Stochastic Theta Method (STM)

• STM is a family of one-step discretizations for stochastic differential equations that parametrizes drift implicitness using a parameter θ in [0,1].
• It encompasses variants like Euler–Maruyama (θ=0), Crank–Nicolson-type (θ=1/2), and backward Euler (θ=1), each balancing stability and accuracy.
• Rigorous theories demonstrate STM’s strong convergence order of 1/2 and weak order of 1, along with invariant measure and ergodicity properties under suitable conditions.

Searching arXiv for recent and foundational papers on the stochastic theta method to ground the encyclopedia entry. arxiv_search.query({ "^^^^^^^^3search_query3 "^^^^^^^^3\3 theta method^^^^^^^^3\3 OR ^^^^^^^^3\3 theta methods^^^^^^^^3\3 "start": ^^^^3search_query3^^^^, "max_results": ^^^^3\3search_query3^^^^, "sortBy": "submittedDate", "sortOrder": "descending" }) The Stochastic Theta Method (STM) is a family of drift-implicit, diffusion-explicit one-step discretizations for stochastic differential equations, parametrized by PRESERVED_PLACEHOLDER_^^^^3search_query3^^^^. In its classical Itô form, the method updates the drift as a convex combination of evaluations at the old and new time levels, so that PRESERVED_PLACEHOLDER_^^^^3\3^^^^ gives Euler–Maruyama, PRESERVED_PLACEHOLDER_^^^^3 OR \3^^^^ gives a Crank–Nicolson-type or midpoint treatment of the drift, and $\theta=1$ gives a backward Euler–type method. Across the recent literature, STM has been studied for monotone SDEs, jump-diffusions, stochastic delay equations, stochastic differential algebraic equations, stochastic partial differential equations, free stochastic differential equations, and equations driven by fractional Brownian motion or time-changed Lévy noise, with analyses covering well-posedness, strong and weak convergence, contractivity, ergodicity, invariant measures, and structure preservation (&&&^^^^3search_query3^^^^&&&, &&&^^^^3\3^^^^&&&, &&&^^^^3 OR \3^^^^&&&, Niu et al., 2024).

^^^^3\3^^^^. Canonical formulation and principal variants

For an Itô SDE $dX(t)=b(X(t))\,dt+\sigma(X(t))\,dW(t)$, the STM with time step $\tau$ is written as

$$X_{j+1}=X_j+(1-\theta)b(X_j)\tau+\theta b(X_{j+1})\tau+\sigma(X_j)\Delta W_j,$$

with $\Delta W_j=W(t_{j+1})-W(t_j)$. The same drift-implicit, diffusion-explicit pattern also appears in the time-dependent form

$$X_{n+1}=X_n+h\big((1-\theta)a(t_n,X_n)+\theta a(t_{n+1},X_{n+1})\big)+b(t_n,X_n)\Delta W_n,$$

and in each case $\theta$ controls the degree of implicitness (&&&^^^^3search_query3^^^^&&&, &&&^^^^3\3^^^^&&&).

In jump-diffusion settings, the STM keeps the jump term explicit. For monotone jump-diffusion stochastic ordinary differential equations,

PRESERVED_PLACEHOLDER_^^^^3\3search_query3^^^^

where the drift is implicit and the diffusion and jump terms are evaluated at PRESERVED_PLACEHOLDER_^^^^3\3\3^^^^. The special case PRESERVED_PLACEHOLDER_^^^^3\3 OR \3^^^^ is the Backward Euler Method (BEM) (&&&^^^^3 OR \3^^^^&&&). Closely related compensated versions for jump SDEs replace raw Poisson increments by compensated ones, yielding the compensated stochastic theta method

PRESERVED_PLACEHOLDER_^^^^3\33^^^^

with PRESERVED_PLACEHOLDER_^^^^3\3^^^^4 (Mukam, 2015).

The same structural idea persists in more specialized models. For random time change equations and piecewise deterministic Markov process representations, the PRESERVED_PLACEHOLDER_^^^^3\3^^^^5-Maruyama scheme combines an implicit drift update with Poisson increments driven by quadrature approximations of integrated intensities (Riedler et al., 2013). For free stochastic differential equations, the free STM is

PRESERVED_PLACEHOLDER_^^^^3\3^^^^6

again retaining drift implicitness and explicit noise evaluation (Niu et al., 2024).

^^^^3 OR \3^^^^. Solvability and structural assumptions

The solvability theory of STM is highly assumption-dependent, but a recurring pattern is monotonicity or one-sided Lipschitz control of the drift combined with explicit regularity assumptions on the noise. For monotone jump-diffusion SODEs, Assumption A^^^^3\3^^^^ imposes monotonicity and coercivity, while Assumption A^^^^3 OR \3^^^^ gives a dissipativity or contractivity inequality of the form

PRESERVED_PLACEHOLDER_^^^^3\3^^^^7

Under A^^^^3\3^^^^ and PRESERVED_PLACEHOLDER_^^^^3\3^^^^8, the STM admits a unique almost surely solution and defines a homogeneous Markov chain (&&&^^^^3 OR \3^^^^&&&).

For nonlinear SDEs and invariant-measure questions, a standard reformulation is PRESERVED_PLACEHOLDER_^^^^3\3^^^^9. Under one-sided Lipschitz conditions such as PRESERVED_PLACEHOLDER_^^^^3 OR \3search_query3^^^^, strict monotonicity of PRESERVED_PLACEHOLDER_^^^^3 OR \3\3^^^^ yields well-definedness; one representative condition is PRESERVED_PLACEHOLDER_^^^^3 OR \3 OR \3^^^^ (&&&^^^^3\3\3^^^^&&&). In mean-square contractivity analysis, a related implicit resolvent estimate appears as

PRESERVED_PLACEHOLDER_^^^^3 OR \33^^^^

for the implicit mapping PRESERVED_PLACEHOLDER_^^^^3 OR \3^^^^4, under a one-sided Lipschitz constant PRESERVED_PLACEHOLDER_^^^^3 OR \3^^^^5 (&&&^^^^3\3 OR \3^^^^&&&).

In structure-preserving index-^^^^3\3^^^^ SDAEs with time-dependent singular matrices, the issue is not only solvability of the drift-implicit step but also preservation of algebraic constraints. There the matrix PRESERVED_PLACEHOLDER_^^^^3 OR \3^^^^6 is singular, the projectors PRESERVED_PLACEHOLDER_^^^^3 OR \3^^^^7 are time-independent, the algebraic Jacobian PRESERVED_PLACEHOLDER_^^^^3 OR \3^^^^8 is assumed invertible with uniformly bounded inverse, and a reconstruction map PRESERVED_PLACEHOLDER_^^^^3 OR \3^^^^9 is defined by

$\theta=1$^^^^3search_query3^^^^

For sufficiently small $\theta=1$^^^^3\3^^^^, the induced fixed-point map for the differential component is a contraction, so the method is well posed and preserves the constraint manifold at every step (&&&^^^^3\3^^^^&&&). An analogous global monotonicity framework under non-globally Lipschitz coefficients yields well-posedness for index-^^^^3\3^^^^ SDAEs with non-constant singular matrices when $\theta=1$^^^^3 OR \3^^^^ and $\theta=1$3 satisfies explicit small-step restrictions (&&&^^^^3\3^^^^4&&&).

In the free-probability setting, solvability is formulated in operator terms. If the drift coefficient is operator Lipschitz in operator norm with constant $\theta=1$4, then $\theta=1$5 is invertible for $\theta=1$6, and the free STM step has a unique adapted solution (Niu et al., 2024).

3. Convergence theory

The strong convergence theory of STM is broad but not uniform across model classes. In classical monotone SDE settings, the free STM converges strongly with order $\theta=1$7 in $\theta=1$8,

$\theta=1$9

under an operator-Lipschitz drift and locally operator-Lipschitz diffusion coefficients (Niu et al., 2024). For random periodic solutions of SDEs under non-globally Lipschitz conditions, the random periodic solution generated by STM with $dX(t)=b(X(t))\,dt+\sigma(X(t))\,dW(t)$^^^^3search_query3^^^^ converges strongly in mean square with order $dX(t)=b(X(t))\,dt+\sigma(X(t))\,dW(t)$^^^^3\3^^^^ for multiplicative noise and order $dX(t)=b(X(t))\,dt+\sigma(X(t))\,dW(t)$^^^^3 OR \3^^^^ for additive noise (&&&^^^^3\3^^^^7&&&).

For index-^^^^3\3^^^^ SDAEs, strong mean-square order $dX(t)=b(X(t))\,dt+\sigma(X(t))\,dW(t)$3 has now been established under global monotonicity conditions. With $dX(t)=b(X(t))\,dt+\sigma(X(t))\,dW(t)$4, the direct SDAE-level STM

$dX(t)=b(X(t))\,dt+\sigma(X(t))\,dW(t)$5

satisfies

$dX(t)=b(X(t))\,dt+\sigma(X(t))\,dW(t)$6

both for non-constant singular matrices under an additional monotonicity condition and, in a simpler form, for constant singular matrices (&&&^^^^3\3^^^^4&&&). For structure-preserving STM applied to index-^^^^3\3^^^^ SDAEs with time-dependent singular matrices, the weak, rather than strong, conclusion is order one: $dX(t)=b(X(t))\,dt+\sigma(X(t))\,dW(t)$7 for $dX(t)=b(X(t))\,dt+\sigma(X(t))\,dW(t)$8 (&&&^^^^3\3^^^^&&&).

The weak convergence theory is particularly sharp for BEM and for time-changed Lévy settings. For monotone jump-diffusion SODEs, the weak error analysis in (&&&^^^^3 OR \3^^^^&&&) is carried out for $dX(t)=b(X(t))\,dt+\sigma(X(t))\,dW(t)$9. Under A3–A4, the finite-time weak error with jumps satisfies

$\tau$^^^^3search_query3^^^^

where the exponent reflects the jump terms and polynomial growth of derivatives. In the jump-free, strongly dissipative case, the same paper proves the time-independent estimate

$\tau$^^^^3\3^^^^

and consequently the invariant measures satisfy

$\tau$^^^^3 OR \3^^^^

That first-order convergence of invariant measures answers a question left open in Z. Liu and Z. Liu, J. Sci. Comput. (^^^^3 OR \3search_query3 OR \3^^^^5) ^^^^3\3search_query33^^^^:87 (&&&^^^^3 OR \3^^^^&&&).

For SDEs driven by time-changed Lévy noise under global Lipschitz and linear growth conditions, the weak convergence order of STM with $\tau$3 is again order one: $\tau$4 The analysis separates the SDE discretization error from the approximation error of the inverse subordinator and uses Kolmogorov backward PIDE techniques together with the duality principle (&&&^^^^3 OR \3 OR \3^^^^&&&). In contrast, the strong convergence of STM for time-changed Brownian or Lévy equations under local Lipschitz conditions is governed by the time-change scaling. One paper proves a strong order $\tau$5 for time-changed Brownian equations (&&&^^^^3 OR \33^^^^&&&), and another gives $\tau$6 for time-changed Lévy equations (&&&^^^^3 OR \3^^^^4&&&).

4. Stability, contractivity, ergodicity, and invariant measures

A central reason for using STM is its stability profile under monotone or dissipative dynamics. In nonlinear mean-square contractivity, the continuous rate is $\tau$7, where $\tau$8 is a one-sided Lipschitz constant for the drift and $\tau$9 is a global Lipschitz constant for the diffusion. For the $$X_{j+1}=X_j+(1-\theta)b(X_j)\tau+\theta b(X_{j+1})\tau+\sigma(X_j)\Delta W_j,$$^^^^3search_query3^^^^-Maruyama method,

$$X_{j+1}=X_j+(1-\theta)b(X_j)\tau+\theta b(X_{j+1})\tau+\sigma(X_j)\Delta W_j,$$^^^^3\3^^^^

with one-step factor

$$X_{j+1}=X_j+(1-\theta)b(X_j)\tau+\theta b(X_{j+1})\tau+\sigma(X_j)\Delta W_j,$$^^^^3 OR \3^^^^

This yields sharp stepsize restrictions for contractivity; when $$X_{j+1}=X_j+(1-\theta)b(X_j)\tau+\theta b(X_{j+1})\tau+\sigma(X_j)\Delta W_j,$$3, implicit Euler–Maruyama is unconditionally mean-square contractive whenever $$X_{j+1}=X_j+(1-\theta)b(X_j)\tau+\theta b(X_{j+1})\tau+\sigma(X_j)\Delta W_j,$$4 (&&&^^^^3\3 OR \3^^^^&&&).

For monotone SODEs driven by nondegenerate multiplicative noise, STM with $$X_{j+1}=X_j+(1-\theta)b(X_j)\tau+\theta b(X_{j+1})\tau+\sigma(X_j)\Delta W_j,$$5 has a unique invariant measure, without growth restriction on the coefficients. The analysis uses new Lyapunov functions involving the coefficients, the stepsize, and $$X_{j+1}=X_j+(1-\theta)b(X_j)\tau+\theta b(X_{j+1})\tau+\sigma(X_j)\Delta W_j,$$6, plus irreducibility and strong Feller arguments. For $$X_{j+1}=X_j+(1-\theta)b(X_j)\tau+\theta b(X_{j+1})\tau+\sigma(X_j)\Delta W_j,$$7, geometric ergodicity is also proved under a mild additional smoothness assumption on $$X_{j+1}=X_j+(1-\theta)b(X_j)\tau+\theta b(X_{j+1})\tau+\sigma(X_j)\Delta W_j,$$8 (&&&^^^^3search_query3^^^^&&&). In the jump-diffusion setting, a discrete Lyapunov function

$$X_{j+1}=X_j+(1-\theta)b(X_j)\tau+\theta b(X_{j+1})\tau+\sigma(X_j)\Delta W_j,$$9

yields exponential ergodicity for $\Delta W_j=W(t_{j+1})-W(t_j)$^^^^3search_query3^^^^. Under A^^^^3 OR \3^^^^, the method has a unique invariant measure $\Delta W_j=W(t_{j+1})-W(t_j)$^^^^3\3^^^^ and exponential mixing with rate $\Delta W_j=W(t_{j+1})-W(t_j)$^^^^3 OR \3^^^^ (&&&^^^^3 OR \3^^^^&&&).

The invariant-measure literature now distinguishes several regimes. For nonlinear SDEs, the existence and uniqueness of the stationary distribution of STM depend on $\Delta W_j=W(t_{j+1})-W(t_j)$3: when $\Delta W_j=W(t_{j+1})-W(t_j)$4, global Lipschitz and explicit step-size conditions are required, while for $\Delta W_j=W(t_{j+1})-W(t_j)$5 the drift may be only one-sided Lipschitz or dissipative, provided the diffusion is Lipschitz (&&&^^^^3\3\3^^^^&&&). For super-linearly growing drift and diffusion coefficients, STM with $\Delta W_j=W(t_{j+1})-W(t_j)$6 admits a unique numerical invariant measure under polynomial local Lipschitz growth, coupled monotonicity, and coercivity assumptions, and the numerical invariant measure converges to the exact invariant measure with

$\Delta W_j=W(t_{j+1})-W(t_j)$7

for sufficiently small $\Delta W_j=W(t_{j+1})-W(t_j)$8 (&&&^^^^3 OR \3^^^^9&&&). This extends earlier invariant-measure results beyond linear-growth diffusion.

Stability inheritance also appears outside classical Brownian Itô calculus. In free stochastic differential equations, if $\Delta W_j=W(t_{j+1})-W(t_j)$9, then the exact equation is exponentially mean-square stable, and the free STM with $$X_{n+1}=X_n+h\big((1-\theta)a(t_n,X_n)+\theta a(t_{n+1},X_{n+1})\big)+b(t_n,X_n)\Delta W_n,$$^^^^3search_query3^^^^ preserves that stability under a step-size restriction, while the free BEM ($$X_{n+1}=X_n+h\big((1-\theta)a(t_n,X_n)+\theta a(t_{n+1},X_{n+1})\big)+b(t_n,X_n)\Delta W_n,$$^^^^3\3^^^^) is unconditionally exponentially mean-square stable for any $$X_{n+1}=X_n+h\big((1-\theta)a(t_n,X_n)+\theta a(t_{n+1},X_{n+1})\big)+b(t_n,X_n)\Delta W_n,$$^^^^3 OR \3^^^^ (Niu et al., 2024). For SDEs driven by fractional Brownian motion with Hurst parameter $$X_{n+1}=X_n+h\big((1-\theta)a(t_n,X_n)+\theta a(t_{n+1},X_{n+1})\big)+b(t_n,X_n)\Delta W_n,$$3, unconditional mean-square stability of STM is proved for the linear test equation when either $$X_{n+1}=X_n+h\big((1-\theta)a(t_n,X_n)+\theta a(t_{n+1},X_{n+1})\big)+b(t_n,X_n)\Delta W_n,$$4 and

$$X_{n+1}=X_n+h\big((1-\theta)a(t_n,X_n)+\theta a(t_{n+1},X_{n+1})\big)+b(t_n,X_n)\Delta W_n,$$5

or $$X_{n+1}=X_n+h\big((1-\theta)a(t_n,X_n)+\theta a(t_{n+1},X_{n+1})\big)+b(t_n,X_n)\Delta W_n,$$6 and $$X_{n+1}=X_n+h\big((1-\theta)a(t_n,X_n)+\theta a(t_{n+1},X_{n+1})\big)+b(t_n,X_n)\Delta W_n,$$7; by contrast, $$X_{n+1}=X_n+h\big((1-\theta)a(t_n,X_n)+\theta a(t_{n+1},X_{n+1})\big)+b(t_n,X_n)\Delta W_n,$$8 does not preserve stability unconditionally (&&&3^^^^3\3^^^^&&&).

5. Structured and generalized settings

STM has been adapted to problems with constraints, infinite-dimensional structure, delay, and noncommutative noise. In index-^^^^3\3^^^^ SDAEs with time-dependent singular matrices, the structure-preserving method

$$X_{n+1}=X_n+h\big((1-\theta)a(t_n,X_n)+\theta a(t_{n+1},X_{n+1})\big)+b(t_n,X_n)\Delta W_n,$$9

preserves the algebraic constraint $\theta$^^^^3search_query3^^^^ because applying $\theta$^^^^3\3^^^^ to the discrete equation yields $\theta$^^^^3 OR \3^^^^. This direct SDAE-level formulation preserves both the constraint manifold and the differential–algebraic splitting (&&&^^^^3\3^^^^&&&). The complementary strong-convergence theory for non-globally Lipschitz index-^^^^3\3^^^^ SDAEs uses a projector-based reduction, algebraic reconstruction via $\theta$3, and monotonicity conditions on the reduced coefficients (&&&^^^^3\3^^^^4&&&).

In SPDEs, the $\theta$4 case becomes a drift-implicit Euler method in Hilbert space. For the stochastic Allen–Cahn equation and related monotone SPDEs driven by infinite-dimensional nondegenerate multiplicative trace-class noise, the Galerkin-based full discretizations inherit unique ergodicity. In particular, for the stochastic Allen–Cahn equation, unique ergodicity holds for any interface thickness $\theta$5 under

$\theta$6

(&&&^^^^3search_query3^^^^&&&). This places STM within the broader theory of drift-implicit space-time discretizations for monotone SPDEs.

Delay equations constitute another extension. For neutral stochastic differential delay equations under non-globally Lipschitz coefficients, the theta–EM scheme satisfies

$\theta$7

in the Brownian case, while for pure-jump NSDDEs the corresponding bound is

$\theta$8

with almost sure convergence rates $\theta$9 and PRESERVED_PLACEHOLDER_^^^^3\3search_query3search_query3^^^^, respectively (Tan et al., 2017). For SDDEs with small noise, the theta EM scheme is further embedded into a multilevel Monte Carlo construction, and the level-difference variance under global Lipschitz assumptions satisfies

PRESERVED_PLACEHOLDER_^^^^3\3search_query3\3^^^^

which sharpens the single-level variance bound in the small-noise regime (Tan et al., 2019).

STM also appears in random periodic, free, and hybrid stochastic systems. For random periodic solutions under non-globally Lipschitz conditions, STM with PRESERVED_PLACEHOLDER_^^^^3\3search_query3 OR \3^^^^ admits a unique numerical random periodic solution, and that solution converges strongly in mean square to the exact random periodic solution, with order PRESERVED_PLACEHOLDER_^^^^3\3search_query33^^^^ for multiplicative noise and PRESERVED_PLACEHOLDER_^^^^3\3search_query3^^^^4 for additive noise (&&&^^^^3\3^^^^7&&&). In stochastic hybrid systems formulated through random time change equations, the PRESERVED_PLACEHOLDER_^^^^3\3search_query3^^^^5-Maruyama family is analyzed by local error expansions, and strong convergence in mean is established for semi-implicit Maruyama-type methods (Riedler et al., 2013).

6. Computational practice and open directions

Across the literature, implementation of STM is dominated by the implicit solve in the drift variable. For monotone jump-diffusions, one solves

PRESERVED_PLACEHOLDER_^^^^3\3search_query3^^^^6

typically by fixed-point iteration or Newton-type methods; monotonicity of PRESERVED_PLACEHOLDER_^^^^3\3search_query3^^^^7 and small PRESERVED_PLACEHOLDER_^^^^3\3search_query3^^^^8 ensure contraction (&&&^^^^3 OR \3^^^^&&&). In structure-preserving SDAEs, Newton’s method is applied to the residual equation

PRESERVED_PLACEHOLDER_^^^^3\3search_query3^^^^9

and constraint preservation follows automatically for PRESERVED_PLACEHOLDER_^^^^3\3\3search_query3^^^^ (&&&^^^^3\3^^^^&&&). In free stochastic differential equations, either fixed-point iteration or direct evaluation of PRESERVED_PLACEHOLDER_^^^^3\3\3\3^^^^ is used, depending on whether the drift is linear or nonlinear (Niu et al., 2024).

Specialized noise models introduce additional numerical layers. For time-changed SDEs, one typically simulates the subordinator on an operational-time grid and approximates the inverse subordinator by first-passage indexing; in the PRESERVED_PLACEHOLDER_^^^^3\3\3 OR \3^^^^-stable case, Chambers–Mallows–Stuck sampling is used for stable increments, while the weak-analysis paper on time-changed Lévy equations employs a first-passage approximation PRESERVED_PLACEHOLDER_^^^^3\3\33^^^^ together with compound Poisson simulation for finite-activity jumps (&&&^^^^3 OR \33^^^^&&&, &&&^^^^3 OR \3 OR \3^^^^&&&). For time-changed Lévy equations with small jumps, the jump increment can be represented through Poisson counts and sampled jump sizes, whereas infinite-activity models motivate truncation or Lévy–Itô splitting (&&&^^^^3 OR \3 OR \3^^^^&&&, &&&^^^^3 OR \3^^^^4&&&).

Several open problems are now explicit in the STM literature. For monotone jump-diffusion SODEs, time-independent weak error analysis with jumps remains open, mainly because uniform moment bounds for STM are not available in that setting; similarly, extending time-independent weak error and invariant-measure accuracy from BEM to general PRESERVED_PLACEHOLDER_^^^^3\3\3^^^^4 is still unresolved (&&&^^^^3 OR \3^^^^&&&). In structure-preserving SDAEs, time-dependent splitting, non-globally Lipschitz coefficients, and higher weak order remain open directions (&&&^^^^3\3^^^^&&&). For monotone SDEs with multiplicative noise, geometric ergodicity is established for PRESERVED_PLACEHOLDER_^^^^3\3\3^^^^5, but not for PRESERVED_PLACEHOLDER_^^^^3\3\3^^^^6 or PRESERVED_PLACEHOLDER_^^^^3\3\3^^^^7 (&&&^^^^3search_query3^^^^&&&). For fractional Brownian motion, the regime PRESERVED_PLACEHOLDER_^^^^3\3\3^^^^8 with PRESERVED_PLACEHOLDER_^^^^3\3\3^^^^9 remains unsettled, even though numerical evidence suggests stability may persist at PRESERVED_PLACEHOLDER_^^^^3\3 OR \3search_query3^^^^ in some cases (&&&3^^^^3\3^^^^&&&).

Taken together, these results place STM not as a single scheme with a single theory, but as a parameterized discretization principle whose precise analytical behavior depends on the interaction among drift implicitness, monotonicity, coercivity, constraint structure, and the regularity of the driving noise. The dominant pattern across the current literature is that PRESERVED_PLACEHOLDER_^^^^3\3 OR \3\3^^^^ is the analytically preferred regime: it strengthens dissipativity, enlarges stability regions, supports ergodicity and invariant-measure approximation, and, in several generalized settings, is the threshold at which the strongest known long-time results become available (&&&^^^^3search_query3^^^^&&&, &&&^^^^3 OR \3^^^^&&&, &&&^^^^3 OR \3^^^^9&&&, &&&^^^^3\3^^^^4&&&).