Tiny differentiable ODE/SDE/DAE/SDAE solvers for JAX: fixed-step Euler/RK4,
adaptive Tsit5, linearly implicit Rodas5P for stiff ODEs and index-1 DAEs,
and Euler–Maruyama for Itô SDEs and semi-explicit index-1 SDAEs. The package
also includes primal, vmap-friendly finite-state DTMC and CTMC simulators with
sequential and associative parallel-prefix execution. Deterministic probability
forecasts are differentiable in the initial mass and include DTMC matrix powers,
dense CTMC exponentials, and matrix-free Arnoldi/Krylov actions over probability
pytrees. The same dense and matrix-free backends are available directly through
solve_linear_ode for any fixed homogeneous linear array or pytree operator;
jvp_linear_ode and vjp_linear_ode apply the exact initial-state tangent and
adjoint exponential actions without differentiating Arnoldi orthogonalization.
Bounded lax.scan loops provide exactly max_steps attempt slots for fixed and
adaptive stepping, so shapes are static, nothing recompiles as tolerances or
curvature change, and every solve is differentiable in both forward and
reverse mode — including reverse-over-forward, the pattern a
Levenberg–Marquardt optimizer with geodesic acceleration needs when it
differentiates through a rollout. Adaptive attempts are grouped into static
chunks, allowing one lax.cond to
skip solver and controller work for an entire padded chunk.
This is a deliberately small, jvp/vjp-friendly package. Rodas5P is a JAX
adaptation of Steinebach's method and follows SciML's
OrdinaryDiffEqRosenbrock
implementation. Use diffrax or
SciML if you need general mass
matrices, fully implicit or higher-index DAEs, events, continuous solution
objects, sparse/Krylov linear solvers for ODE/DAE stages, or specialized
adjoints. Initial DAE
consistency and explicit DAE stages use nlls-gram.
The linear exponential-action API follows SciML
ExponentialUtilities.expv.
It includes fixed and residual-controlled adaptive matrix-free time slicing;
the latter keeps the Krylov dimension static for predictable JAX compilation.
SciML's
ExponentialIntegrators.jl
is the reference for the broader nonlinear exponential-integrator family.
uv add tinydiffeqFor GPU use, install the JAX accelerator build that matches your hardware, for example:
uv add tinydiffeq "jax[cuda13]"The vector field may take (x), (x, t), (x, t, args), or
(x, t, args, p) — always in that order. args is pass-through data (not an
AD target by convention); p holds differentiable parameters (any pytree).
The state may also be any JAX pytree. It must contain at least one leaf, and
every leaf must be a nonempty real floating array with the same dtype; vector
fields and project preserve that structure. Output keeps the structure and
adds the saved-time axis to each
leaf.
import jax
import jax.numpy as jnp
from tinydiffeq import solve_ode, Tsit5, IController, SaveAt
jax.config.update("jax_enable_x64", True) # your call — the library never sets it
def f(x, t, args, p):
return -p * x
sol = solve_ode(
f, Tsit5(), 0.0, 2.0, jnp.asarray(1.0),
p=jnp.asarray(1.3),
dt_0=0.1,
controller=IController(rtol=1e-8, atol=1e-10),
max_steps=512,
save_at=SaveAt(ts=jnp.linspace(0.0, 2.0, 21)), # fixed output shape,
) # however many steps adapt
print(sol.xs) # states on the grid
print(sol.ok) # reached t_1 with every requested output valid?IController() and PIController() choose tolerances from x_0.dtype:
rtol=1e-4, atol=1e-6 for float32 and rtol=1e-7, atol=1e-9 for
float64. Pass explicit values when tolerances are part of your model's
scientific specification. The default dt_min is
10 * finfo(dtype).eps * max(1, abs(t_1)).
max_steps is the total internal attempt budget: accepted steps plus
rejections. It is not normally the number of returned times. Endpoint mode
returns one time/state, SaveAt(ts=...) returns the requested grid, and
SaveAt(steps=True) returns the initial state and accepted internal steps as
a contiguous prefix of max_steps + 1 rows. The remaining rows repeat the
last accepted state by default; sol.accepted distinguishes data from
padding. Rejected attempts never appear in the returned trajectory.
SaveAt(ts=...) also accepts a Python sequence. These are observation times:
the adaptive controller still chooses its own internal mesh. Explicit methods
use cubic Hermite interpolation; Rodas5P uses its published stiff-aware
fourth-order continuous extension.
For a square index-1 system dy/dt = f(y, z, t, args, p) and
0 = g(y, z, t, args, p):
from tinydiffeq import IController, Rodas5P, Tsit5, solve_semi_explicit_dae
def dae_f(y, z, t, args, p):
dy = p * z
return dy, {"flow": dy}
def dae_g(y, z, t, args, p):
return z - y
dae_sol = solve_semi_explicit_dae(
dae_f, dae_g, Tsit5(), 0.0, 1.0,
jnp.asarray(1.0), jnp.asarray(0.5),
p=jnp.asarray(2.0), dt_0=0.1,
controller=IController(), max_steps=128,
)
print(dae_sol.ys, dae_sol.zs, dae_sol.aux["flow"])
# One initial nonlinear consistency solve, then linear Rodas5P stages.
stiff_dae_sol = solve_semi_explicit_dae(
dae_f, dae_g, Rodas5P(), 0.0, 1.0,
jnp.asarray(1.0), jnp.asarray(0.5),
p=jnp.asarray(2.0), dt_0=0.1,
controller=IController(), max_steps=128,
)z_0 is a guess and is made consistent automatically. RK4 and Tsit5 restore
the algebraic root at every stage. Rodas5P performs no nonlinear solves after
initialization: it advances the corresponding block mass-matrix system using
one reused LU factorization per attempt. Differential fields may return a
floating saved-aux pytree stored at accepted nodes and interpolated on requested
deterministic grids. Algebraic equations may separately return internal context
passed to the dynamics. JVP, VJP, and reverse-over-forward propagate through both
implicit initialization and the time integrator. See the
DAE documentation
for root controls, SaveAt, and scope limits.
Fixed-step semi-explicit Itô SDAEs use the corresponding
solve_semi_explicit_sdae interface with EulerMaruyama, a PRNG key, and
n_steps; see the SDAE documentation.
def endpoint(p):
return solve_ode(
f, Tsit5(), 0.0, 2.0, jnp.asarray(1.0), p=p,
dt_0=0.1, controller=IController(rtol=1e-10, atol=1e-12),
max_steps=512,
).xs
jax.grad(endpoint)(jnp.asarray(1.3)) # reverse mode
jax.jvp(endpoint, (jnp.asarray(1.3),), (jnp.asarray(1.0),)) # forward mode
jax.grad(lambda p: jax.jvp(endpoint, (p,), (jnp.asarray(1.0),))[1])(
jnp.asarray(1.3)
) # reverse-over-forwardThe step-size controller is wrapped in stop_gradient (accept/reject is
non-differentiable either way, and the error-ratio power blows up at exactly
zero error); the states differentiate fully through the solver stages. See the
docs for the design
contracts: static shapes and SaveAt, AD through adaptive stepping, SDE key
semantics, and the package API.
MIT