A mass on a spring, a pendulum at small angles, a gust-loaded wing flexing and the pitch oscillation of an aircraft are all described by a second derivative equal to a combination of lower derivatives. In 3.4 Linear First-Order Equations & Integrating Factors the first-order linear equation was solved in one stroke by an integrating factor. A second derivative needs two initial conditions, and the single integrating factor is not enough. What replaces it is a piece of structure: the solutions form a small vector space, and everything in this module (5.2 Constant Coefficients & the Characteristic Equation, vibrations, forcing, series) is a way of finding a basis of it. This lesson builds that structure without solving any particular family of equations; the pay-off is that once you have two good solutions you have all of them.
Transcript
- L takes y to y double prime plus p y prime plus q y. It is linear, since derivatives pass through sums and constants, so combinations of solutions are solutions. Test on y double prime equals y: e to the t, e to the minus t and two e to the t minus e to the minus t all leave residual zero.
- Existence and uniqueness: each starting value and slope has exactly one solution. So the map from solutions to pairs is linear, onto, and sends only the zero function to zero. The solutions form a plane: two dimensional. A basis is a fundamental set.
- Stack the value-and-slope vectors of two solutions. At time zero, e to the t gives one, one and e to the minus t gives one, minus one. Their determinant, the Wronskian, is one times minus one, minus one times one: minus two. A signed area, two, clockwise.
- Let time run. For cosine and sine the area stays exactly one. For cosine and twice cosine the arrows always share a line: area zero, so combinations reach only that line, and the target one, three is out of reach.
- Match value one and slope three: c one plus c two equals one, c one minus c two equals three. Cramer's rule with W equal minus two gives c one two, c two minus one: two of v one, minus one of v two, reaches the target. So y is two e to the t minus e to the minus t, 5.068684 at t equals one.
- Differentiate W: the first-derivative products cancel. Substitute the equation for each second derivative and the q terms cancel too, leaving W prime equals minus p W. So W equals W at t zero times the exponential of minus the integral of p: zero everywhere or nowhere.
- Check on y double prime plus two y prime plus five y, solved by e to the minus t cosine two t and sine two t. At zero W is two and p is two, so Abel says two e to the minus two t: 0.270671 at t equals one, 0.036631 at two, matching the product-rule values. The area shrinks, and q never entered.
- Reduction of order: let y two be v times y one. The v terms multiply L of y one, which is zero, leaving v prime and v double prime. For the repeated-root example, with y one e to the t, the v prime coefficient vanishes, so v is c one plus c two t and y two is t e to the t. Its Wronskian, e to the two t, agrees with Abel.
- To compute, convert: let x be y and y prime, so x prime is A x with trace minus p. RK4 on y double prime plus two y prime plus five y shows errors at t equals three shrinking by 16.66, then 16.50, when h halves: fourth order.
- Place two solutions as columns of X; its determinant is the Wronskian, and from memory its derivative is the trace times itself, Abel's law. RK4 returns 0.13533593 at t equals one against e to the minus two, 0.13533528. The undamped case keeps 0.99999861, the t squared case gives 9.00000113 against nine.
- Where it stops: for y double prime equals y squared, six over t squared solves it, but twice it does not: seventy-two against one hundred forty-four at t equals one. The flow agrees: from value 0.8, slope minus 0.6, the superposition gap is zero for the linear equation, and about minus 0.217, minus 0.585 here.
- Last warning: t squared and t times absolute t have Wronskian zero, yet their ratio is minus one at t equals minus one and plus one at t equals one. No linear equation with continuous p and q has both as solutions, so the zero means nothing.
- So: the solutions form a two-dimensional space, the Wronskian is a signed area that tests independence, and Abel's law makes it zero everywhere or nowhere. Compute by converting to a system; nonlinear equations break all this. Next: constant coefficients.
The operator L and superposition
Fix an interval of time and two functions , . Define a machine that takes a (twice differentiable) function and returns another function:
The equation is second-order linear homogeneous. “Linear” refers to how enters: , and appear only to the first power, multiplied by functions of alone. The machine is linear in the sense of 1.2 Matrices as Transformations: for functions and numbers ,
The middle step uses only that differentiation passes through sums and constants () and that multiplying by and does too. So if and then . This is the superposition principle: any combination of solutions is a solution.
Here is the matrix picture. For a matrix , the kernel of 1.3 Linear Systems, Kernel & Column Space is the set of inputs that sends to , and it is closed under combinations for exactly the reason above. The operator is the same idea with functions as the inputs: its kernel is the set of all functions that sends to the zero function. That kernel is the solution set. A matrix with 4 columns and 2 independent rows has a 2-dimensional kernel because 2 of the 4 unknowns are free. The question for is how many free choices a solution has.
Why there are exactly two: existence and uniqueness
The answer comes from 3.7 Existence & Uniqueness in its linear form (stated here, not proved). If and are continuous on an interval and , then for any two numbers there is exactly one solution of with and , and it exists on all of . The last clause differs from the nonlinear case, where solutions can blow up early (see “Where this stops” below).
Now count. Define the map that takes a solution to the pair .
- is linear (value and derivative of a sum are sums).
- is onto: existence says every pair is hit.
- has kernel zero: if , then and the zero function both solve the equation with the same data, so uniqueness says .
A linear map that is onto and one-to-one identifies the solution space with . So the kernel of is 2-dimensional: a solution is determined by two free numbers, its starting value and starting slope, which is the “two free unknowns” of the matrix picture. A fundamental set is a pair of solutions that forms a basis, so that every solution is for exactly one pair . Finding a fundamental set is the whole game; the tool that recognizes one is next.
The Wronskian: a determinant, an area, a test
Suppose are solutions and you want the combination with and . Write down both conditions:
This is a linear system for with matrix whose columns are and at . By 1.4 Determinants as Volume Scaling it has exactly one solution for every if and only if the determinant is nonzero. That determinant is the Wronskian,
Geometrically, and are two arrows in the plane (the phase plane of 4.7 Energy & Symplectic Integrators), and is the signed area of the parallelogram they span. If the area is zero the arrows lie on one line, and combinations of them can only reach that line: most targets cannot be matched. If the area is nonzero the arrows span the plane and every target is reached, with the coefficients given by Cramer’s rule,
Worked example. has solutions and (differentiate twice: each returns itself). At the vectors are and and . For , the system is , , so adding gives and , . Check with Cramer: . So . The playground below computes from it, and the sim’s option with , reproduces , (the minus sign in only says the pair is oriented clockwise; the area is ).
The same determinant tests independence. If for all , differentiating gives , so the matrix times is zero. If at even one point, the matrix is invertible there and : the functions are independent. Conversely, for solutions of the same equation, means some nonzero makes . Then is a solution with zero value and zero slope at , and uniqueness forces it to be the zero function: the two are dependent. The uniqueness step is where “solutions of the same linear equation” matters, and the last section shows the test failing without it.
Abel’s theorem
The sim’s middle panel draws for each preset together with a dotted curve . They coincide, and here is why. Differentiate with the product rule:
The terms with two first derivatives cancel. Now use the equation to replace the second derivatives, :
The terms cancel, so the result does not depend on at all. The equation is separable (3.3 Separable Equations) with solution
Abel’s theorem. An exponential is never zero, so either and then on the whole interval, or and is never zero there. One look at one point settles independence everywhere. The first-derivative coefficient is the whole story: it is the rate at which the parallelogram of the previous section grows or shrinks, an area-expansion rate of . With the area is constant, the property of 4.7 Energy & Symplectic Integrators in Hamiltonian systems. With positive (friction, as in 5.3 Mechanical Vibrations: Spring–Mass–Damper) the area decays.
Three checks, all in the lab and the sim:
- with : , so is constant, and the playground prints at . For it prints at each time.
- has and (you can check them by substitution; 5.2 Constant Coefficients & the Characteristic Equation derives them). At the vectors are and and . Abel with predicts , and the playground prints at and at from the product-rule formula, the same as . In the sim’s option the readout shows, at , both the direct value and the Abel value as .
- on has , (substitute: and ). Here , so and . Directly , which is at in the playground and at in the sim’s option. Note the pair is independent on but . Abel does not forbid this: the equation divided by has , which is not continuous at , so is not in the interval.
Reduction of order
Suppose you know one solution with no zeros and want a second. Look for it as with a function to find. Then and . Substitute into :
because the bracket multiplying is . The unknown itself has vanished from the equation; only and remain. Let and the condition becomes the first-order equation , which is separable:
That is the mechanism: a second-order problem whose lower-order piece is already solved collapses to two first-order integrations. (Compare Abel: , so , consistent.)
Worked example. has the obvious solution (every derivative is , and ). Here and , so the reduced equation is , hence and . Taking , gives . Check: , , and . The Wronskian is , and Abel’s agrees. This is the repeated-root case that 5.2 Constant Coefficients & the Characteristic Equation will need: the factor is not a lucky guess; it is what the algebra produces.
From one second-order equation to a first-order system
You have been given a numerical-methods toolbox that eats first-order systems, so convert. Let the state be . Then and :
This is the form 4.2 The Euler Family: Forward, Backward, Heun & Midpoint, 4.3 Runge–Kutta Methods & Butcher Tableaux and 4.4 Adaptive Step Size & RK45 are written for. To feed it to RK4, f(t, x) returns [x[1], -p(t)*x[1] - q(t)*x[0]] and the rest is the stepper you already have. There is no second-order RK4, because the state of a second-order equation is a pair.
Two solutions give two columns and , and putting them side by side gives a matrix with . Its determinant is the Wronskian. For a matrix ODE, (a standard identity, Liouville’s formula, from memory), and : Abel’s theorem is the trace formula. For the trace vanishes, the area is preserved, and the flow is the Hamiltonian flow of 4.7 Energy & Symplectic Integrators (with constant, ).
Block D of the playground integrates the system for with , , whose exact solution is (match with the damped pair: gives , then gives ). At the exact value is and RK4 with has errors , , . The ratios are and , close to the of a fourth-order method (4.1 Local Error, Global Error & Order). Integrating from and taking the determinant, RK4 returns at against Abel’s ; for up to with it returns against , a small RK4 area loss of the kind seen in 4.7 Energy & Symplectic Integrators; and for from to it returns against .
Where this stops: nonlinear, inhomogeneous, not solutions
The structure above used three facts: is linear, the right-hand side is , and solve the same equation. Dropping each one breaks something.
Nonlinear. Take . The function solves it for : and . But gives while : at the playground prints against . Doubling a solution does not give a solution. The flow shows the same thing: block E starts RK4 at and compares with . For the gap at is in both components; for it is . And the solution blows up at even though nothing in the equation looks singular there, which is why the nonlinear theory only promises solutions near the start.
Inhomogeneous. For the function is a solution, but gives . The sum of two solutions of solves , not . What does survive: the difference of two solutions solves the homogeneous equation, so everything is “one particular solution plus the kernel of ”, the structure used in 5.4 Nonhomogeneous Equations: Undetermined Coefficients.
Functions that are not solutions of one linear equation. The functions and are differentiable with derivatives and , and the playground prints at (it is zero for all ). Yet they are not multiples of each other: the ratio is at and at . No linear equation with continuous and has both as solutions, so the uniqueness step of the independence argument is unavailable, and means nothing. The Wronskian test is a statement about solutions, not about arbitrary pairs of functions.
Lab
Block A computes Wronskians, B matches initial conditions, C checks the counterexamples, D integrates the system with RK4 and compares with Abel, and E tests superposition on the flow. The code runs in the browser in a few seconds.
Show that and solve , compute their Wronskian, and find the solution with , .
Show worked solution
and , so both satisfy . . Since they are a fundamental set. At : , , , , so from the first condition and from the second, . The solution is . Check: and . Abel agrees with constant because .
solves on . Use reduction of order to find a second solution, and confirm your Wronskian with Abel’s theorem.
Show worked solution
Put . Then and . Substitute:
The terms in cancel as they must. Divide by and set : , so , and . Taking and gives and . Check: , , and . Wronskian: . Abel: , so , and gives . They agree.
On the equation has a fundamental set with , where is continuous. (a) Find and . (b) Write the equation as and give . (c) A student claims two solutions of this equation have . Can that be true if they are independent?
Show worked solution
(a) gives , so and . The unknown never appeared.
(b) and , so , the same equation as Abel’s.
(c) No. is positive on all of , including where it is . If two solutions had , Abel would force on , so they would be dependent. The singular point is outside the interval, where is continuous.
Exercise: the system form, RK4 and a numerical Wronskian
Build the pipeline of this lesson. Implement an RK4 step for a vector state, turn into the system function for , integrate, and finally compute the Wronskian at time from two numerically integrated solutions started at and at . Abel says the answer is whatever is.
Key Takeaways
- is a linear operator, so : combinations of solutions of are solutions (superposition).
- The solution set is the kernel of . Existence and uniqueness make a one-to-one onto linear map, so the kernel is 2-dimensional and a fundamental set is a basis of it.
- The Wronskian is the determinant of the matrix with columns : the area of the parallelogram they span in the plane. is exactly what lets you match any by solving a system.
- Abel’s theorem, , gives : the Wronskian of solutions is zero everywhere or nowhere, and it does not depend on .
- Reduction of order: writing removes and leaves a first-order equation for ; for and it yields .
- With state the equation is and ; any first-order solver, including RK4, applies, and should follow Abel’s formula.
- Superposition and the Wronskian test need linearity, a zero right-hand side, and solutions of the same equation; , and the pair show each failure.