives the exact identities
\[
p_v(w)\,\partial_z\pi_\lambda(z,w)=\rho(z,-w),\qquad
p_v(w)\,\pi_\lambda(0,w)=f_0(-w)=f_\lambda(w),
\]
which imply the marginal relations
\[
\partial_z\pi_\lambda(z)=\rho(z),\qquad \pi_\lambda(0)=f_w.
\]
Thus the stationary density is the spatial derivative of the splitting probability, and the fraction of dynamically adsorbed particles at a wall equals the wall splitting probability [2606.28709].

A related confined-active-particle study formulates the right-exit probability as \(p_R(x,\theta)\) in a 1D interval and as \(p_R(x,y,\theta)\) in a 2D corrugated channel, both satisfying backward Fokker–Planck equations with absorbing exit boundaries and reflecting side walls [2603.13621]. In the 1D interval,
\[
Pe\cos\theta\,\partial_x p_R+\partial_x^2 p_R+\chi\,\partial_\theta p_R+\gamma\,\partial_\theta^2 p_R=0,
\]
with \(p_R(-1,\theta)=0\) and \(p_R(1,\theta)=1\). The orientation average obeys a perturbative expansion
\[
\langle p_R\rangle(x)=\frac{x+1}{2}+Pe^2\,C_0(x)+o(Pe^2)
\]
for weak activity, while in the strong-activity regime the averaged splitting probability tends to \(1/2\) away from boundary layers. In narrow corrugated channels, a Fick–Jacobs reduction replaces \(\partial_x^2\) by
\[
\frac1{h(x)}\frac{d}{dx}\!\left(h(x)\frac{d}{dx}\right),
\]
so geometry enters through an entropic drift-diffusion operator.

Memory effects generalize the same theme beyond Markovian dynamics. For one-dimensional isotropic non-Markovian Gaussian processes with stationary increments and two targets at \(0\) and \(L\), the splitting probabilities \(\pi_1,\pi_2\) are governed by post-first-passage average trajectories \(\mu_1(t)\) and \(\mu_2(t)\), through
\[
x_0=\lim_{t\to\infty}\big[\pi_1\mu_1(t)+\pi_2\mu_2(t)\big].
\]
The paper’s central claim is that splitting probabilities are controlled by out-of-equilibrium trajectories observed after the first passage. For scale-invariant mean-square displacement \(\psi(t)\simeq \kappa t^{2H}\), the far-target probability obeys
\[
\pi_2(x_0)\simeq A_H\left(\frac{x_0}{L}\right)^{1/H-1}\qquad (x_0\ll L),
\]
so memory can steepen or flatten the dependence on initial position depending on the Hurst exponent \(H\) [2402.05005].

These works collectively show that “conditional” in first-passage splitting is often irreducibly state-enriched. Conditioning only on position is insufficient for active particles because the relevant Markov state includes velocity or orientation; for non-Markovian Gaussian processes, even that is not enough, and the effective conditioning is encoded in post-first-passage trajectory statistics.

## 5. Conditional splitting in monitored quantum walks

For monitored continuous-time quantum walks with two targets, splitting probabilities are the probabilities of eventual first detection at the left or right target under repeated projective measurements. The setup is a finite-dimensional Hilbert space with unitary evolution \(U(\tau)=e^{-iH\tau}\), measurements at times \(n\tau\), and a survival operator
\[
\mathcal S=\big(\mathbbm1-\ket{x_L}\bra{x_L}-\ket{x_R}\bra{x_R}\big)U(\tau).
\]
The first-detection amplitude at target \(\alpha\in\{L,R\}\) on the \(n\)-th measurement cycle is
\[
\phi_n^{(\alpha)}=\bra{x_\alpha}U(\tau)\mathcal S^{\,n-1}\ket{\psi_0},
\]
and the corresponding splitting probability is
\[
P_\alpha(\psi_0)=\sum_{n=1}^\infty |\phi_n^{(\alpha)}|^2
\]
[2601.16111].

A key structural result is the mapping of the two-target problem onto two single-target detection problems using the symmetric and antisymmetric target states
\[
\ket{d_\pm}=\frac1{\sqrt2}\big(\ket{x_L}\pm\ket{x_R}\big).
\]
For parity-symmetric Hamiltonians,
\[
\phi_n^{(L)}=\frac{\chi_n^{(+)}+\chi_n^{(-)}}{\sqrt2},\qquad
\phi_n^{(R)}=\frac{\chi_n^{(+)}-\chi_n^{(-)}}{\sqrt2},
\]
so the left/right splitting probabilities become
\[
P_L(x_0)=\frac12+\xi(x_0,N,\tau),\qquad P_R(x_0)=\frac12-\xi(x_0,N,\tau),
\]
with \(\xi\) an interference term constructed from the two auxiliary single-target processes.

This produces a specifically quantum notion of conditional splitting. The outcome is conditioned not only on the initial site but on the measurement cadence \(\tau\), which reorganizes the spectrum of the survival operator. For a tight-binding chain, the critical sampling time is
\[
\tau_c=\frac{2\pi}{\Delta E},
\]
with \(\Delta E\) the bandwidth. For large \(N\) and \(0<\tau\le \tau_c\), the paper finds a universal regime
\[
P_L(x_0)\approx P_R(x_0)\approx \frac12
\]
for all bulk initial conditions \(1<x_0<N\). For \(\tau>\tau_c\), a nonuniversal regime appears in which splitting probabilities deviate from \(1/2\) and develop pronounced peaks and dips depending on \(\tau\) and on the initial condition. At resonant sampling times satisfying
\[
(E_k-E_\ell)\tau=0 \pmod{2\pi},
\]
dark states can occur, so \(P_L+P_R<1\). This sharply distinguishes monitored quantum splitting from classical absorbing-boundary problems, where total absorption probability on a finite graph is 1 [2601.16111].

## 6. Rare-event simulation, conditional sampling, and abstract splitting formalisms

In rare-event computation, the expression usually means exactly what the name suggests: conditional probabilities associated with successive splits of a rare event into less rare nested events. For multilevel splitting or subset simulation with nested sets
\[
F=F_L\subset F_{L-1}\subset \cdots \subset F_1\subset F_0,
\]
the target probability factors as
\[
p=\mathbb P(F_L)=\prod_{i=1}^L \mathbb P(F_i\mid F_{i-1}).
\]
The Moving Particles framework summarizes earlier results that optimal variance at fixed cost is achieved when these conditional probabilities are all equal, \(p_1=\cdots=p_L=p_0\), citing Cérou et al.; it also recalls Guyader et al.’s limit \(p_0=1-1/N\) and then replaces explicit level design by a particle-moving scheme whose number of moves to reach the rare set is Poisson with parameter \(\log(1/p)\) for one particle and \(N\log(1/p)\) for \(N\) particles, rather than order \(1/p\) as in naive Monte Carlo [1405.2800].

Large-deviation analysis casts the same factors as hitting probabilities between successive level sets of an importance function. In that formulation, the rare event is \(\{\tau_B<\tau_A\}\), and the probability is written as a product of level-to-level terms of the form
\[
P(\text{reach level }k \text{ before }A \mid \text{reached level }k-1 \text{ before }A).
\]
The paper shows that subexponential particle growth and asymptotic variance control are equivalent to the existence of a suitable subsolution for the associated calculus-of-variations problem [0711.2037]. In the idealized exact-resampling analysis of adaptive multilevel splitting, the algorithm enforces an empirical conditional survival fraction \(1-k/n\) at each iteration by killing the \(k\) least-adapted particles among \(n\) and resampling them from the exact conditional law above the current level; the resulting rare-event estimator is unbiased whatever \(k\), and the paper derives large-\(n\) expansions for both variance and cost [1405.1352].

The same logic extends from probability estimation to sampling from conditional rare-event laws. Generalized splitting targets
\[
q(x)=f(x)\,\frac{\mathbf 1\{S(x)\ge \gamma\}}{\ell(\gamma)},
\qquad \ell(\gamma)=P(S(X)\ge \gamma),
\]
by using levels \(-\infty=\gamma_0<\gamma_1<\cdots<\gamma_\tau=\gamma\) with
\[
\rho_\ell=P(S(Y)\ge \gamma_\ell \mid S(Y)\ge \gamma_{\ell-1})\approx 1/s,
\]
Markov kernels with stationary distributions \(q_\ell(y)=f(y)\mathbf 1\{S(y)\ge \gamma_\ell\}/P(S(Y)\ge \gamma_\ell)\), and repeated non-empty splitting trials. The paper’s main conclusion is that approximation error depends crucially on the relative variability of the number of points \(M\) produced in one successful run, and it provides explicit total-variation and mean-absolute-error bounds in terms of low-order moments of \(M\) [1909.03566].

Several adjacent frameworks recast the same conditional-splitting idea in different algebraic forms. Weighted ensemble (WE) does not write explicit level probabilities \(q_k\), but its bin-conditional offspring counts and weights preserve the target law \(\mu\); the paper proves that WE is the only splitting and killing method that gives asymptotically consistent long-time MCMC estimates when total weight is preserved pathwise, and it derives a lower bound on asymptotic variance together with near-optimal designs based on binwise conditional variances [2011.13899]. Split sampling writes
\[
Z=\mathbb E_\pi[L(\mathbf x)]=\int_0^\infty Z(m)\,dm,\qquad Z(m)=\mathbb P_\pi(L(\mathbf x)>m),
\]
introduces an auxiliary variable \(m\), and shows that with choices such as \(\omega_t=1/Z_t\) or \(\Omega_t=1/Z_t\) the method reproduces the long-run sampling distribution of product estimators and relates directly to nested sampling [1212.0534]. Probabilistic divide-and-conquer with deterministic second half factors an exact conditional law \(L(X\mid X\in E)\) over a product decomposition \(X=(A,B)\); when each \(a\in A\) has a unique completion \(b_a\in B\) satisfying the constraint, the second conditional law collapses to \(L(B\mid E,A=a)=\delta_{b_a}\), so all nontrivial conditional splitting resides in the first half [1411.6698]. At the highest level of abstraction, categorical probability interprets idempotent Markov kernels as conditional-type projections that split through an intermediate space \(T\), \(X\xrightarrow{\pi}T\xrightarrow{\iota}X\), and proves that every idempotent measurable Markov kernel between standard Borel spaces splits in this sense [2308.00651].

Across these literatures, the unifying idea is not a single formula but a recurring operation: a hard global object is replaced by conditional pieces whose structure is easier to analyze or simulate. What varies is the object being split—support lattices, parse trees, phase-space trajectories, nested rare-event sets, or probability kernels themselves.

Source: https://www.emergentmind.com/topics/conditional-splitting-probabilities