Motivation
A higher-order time integrator is valuable because one accurate step may replace many small first-order steps. The extra slopes must, however, be combined for a reason. Averaging four arbitrary slopes does not produce a fourth-order method. The stages sample a changing vector field, and their interactions must reproduce the different derivatives appearing in the exact solution. This is why a scalar test equation can verify a stability polynomial while still missing conditions needed for a nonlinear equation.
We build a manageable part of the general theory: explicit Runge–Kutta conditions through order four. The derivation makes classical RK4 understandable without requiring a full theory of rooted trees. We then examine an embedded third/second-order pair. Two formulas sharing stages give an inexpensive disagreement that can guide the next step size. That disagreement is evidence about local error under smoothness and asymptotic assumptions; it is not an automatic bound on the final global error.
Earlier notes supply Lipschitz error propagation and scalar stability; here the new work is nonlinear order matching and the meaning of an error controller.
Definitions
Definition 2.2.1
(Explicit Runge–Kutta method)
For an autonomous system in , an explicit -stage Runge–Kutta method has a strictly lower triangular matrix , weights , stage abscissae , and equations
For a nonautonomous equation, use . The relation permits the same autonomous calculation after adjoining . Powers such as below mean entrywise powers; means matrix multiplication. Write .
Order means that the unscaled defect of an exact-start step is , uniformly on the region and time interval under consideration. A global order conclusion also needs stability of the step map and sufficiently accurate initial data. A tableau records and to the left/above a horizontal rule and below it; the tableau is a compact specification of all stage inputs, rather than an independent definition of accuracy.
Definition 2.2.2
(Embedded pair and scaled indicator)
An embedded pair uses the same stages with two weight vectors and , producing and . Their difference is the error indicator. With component scales , the scaled indicator is .
This maximum is a common dimensionless choice. Absolute tolerances prevent a component near zero from demanding a meaningless relative accuracy.
Theorem/proposition
Theorem 2.2.3
(Order conditions through degree four)
Suppose has continuous bounded derivatives through order four on a neighborhood containing all exact and stage curves for sufficiently small steps. With , the following eight conditions are sufficient for a defect :
The first two give order two; the first four give order three. If also has a uniform Lipschitz bound on a convex stage region containing the compared trajectories, a fixed explicit tableau satisfying all eight conditions has global error on a fixed finite interval, assuming exact initial data and that all steps remain in that region.
Proposition 2.2.4
(What an embedded third/second pair estimates)
Suppose an exact-start order-three formula and an order-two formula satisfy uniform expansions
Then . When its leading scaled coefficient is nonzero and the scales vary smoothly, replacing by changes the leading indicator by . Thus is an asymptotic local controller, with safety factor . This statement does not bound global error or apply when the leading coefficient vanishes.
Concept lensCombinatorial
Order conditions indexed by rooted trees
Each derivative expression corresponds to a rooted tree: every vertex is a factor , and a vertex with children carries an -th derivative. There are trees with one to four vertices, hence eight conditions. Each condition reads , where multiplies, over all vertices, the size of the subtree rooted there. The path of four vertices has , giving ; the root with three leaves has , giving .
Proof sketch or proof idea
Proof
(Matching Taylor coefficients of exact and numerical steps)
All derivatives in the next calculation are evaluated at the starting state . Write for the derivative applied to and for the bilinear second derivative. Repeatedly differentiating the ODE gives
The coefficient three has a concrete origin: differentiating the two arguments of gives two copies, and differentiating the outer derivative in gives one more. A nonlinear vector field distinguishes these expressions even though several collapse to the same monomial for a scalar linear equation.
To expand a stage, first observe that its argument displacement begins with . Substituting the lower-stage expansions successively yields
Multiplying by and summing now exposes each condition. For example the coefficients of are numerically and exactly. The coefficients of are and . Equating all displayed coefficients gives the eight equations. Taylor's theorem and the bounded derivatives make the omitted terms uniform, completing the local proof.
Proof
(From a fifth-degree defect to global fourth order)
For propagation, let two inputs differ by . The first stage difference is at most . Inductively each later stage difference is bounded by , where the earlier constants are bounded for . Hence the complete step map has Lipschitz factor at most , with independent of the step. Its error recurrence is . Summing over at most steps gives . This argument separates the algebraic order conditions from the analytic region assumptions needed to use them.
Proof X-ray
Which assumptions do the work?
The stage expansion can be computed at all because is strictly lower triangular: stage depends only on stages , so successive substitution is an induction that terminates after stages. The row-sum relation is used at the first move, where the displacement of stage begins with . The non-obvious point is that the four fourth-order expressions , , and are different for a general nonlinear field, so each needs its own equation; on three of them vanish and the conditions collapse. Smoothness enters only to make the omitted Taylor terms uniformly on the stage region. The Lipschitz bound then turns the one-step defect into a recurrence , and one power of is lost because about defects are summed while stays bounded. None of these jobs is supplied by recognizing the name “RK4.”
Proof
(Subtracting the two expansions)
The embedded proposition follows by subtraction: equals the lower formula's error minus the higher formula's error. Its leading term has degree three. Taking a maximum of finitely many scaled component magnitudes preserves this degree when the leading scaled vector is nonzero. Solving gives . A safety factor allows for neglected terms; it does not turn the expansion into a rigorous enclosure.
Worked examples
Worked example 2.2.5
(Checking classical RK4 rather than memorizing it)
Classical RK4 uses
Thus , and . Direct weighted sums give , , , , , , and . These verify nonlinear fourth order under the theorem's hypotheses.
For , , , the slopes are , , , . The update is . The exact value is , so the one-step error is about . The two middle stages use the same time but different states; merging them changes the method.
Worked example 2.2.6
(A shared-stage third/second-order pair)
Use stages at with nonzero rows , , and . Take
The higher weights satisfy the four conditions through order three: in particular and . The lower weights have sum one and , but , so they are not generally order three.
For , , , the stages are , , , . The higher update is , the lower is , and . The true higher error is about . The disagreement estimates the leading lower-order error, rather than being exactly the higher-order error. Local extrapolation means accepting the higher update while using this indicator to select a step.
Worked example 2.2.7
(Rejection and a meaningful next step)
Suppose the scaled embedded indicator is for a trial step , and use . The third-power law suggests . A rejected trial does not advance time: recompute all needed stages from the last accepted state. If , the same rule proposes ; a practical controller also caps growth so a surprisingly small indicator cannot cause an excessive jump.
When , the displayed power formula is undefined. One can propose a bounded increase and recompute the next indicator. Zero disagreement can occur for a constant vector field, where both formulas are exact, or by accidental cancellation, where it says much less. It does not justify accepting an arbitrarily long future step across a rapidly changing or nonsmooth part of the equation.
Step doubling and the cost of a trustworthy comparison
Another error indicator uses the same order-four method twice: once with step , then twice with step . Suppose the exact-start coarse result is . Smooth dependence on the state implies that the two half-step defects have the same leading coefficient up to changes; their combined contribution is . Thus the fine result satisfies , and . The fine result's signed error is therefore . The denominator is fifteen, rather than sixteen, because the comparison subtracts two imperfect results. This argument concerns one exact-start macrostep, not two complete simulations with unrelated accumulated errors.
Step doubling costs more evaluations than a purpose-designed embedded pair. Its advantage is that it can be added to an existing one-step method without deriving a new shared-stage tableau. Its assumptions still matter: a discontinuity inside one half interval can destroy the common leading-coefficient expansion. The comparison should be interpreted together with smoothness, stability and the problem's known events.
The shared-stage pair above has a further practical feature: its fourth stage evaluates the vector field at the accepted higher-order endpoint. After acceptance, that slope can serve as the first slope of the next step, even when the next step length changes, because the first stage uses only the accepted time and state. After rejection, the endpoint is discarded and that reuse is invalid. Reusing the slope is a computational saving, not an extra accuracy condition. Dense output between accepted endpoints likewise needs its own interpolation formula and error reasoning; joining endpoints by arbitrary straight segments does not inherit the third-order accuracy of the propagated formula.
The norm used by a controller also affects its meaning. A component maximum asks every scaled coordinate to meet the same threshold, whereas a root-mean-square indicator allows one component to exceed it if others are small. Neither is invariant under arbitrary changes of coordinates. Choose scales from the physical or mathematical quantities whose accuracy matters, and record that choice when comparing computations. A small indicator for a negligible component should not obscure a poorly resolved quantity of interest.
The word "uniform" in an order statement also deserves attention. If the local defect bound has the form and the computation approaches a place where derivatives become unbounded, there may be no fixed . For example, when a solution approaches finite-time blow-up, every early short interval is smooth, yet a fixed-interval theorem cannot be carried across the blow-up. A step-size controller may then shrink the step drastically; this signals something about the problem itself or about the interval on which the model is valid, and should not be read merely as inefficiency of the method.
When reading the stage expansion, it helps to label each term by whether it comes from the first-, second- or third-degree Taylor term of the displacement, and then to trace how earlier stages enter that displacement. This shows that the matrix products record chains of dependence, rather than leaving the eight conditions as unrelated numbers. For high-dimensional systems, the order in which derivatives act must not be interchanged carelessly.
Common mistakes
Matching to checks only the chain expressions on . Every higher derivative of this vector field vanishes. Conditions involving and therefore cannot be inferred from that test.
Counterexample mode
A correct stability polynomial with the wrong order
Claim: if agrees with through , the method has order three. Take three stages with , , , and ; the rows sum to . Then , , so , and ; hence . But . On , , the stages are , so one step gives , whereas . The defect is only : the method has order two for this smooth equation, which becomes a nonlinear autonomous system after adjoining . The repair is to impose all four conditions of degree at most three, including .
Common mistake
(Integrating a smooth formula across an unlocated event)
A formula designed for smooth vector fields may lose its nominal order at an unlocated discontinuity or event. Splitting at a known event restores a smooth-interval setting; applying a smooth Taylor theorem across a jump does not.
Common mistake
(Reading a local tolerance as a final-time tolerance)
An accepted local tolerance is not the same as a final-time tolerance. Errors can accumulate or be amplified by the differential equation. An adaptive code must also respect stability restrictions, finite arithmetic and nonlinear stage solves where relevant. The current explicit examples require no nonlinear solves, but they do not become suitable stiff solvers merely because their error indicators are small.
Summary
Runge–Kutta order comes from matching distinct derivative expressions. Classical RK4 satisfies eight conditions through order four; smoothness and controlled propagation then convert a fifth-degree local defect into fourth-order global accuracy. An embedded pair saves stage work, and the lower order determines the leading power used by its disagreement controller. The controller remains a local asymptotic tool.
Exercises
Quick check
(Show that any explicit two-stage tableau with has order two when and . Can it satisfy ?)
Compute before attempting a higher-order claim.
Solution · Answer
The weights sum to one and , giving order two. But and the only nonzero matrix entry multiplies , so . Thus : no explicit two-stage tableau reaches order three for general smooth systems. Adjusting the weights cannot fix a missing stage interaction.
Quick check
(Apply classical RK4 with to , . Compare the numerical value with the first five terms of the exponential series.)
Keep the fractions exact.
Solution · Answer
The stability polynomial is . At it is . These are precisely the exponential terms through degree four. Since all omitted terms are positive, this one-step value is below . This checks the linear calculation, while the classical RK4 worked example supplies the additional nonlinear conditions.
Quick check
(For an embedded third/second pair, let , and . Find the proposed replacement step and explain whether time advances on rejection.)
The indicator has leading degree three.
Solution · Answer
The cube root of is , so the new step is . Rejection retains the old accepted time and state; the trial is repeated with the shorter step. Advancing with a rejected value would invalidate the intended accuracy policy and change the initial state of the retry.
Quick check
(Prove that the shared-stage third/second-order pair gives on , where . What happens at ?)
Compute both amplification polynomials; a zero indicator need not mean zero error.
Solution · Answer
The higher polynomial is . The lower polynomial is . Subtraction gives the claimed expression. At both factors equal , so , although the exact factor is . This is a concrete cancellation outside a reliable small-step interpretation: a zero estimate does not prove exactness.