Dual Tamed Scheme for Index-1 SDAEs

• Dual Tamed Scheme is an equivalent reformulation of a semi-implicit tamed discretization for index-1 SDAEs, expressed in reduced coordinates to align with the inherent SDE dynamics.
• The approach uses projector-based reduction to decouple the algebraic constraints, enabling stability analysis and strong convergence order 1/2 under superlinear drift conditions.
• It distinguishes itself from standard tamed schemes by implicitly handling the linear drift and explicitly taming the nonlinear component, providing a rigorous framework for constrained stochastic dynamics.

A dual tamed scheme is an equivalent reformulation of a semi-implicit tamed discretization for index-1 stochastic differential algebraic equations (SDAEs) under non-global Lipschitz coefficients. In the usage documented in the recent SDAE literature, the term does not denote a second independent numerical algorithm; rather, it denotes the same semi-implicit tamed method rewritten in reduced variables so that the analysis aligns with the inherent stochastic differential equation obtained after eliminating the algebraic constraints. The central setting is the index-1 SDAE with singular matrix structure, superlinear drift growth, and strong convergence order $1/2$ for the resulting approximation (Tsafack et al., 10 Sep 2025).

1. Terminology and scope

The phrase “dual tamed scheme” is used explicitly for SDAEs with singular matrix $A(t)$, where the original numerical method is first posed in the full variable $X_n^N$ and then rewritten in reduced coordinates $(u_n,\hat v^N)$. In that setting, “dual” means that the numerical scheme is expressed in the coordinates induced by projector-based elimination of the algebraic constraints, so that it can be compared directly with the reduced exact dynamics $u(t)$ (Tsafack et al., 10 Sep 2025).

This terminology is narrower than the broader literature on tamed discretizations. In related work on McKean–Vlasov equations driven by Lévy noise, the core method is a tamed Euler scheme for an interacting particle system, and the paper states explicitly that it does not study a “Dual Tamed Scheme” in the sense of a separate duality-based numerical method; the approximation is instead two-stage, consisting of particle approximation and time discretization (Neelima et al., 2020). Likewise, the “semi-tamed Euler scheme” for jump–diffusion SDEs and the “tamed $\theta$-Euler-Maruyama scheme” for neutral SDDEs are distinct constructions, not dual schemes in this specific SDAE sense (Tambue et al., 2015, Tan et al., 2016).

A plausible implication is that the term should be reserved for contexts where a constrained stochastic system is first reduced to an inherent SDE and the numerical method is then rewritten in those reduced variables. Outside that setting, the literature more commonly distinguishes between tamed, semi-tamed, modified tamed, and $\theta$-tamed schemes rather than dual schemes.

2. SDAE setting and reduction to an inherent SDE

The underlying model is the SDAE

$$A(t)dX(t)=\left[B(t)X(t)+f(t,X(t))\right]dt + g(t,X(t))dW(t), \quad t\in \left[0,T\right] ,\quad X(0)=X_0,$$

where $A(t)\in\mathbb{R}^{d\times d}$ is singular, $B(t)\in\mathbb{R}^{d\times d}$, $A(t)$0 is the drift, $A(t)$1 the diffusion, and $A(t)$2 is an $A(t)$3-dimensional Wiener process (Tsafack et al., 10 Sep 2025).

Because $A(t)$4 is singular, the system is not a standard SDE. The analysis proceeds by using Itô’s formula on $A(t)$5, which yields

$A(t)$6

To define an index-1 SDAE, the paper requires that the noise does not enter the algebraic constraints,

$A(t)$7

and that the constraint equations are globally uniquely solvable (Tsafack et al., 10 Sep 2025).

The reduction uses projector objects associated with $A(t)$8: the pseudo-inverse $A(t)$9, the projector $X_n^N$0, and $X_n^N$1 with $X_n^N$2. Under these projectors, the SDAE is reduced to an equivalent inherent SDE in a reduced variable $X_n^N$3, with the algebraic component recovered through $X_n^N$4. The reduced form is

$X_n^N$5

where

$X_n^N$6

This decomposition is the exact structural template that the dual tamed scheme mirrors (Tsafack et al., 10 Sep 2025).

3. Semi-implicit tamed discretization and its dual form

The original numerical method is the semi-implicit tamed scheme. Its design principle is stated explicitly: the linear component of the drift term is approximated implicitly, whereas its nonlinear component is tamed and approximated explicitly. The scheme is

$X_n^N$7

or equivalently

$X_n^N$8

with $X_n^N$9 and $(u_n,\hat v^N)$0 (Tsafack et al., 10 Sep 2025).

The explicit decomposition of the update is central. The linear drift term $(u_n,\hat v^N)$1 is handled implicitly, the nonlinear drift $(u_n,\hat v^N)$2 is tamed through

$(u_n,\hat v^N)$3

and the diffusion remains explicit (Tsafack et al., 10 Sep 2025).

The dual tamed scheme is then stated in reduced coordinates. In Lemma 7, it is written as

$(u_n,\hat v^N)$4

where $(u_n,\hat v^N)$5 is globally Lipschitz in $(u_n,\hat v^N)$6 (Tsafack et al., 10 Sep 2025).

This dual form is the numerical counterpart of the exact decomposition $(u_n,\hat v^N)$7. The paper states that the original semi-implicit scheme in $(u_n,\hat v^N)$8 is equivalent to the dual scheme in $(u_n,\hat v^N)$9, provided $u(t)$0 is nonsingular. This is the precise sense in which the method is dual rather than separate (Tsafack et al., 10 Sep 2025).

4. Assumptions, stability structure, and strong convergence

The coefficient assumptions for the SDAE formulation are: $u(t)$1

$u(t)$2

$u(t)$3

$u(t)$4

These are the non-global Lipschitz hypotheses under which the method is analyzed (Tsafack et al., 10 Sep 2025).

The matrix and projector assumptions are equally important. Assumption 2 requires that $u(t)$5 are bounded and Lipschitz; that $u(t)$6 and $u(t)$7 are nonsingular with bounded inverses; and that

$u(t)$8

is nonsingular with

$u(t)$9

where

$\theta$0

This algebraic structure is what makes the reduced formulation mathematically tractable (Tsafack et al., 10 Sep 2025).

The main convergence theorem states that, for all $\theta$1, there exists a constant $\theta$2 such that

$\theta$3

Thus the scheme converges strongly with order $\theta$4 (Tsafack et al., 10 Sep 2025).

The proof structure combines several intermediate results. Theorem 2 gives uniform moment bounds for the numerical solution,

$\theta$5

Lemma 10 gives interpolation estimates for the reduced variable,

$\theta$6

and Lemma 9 controls the algebraic component through

$\theta$7

The final estimate for $\theta$8 is then transferred to $\theta$9 (Tsafack et al., 10 Sep 2025).

5. Relation to other tamed constructions

The dual tamed scheme belongs to a broader family of explicit or semi-implicit stabilizations for stochastic systems with superlinear drift, but its defining feature is the projector-based reduced-coordinate reformulation for SDAEs.

For jump–diffusion SDEs, one related construction is the semi-tamed Euler scheme, where the drift is split as $\theta$0 and only the nonglobally Lipschitz part is tamed: $\theta$1 The same paper also studies the fully tamed drift scheme

$\theta$2

and compares their linear and nonlinear mean-square stability under non-global Lipschitz conditions (Tambue et al., 2015). This is a duality only in the informal sense of comparing two taming placements, not in the reduced-coordinate SDAE sense.

For neutral stochastic delay differential equations, the tamed $\theta$3-Euler-Maruyama scheme is semi-implicit in the drift through a $\theta$4-weight: $\theta$5 Its tamed coefficients satisfy bounds such as

$\theta$6

and the strong convergence estimate

$\theta$7

under global monotonicity assumptions (Tan et al., 2016). Again, this is not presented as a dual scheme.

For McKean–Vlasov SDEs driven by Lévy noise, the numerical method is a tamed Euler scheme applied to an interacting particle system,

$\theta$8

with taming exemplified by

$\theta$9

The paper states that there is no dual scheme there; the relevant viewpoint is only the two-level approximation by particle approximation and time discretization (Neelima et al., 2020).

A further related variant is the modified tamed scheme

$$A(t)dX(t)=\left[B(t)X(t)+f(t,X(t))\right]dt + g(t,X(t))dW(t), \quad t\in \left[0,T\right] ,\quad X(0)=X_0,$$0

where the drift is unchanged whenever $$A(t)dX(t)=\left[B(t)X(t)+f(t,X(t))\right]dt + g(t,X(t))dW(t), \quad t\in \left[0,T\right] ,\quad X(0)=X_0,$$1, and the paper proves strong order $$A(t)dX(t)=\left[B(t)X(t)+f(t,X(t))\right]dt + g(t,X(t))dW(t), \quad t\in \left[0,T\right] ,\quad X(0)=X_0,$$2 and weak order $$A(t)dX(t)=\left[B(t)X(t)+f(t,X(t))\right]dt + g(t,X(t))dW(t), \quad t\in \left[0,T\right] ,\quad X(0)=X_0,$$3 for Euler under the stated assumptions (Ju et al., 13 Jul 2025). This suggests that contemporary taming research distinguishes methods mainly by where and when the damping acts: on the full drift, on a decomposed nonlinear component, on an implicit-explicit split, or on a reduced coordinate system.

6. Example, interpretation, and limitations

The representative numerical example for the dual tamed scheme is a three-dimensional SDAE,

$$A(t)dX(t)=\left[B(t)X(t)+f(t,X(t))\right]dt + g(t,X(t))dW(t), \quad t\in \left[0,T\right] ,\quad X(0)=X_0,$$4

with

$$A(t)dX(t)=\left[B(t)X(t)+f(t,X(t))\right]dt + g(t,X(t))dW(t), \quad t\in \left[0,T\right] ,\quad X(0)=X_0,$$5

$$A(t)dX(t)=\left[B(t)X(t)+f(t,X(t))\right]dt + g(t,X(t))dW(t), \quad t\in \left[0,T\right] ,\quad X(0)=X_0,$$6

The corresponding projectors satisfy

$$A(t)dX(t)=\left[B(t)X(t)+f(t,X(t))\right]dt + g(t,X(t))dW(t), \quad t\in \left[0,T\right] ,\quad X(0)=X_0,$$7

and the constraint equation yields

$$A(t)dX(t)=\left[B(t)X(t)+f(t,X(t))\right]dt + g(t,X(t))dW(t), \quad t\in \left[0,T\right] ,\quad X(0)=X_0,$$8

showing that the SDAE is index-1 (Tsafack et al., 10 Sep 2025).

The paper further verifies the one-sided Lipschitz property

$$A(t)dX(t)=\left[B(t)X(t)+f(t,X(t))\right]dt + g(t,X(t))dW(t), \quad t\in \left[0,T\right] ,\quad X(0)=X_0,$$9

and the matrix estimate

$A(t)\in\mathbb{R}^{d\times d}$0

The simulation uses $A(t)\in\mathbb{R}^{d\times d}$1, $A(t)\in\mathbb{R}^{d\times d}$2, $A(t)\in\mathbb{R}^{d\times d}$3, and 300 Monte Carlo samples; the resulting log-log plot shows strong convergence with slope $A(t)\in\mathbb{R}^{d\times d}$4, matching the theorem (Tsafack et al., 10 Sep 2025).

The principal limitation is terminological rather than mathematical. In the available literature, “dual tamed scheme” is tied specifically to the projector-based reduction of index-1 SDAEs. By contrast, in other settings the literature either uses a tamed Euler scheme without any dual formulation, or introduces different modifiers such as semi-taming, $A(t)\in\mathbb{R}^{d\times d}$5-taming, or selective cut-off taming (Neelima et al., 2020, Tambue et al., 2015, Tan et al., 2016, Ju et al., 13 Jul 2025). This suggests that the term names an analytical representation of a scheme adapted to constrained stochastic dynamics, rather than a universal class of tamed integrators.