Differential-algebraic equations and index

A semi-explicit DAE has differential variables x and algebraic variables z, with $x'=f(t,x,z)$ and $0=g(t,x,z)$. If the Jacobian of g with respect to z is nonsingular near a consistent point, the implicit function theorem locally determines z from t and x. Substitution then defines a local ODE for x. Differentiating the constraint once determines z prime. Under the usual differentiation-index convention, this regular semi-explicit form is index one.

Consistent initial data must satisfy the algebraic constraint, and the initial derivative must satisfy its differentiated relation. Assigning arbitrary values to both x and z usually violates those requirements. A residual-form solver instead accepts an equation involving time, state, and derivative, but its inputs must still be mutually consistent. A regular implicit ODE whose derivative Jacobian is invertible is locally solvable directly for the derivative and is index zero under this convention; implicit notation by itself does not make a DAE high index.

A Cartesian pendulum illustrates why checking only the position constraint is insufficient. Position must remain on a circle, so differentiating its squared-length constraint requires velocity to be tangent to that circle. Differentiating again relates acceleration and speed, allowing the constraint multiplier to be determined under regularity. The velocity condition is a hidden constraint: a position can lie on the circle while an arbitrarily assigned outward velocity makes the initial state inconsistent.

In the position-constrained first-order pendulum formulation, two differentiations reveal the multiplier and a third determines its derivative. It is conventionally an index-three system when the multiplier is included among the unknowns. Index is tied to the formulation and to regularity assumptions, not simply to the number of equations written with an equals sign. Eliminating variables, using independent coordinates, or introducing derivative constraints can change the index while also introducing numerical issues such as drift from the original constraint.

For a regular index-one problem, an implicit step can solve differential and algebraic residuals together, enforcing the algebraic constraint at accepted endpoints up to the solve tolerance. That endpoint condition does not automatically give an equally accurate dense-output constraint between steps. Higher-index systems may need index reduction, specialized formulas, or projection onto the relevant constraint manifolds. Projection can reduce drift, but its effect on order, momentum, and geometry must be analyzed rather than assumed beneficial in every respect.

Algebraic variables can be sensitive to differential errors when their constraint Jacobian is poorly conditioned. Scaling and Jacobian structure therefore matter for both initialization and time stepping. Modern DAE solvers provide consistent-initialization routines for specified classes, not arbitrary cures for every singular model. A structural singularity may indicate redundant constraints, a changing constraint rank, or an ill-posed physical formulation. Tightening a tolerance alone cannot resolve such a modeling problem, and a successful residual solve does not by itself establish a valid trajectory.