Moved from SciML/StochasticDelayDiffEq.jl#24.
When setting the delay to zero in an SDDE, non-EM solvers (EulerHeun, RKMil, LambaEM, etc.) produce incorrect solutions, while EM is unaffected. The reporter observed this when checking coherence convergence in a network of oscillators — the analytic result was only matched by EM.
This suggests the delay implementation (history function evaluation, fixed-point iteration, or interpolation) introduces artifacts for higher-order methods when the delay is very small or zero.
This is a correctness issue that should be investigated with the new DelayDiffEq-based SDDE implementation. See original issue for the full oscillator model reproducer and discussion (8 comments with detailed analysis).
Moved from SciML/StochasticDelayDiffEq.jl#24.
When setting the delay to zero in an SDDE, non-EM solvers (EulerHeun, RKMil, LambaEM, etc.) produce incorrect solutions, while EM is unaffected. The reporter observed this when checking coherence convergence in a network of oscillators — the analytic result was only matched by EM.
This suggests the delay implementation (history function evaluation, fixed-point iteration, or interpolation) introduces artifacts for higher-order methods when the delay is very small or zero.
This is a correctness issue that should be investigated with the new DelayDiffEq-based SDDE implementation. See original issue for the full oscillator model reproducer and discussion (8 comments with detailed analysis).