Adaptive Diagonally Implicit Runge-Kutta Methods Devoid of Order Reduction for Semilinear ODEs

1. Introduction

• Stiff semilinear problems can require step sizes larger than the fastest time scale, where traditional Runge–Kutta methods may converge below their classical order.Such problems often arise from method-of-lines discretizations of PDEs.Source: §1 · p. 2

1.1. Runge–Kutta Methods: Classical Order Versus Uniform Accuracy

• Classical Taylor-based local-error expansions can be unreliable in stiffness because powers of J may dominate, while the alternative expansion remains uniformly bounded in Z = hJ.Enforcing stage-order conditions can remove local-error terms, but stage order is limited to two for DIRK methods; semilinear conditions are less restrictive and admit DIRK structures.Source: §1.1 · pp. 3–4

1.2. Order Reduction and Work in Context

• Prior work on order reduction has largely focused on linear problems, modified equations, or methods tied to stage order, while adaptive error estimation and embeddings have received little attention.The paper positions its semilinear order-condition theory as a companion to.Source: §1.2 · p. 4

1.3. Contributions and Outline of the Paper

• The paper reformulates stiff semilinear order conditions as algebraic matrix equations and constructs fourth- and fifth-order DIRK methods with optimized errors and stability properties.Experiments assess predicted convergence orders and order reduction, including the effect on embedded error estimates and adaptive stepping.Source: §1.3 · pp. 4–5

Assumptions

• The theorem assumes bounded derivatives of f and g and stability properties of the Runge–Kutta scheme; the function bounds may instead hold in a tubular neighborhood with additional step-size restrictions.The scheme assumptions include nonsingularity of I − zA and uniform boundedness conditions in the closed left half-plane.Source: §2 · p. 5

Semilinear order conditions

• Semilinear order pSL is defined by satisfying the listed conditions through order pSL, with eight conditions through order four and three higher-order conditions redundant.Each condition must hold over all powers of A; setting its indices to zero recovers classical order conditions.Source: §2 · p. 5, Table 1 · p. 6

Uniform-in-J error estimates

• Under the stated smoothness, boundedness, and scheme-stability assumptions, local and global error bounds are uniform in J and Z.For equal step sizes, the global error bound is also independent of J and Z.Source: §2 · p. 6

Superconvergence

• With additional differentiability, classical order pSL + 1 conditions, and a technical stability condition, the global order improves to pSL + 1.The technical condition is expressed through the stability function R(z).Source: §2 · p. 6

Matrix formulation

• The order conditions are interpreted as orthogonality relations, yielding a matrix-equation formulation for constructing schemes up to semilinear order four.The formulation is intended to make scheme construction tractable.Source: §3 · p. 7

Order-four conditions

• For semilinear order four, B(4) and a solution to system (18) are sufficient to satisfy the order conditions in Table 1.The construction uses matrices Λ, X, W, and Y, with dimensions selected through d1 and d2.Source: §3 · pp. 7–8

Order-three conditions

• Semilinear order three can be established by B(3) together with a subset of matrix equations, and this case requires no block structure on Λ, X, or W.The stated subset uses Λ, X, and W.Source: §3 · p. 8

4. Construction of methods with high semilinear order

• The methods satisfy semilinear conditions up to order four and can attain fifth-order global error through superconvergence.The schemes are designed with stability and accuracy properties for stiff semilinear ODEs.Source: §4 · p. 8
• The paper omits new semilinear-order-two schemes because many optimized methods of that type already exist, and develops ESDIRK methods with stage order two and EDIRK methods with stage order one.The two EDIRK schemes are less optimized and primarily of theoretical interest.Source: §4 · p. 9

4.1. Construction of ESDIRK-(8,4,3) and ESDIRK-(10,5,4)

• Symbolic computation produces stiffly accurate, L-stable ESDIRK-(8,4,3) and ESDIRK-(10,5,4) schemes with embeddings that satisfy classical and semilinear order conditions.For ESDIRK-(8,4,3), the penultimate stage supplies an embedded solution of classical order three and semilinear order two.Source: §4.1 · pp. 9–10
• Eight stages provide extra freedom for ESDIRK-(8,4,3), improving principal error and stiff-limit internal stability relative to the seven-stage construction.The authors did not find L-stable ESDIRK-(s,4,3) schemes for s = 5 or 6, but found one for s = 7.Source: §4.1 · p. 9

4.2. Construction of EDIRK-(7,4,4) and EDIRK-(19,5,4)

• EDIRK-(7,4,4) is obtained by numerical optimization subject to order, semilinear-order, positivity, and sampled A-stability constraints, with A-stability verified afterward.Its high-dimensional search makes it suboptimal in principal error and other properties, and it has no embedded method.Source: §4.2 · pp. 10–11
• Richardson extrapolation of EDIRK-(7,4,4) yields the 19-stage EDIRK-(19,5,4), raising classical order to five while retaining semilinear order four and stage order one.The authors do not recommend it for practical computations because of its stage count.Source: §4.2 · p. 11

5. Numerical results: Convergence

• The test suite progresses from semilinear ODE and PDE problems to nonlinear PDE and ODE cases that increasingly depart from the theory’s assumptions.The sequence includes advection-reaction, 2D Allen–Cahn, viscous Burgers’, and fully nonlinear Van der Pol tests.Source: §5 · p. 11

5.1. A semilinear variant of the Prothero–Robinson problem

• On the semilinear Prothero–Robinson test, new methods avoid the classical L-shaped error curves near λh ≈1, though convergence rates need not be uniform across step sizes.For sufficiently small step sizes, their errors lie below a line of slope min(p, pSL +1), independent of λ.Source: §5.1 · pp. 12–13, Figure 1 · p. 12

5.2. Methodology for PDE test problems

• The PDE tests use spatial discretizations designed to make spatial errors negligible, isolating temporal error until the curves approach the spatial-error plateau.Stage systems are solved by Newton iteration to high accuracy so solver error does not affect temporal convergence curves.Source: §5 · p. 11, §5.2 · p. 13

5.3. Semilinear advection-reaction test problem

• In the semilinear advection-reaction test, methods with pSL ≥ p −1 reach full order p, while classical methods fall to order 2.5 in the stiff regime.The new methods generally need only pSL = p −1 to achieve full order.Source: Figure 2 · p. 14, §5.3 · p. 15

5.4. Allen–Cahn equation in two space dimensions

• For 2D Allen–Cahn, classical methods converge at 2.25, while new methods with pSL = p−1 achieve rates very close to p.The results extend the observed mitigation of order reduction to a higher-dimensional PDE.Source: §5.4 · p. 15, Figure 3 · p. 16

5.5. Viscous Burgers’ equation

• In viscous Burgers’ equation, the new methods converge at orders 4.0 and 4.5, versus 2.25 for classical methods.The fifth-order method’s observed order remains below its formal order.Source: §5.5 · p. 16, Figure 4 · p. 17

5.6. Van der Pol equation

• For Van der Pol, which lies outside the semilinear theory, new methods consistently improve convergence rates but do not eliminate order reduction.The results cover mildly stiff ε = 10−3 and strongly stiff ε = 10−6 regimes.Source: §5.6 · p. 17, Figure 5 · p. 18

6. Numerical results: Adaptive time-stepping

• The adaptive tests use a semilinear advection-reaction problem with changing spatiotemporal scales and compare method and embedding errors against an exact one-step reference.Adaptive control uses the discrepancy between the two approximations as an error indicator.Source: §6 · pp. 18–19
• With inflow boundaries, the (5,4,1) estimator poorly predicts local errors, so meeting its tolerance does not ensure actual errors meet it; (8,4,3) behaves as desired.The reference method also incurs more accepted and rejected steps in this case.Source: §6 · p. 19, Figure 6 · p. 20
• The (5,4,1) method develops an inflow boundary layer and error artifacts; periodic boundaries or the (8,4,3) method avoid boundary layers.Under order reduction, the max-norm error scaled by the mean step size to power p is an order of magnitude worse than without order reduction.Source: Figure 7 · p. 21, §6 · p. 22
• At low accuracy the (5,4,1) method is preferable, but at higher accuracy the (8,4,3) method is notably more effective.The comparison uses work-precision diagrams reporting global error and computational cost.Source: Figure 8 · p. 22, §6 · p. 23
• For fifth-order tests, order reduction increases step counts by an order of magnitude and produces scaled max errors four orders larger, while local errors miss the tolerance.The new (10,5,4) method is clearly favorable over (6,5,1) except at very loose tolerances that incur only about a dozen steps.Source: §6 · p. 23
• The fifth-order (10,5,4) method is clearly favorable without order reduction, except at loose tolerances requiring only about a dozen steps.Its slightly fewer steps than (6,5,1) under periodic boundary conditions also favor its work-precision results.Source: §6 · p. 23

7. Conclusion

• The authors recommend ESDIRK-(8,4,3) and ESDIRK-(10,5,4) for L-stability, optimized principal error, and quality embeddings.The paper presents the first DIRK methods of fourth and fifth order without order reduction for an important class of stiff semilinear ODEs.Source: §7 · p. 23
• For (6,5,1) with inflow boundaries, a significant boundary layer yields a scaled reference error four orders of magnitude worse than the other three tested situations without order reduction.Source: Figure 10 · p. 25
• The work-precision comparison shows significant efficiency separation between (6,5,1) and (10,5,4); all but the three rightmost (6,5,1) points are in the stiff regime.Source: Figure 11 · p. 25
• The less-optimized EDIRK-(7,4,4) and EDIRK-(19,5,4) are mainly of theoretical interest, showing that stage order one can achieve semilinear order four.Source: §7 · p. 26
• New schemes attain predicted orders on semilinear problems and partly mitigate order reduction beyond the theory; low-semilinear-order methods can underestimate local error and exceed tolerances.Such methods can also overestimate the optimal step size.Source: §7 · p. 26
• For families of spatial discretizations, the error bounds may depend on mesh size N through M, L, and D, although the numerical PDE tests showed no overall-error dependence on mesh size.The possible dependence arises primarily because the bounds use a discrete L2 norm.Source: §7 · p. 26
• The implicit Runge–Kutta framework generally requires nonlinear solves, motivating study of methods such as Rosenbrock or implicit-explicit Runge–Kutta schemes.Source: §7 · p. 26

Appendix A. Table of method properties

• Table A.2 lists properties of the paper's methods and comparable methods from the literature.Its naming convention is TYPE-(s, p, pSL), where s is stages, p classical order, and pSL semilinear order.Source: Table A.2 · p. 27
• The principal error norm is the 2-norm of the classical order p+1 residuals.Source: Appendix A · p. 27
• Embedding quality is determined by B(p+1), C(p+1), and E(p+1), while D denotes the maximum coefficient magnitude.Source: Appendix A · p. 27

Appendix B. New method coefficients

• The appendix provides coefficients accurate to 34 digits for three new methods.Source: Appendix B · p. 28
• All three methods are diagonally implicit and stiffly accurate, with a_ij = 0 for 1 ≤ i < j ≤ s and b_i = a_s,i.Source: Appendix B · p. 28

Appendix B.1. Coefficients for ESDIRK-(8,4,3)

• ESDIRK-(8,4,3) has eight stages, fourth classical order, and semilinear order three.Source: Appendix B.1 · p. 28
• Its embedded weights are defined by the seventh row of coefficients, b̂_i = a_7,i for i = 1, …, 8.Source: Appendix B.1 · p. 28

• EDIRK-(7,4,4) is fourth-order accurate.Source: Appendix B.2 · p. 28
• EDIRK-(7,4,4) has semilinear order four.Source: Appendix B.2 · p. 28
• The method is primarily of academic interest and does not include an embedding.Source: Appendix B.2 · p. 28

• ESDIRK-(10,5,4) is fifth-order accurate.Source: Appendix B.3 · p. 29
• ESDIRK-(10,5,4) has semilinear order four.Source: Appendix B.3 · p. 29