Additional Mathematics | How to Reconstruct a Shifted Exponential Model from Three Observations

A shifted exponential model has the form P(t) = C + Ab^t, where C is a fixed baseline, A is a non-zero scale factor and b is a positive multiplier different from 1. Three equally spaced exact observations can determine all three parameters: subtract consecutive outputs, divide the second difference by the first to obtain the multiplier, then recover A and C.

For example, a hypothetical quantity is 14 at t = 0, 20 at t = 1 and 32 at t = 2. Its successive differences are 6 and 12, so b = 12/6 = 2. Then A = 6 and C = 8, giving P(t) = 8 + 6(2^t). This guide shows why the difference-ratio method works, how to handle nonzero starting times and decay, and when the data cannot determine a valid model.

Why a baseline changes the exponential problem

The simple model P(t) = Ab^t has two parameters, A and b. In the shifted model P(t) = C + Ab^t, a third parameter C sets a baseline from which the exponential component grows or decays.

At t = 0 the value is C + A, not A unless C is zero. Therefore reading the first observation as the exponential coefficient is generally wrong.

Adding a baseline also prevents us from using the raw observation ratio P(t + 1)/P(t) to obtain b. Those ratios are affected by C and are not generally constant.

The key idea is to subtract consecutive observations. The baseline disappears when outputs are subtracted, leaving a pure exponential difference whose ratio is the true multiplier. Recognising what cancels is the central algebraic insight.

Write all three observations before solving

Let P(0) = p₀, P(1) = p₁ and P(2) = p₂. With P(t) = C + Ab^t, the equations are p₀ = C + A, p₁ = C + Ab and p₂ = C + Ab².

These are three equations for the three unknown parameters C, A and b. Direct simultaneous solution is possible, but subtracting eliminates C immediately.

Define the first difference Δ₀ = p₁ − p₀ and the second difference Δ₁ = p₂ − p₁. These are changes over equal time intervals, not absolute ratios of observations.

For the illustrative data 14, 20 and 32, the differences are Δ₀ = 6 and Δ₁ = 12. Recording them explicitly prevents confusing a raw ratio of 20/14 with the multiplier that belongs to the underlying shifted exponential term.

Derive the difference-ratio identity

Subtract the first two model equations: Δ₀ = (C + Ab) − (C + A) = A(b − 1).

Subtract the next pair: Δ₁ = (C + Ab²) − (C + Ab) = Ab(b − 1). The same factor A(b − 1) appears in both differences, with an extra factor b in the second.

Provided Δ₀ is not zero, dividing gives Δ₁/Δ₀ = b. The baseline C has disappeared, as has the unknown scale A.

Thus b = (p₂ − p₁)/(p₁ − p₀) for three equally spaced measurements with a nonzero first difference. This formula is justified by the assumed model; it is not a universal rule for arbitrary three data points.

Find the multiplier in the worked example

For p₀ = 14, p₁ = 20 and p₂ = 32, calculate Δ₀ = 20 − 14 = 6 and Δ₁ = 32 − 20 = 12.

The exponential multiplier is b = Δ₁/Δ₀ = 12/6 = 2. The exponential component doubles over each one-unit interval.

Notice that the outputs themselves do not double: 14 becomes 20, then 32. This is consistent with an exponential part that doubles above a constant baseline.

A student who instead sets b = 20/14 or b = 32/20 will obtain two different factors and may incorrectly conclude that no exponential model exists. The shifted model requires comparison of differences, not direct output ratios.

Recover the scale and baseline

The first difference satisfies Δ₀ = A(b − 1). With Δ₀ = 6 and b = 2, we have 6 = A(1), so A = 6.

The initial observation is p₀ = C + A = 14, giving C = 14 − 6 = 8.

Therefore the completed model is P(t) = 8 + 6(2^t). The baseline is 8, while the exponential excess above it begins at 6.

All parameters have now been determined. This two-stage reconstruction—differences for b, then an anchor for A and C—is generally clearer than substituting every numerical observation into a system containing powers of an unknown b.

Verify each of the three observations

At t = 0, the model gives 8 + 6(1) = 14. At t = 1, it gives 8 + 6(2) = 20. At t = 2, it gives 8 + 6(4) = 32.

Every supplied observation is reproduced exactly. The increments are 6 and 12, which are related by the recovered multiplier 2.

A model-check should use the original observations rather than only rechecking Δ₁/Δ₀. The latter confirms the change pattern but cannot detect an incorrectly chosen vertical baseline.

The three direct substitutions establish that the mathematical reconstruction fits the three observations under the assumed model. They do not yet prove that the same exponential law will continue indefinitely in a real process.

Predict the next measurement without refitting

With P(t) = 8 + 6(2^t), the next predicted value is P(3) = 8 + 6(8) = 56.

The next difference is 56 − 32 = 24, twice the preceding difference of 12. The subsequent predicted value is P(4) = 8 + 6(16) = 104, with a difference of 48.

The sequence of outputs is 14, 20, 32, 56, 104, while the successive differences are 6, 12, 24, 48. The differences, not the outputs, follow a geometric progression.

That pattern is a useful independent check on the role of the baseline. It also gives an efficient prediction method for a fourth equally spaced time point when a shifted exponential assumption is appropriate.

A shifted decay model uses the same method

Suppose a hypothetical quantity is 90 at t = 0, 50 at t = 1 and 30 at t = 2. The differences are −40 and −20.

Their ratio is b = (−20)/(−40) = 1/2. The exponential multiplier lies between zero and one, so the exponential component decays toward a baseline.

Use Δ₀ = A(b − 1): −40 = A(−1/2), giving A = 80. Then C = 90 − 80 = 10.

The model is P(t) = 10 + 80(1/2)^t. Check P(1) = 10 + 40 = 50 and P(2) = 10 + 20 = 30. The baseline is approached from above rather than crossed in this positive-scale decay example.

A negative scale can approach the baseline from below

A shifted exponential can also have A < 0, provided the interpretation allows the resulting outputs. For example P(t) = 20 − 12(1/2)^t produces values 8, 14 and 17 at t = 0, 1 and 2.

Its differences are +6 and +3. Their ratio is 1/2, so b = 1/2 despite the quantity increasing. The reason is that a negative exponential component is becoming less negative.

Recover A from Δ₀ = A(b − 1): 6 = A(−1/2), giving A = −12. Then C = 8 − (−12) = 20.

This shows why the base b alone does not decide whether the total shifted model increases or decreases. The sign of A and whether b is above or below one both matter.

The baseline is not always the first observation

For P(t) = 10 + 80(1/2)^t, the first observation is P(0) = 90, while the baseline is C = 10.

The baseline describes the value approached as t becomes large because (1/2)^t tends to zero. It is not equal to the initial value or the first listed measurement.

For the growth example P(t) = 8 + 6(2^t), the constant C = 8 is a vertical offset in the formula, but the model grows without bound for increasing t. It does not approach 8 as t tends to positive infinity.

Students should interpret C through the entire expression and its direction of time, not automatically label it a long-run equilibrium in every shifted exponential regardless of b.

Measurements may begin at a nonzero time

Suppose the observations occur at t = 2, 3 and 4, with values 32, 56 and 104. The differences are 24 and 48, giving b = 2.

The first difference now satisfies 24 = A b²(b − 1) = A(4)(1), so A = 6. Using P(2) = 32 gives C + 6(4) = 32, hence C = 8.

The same model P(t) = 8 + 6(2^t) is recovered. But setting A equal to the difference 24 would be wrong because that difference occurs after two exponent steps.

For observations starting at t = t₀, the first difference over one unit is A b^t₀(b − 1). The time origin must be included when recovering the scale.

Changing the time origin gives an equivalent model

In the previous example, define s = t − 2 so that s = 0 occurs at the first observed measurement. Then P becomes 8 + 24(2^s).

This is equivalent to 8 + 6(2^t), because t = s + 2 and 6(2^(s+2)) = 24(2^s).

The baseline C = 8 and per-step multiplier b = 2 remain unchanged under a shift of time origin. The scale coefficient changes from 6 to 24 because it refers to the exponential component at the new origin.

Both expressions are correct when their time variables are stated consistently. Problems involving dates or years can become easier by choosing a convenient origin, but the final interpretation must remain clear.

Equally spaced observations do not have to be one unit apart

Suppose outputs occur at t = 0, 2 and 4. The model is still P(t) = C + Ab^t, but each observation gap is two time units.

The successive differences have ratio b², not b. For the original growth model, the observed values are P(0) = 14, P(2) = 32 and P(4) = 104. The differences are 18 and 72, with ratio 4.

Therefore b² = 4, and since b > 0, b = 2. We then recover A = 6 and C = 8 from the original observation equations.

For equally spaced measurements separated by h > 0 time units, the ratio of successive differences gives b^h. Recover the per-unit multiplier by taking the positive h-th root. Do not mistake a per-two-unit multiplier for a per-one-unit multiplier.

Why three arbitrary positive data points may be incompatible

The ordinary shifted exponential with A ≠ 0 and b > 0, b ≠ 1 has successive nonzero differences whose ratio equals b. Therefore those differences must have the same sign and a positive ratio.

Suppose three equally spaced outputs are 10, 20 and 15. Their differences are +10 and −5, whose ratio is −1/2. No positive b can produce that ratio under this model.

The output has changed direction between the observations, while a simple shifted exponential with fixed real A and positive b is monotonic over time.

The correct conclusion is that these exact observations do not fit the assumed model. Forcing a negative base would create a different structure that may not be real-valued at all real times.

Equal nonzero differences are another special case

Suppose equally spaced outputs are 10, 15 and 20. The differences are both 5. Their ratio is 1.

The difference-ratio formula would give b = 1, but if b = 1 then C + Ab^t = C + A is constant and all differences must be zero. This contradicts the nonzero differences.

Therefore no exact shifted exponential of the stated nondegenerate form fits 10, 15 and 20. These values instead follow a linear pattern.

This example is particularly useful because a student might blindly obtain b = 1 and continue dividing by b − 1 = 0. The algebraic division warning reveals that the proposed exponential model is inconsistent with the observations.

A constant data sequence is not uniquely identifiable

Suppose the three measurements are 10, 10 and 10. Both differences vanish, so the ratio Δ₁/Δ₀ is 0/0 and is undefined.

There are many representations of a constant function in the broad form C + Ab^t if degenerate parameters are permitted. For instance A = 0, C = 10 with arbitrary positive b gives the constant function 10. Alternatively b = 1 and C + A = 10 also gives it.

Thus the data do not uniquely determine all three parameters. Under the stricter assumptions A ≠ 0 and b ≠ 1, no such constant sequence is possible.

Never divide by an observed difference until checking that it is nonzero. Special cases can change the entire answer from a unique fit to no admissible fit or an underdetermined family.

Two observations cannot generally determine three parameters

If the only data are P(0) = 14 and P(1) = 20, the equations are C + A = 14 and C + Ab = 20.

Subtracting yields A(b − 1) = 6. For every positive b ≠ 1, we can set A = 6/(b − 1) and C = 14 − A. That gives a different shifted exponential fitting the two observations.

For b = 2, the model is 8 + 6(2^t). For b = 3, the model is 11 + 3(3^t). Both give 14 and 20 at the first two times, but at t = 2 they predict 32 and 38 respectively.

The third measurement distinguishes the models. This is why three independent observations, not merely a desire for a more complicated formula, are needed for the three-parameter reconstruction.

A fourth observation checks rather than redefines the fit

After fitting three exact observations, a fourth observation can test the model’s consistency. For P(t) = 8 + 6(2^t), the predicted value at t = 3 is 56.

If an actual fourth observation were 56, it would agree exactly with the model. If it were 60, the three-parameter model fitted from the first observations would not reproduce all four measurements.

Real measurements may contain rounding or noise, so a statistical fit can be appropriate in a more advanced modelling context. But the exact algebraic task and the noisy-data task are different.

Do not silently adjust the baseline merely to fit the fourth point while pretending that the first three are still matched exactly. The model assumptions and accepted accuracy must be stated.

Use the reconstructed model to solve a threshold

Suppose P(t) = 8 + 6(2^t), with t ≥ 0, and ask when P first equals 50. Solve 8 + 6(2^t) = 50.

Subtract the baseline: 6(2^t) = 42, so 2^t = 7. Therefore t = ln 7/ln 2, approximately 2.807 time units.

If the question asks for the first whole-number t with P(t) ≥ 50, check t = 2 and 3. We have P(2) = 32 and P(3) = 56, so the first whole t is 3.

The baseline must be subtracted before using logarithms. Taking the logarithm of an arbitrary sum C + Ab^t as though it split into separate logarithms would be invalid.

A baseline changes the meaning of doubling time

For P(t) = 8 + 6(2^t), the exponential component P(t) − 8 doubles every time unit. But the total output does not double from t = 0 to 1: it increases from 14 to 20, not to 28.

Therefore “doubling time” must specify what is doubling. If the question concerns the excess above baseline, the time is one unit in this model. If it concerns the total output, the time depends on the starting time and baseline.

For example, to double the initial total P(0) = 14, solve 8 + 6(2^t) = 28, so 2^t = 10/3 and t = ln(10/3)/ln 2.

This distinction prevents a student from applying a standard unshifted exponential formula to the wrong quantity.

Interpret a fitted baseline cautiously

A fitted C is a mathematical parameter. Whether it represents a physical floor, residual amount, background signal or measurement offset depends on the actual context.

For the decay model P(t) = 10 + 80(1/2)^t, C = 10 is approached as t grows, so interpreting it as a long-run level can be reasonable if the model remains valid.

For a growing model with b > 1, C is still a vertical offset but is not the long-run positive-time limit of P(t). An increasing exponential term dominates it.

Two or three observations do not prove a physical mechanism. The mathematical model should be tested on additional data and not extrapolated far beyond the observed interval without justification.

A compact reconstruction routine

Start by verifying that the three observations are equally spaced in time, with interval h. Write the outputs in increasing time order and calculate their successive differences.

If the first difference is nonzero, find the ratio of differences. Under the model, that ratio equals b^h and must be positive. For h = 1, this is b directly.

Recover A using one difference and the appropriate starting-time factor, then recover C using one observed output. Substitute all three observations into the completed formula to verify them.

Finally inspect degenerate cases such as zero differences, a ratio equal to one with nonzero differences, or a negative ratio, and state when the assumed model has no valid unique fit. This sequence is concise but mathematically complete.

Independent practice with checked answers

Question one: equally spaced outputs at t = 0, 1, 2 are 14, 20, 32. Answer: P(t) = 8 + 6(2^t).

Question two: outputs at t = 0, 1, 2 are 90, 50, 30. Answer: P(t) = 10 + 80(1/2)^t.

Question three: outputs at t = 0, 1, 2 are 8, 14, 17. Answer: P(t) = 20 − 12(1/2)^t.

Question four: outputs at t = 2, 3, 4 are 32, 56, 104. Answer: P(t) = 8 + 6(2^t).

Question five: outputs at t = 0, 1, 2 are 10, 15, 20. Does a nondegenerate shifted exponential with positive b fit them exactly? No.

Question six: P(t) = 8 + 6(2^t). First whole t ≥ 0 for P(t) ≥ 50? Answer: t = 3.

For each, show the difference ratio and all three observation checks.

Reconstructing a three-parameter shifted exponential connects indices, equations, growth and decay, and careful interpretation of measurements. It can be useful enrichment beyond a two-parameter model; the exact assessed scope depends on the learner’s school and examination syllabus.

For Singapore, refer to SEAB’s applicable 2026 O-Level Additional Mathematics list, 2027 SEC G3 list or 2027 SEC G2 list.

Parents can enquire about suitable eduKate Singapore three-student tutorials near Sixth Avenue MRT. The goal is to identify what three independent observations actually determine, then check whether the chosen model remains meaningful. Properly taught kids shine a bright light into the future.