---
title: Stochastic Sewing in Analysis
url: https://www.emergentmind.com/topics/stochastic-sewing
type: topic
---

# Stochastic Sewing in Analysis

Searching arXiv for recent and foundational papers on stochastic sewing.
Stochastic sewing is a probabilistic extension of the classical sewing lemma that reconstructs a one-parameter process from a two-parameter family of random increments \(A_{s,t}\) by proving convergence of Riemann-type sums \(\sum_{[s,t]\in\pi}A_{s,t}\) as the mesh \(|\pi|\to0\). Its defining feature is that it replaces purely pathwise control of the defect
\[
\delta A_{s,u,t}:=A_{s,t}-A_{s,u}-A_{u,t}
\]
by a mixed estimate combining \(L_m\)-control of the defect with stronger control of its conditional expectation. In this way, stochastic sewing exploits adaptedness, martingale cancellations, and Burkholder–Davis–Gundy-type estimates to lower the effective regularity threshold from the deterministic \(>1\) regime to a stochastic threshold in which the random part only needs to be better than order \(1/2\) [1810.10500]. Subsequent work has developed semigroup-adapted mild formulations for SPDEs, Banach-valued variants, extensions based on conditioning at earlier times to capture asymptotic decorrelation, multiparameter versions for random fields, and a Wasserstein-space analogue for flows of probability measures [2108.05573] [2105.09364] [2206.01686] [2307.11527] [2410.17096].

## 1. Foundational formulation and probabilistic mechanism

The basic stochastic sewing problem begins with an adapted two-parameter process \(A_{s,t}\), typically \(\mathcal F_t\)-measurable, and asks when the random Riemann sums
\[
\sum_i A_{t_i,t_{i+1}}
\]
converge in \(L_m(\Omega)\) or in probability to an increment \(A_t-A_s\). In the one-parameter stochastic sewing lemma introduced by Lê, the central hypotheses are estimates on the defect \(\delta A_{s,u,t}\) and on its predictable projection with respect to \(\mathcal F_s\) [1810.10500].

In the form summarized in the lecture notes, one assumes exponents \(\gamma_1>\tfrac12\), \(\gamma_2>1\) and constants \(\Gamma_1,\Gamma_2\ge0\) such that
\[
\|\delta A_{s,u,t}\|_{L_m(\Omega)}\le \Gamma_1|t-s|^{\gamma_1},
\qquad
\|E^s\delta A_{s,u,t}\|_{L_m(\Omega)}\le \Gamma_2|t-s|^{\gamma_2}.
\]
Under these bounds, the sewn increment exists and satisfies
\[
\|A_T-A_S\|_{L_m(\Omega)} \le \|A_{S,T}\|_{L_m(\Omega)} + C\Gamma_1|T-S|^{\gamma_1} + C\Gamma_2|T-S|^{\gamma_2}
\]
[2510.12165].

The probabilistic improvement over deterministic sewing is tied to the decomposition of local errors into a predictable part and a martingale-difference part. Deterministic sewing requires direct summability of the full defect and thus a threshold \(>1\). Stochastic sewing instead estimates the centered random part by BDG, so square-summability suffices; this is the source of the characteristic threshold \(>\tfrac12\) for the random component [1810.10500] [2510.12165]. The lecture notes emphasize that in stochastic problems one often does not need pathwise cancellation of \(\delta A\); conditional expectation is enough because randomness creates cancellations that are invisible in a deterministic framework [2510.12165].

A useful simplification recorded in the notes is that, for the \(\gamma_1\)-control, it is enough to show
\[
\|A_{s,t}\|_{L_m(\Omega)}\le \Gamma_1|t-s|^{\gamma_1},\qquad \gamma_1>\tfrac12
\]
[2510.12165]. This reflects the fact that the random part can be propagated through the dyadic proof architecture using martingale inequalities rather than direct Hölder summability.

## 2. Lê’s stochastic sewing lemma and the Doob–Meyer-type structure

Lê’s 2018 paper gives the canonical one-parameter theorem. For an adapted \(L^m\)-integrable process \(A_{s,t}\), with \(m\ge2\), suppose there exist \(T_1,T_2\ge0\) and \(\varepsilon_1,\varepsilon_2>0\) such that for every \(S<s<u<t<T\),
\[
\big\|\mathbb E_s \delta A_{s,u,t}\big\|_{L^m}\;<\;T_1|t-s|^{1+\varepsilon_1},
\]
and
\[
\big\|\delta A_{s,u,t}-\mathbb E_s\delta A_{s,u,t}\big\|_{L^m}\;<\;T_2|t-s|^{\frac12+\varepsilon_2}.
\]
Then there exists a unique adapted \(L^m\)-integrable process \((A_t)\) with \(A_S=0\) such that \(A_t^\pi:=\sum_i A_{t_i,t_{i+1}}\) converges to \(A_t\) in \(L^m\) as \(|\pi|\to0\), together with increment estimates of the form
\[
\|A_t-A_s-A_{s,t}\|_{L^m}
\le C_1|t-s|^{1+\varepsilon_1}+C_2|t-s|^{\frac12+\varepsilon_2}
\]
and
\[
\big\|\mathbb E_s(A_t-A_s-A_{s,t})\big\|_{L^m}
\le C_1|t-s|^{1+\varepsilon_1}
\]
[1810.10500].

The theorem also admits a simplification in which the centered estimate is replaced by the stronger but easier condition
\[
\|\delta A_{s,u,t}\|_{L^m}\le T_2|t-s|^{\frac12+\varepsilon_2}
\]
up to changing constants [1810.10500]. No continuity of \((s,t)\mapsto A_{s,t}\) is assumed; adaptedness is essential.

A distinctive aspect of Lê’s formulation is its Doob–Meyer-type decomposition. Under an additional estimate
\[
\big\|\mathbb E_s A_{u,t}-\mathbb E_u A_{u,t}\big\|_{L^m} < T_3|t-s|^{\frac12+\varepsilon_3},
\]
the sewn process decomposes uniquely as
\[
A_t=M_t+J_t,
\]
where \(M\) is a martingale and \(J\) is an adapted remainder satisfying quantitative increment estimates [1810.10500]. This is not a finite-variation decomposition in the classical sense; rather, it is a sewing-based martingale/predictable splitting. The same paper shows that the corresponding partial Riemann sums
\[
M_t^\pi=\sum_i \big(A_{t_i,t_{i+1}}-\mathbb E_{t_i}A_{t_i,t_{i+1}}\big),
\qquad
J_t^\pi=\sum_i \mathbb E_{t_i}A_{t_i,t_{i+1}}
\]
converge in \(L^m\) to \(M_t\) and \(J_t\) [1810.10500].

This framework subsumes standard objects from Itô calculus. For \(A_{s,t}=f(B_s)\cdot(B_t-B_s)\), the stochastic sewing lemma yields the Itô integral as the limit of left-point Riemann sums. For \(A_{s,t}=M_{s,t}\otimes M_{s,t}\), it constructs quadratic variation as a sewn limit. It also recovers Itô’s formula by sewing the Taylor expansion increments of \(f(M_t)\) [1810.10500]. The relevance of stochastic sewing is therefore not confined to abstract convergence theory; it supplies an increment-based architecture in which Itô integration, quadratic variation, and stochastic Taylor expansions appear as instances of a single principle.

## 3. Structural generalizations: Banach, mild, multiparameter, and Wasserstein formulations

Subsequent developments preserve the core idea while changing the ambient geometry.

### Banach-space stochastic sewing

The Banach-valued extension replaces Hilbert or Euclidean target spaces by a separable Banach space \(V\) with martingale type \(p\in(1,2]\). In this setting the stochastic threshold changes from \(1/2\) to \(1/p\). For a measurable \(A:\Omega\times\Delta\to V\) with \(A_{s,t}\) \(\mathcal F_t\)-measurable, one assumes
\[
\big\|E_s[\delta A_{s,u,t}]\big\|_{V;n} \le \Gamma_1\,w(s,t)^{1+\varepsilon_1},
\]
and
\[
\big\|\big\|\delta A_{s,u,t}-E_s\delta A_{s,u,t}\,\big|_s\big\|_{V;m}\big\|_n \le \Gamma_2\,w(s,t)^{\frac1p+\varepsilon_2},
\]
for a control \(w\). Then the Riemann sums converge to an adapted process \(\delta_t\), with quantitative analogues of the scalar estimates [2105.09364]. In Hilbert spaces, martingale type \(2\) recovers the classical exponent \(1/2\). This extension makes stochastic sewing available in Besov, Sobolev, and Triebel–Lizorkin settings [2105.09364].

### Mild stochastic sewing

For semigroup-driven evolution equations, the natural increment is not \(x_t-x_s\) but
\[
\hat\delta x_{s,t}:=x_t-S_{t-s}x_s,
\]
where \(S\) is a semigroup generated by an operator \(A\). The mild stochastic sewing lemma adapts the coboundary to
\[
\hat\delta \Xi_{r,s,t} :=\Xi_{r,t}-\Xi_{s,t}-S_{t-s}\Xi_{r,s},
\]
and proves convergence of
\[
\sum_{[u,v]\in P} S_{t-v}\Xi_{u,v}
\]
under \(L^p\)-size and conditional-defect estimates in semigroup interpolation spaces [2108.05573]. A distinctive feature is the quantitative estimate
\[
\big\|\Xi_{s,t}^{\mathrm{sewn}}-\Xi_{s,t}\big\|_{L^p(E_{\kappa+\varpi})} \lesssim C\,|t-s|^{\bar\eta-\varpi},
\]
which trades time regularity for spatial regularity through semigroup smoothing [2108.05573]. The authors explicitly note that this quantitative bound does not follow directly from earlier stochastic sewing results because the mild formulation depends essentially on interpolation and semigroup structure [2108.05573].

### Multiparameter stochastic sewing

In several parameters, one no longer controls a single defect. The multiparameter lemma for random fields on rectangles in \([0,T_1]\times\cdots\times[0,T_d]\) requires control of all “partial defects” obtained by cutting in subsets of coordinate directions. The theorem is formulated relative to a commuting multiparameter filtration and a function space \(C_2^{\alpha,\beta,\gamma}L_m\). If \(\Xi\in C_2^{\alpha,\beta,\gamma}L_m\), with \(m\ge2\), \(\alpha\in(0,1)\), \(\gamma\in(0,\tfrac12)\), \(\beta\in(\tfrac12,\infty)\), and \(\beta+\gamma>1\), then the grid Riemann sums converge to sewing operators \(I^\theta\Xi\) indexed by subsets \(\theta\subset[d]\), uniquely characterized by additivity in the directions of \(\theta\) and an alternating-remainder estimate [2307.11527].

The conceptual novelty is that multiple directions force a lattice of defects and conditional expectations. This distinguishes stochastic sewing in several parameters from deterministic hypercube sewing, such as the earlier multi-parameter deterministic theory on hypercubes [1806.09915]. A plausible implication is that multiparameter stochastic sewing should be viewed not merely as a higher-dimensional reformulation but as a genuinely new algebra of conditional cancellations.

### Wasserstein-space stochastic sewing

A further generalization lifts sewing from random variables to maps on probability measures
\[
\Theta_{s,t}:\mathcal P_p(H)\to \mathcal P_p(H).
\]
Here there is no intrinsic martingale structure at the level of measures, so the construction proceeds through a probabilistic representation by coupling operators \(\mathcal X_{s,r,t}\). Under a \(\beta\)-coupling family satisfying representation, growth, and stochastic sewing estimates, one obtains a flow \(\theta_{s,t}\) on \(\mathcal P_p(H)\) with
\[
\theta_{s,t}=\theta_{r,t}\circ\theta_{s,r},
\]
together with convergence rates of partition schemes in \(W_p\) [2410.17096]. The paper emphasizes that the genuinely new regime is \(\beta\in(0,1/2]\), where classical sewing on Wasserstein space is not sufficient but the stochastic decomposition still yields convergence [2410.17096].

## 4. Earlier-time conditioning and asymptotic decorrelation

A major extension is the 2022 result that replaces conditioning on the present \(\mathcal F_s\) by conditioning on a strictly earlier \(\mathcal F_v\), \(v<s\), in order to exploit asymptotic decorrelation. The motivating observation is that some fractional Brownian increments become weakly correlated with the distant past in a way that is invisible if one conditions only at time \(s\) [2206.01686].

The generalized theorem assumes \(A_{s,t}\) is \(\mathcal F_t\)-measurable, \(A_{s,s}=0\), and that for every \(0\le t_0<t_1<t_2<t_3\le T\),
\[
\bigl\|\mathbb E[\delta A_{t_1,t_2,t_3}\mid\mathcal F_{t_0}]\bigr\|_{L_m(\mathbb P)} \le \Gamma_1 (t_1-t_0)^{-\alpha}(t_3-t_1)^{\beta_1},
\quad\text{if } M(t_3-t_1)\le t_1-t_0,
\]
and
\[
\|\delta A_{t_0,t_1,t_2}\|_{L_m(\mathbb P)} \le \Gamma_2 (t_2-t_0)^{\beta_2}.
\]
Under
\[
\beta_1>1,\qquad \beta_2>\tfrac12,\qquad \beta_1-\alpha>\tfrac12,
\]
there exists a unique adapted process \((A_t)\) such that the Riemann sums converge in \(L_m(\mathbb P)\), with conditional and moment sewing estimates
\[
\bigl\|\mathbb E[A_{t_2}-A_{t_1}-A_{t_1,t_2}\mid\mathcal F_{t_0}]\bigr\|_{L_m}
\le C_1 |t_1-t_0|^{-\alpha}|t_2-t_1|^{\beta_1},
\]
and
\[
\|A_{t_2}-A_{t_1}-A_{t_1,t_2}\|_{L_m}
\le C_2 |t_2-t_1|^{\beta_1-\alpha}+C_3 |t_2-t_1|^{\beta_2}
\]
[2206.01686].

The singular factor \((t_1-t_0)^{-\alpha}\) is the key new ingredient. It quantifies dependence on the past and turns asymptotic decorrelation into a sewing estimate. The theorem also yields a convergence rate
\[
\|A_\tau-A^\pi_\tau\|_{L_m(\mathbb P)} \lesssim (\Gamma_1+\Gamma_2)\,|\pi|^\varepsilon
\]
for some \(\varepsilon>0\) [2206.01686]. A companion corollary allows singularities at the origin, with factors such as \(t_1^{-\gamma_1}\) and \(t_0^{-\gamma_2}\), which is needed in local-time applications [2206.01686].

The proof retains the standard dyadic/BDG architecture but estimates the predictable part by iteratively reapplying the conditional estimate across dyadic levels, rather than by a direct triangle inequality. This iterative use of earlier-time conditioning is precisely what removes an extra technical restriction that would otherwise appear [2206.01686].

For fractional Brownian motion \(B\), represented through Mandelbrot–Van Ness,
\[
B_t = B(0)+\int_{-\infty}^t K_H(t,s)\,dW_s,
\]
the paper derives estimates such as
\[
\bigl\|\mathbb E[\delta A_{s,u,t}\mid \mathcal F_v]\bigr\|_{L_m} \lesssim_H \Bigl(\frac{t-s}{s-v}\Bigr)^{1-H}(t-s)
\]
for \(A_{s,t}=|B_t-B_s|^{1/H}\) [2206.01686]. This is the template instance of a defect that fails the original \(\mathcal F_s\)-condition but becomes tractable once conditioned on a remote past \(\mathcal F_v\). A plausible interpretation is that this result enlarges stochastic sewing from a local-adapted principle to a framework capable of encoding long-memory Gaussian structure.

## 5. Principal applications

The applications of stochastic sewing are broad but structurally unified: one identifies a local increment whose defect has a small predictable component and a square-summable random component, then uses the sewn limit as an integral, additive functional, or effective flow.

### Stochastic integrals and rough Gaussian drivers

For Brownian motion, stochastic sewing reconstructs the Itô integral from left-point Riemann sums and yields quadratic variation and Itô’s formula [1810.10500]. The same technique extends to fractional Brownian motion in regimes where deterministic sewing is insufficient. In the 2022 extension, if \(H>\tfrac12\) and \(f\) is bounded measurable, then for
\[
A_{s,t}=f(B_s)\cdot(B_t-B_s),
\]
the theorem applies with
\[
\alpha=H,\qquad \beta_1=2H,\qquad \beta_2=H,\qquad M=1,
\]
and the sums
\[
\sum_{[s,t]\in\pi} f(B_s)\cdot(B_t-B_s)
\]
converge in \(L_m\) for every \(m<\infty\), with
\[
\Bigl\|\int_s^t f(B_r)\,dB_r\Bigr\|_{L_m} \lesssim \|f\|_\infty |t-s|^H
\]
[2206.01686]. The significance is that no Hölder smoothness of \(f\) is required.

For Stratonovich-type sums along fBM with \(H\in(\tfrac16,\tfrac12)\), the same paper considers
\[
A_{s,t}=(f(B_s)+f(B_t))\cdot B_{s,t},
\]
proves convergence of the midpoint/trapezoidal sums in \(L_m\), and derives
\[
\Bigl\|\int_s^t f(B_r)\circ dB_r-\frac{f(B_s)+f(B_t)}2\cdot B_{s,t}\Bigr\|_{L_m} \lesssim \|f\|_{C^\gamma}|t-s|^{(\gamma+1)H}
\]
under \(\gamma>\frac{1}{2H}-1\), plus the symmetry condition \(\partial_i f^j=\partial_j f^i\) when \(H\le\tfrac14\) and \(d>1\) [2206.01686].

### Additive functionals and regularization by noise

Lê’s original paper applies stochastic sewing to additive functionals of Markov processes and SDEs with irregular drift, including Brownian and fractional Brownian settings [1810.10500]. The lecture notes revisit this theme as one of the central applications. For Brownian motion, they state the averaging estimate
\[
\Bigl\|\int_S^T (b(W_r+x)-b(W_r+y))\,dr\Bigr\|_{L_m(\Omega)}
\le C\|b\|_{C^\gamma}|x-y|\,|T-S|^{\frac12+\frac\gamma2},
\]
obtained by choosing the germ
\[
A_{s,t}:=E^s\int_s^t \bigl(b(W_r+x)-b(W_r+y)\bigr)\,dr
\]
and applying stochastic sewing together with heat-semigroup smoothing [2510.12165]. For fractional Brownian motion, the notes record
\[
\Bigl\|\int_S^T (b(B^H_r+x)-b(B^H_r+y))\,dr\Bigr\|_{L_m}
\le C\|b\|_{C^\gamma}|x-y|^\rho |T-S|^{1+(\gamma-\rho)H},
\]
for \(\rho<\gamma+\frac1{2H}\), linking stochastic sewing to Catellier–Gubinelli-type regularization by noise [2510.12165].

The Banach-valued theorem is designed precisely for singular additive functionals. For
\[
I[f]_t(x):=\int_0^t f_r(B_r+x)\,dr,
\]
with \(f\) distributional in space, the paper proves that for suitable \(\gamma\),
\[
I: L^\theta\big([0,T];B^\alpha_{p,1}\big) \to C\big([0,T];L^m(\Omega;B^{\alpha+\gamma}_{p,q})\big),
\]
and gives the estimate
\[
\big\|\big\|\delta I[f]_{s,t}\,|_s\big\|_{B^{\alpha+\gamma}_{p,q};m}\big\|_\infty \lesssim |1_{[s,t]}f|_{L^\theta B^\alpha_{p,\infty}}\,|t-s|^{1-H\gamma-\frac1\theta}
\]
[2105.09364]. This yields a regularization-by-noise principle in Besov scales.

### Local times

The 2022 fractional stochastic calculus paper gives a discretization formula for the local time \(L_T(a)\) of one-dimensional fBM. For
\[
A_{s,t}(a):=(t-s)^{1-(1+\gamma)H}|B_t-B_s|^\gamma\mathbf 1_{\{B_s<a<B_t\}},
\]
the Riemann sums converge in \(L_m\) to
\[
c_{H,\gamma}L_T(a)
\]
[2206.01686]. It also provides an integrated convergence statement in \(x\) and a crossing-count representation of local time [2206.01686]. In several parameters, the multiparameter stochastic sewing lemma is used to control Fourier transforms of occupation measures of Gaussian sheets, leading to Sobolev/Bessel regularity of local times and then to regularization-by-noise results for planar SDEs [2307.11527].

### SPDEs and averaging

The mild stochastic sewing lemma is applied to SPDEs
\[
dX_t = AX_t\,dt + f(X_t,Y_t)\,dt + g(X_t,Y_t)\,dB_t,
\]
with \(A\) generating a bounded analytic semigroup, \(B\) a trace-class fBM with \(H>\tfrac12\), and \(Y\) a fast random environment [2108.05573]. For the stochastic convolution, the paper studies
\[
\Xi_{s,t}:=\int_s^t S_{t-r}g(r,X_s)\,dB_r
\]
and proves an \(L^p\)-estimate for the mild integral
\[
\left\|\int_s^t S_{t-r}g(r,X_r)\,dB_r\right\|_{L^p(E_{\kappa+\alpha})}
\lesssim
\mathfrak g_{-\delta,\gamma}^{p/\gamma}(L(E_{\kappa+\alpha}))
\bigl(1+\|X\|_{\hat C^\alpha_p(E_{\kappa+\alpha})}^\gamma\bigr)
|t-s|^{\bar\eta-\alpha}
\]
[2108.05573]. Together with the corresponding drift estimate, this yields uniform \(\hat C^\alpha_p\)-bounds, stability under perturbation, and a fractional averaging principle for SPDEs in random environments [2108.05573].

### Law-dependent jump SDEs

The Wasserstein-space theorem is applied to a McKean–Vlasov jump SDE on a Hilbert space, yielding a flow \(\theta_{s,t}\) on \(\mathcal P_2(H)\), weak existence of solutions, and uniqueness of marginal laws among solutions with \(W_2\)-continuous marginals [2410.17096]. The one-step Euler map satisfies the abstract \(\beta\)-sewing property with
\[
\beta=\frac12,
\]
which is sufficient for the stochastic-sewing construction of the law flow [2410.17096].

## 6. Proof architecture, related sewing paradigms, and conceptual scope

Across formulations, the proof strategy is remarkably stable. The standard ingredients are dyadic refinement, decomposition of partition errors into localized defect terms, a martingale inequality of BDG or Burkholder type for the centered component, and direct control of the predictable part by conditional expectations [1810.10500] [2108.05573] [2206.01686] [2510.12165]. In the Banach-valued theory, BDG is replaced by martingale type \(p\), changing the random threshold from \(1/2\) to \(1/p\) [2105.09364]. In several parameters, repeated one-dimensional BDG arguments require a commuting multiparameter filtration [2307.11527]. In the mild theory, semigroup smoothing provides the additional “seesaw” gain from time regularity to spatial regularity [2108.05573]. In the earlier-time extension, the predictable estimate is iterated across dyadic levels to exploit the singular factor \((t_1-t_0)^{-\alpha}\) [2206.01686].

The broader sewing landscape includes deterministic hypercube sewing and multiplicative sewing. The hypercube theory extends the deterministic lemma to fields on \([0,1]^k\), leading to multi-parameter Young integration and hyperbolic equations driven by Hölder fields [1806.09915]. The multiplicative sewing lemma, by contrast, constructs an evolution system from an almost multiplicative family of operators \(\mu_{s,t}\), with quantitative weak-approximation applications to SDEs [2206.10297]. These theories are adjacent rather than identical. Stochastic sewing is additive at the level of increments, but its proof and applications often interface with semigroup or operator structures that resemble multiplicative sewing. This suggests a common algebraic motif: local consistency plus a summability mechanism produce a global object. What differs is the summability mechanism—pathwise Hölder regularity in deterministic sewing, stochastic cancellation in stochastic sewing, and near-multiplicativity in multiplicative sewing.

A recurrent misconception is that stochastic sewing is merely a probabilistic proof of the deterministic sewing lemma. The literature indicates a stronger claim. The method is specifically designed for situations in which deterministic sewing fails because the defect is not regular enough pathwise, yet the conditional expectation or martingale structure is sufficiently regular to force convergence [1810.10500] [2510.12165]. Another misconception is that the framework is limited to semimartingales. The applications to fractional Brownian motion, Gaussian sheets, SPDEs with semigroup structure, and Wasserstein flows show that stochastic sewing is effective precisely in non-semimartingale and non-Markovian settings where conditional cancellations remain exploitable [2206.01686] [2108.05573] [2307.11527] [2410.17096].

In that sense, stochastic sewing functions as a unifying technique for converting local conditional smoothing into global integral, flow, or averaging estimates. The lecture notes place under this umbrella regularization by noise for Brownian and fractional Brownian SDEs, well-posedness of SPDEs with irregular drift, averaging operators and local times, and numerical algorithms such as Euler schemes [2510.12165]. A plausible synthesis is that stochastic sewing has evolved from a lemma about convergence of random Riemann sums into a general methodology for stochastic analysis whenever local increments are irregular pathwise but regular after conditioning.

Source: https://www.emergentmind.com/topics/stochastic-sewing