Full-course synthesis workshop

Every numerical-analysis course teaches its tools one at a time — finite differences in one article, interpolation in the next, then integration, then root-finding. Real applied problems never arrive pre-sorted like that. A sensor logs a handful of readings, and you must decide for yourself: is this a "what is the value here" question, a "what is the running total" question, or a "where does this cross zero" question? This capstone article is the wrap-up of the "Numerical Analysis" block of MA10002. It introduces no new technique — every method used here was already built and justified in earlier articles of this chapter, and numerical linear algebra (Gauss elimination, Gauss-Seidel) is used strictly as a black-box tool from the "Matrices" chapter. The goal is pipeline literacy: given a table of numbers and a question in English, know instantly which tool answers it, and how to hand values cleanly from one stage to the next without losing accuracy along the way.

The core insight

A tabulated data problem is solved by reading finite differences to check smoothness, then routing the question — value at a point, cumulative total, or threshold crossing — to interpolation, numerical integration, or root-finding respectively, with numerical linear algebra as a supporting black box when several unknowns must be solved together.

The core idea

A typical applied numerical-analysis pipeline has a fixed shape, and recognizing the shape is the actual skill being tested here — not re-deriving any formula.

Stage 1 — start from tabulated data. Sensor readings, experimental measurements, or any function sampled at equally spaced points $x_0, x_1, \dots, x_n$ with common spacing $h$ .

Stage 2 — build the difference table. Compute forward differences $\Delta y_i = y_{i+1} - y_i$ , then $\Delta^2 y_i = \Delta y_{i+1} - \Delta y_i$ , and so on. If some order of difference settles to (nearly) constant, the data behaves like a polynomial of that degree — a diagnostic that both validates the data and simplifies everything downstream.

Stage 3 — interpolate for a value. To estimate $y$ at an $x$ not in the table, pick Newton's forward formula near the start of the table, Newton's backward formula near the end, or a central formula (Gauss/Stirling) near the middle — the choice is about where the target point sits, not about the data itself. With $p = (x - x_0)/h$ :

y(x) \approx y_0 + p\,\Delta y_0 + \frac{p(p-1)}{2!}\Delta^2 y_0 + \frac{p(p-1)(p-2)}{3!}\Delta^3 y_0 + \dots

where: $y_0$ is the table value at the base point $x_0$ , $p$ is the fractional distance to the target measured in steps of $h$ , and each $\Delta^k y_0$ is a $k$ -th forward difference read straight off the table.

Stage 4 — integrate for a cumulative quantity. When the question is "what is the total," use the trapezoidal rule for any even or odd number of intervals, or Simpson's $1/3$ rule (needs an even number of intervals $n$ , i.e. an odd number of points) for higher accuracy:

\int_{x_0}^{x_n} y\,dx \approx \frac{h}{3}\Big[(y_0+y_n) + 4(\text{odd-indexed } y) + 2(\text{even-indexed, interior } y)\Big]

Stage 5 — solve for a root when the question is inverted. "At what point does this quantity cross a threshold" is answered by bisection, regula falsi, or Newton-Raphson applied to the raw data function or a fitted interpolating polynomial, never by more interpolation.

Stage 6 — call numerical linear algebra when several unknowns appear together, e.g. fitting multiple coefficients simultaneously or solving simultaneous balance equations. Gauss elimination, Gauss-Jordan, or Gauss-Seidel are used exactly as learned in the Matrices chapter — no new derivation is needed here. The unifying thread: each technique converts one representation of information — a table, a fitted polynomial, an equation — into the representation the next question actually needs.

Figure

A pipeline diagram showing tabulated data flowing through finite differences into a fork that leads to interpolation for a point value integration for a cumulative total and root finding for a threshold crossing with a side box for numerical linear algebra

Derivation ASimpson's 1/3 rule applied to a distance table, then reading exactness off constant second differences

The goal is to compute the total distance a car travels over an eight-second interval from five tabulated velocity readings using Simpson's $1/3$ rule, and to show — from the difference table alone, without knowing the algebraic form of $v(t)$ in advance — that the tabulated velocity is exactly quadratic in $t$ , which certifies that Simpson's rule (exact for polynomials up to degree three) returns the true distance rather than an approximation.

Step 1. Lay out the table. The velocity $v(t)$ in m/s is recorded at $t = 0, 2, 4, 6, 8$ s as $v = 0, 8, 20, 36, 56$ , so $h = 2$ and there are $n = 4$ intervals (even, as Simpson's $1/3$ rule requires).

Step 2. Build the forward-difference table to check regularity. $\Delta v_0 = 8, \Delta v_1 = 12, \Delta v_2 = 16, \Delta v_3 = 20$ ; second differences $\Delta^2 v_0 = 4, \Delta^2 v_1 = 4, \Delta^2 v_2 = 4$ — constant; third difference $\Delta^3 v_0 = 0$ . A constant second difference with a vanishing third difference is the signature of an exactly quadratic function, so $v(t)$ is a degree-2 polynomial in $t$ , not merely close to one.

Step 3. Apply Simpson's $1/3$ rule to the same table for the cumulative distance. With $h = 2$ : $\int_0^8 v\,dt \approx \frac{h}{3}\big[(v_0+v_4) + 4(v_1+v_3) + 2v_2\big] = \frac{2}{3}\big[(0+56) + 4(8+36) + 2(20)\big] = \frac{2}{3}\big[56 + 176 + 40\big] = \frac{2}{3}(272) = 181.33\ \text{m}$ .

Step 4. Justify exactness, not just accuracy. Simpson's $1/3$ rule is derived by fitting a quadratic through each pair of intervals and integrating that quadratic exactly; it is therefore exact for any polynomial of degree three or lower. Since Step 2 already certified $v(t)$ is exactly quadratic (degree 2), the fitted quadratic on each panel is the true velocity function, so the $181.33$ m answer is the exact distance travelled, with zero truncation error — a conclusion available only because the difference table was checked first. Result: distance $= 181.33$ m, exact because $v(t)$ is confirmed exactly quadratic by constant second differences.

Worked example

A car's velocity $v(t)$ (m/s) is tabulated at $t = 0, 2, 4, 6, 8$ s as $v = 0, 8, 20, 36, 56$ . Question: at what time does the car's displacement first reach $100$ m, given $s(t) = \dfrac{2t^2}{3} + \dfrac{t^3}{3}$ (found by integrating the tabulated $v(t)$ )?

Step 1 — bracket the root. Define $g(t) = \dfrac{2t^2}{3} + \dfrac{t^3}{3} - 100$ . Then $g(6) = 24 + 72 - 100 = -4$ and $g(8) = 42.67 + 170.67 - 100 = 113.33$ . The sign change confirms a root in $[6, 8]$ , so Newton-Raphson from $t_0 = 7$ is a reasonable choice.

Step 2 — differentiate. $g'(t) = \dfrac{4t}{3} + t^2$ .

Step 3 — first iteration. $g(7) = 32.67 + 114.33 - 100 = 47.00$ ; $g'(7) = 9.33 + 49 = 58.33$ .

t_1 = 7 - \frac{47.00}{58.33} = 7 - 0.8057 = 6.1943

Step 4 — second iteration. $g(6.1943) = 25.58 + 79.22 - 100 = 4.80$ ; $g'(6.1943) = 8.259 + 38.37 = 46.63$ .

t_2 = 6.1943 - \frac{4.80}{46.63} = 6.0914

Step 5 — third iteration and stop. $g(6.0914) \approx 24.74 + 75.30 - 100 = 0.04$ , close enough to zero to stop. The car has covered exactly $100$ m at $t \approx 6.09$ s.

Competitive stretch

Notice how the two worked pieces chain: the difference table (Stage 2) certified the data was exactly quadratic, which let Simpson's rule (Stage 4) return an exact $181.33$ m over $[0,8]$ s with zero error; separately, inverting the "when does distance equal $100$ m" question routed to Newton-Raphson (Stage 5) rather than more integration. A synthesis question often hides a second sub-question that needs a different tool from the first — always re-ask "what kind of question is this" at each stage rather than reusing the last tool used.

Common traps

Jumping straight to root-finding or integration without first reading the difference table — constant differences can certify an exact answer (as in the worked Simpson's example) rather than merely an approximate one, and skipping this check throws away that certainty.
Confusing which technique answers which question type: interpolation answers "what is $f$ here," integration answers "what is the cumulative total," and root-finding answers "where does $f$ cross a given value" — misreading the question sends you to the wrong tool entirely.
Applying Simpson's $1/3$ rule with an odd number of intervals; it requires an even $n$ (odd number of data points), unlike the trapezoidal rule, which works for any $n$ .
Treating Gauss elimination or Gauss-Seidel as unrelated to this chapter — when a synthesis problem needs several unknowns solved together, those are the correct tool, used exactly as learned, not re-derived from scratch.

Ask the tutor: Describe your own tabulated scenario — sensor data, an experiment, anything with equally spaced readings — and work through with the AI tutor which tool (differences, interpolation, integration, root-finding, or linear algebra) each part of your question actually calls for.

Why check the difference table before choosing a method — isn't that just extra work?

It pays for itself. If a table's differences settle to a constant at some order, the data is exactly (not approximately) a polynomial of that degree, which means later steps — an interpolation, an integration, even a root search on a fitted polynomial — inherit that same exactness rather than an unknown truncation error. In the worked Simpson's example here, checking the second differences first is what let us claim the $181.33$ m answer was exact rather than merely a good estimate. Skipping the check still gives you a numeric answer, but you lose the ability to say how trustworthy it is.

How do I decide between Newton's forward, backward, and central interpolation formulae in a synthesis problem?

Purely by where your target point sits in the table, not by the data's shape. If the point is near the first row, forward differences (built from the top of the table) converge fastest, so use Newton's forward formula, as in the $t = 1$ example here where $t_0 = 0$ was the natural base. Near the last row, use the backward formula built from bottom-up differences. Near the middle of a long table, a central formula (Gauss's or Stirling's) usually gives better accuracy because it uses differences from both sides of the target symmetrically.

If a synthesis problem needs several unknowns solved at once, do I have to re-derive Gauss elimination or Gauss-Seidel here?

No. This capstone article deliberately treats numerical linear algebra as a finished, black-box tool inherited from the "Matrices" chapter of MA10002. If a scenario needs, say, three unknown coefficients fit simultaneously from three equations, you set up the linear system exactly as you would in that chapter and solve it with Gauss elimination (direct, exact arithmetic) or Gauss-Seidel (iterative, useful for large diagonally dominant systems) — using the method as-is, with no new derivation expected in a synthesis question.

Full-course synthesis workshop

The core idea

Worked example

Competitive stretch

Common traps

Before this