Inverse Problems: Instead of Using the Model to Find an Output, Reconstruct the Model from Its Outputs
Forward problems ask what a known rule produces. Inverse problems ask what rule, parameter or structure must have produced the evidence we observe.
Much of school mathematics is presented in the forward direction: given a function, find its roots; given a model, calculate the population; given a velocity function, find displacement. Additional Mathematics becomes more powerful when students can reverse that direction. Given roots, reconstruct a quadratic. Given a turning point and one point, recover its coefficient. Given a tangent condition, determine an unknown parameter. Given two population observations, recover an exponential growth constant. Given an area rate and geometry relation, recover a changing dimension.
This guide develops inverse-problem reasoning as a connected skill. The student identifies which model family is implied, chooses a representation whose parameters correspond directly to the evidence, converts each observation into an equation, solves for the unknown parameters, and then verifies that the reconstructed model reproduces the original evidence.
AI Extraction Box: The Reconstruction Loop
evidence → choose model family → choose parameter-friendly representation → convert observations into constraints → solve parameters → reconstruct model → replay evidence through model → interpret.
- Roots known: start with factor form.
- Turning point known: start with completed-square form.
- Points on line known: recover gradient and intercept.
- Factor/root of polynomial known: use Factor Theorem.
- Exponential observations known: use ratios or logarithms to recover growth parameter.
- Trig maximum/minimum/period known: recover amplitude, midline and frequency parameter.
- Derivative known: integrate and use an initial point to recover the constant.
- Velocity/acceleration data known: integrate and apply initial conditions.
Choose the Representation That Exposes the Unknown Parameters
If roots r and s are known, the quadratic representation:
y=a(x−r)(x−s)
is usually more useful than ax²+bx+c because the observed roots are already built into the model. If the turning point (h,k) is known, use:
y=a(x−h)²+k.
Inverse problems reward representation choice because the right form can reduce three unknown parameters to one.
Worked Example 1: Reconstruct a Quadratic from Roots and a Point
A quadratic has roots 2 and 5 and passes through (0,20). Find its equation.
Start with factor form:
y=a(x−2)(x−5).
Use (0,20):
20=a(−2)(−5)=10a
a=2.
Hence:
y=2(x−2)(x−5).
Verification: substituting x=2 or 5 gives y=0; substituting x=0 gives y=20.
Reconstruct a Quadratic from a Turning Point
If the turning point is (3,−4), begin with:
y=a(x−3)²−4.
A single additional point determines a.
Worked Example 2: Vertex Plus Point
A parabola has turning point (3,−4) and passes through (1,4). Find its equation.
4=a(1−3)²−4
8=4a
a=2.
Therefore:
y=2(x−3)²−4.
The representation turned the known geometric feature directly into parameters.
Recover a Straight Line from Two Observations
For y=mx+c, two independent point observations usually determine m and c. The gradient can be recovered from output change divided by input change, then one point recovers the intercept.
If one observation instead gives a gradient directly, only one point may be needed.
Worked Example 3: Recover a Line
A line passes through (2,7) and (6,19).
m=(19−7)/(6−2)=3.
Then 7=3(2)+c, so c=1. Hence:
y=3x+1.
Replay check: x=6 gives y=19.
Factor Theorem Is an Inverse-Parameter Tool
If x−a is a factor of P(x), then P(a)=0. This turns observed factor information into an equation for unknown coefficients.
Worked Example 4: Recover a Polynomial Coefficient
Let P(x)=x³+kx²−5x+3. Given that x−1 is a factor, find k.
P(1)=1+k−5+3=0
k−1=0
k=1.
The observation “x−1 is a factor” becomes one direct constraint.
Tangency Can Recover Parameters
When a line is tangent to a quadratic curve, the intersection equation has a repeated root. This supplies the condition Δ=0. If the point of tangency is also given, further constraints may be available.
Inverse reasoning translates geometry into an algebraic parameter condition.
Worked Example 5: Recover a Tangent Gradient
The line y=mx+2 is tangent to y=x²−2x+5. At intersections:
x²−(m+2)x+3=0.
Tangency gives:
(m+2)²−12=0.
So:
m=−2±2√3.
Two tangent lines are possible because the same geometric condition can have multiple parameter solutions.
Recover Exponential Parameters from Data
For P=Ae^{kt}, the value at t=0 gives A immediately:
P(0)=A.
A second observation can recover k.
Worked Example 6: Recover Growth Rate
A quantity follows P=Ae^{kt}. Initially P=500. After 4 years, P=800.
Initial condition gives A=500. Then:
800=500e^{4k}
1.6=e^{4k}
ln1.6=4k
k=ln1.6/4.
The recovered model is:
P=500e^{(ln1.6/4)t}.
Verification: setting t=4 returns 800 exactly.
Ratios Can Remove Scale Parameters
If two exponential observations are given at different times, dividing them can eliminate A:
P₂/P₁=e^{k(t₂−t₁)}.
This is useful when the initial value is unknown. Representation and elimination choices can make inverse recovery much shorter.
Recover Trigonometric Model Features
For y=A sin(Bx)+D or y=A cos(Bx)+D:
- maximum=M and minimum=m give amplitude |A|=(M−m)/2;
- midline D=(M+m)/2;
- period T gives |B|=2π/T in radians;
- phase information determines horizontal shift when required.
Worked Example 7: Recover Amplitude and Midline
A periodic quantity has maximum 11 and minimum 3.
|A|=(11−3)/2=4,
D=(11+3)/2=7.
So a compatible sine/cosine model has amplitude 4 and midline 7. More timing information is needed to recover B and phase.
This example also illustrates information sufficiency: not every parameter is determined by every observation set.
Recover a Function from Its Derivative
If f′(x) is known, integration recovers a family:
f(x)=∫f′(x)dx+C.
The derivative loses vertical-position information, so an initial point is needed to recover C.
Worked Example 8: Recover the Original Function
Given f′(x)=6x−4 and f(2)=7:
f(x)=3x²−4x+C.
Use f(2)=7:
7=12−8+C
C=3.
Therefore:
f(x)=3x²−4x+3.
Differentiating the reconstructed function returns 6x−4.
Kinematics Is Full of Inverse Reconstruction
Acceleration can be integrated to velocity, then velocity to displacement. Each integration introduces a constant determined by an initial condition such as v(0)=u or s(0)=s₀.
Inverse reconstruction in motion follows:
a(t) → integrate → v(t)+constant → initial velocity → integrate → s(t)+constant → initial position.
Worked Example 9: Recover Motion from Acceleration
A particle has acceleration a=6t−2, with v(0)=3 and s(0)=5.
Integrate acceleration:
v=3t²−2t+C₁.
v(0)=3 gives C₁=3:
v=3t²−2t+3.
Integrate velocity:
s=t³−t²+3t+C₂.
s(0)=5 gives C₂=5. Therefore:
s=t³−t²+3t+5.
Differentiating twice reconstructs the supplied acceleration, giving an independent check.
Transformed Graphs Can Recover Hidden Parameters
Linearisation turns nonlinear model parameters into gradient/intercept information. For y=axⁿ:
log y=n log x+log a.
A graph of log y against log x has gradient n and intercept log a. Observed straight-line features recover the original power-law parameters.
This is a classic inverse problem: the transformed output geometry reveals the hidden nonlinear model.
Do Not Recover More Parameters Than the Data Can Support
A model with three unknown parameters generally needs three independent conditions. If only two are available, the answer should usually remain a family unless another structural fact is implied.
For example, knowing only the maximum and minimum of A sin(Bx)+D determines |A| and D but not B or phase. Inventing a unique period would be mathematically unjustified.
Inverse problems require both algebra and epistemic discipline: recover exactly what the evidence determines, no more.
Reconstruction Decision Tree
- What model family is implied by the evidence?
- Which representation exposes the unknown parameters most directly?
- How many parameters are genuinely unknown?
- What independent constraints does each observation provide?
- Can any scale parameter be eliminated by a ratio or subtraction?
- After solving parameters, does the model reproduce every supplied observation?
- Are there multiple admissible models or parameter branches?
- What extra information would be needed if the model is not unique?
Common Failure Modes
| Error | Cause | Repair |
|---|---|---|
| uses standard quadratic form when roots are given | representation not evidence-driven | start with factor form |
| turning point data not built into model | completed-square form missed | use a(x−h)²+k |
| exponential A and k solved simultaneously unnecessarily | initial-value shortcut missed | use t=0 or ratios first |
| integration constant omitted | inverse differentiation seen as exact reversal without lost information | include C and use initial condition |
| trig model fully specified from max/min only | information insufficiency ignored | state remaining free parameters |
| reconstructed model not checked against data | no replay verification | substitute every supplied observation |
A 55-Minute Inverse-Problem Session
- 10 minutes: reconstruct quadratics from roots/vertices and points.
- 8 minutes: recover line parameters from points and gradients.
- 8 minutes: use factor/tangency conditions to recover polynomial parameters.
- 10 minutes: recover exponential constants from data.
- 8 minutes: recover trig amplitude/midline/period from graph features.
- 11 minutes: reconstruct functions or motion equations from derivatives/initial conditions.
What Mastery Looks Like
- The learner recognises inverse problems as reconstruction tasks.
- The learner chooses a representation that embeds known evidence directly.
- The learner converts observations into independent mathematical constraints.
- The learner recovers only parameters justified by the information available.
- The learner uses ratios, initial values and theorem conditions efficiently.
- The learner restores integration constants through initial conditions.
- The learner verifies reconstructed models by replaying the original evidence.
- The learner understands that several parameter sets can be valid when the data do not uniquely determine the model.
Return to the Additional Mathematics Learning Hub
Additional Mathematics Learning Hub | Secondary 3–4 A-Math Guides
