The Gold Standard of Numerical Analysis is a computed answer accompanied by a defensible account of its accuracy, sensitivity and limitations. It explains why an approximation should work, how errors can enter, and what checks make the result useful for the problem being solved.
Numerical analysis studies methods for obtaining approximate solutions to mathematical problems. It includes root finding, interpolation, integration, differential equations, linear systems and other computational tasks. Its central question is not simply whether a computer returns a number. It is whether the method preserves enough meaningful information, at an appropriate cost, for the intended use.
Consider a calculator displaying twelve decimal places. Those digits describe a representation. They do not establish that the input measurements, mathematical model or algorithm justify twelve places of confidence. A short, carefully bounded approximation can be more useful than a long display with no error account.
This guide connects algorithms with the reasoning that makes them dependable. The worked examples are original and deliberately small enough to inspect by hand. They illustrate principles, rather than validate any actual engineering, financial or scientific system. The educational link is learning that transfers into life: a learner should recognise the same issue when it appears in a spreadsheet, simulation or unfamiliar mathematical problem.
Find the question you want to answer
Follow the route from a mathematical question to a computed answer with justified checks.
Chapters 14–18
Check the whole computation
Connect differentiation, equations, verification and practice.
Full chapter index · Practice and explained answers · How X Works
Chapter contents
Define accuracy · Chapters 1–4
Understand finite computation · Chapters 5–7
Solve and approximate · Chapters 8–13
Check the whole computation · Chapters 14–18
CHAPTER 1 OF 22 · Define accuracy
1. Begin with the mathematical problem and its purpose
Before selecting an algorithm, state what must be found. Is it a zero of a function, an area, a trajectory, a solution to simultaneous equations or a fitted relationship? These questions can share computational tools while requiring different error checks.
Then state why the answer is needed. A rough estimate for exploring a pattern can have different requirements from a value close to a consequential decision boundary. Accuracy is meaningful relative to the quantity, scale and use. “As accurate as possible” leaves practical questions unanswered.
MIT’s Introduction to Numerical Analysis maps major topics including roots, interpolation, integration, differential equations and linear algebra. The course is a useful formal continuation, not a prerequisite for following the small examples here.
Write a short problem account: inputs, desired output, units, relevant assumptions and required accuracy. That account becomes the basis for checking the computation. Without it, an algorithm can solve a precisely specified mathematical task while the user assumes it answered a different real-world question.
CHAPTER 2 OF 22 · Define accuracy
2. Separate model error, data error and numerical error
A mathematical model simplifies a real system. Input data may be uncertain. A numerical method introduces approximations and finite arithmetic. These sources of error belong to different parts of the chain, even when their effects appear together in the final output.
Suppose a fictional model assumes a constant rate, while the real rate changes over time. Solving the constant-rate model with more digits does not remove the modelling difference. Likewise, a highly accurate solver cannot reconstruct an input value that was never measured precisely.
Numerical error concerns the gap introduced when solving the mathematical problem computationally. Truncating a series, replacing a derivative with a finite difference or stopping an iteration all require an approximation account. Rounding during finite arithmetic adds another contribution.
The existing numerical stability and conditioning guide develops these distinctions. A useful diagnostic asks where the uncertainty entered. If the problem is inaccurate input, investigate measurement. If the model is unsuitable, revise the representation. If the algorithm is unstable, choose a better computational route. Treating all three as “rounding” prevents the right repair.
CHAPTER 3 OF 22 · Define accuracy
3. Absolute and relative error answer different questions
Absolute error is the magnitude of the difference between an approximation and the reference value. Relative error compares that difference with a suitable scale, commonly the magnitude of a nonzero reference value. The comparison is useful only when the reference and denominator are clearly defined.
An error of 0.01 has different significance for a quantity near 100 and one near 0.001. For the former it is small in relative terms; for the latter it is much larger than the quantity. Units matter for absolute error, while the relative ratio is dimensionless when compatible quantities are compared.
Near zero, a pure relative-error requirement can be inappropriate or undefined. Computational software often combines an absolute tolerance with a relative tolerance. The absolute part supplies a meaningful scale near zero, while the relative part scales with larger values.
A solver reports that its answer changed by less than 10⁻⁶. What is still missing? We need to know the quantity’s scale, the relevant units and whether small successive changes actually imply a small solution error. An iteration can stagnate without reaching an acceptable answer. A tolerance is part of a checking strategy, rather than a certificate attached to a number.
For example, a position tolerance of 0.001 metre corresponds to 1 millimetre. Changing a program’s coordinate unit from metres to millimetres while leaving a bare numeric tolerance of 0.001 unchanged makes the physical tolerance a thousand times smaller. The solver setting did not preserve its meaning through the unit change. A reliable interface records both the value and the scale it represents.
CHAPTER 4 OF 22 · Define accuracy
4. Conditioning describes the problem’s sensitivity
A problem is sensitive when small changes in its inputs can cause large changes in the desired output. Conditioning concerns that mathematical relationship. It is distinct from whether a particular algorithm handles the relationship well.
Consider two nearly parallel lines. Their intersection can move substantially when one line’s position or slope changes slightly. An exact algebraic calculation of the supplied lines does not remove uncertainty in the lines themselves. The sensitivity belongs to the problem.
Condition numbers formalise sensitivity for particular problems and norms. Their interpretation depends on what inputs and outputs are being compared. A large condition number warns that small relative input changes can be amplified; it is not a universal percentage error attached to every result.
A learner can begin with a simpler question: if I perturb a relevant input by an amount consistent with its uncertainty, how much does the answer move? This is not always a substitute for a formal analysis, but it makes the idea concrete. Sensitivity checks help distinguish unstable computation from a problem whose answer is intrinsically difficult to determine from the available data.
CHAPTER 5 OF 22 · Understand finite computation
5. Stability describes the computational method
Numerical stability concerns how an algorithm handles errors introduced through finite computation. A stable method avoids unnecessary amplification relative to an appropriate problem account. The precise mathematical definition depends on the class of algorithm and the error framework being used.
Backward error asks whether the computed result can be interpreted as an exact solution to a nearby problem. Forward error asks how far the computed result lies from the solution of the original problem. A small backward error is valuable, but an ill-conditioned problem can still have a substantial forward error.
This distinction explains why two reassuring statements can coexist: the algorithm behaved well, yet the final answer remains sensitive. The correct response is to preserve both facts. Blaming the computer for the sensitivity hides the input problem; praising the algorithm without mentioning sensitivity overstates the answer.
MIT’s numerical-methods course develops floating-point arithmetic, backward error, conditioning and stability. The existing focused stability article offers an accessible bridge. When judging a computation, ask two questions separately: how sensitive is the problem, and how much additional error does this method introduce?
CHAPTER 6 OF 22 · Understand finite computation
6. Finite representation changes the arithmetic environment
Computers store numbers in finite representations. Many familiar decimal fractions do not have finite binary expansions, so a binary floating-point format stores a nearby representable value. Intermediate operations can also be rounded. This is a designed arithmetic system with rules, rather than random machine behaviour.
Addition is not generally associative after rounding: regrouping a sum can change the final represented value. Values can also overflow the available range or become too small for the representation to preserve normally. The algorithm needs to account for scale and operation order.
Decimal arithmetic can suit tasks governed by decimal rules, but it does not eliminate all numerical issues. A finite decimal representation still needs rounding and range policies. The appropriate representation depends on the problem and its requirements.
The calculator and rounding guide and spreadsheet error-checking guide connect this with familiar tools. For learners, compare an estimate with the displayed output, preserve units and avoid interpreting every printed digit as independently established knowledge. A display is an interface to arithmetic; it is not an explanation of accuracy.
CHAPTER 7 OF 22 · Understand finite computation
7. Equivalent formulas can preserve different amounts of information
Subtraction of nearly equal approximate quantities can expose errors already present in those quantities. The resulting small difference may have poor relative accuracy. A mathematically equivalent formula can sometimes avoid that loss.
Consider f(x) = √(x + 1) − √x for positive x. Multiply numerator and denominator by the sum of the square roots. The exact algebra gives f(x) = 1/[√(x + 1) + √x]. For large x, the first form subtracts nearby values; the second avoids that subtraction.
The formulas are equal in exact arithmetic, but their finite computational paths differ. This does not mean every subtraction is dangerous. Subtraction can be exact for suitable represented operands; the concern is how errors in approximate inputs affect the information in a small result.
Ask the learner to explain why the rationalised numerator is one, then predict which expression is preferable at large x. Finally change the domain: at x = 0 the expression remains defined, but negative values require a separate domain account if working over the reals. Good numerical reasoning keeps algebra, arithmetic and domain conditions together.
For a deliberately simplified arithmetic thought experiment, suppose an instrument rounds each square root to four decimal places before subtracting. At a sufficiently large x, the two rounded values can coincide, making their displayed difference zero. That zero describes the information lost in the intermediate representation; it does not prove the exact mathematical difference is zero. Evaluating the rationalised expression can preserve a useful small value because it avoids that particular subtraction. The choice of operation sequence has become part of the answer.
CHAPTER 8 OF 22 · Solve and approximate
8. Bisection trades speed for a visible guarantee
For a continuous real function with opposite signs at the endpoints of an interval, at least one root lies between them. Bisection evaluates the midpoint and retains a half-interval that preserves a valid sign-changing bracket, unless the midpoint is itself a root.
Each halving reduces the bracket width by a factor of two. After n halvings from initial width W, the width is W/2ⁿ. The midpoint of the resulting bracket is at most half that width from a root contained in the bracket, assuming the evaluations and continuity account support the argument.
This is a useful guarantee because it is visible. The method can be slower than methods using local slope information, but it does not need a derivative. Its reliability belongs to the maintained bracket and assumptions.
An important misconception is that opposite signs always prove a root. Continuity matters. A function with a discontinuity can change sign across a point without taking the value zero there. A numerical routine cannot repair a missing mathematical condition simply by halving the interval many times.
CHAPTER 9 OF 22 · Solve and approximate
9. Work through a bracket for the square root of two
Use f(x) = x² − 2 on [1, 2]. The function is continuous, f(1) is negative and f(2) is positive. The first midpoint is 1.5, with positive function value, so the bracket becomes [1, 1.5].
The next midpoint is 1.25, with negative function value, giving [1.25, 1.5]. Then 1.375 is negative, giving [1.375, 1.5]. The fourth midpoint is 1.4375, which is positive, leaving [1.375, 1.4375]. The width is now 0.0625, equal to 1/16.
The midpoint of that final bracket is 1.40625. Its guaranteed distance from the contained root is at most 0.03125 under the stated exact-evaluation model. The actual error is smaller, but the guarantee comes from the bracket, not from secretly consulting the answer first.
Change the requirement to a midpoint error no larger than 0.001. We need 1/2ⁿ⁺¹ ≤ 0.001 for the initial width one, so nine halvings suffice because 1/1024 is below 0.001. This example checks the meaning of the bound as well as the arithmetic.
CHAPTER 10 OF 22 · Solve and approximate
10. Newton’s method uses a local linear model
Newton’s method replaces the function locally by its tangent-line approximation. Starting from xₙ, the next iterate is xₙ₊₁ = xₙ − f(xₙ)/f′(xₙ), provided the derivative and expression are suitable. The method can converge very rapidly near an appropriate root under appropriate smoothness and starting conditions.
For f(x) = x² − 2, the update becomes xₙ₊₁ = (xₙ + 2/xₙ)/2. Starting at one gives 1.5, then approximately 1.4166667, then approximately 1.4142157. These are original hand-checkable approximations, not a claim about every Newton iteration.
A zero or very small derivative, an unsuitable initial point or a complicated function can lead to failure, cycling or convergence to an unintended root. The method’s local success should not become a global guarantee. A practical solver may combine faster steps with a maintained bracket or other safeguards.
Compare the explanation with bisection. One method’s strength is a preserved interval; the other’s is a local model that can give rapid improvement. Selecting between them requires understanding the function and the required reliability, rather than treating a faster-looking sequence as automatically superior.
CHAPTER 11 OF 22 · Solve and approximate
11. Residuals and stopping rules need interpretation
A residual measures how well an approximate answer satisfies the defining equation. For a root problem, a common residual is f(x). For a linear system Ax = b, the residual is b − Ax, subject to a chosen sign convention. A small residual is useful evidence but is not always a small solution error.
The connection depends on sensitivity. If a function is very flat near a root, a small function value can coexist with a substantial distance from the root. If a linear system is ill-conditioned, a small residual may coexist with meaningful uncertainty in the solution.
A stopping strategy can combine residual checks, step-size checks, a maintained bound where available and a maximum iteration count. The scales should match the problem. Reaching the maximum count should produce a clear status, not silently relabel the last iterate as a successful solution.
This transfers into software interfaces. “Completed” can mean the program stopped, whereas “converged under the specified criteria” means something more particular. A reliable report preserves the difference. The user should be able to tell whether the computation met its criteria, encountered a boundary or needs further investigation.
CHAPTER 12 OF 22 · Solve and approximate
12. Interpolation estimates within a represented relationship
Interpolation constructs a function that matches supplied data at selected points. Linear interpolation between two points uses a straight-line relationship between them. Higher-order polynomial or spline methods offer different ways to represent the interval.
Passing through every data point does not establish that the interpolating function represents the true process between them. The data may be noisy, the function may vary in a way the samples miss, and some high-degree polynomial choices can oscillate substantially. More complexity is not automatically a better model.
Extrapolation extends beyond the represented input range and can require much stronger assumptions. A smooth-looking curve on a chart does not make those assumptions true. The distinction matters in forecasting and reading scientific graphs.
The NIST Digital Library of Mathematical Functions interpolation chapter provides a formal reference for methods and error relationships. The classroom question is more immediate: what evidence supports the relationship between these points, and what changes outside them? A learner should be able to identify when the task is interpolation, when it becomes extrapolation, and which uncertainty the chosen curve leaves unresolved.
An invented interpolation case makes the boundary concrete. Suppose q(2) = 10 and q(6) = 18. Linear interpolation gives slope (18 − 10)/(6 − 2) = 2 and predicts q(4) = 14. The quadratic q(t) = 2t + 6 + (t − 2)(t − 6) matches the same endpoint values but gives q(4) = 10. The two data points alone therefore do not uniquely establish the middle value. Linear interpolation supplies a model-based estimate, not information logically forced by the endpoints. To choose between these relationships, we need additional data or a justified structural assumption about the process.
CHAPTER 13 OF 22 · Solve and approximate
13. Quadrature turns area into a controlled approximation
Numerical integration, often called quadrature, approximates an integral using values of the integrand. The trapezoidal rule joins adjacent values with straight segments and sums the resulting areas. Its accuracy depends on the function and how the interval is divided.
For the original exercise ∫₀¹ x² dx, the exact value is 1/3. With two equal subintervals, the trapezoidal approximation is 0.5[(0 + 1)/2 + 0.25] = 0.375. The absolute error is approximately 0.0416667. With four equal subintervals, the result is 0.34375, with error approximately 0.0104167.
The error decreases by a factor of four in this example when the step size halves. That illustrates the expected second-order behaviour for this smooth setting; it is not a guarantee for every integrand or every implementation.
The NIST quadrature chapter develops the formal landscape. The existing Gauss–Kronrod guide develops adaptive methods and estimated error. The practical habit is to compare refinements and inspect difficult features of the integrand, rather than assume a dense grid has captured everything important.
Consider another invented function that is zero at every chosen sampling node but has a narrow positive peak between two nodes. A rule using only those samples can return zero even though the integral is positive. Two grids that both miss the peak can agree reassuringly. This is why prior knowledge of sharp features, singularities or rapid variation matters when selecting and checking quadrature. Refinement is evidence only when it actually probes the relevant behaviour; matching outputs do not by themselves establish that the behaviour was seen.
CHAPTER 14 OF 22 · Check the whole computation
14. Differentiation balances approximation against roundoff
A forward difference approximates f′(x) by [f(x + h) − f(x)]/h. For a sufficiently smooth function, reducing h initially reduces the truncation error associated with replacing a limit by a finite step. But very small h can make subtraction and rounding increasingly important.
The two error tendencies can compete. Under a simple model, a forward-difference truncation contribution scales with h, while a roundoff-related contribution can scale with machine precision divided by h. The exact behaviour depends on the function, evaluation accuracy and arithmetic.
For f(x) = x² at x = 1, exact arithmetic gives [(1 + h)² − 1]/h = 2 + h. The approximation error is h before finite arithmetic effects. This small example makes the truncation part visible without pretending that a real computer evaluates every intermediate exactly.
The signal smoothing and numerical-derivative guide offers a connected application. When data are noisy, differentiation can amplify that noise, making the data problem important as well. A smaller step is therefore a question to investigate, not an instruction that always improves the result.
The square-function exercise also lets a learner compare specific steps. At h = 0.1, the exact-arithmetic forward difference is 2.1; at h = 0.01 it is 2.01. The truncation error falls from 0.1 to 0.01, just as the algebra predicts. Before making h extremely small, ask which part of this reasoning assumed exact function evaluations. That question identifies why a real machine can eventually depart from the simple trend. It also distinguishes an analytical prediction from an experiment that has actually been run.
CHAPTER 15 OF 22 · Check the whole computation
15. Differential equations connect local change with a whole path
A differential equation specifies a relationship involving a function and its derivatives. Numerical methods advance an approximate solution through steps, using the model and initial or boundary information. A sequence of locally reasonable steps still needs a global accuracy and stability account.
Euler’s method uses the current slope to make a forward step. For the teaching problem y′ = y with y(0) = 1, a step h = 0.5 gives y(0.5) ≈ 1.5 and y(1) ≈ 2.25. The exact value at one is e, approximately 2.71828. The discrepancy illustrates the accumulated approximation for this chosen step size.
Reducing step size can improve the approximation when the method and problem support convergence. Some problems impose restrictive stability requirements on explicit methods, and adaptive or implicit approaches may be appropriate. The correct choice depends on the equation’s behaviour, rather than on a universal preference for one solver.
This connects with simulations of flow, motion and changing populations. Keep the model’s validity separate from the solver’s performance. A converged numerical solution can accurately solve an equation that inadequately represents the actual system.
CHAPTER 16 OF 22 · Check the whole computation
16. Linear systems require attention to scale and dependence
Many computational problems lead to Ax = b: a matrix of coefficients, a vector of unknowns and a vector of specified values. Direct methods transform the system through a finite sequence of operations; iterative methods build successive approximations. Appropriate choices depend on size, structure and accuracy requirements.
Nearly dependent equations can make the solution sensitive. Consider x + y = 2 and x + 1.001y = 2.001. Subtracting gives 0.001y = 0.001, so x = y = 1. If the second right-hand side changes slightly to 2.002, the solution becomes y = 2 and x = 0.
The input change is small relative to the numbers near two, while the solution changes substantially. This original example exposes conditioning. It does not prove that every subtraction algorithm is unstable; even exact arithmetic sees the changed problem’s different solution.
The matrix-product verification guide develops a related checking question. For practical systems, scaling, pivoting and suitable library methods can improve computation, but cannot remove intrinsic sensitivity. A responsible answer carries both its residual and an appropriate account of the problem’s condition.
CHAPTER 17 OF 22 · Check the whole computation
17. Verification asks whether the computation solves the model
Verification examines whether the computational implementation correctly solves the specified mathematical problem to an appropriate accuracy. Validation examines whether the model adequately represents the intended real system for its use. These questions connect, but they need different evidence.
Begin verification with cases whose answers or structural properties are known. Check units, limiting cases, conservation relationships where applicable, and comparisons with independent methods. A simple manufactured case can reveal implementation errors that a visually convincing simulation conceals.
Refinement studies examine how results change as a step size or mesh is adjusted. Agreement across refinements can be informative, but it is not an unconditional proof. The tested quantity may miss a local feature, all runs may share the same error, or the refinements may not yet be in the expected regime.
The finite-element convergence guide and structural modelling application illustrate this distinction. A colourful output is a representation of a computation. Its trustworthiness depends on the model, implementation, data and checks supporting it. Reading those supports is part of numerical literacy.
CHAPTER 18 OF 22 · Check the whole computation
18. Professional computation preserves a reproducible account
A useful computational record includes the mathematical model, input data, algorithm, relevant settings, software environment and checks. The necessary detail scales with the consequence of the result. The purpose is to let another competent person understand what was solved and why the output deserves confidence.
Parallel execution, library choices or operation order can produce small numerical differences. Bitwise reproducibility means identical stored bits under a defined account. Numerical agreement means differences remain within a justified tolerance or do not alter the relevant conclusion. Teams should state which goal matters.
Automated output should preserve status information. A failed convergence check, singular system or exceeded range is evidence about the computation, not an inconvenience to hide. Returning a plausible-looking number without that status can make a failed calculation appear successful.
The roots Computing, Software and Optimisation connect the professional setting. Numerical analysis contributes a particular discipline: quantify approximation, challenge sensitivity and keep the evidence attached to the answer. It complements domain expertise rather than replacing the knowledge needed to decide whether the mathematical question was appropriate.
CHAPTER 19 OF 22 · Learn and explain
19. Practice with changed scales and assumptions
A narrower bracket
A valid root bracket has width 0.08. What error bound follows for its midpoint? At most 0.04 from a contained root under the maintained bracket assumptions. After one further valid halving, the midpoint bound is at most 0.02. State continuity and valid evaluations rather than presenting the bound without conditions.
A small residual
A root solver reports a tiny function value. Must the root position be equally accurate? No. A flat function can produce small residuals far from the root on the relevant scale. Sensitivity or a valid bound must connect residual with solution error.
More displayed digits
A spreadsheet prints fifteen digits from input data known only approximately. Has it gained information? No. It has displayed the numerical representation more fully. Input uncertainty and model limitations remain.
A refined simulation
Two fine meshes give similar outputs. Does that prove the model represents reality? No. It can provide evidence about numerical refinement for the tested quantity. Validation needs suitable comparison with the intended real system and purpose.
CHAPTER 20 OF 22 · Learn and explain
20. Learn by predicting failure before pressing run
A strong learning sequence starts with a small problem whose relationship can be inspected. Ask the learner for an estimate, a method and an expected check before obtaining the computed answer. That makes the program’s output something to evaluate, rather than the authority that defines correctness.
When a result is wrong, identify the first unstable connection. Is the function entered incorrectly? Is the domain wrong? Was a residual mistaken for an error bound? Is the problem sensitive? Did the method fail its assumptions? These require different repairs.
Use worked examples, partially completed cases and independent tasks. Then mix root finding, interpolation and sensitivity questions so that the learner must choose the method. A later task should change the scale or surface context without announcing the chapter name.
The roots Mathematics, Calculus and Education connect this progression. Progress is visible when the student can explain which result is reliable, which check is missing and what the next calculation should resolve. That is a stronger capability than producing a correct number only when the algorithm is already specified.
Is numerical analysis just programming?
No. Programming implements methods, while numerical analysis studies approximation, error, convergence, conditioning and computational reliability. Each informs the other.
Is an approximate answer necessarily inferior?
No. Many problems require approximation, and a controlled approximation can answer a practical question very well. The important issue is the error account and intended use.
Does a stable method fix an ill-conditioned problem?
It avoids unnecessary numerical damage. It cannot remove the problem’s inherent sensitivity to uncertain inputs.
Is Newton’s method always better than bisection?
No. Newton can be fast under suitable conditions; bisection offers a visible bracket-based guarantee. The best route depends on the function and reliability requirements.
Does smaller step size always improve a calculation?
No. It can reduce some approximation errors while increasing cost or exposing roundoff and data-noise effects. Convergence and stability need checking.
What is the everyday benefit?
The subject helps readers judge calculator outputs, spreadsheets, fitted curves and simulation claims. It teaches that a number’s usefulness depends on the relationship and checks supporting it.
CHAPTER 22 OF 22 · Learn and explain
22. Teaching guide: make reliability part of the answer
Start with a familiar calculation and ask what evidence would make its output believable. Establish an estimate, units and a simple check. Then show two algebraically equivalent computational routes, such as the square-root difference, and ask why their finite behaviour can differ.
Teach methods within clear boundaries. For bisection, use a continuous function and a valid bracket. For Newton, make the local linear model visible and show a case where its conditions become questionable. For quadrature, begin with a function whose exact integral can be checked before moving to unfamiliar functions.
Use a compact error record: representation, algorithm, stopping rule, conditioning, input or model. After correction, ask for a changed task that requires the learner to identify the issue independently. An explanation should name what changed and which guarantee remained available.
Finish with a computational handover: problem, method, result, check and limitation. Return to that structure in a spreadsheet or a scientific example. The learner has crossed the useful threshold when they stop asking only “What number did I get?” and begin asking “What justifies using this number for the question in front of me?”
For an application connecting calibration, preparation and measurement uncertainty, continue with The Gold Standard of Analytical Chemistry.
Numerical Analysis: connected reading directory
Choose the question closest to your own. These articles develop particular mechanisms, applications and learning tasks; the short descriptions explain why each belongs here.
Foundations and reliable arithmetic
| Read next | The question this article develops |
|---|---|
| Why Mathematics? | Numerical Stability, Conditioning, Cancellation and Reliable Computation | Distinguish conditioning, cancellation and algorithm stability. |
| Why Mathematics? | Scientific Calculators, Significant Figures, Numerical Methods and Rounding | Interpret calculator digits, significant figures and rounding. |
| Why Mathematics? | Spreadsheets, Cell References, Formulas, Floating-Point and Error Checking | Check spreadsheet formulas and numeric meaning. |
| How Lossy Works | Rounding — When Precision Is Cut to Fit | Understand what rounding removes. |
| The Gold Standard Of Mathematics | Place computation within mathematical reasoning. |
| The Gold Standard of Calculus | Connect derivatives, integrals and numerical approximations. |
Specialist numerical methods
| Read next | The question this article develops |
|---|---|
| Why Mathematics? | Gauss–Kronrod Quadrature, Nested Nodes and Adaptive Error Estimates | Explore adaptive quadrature and estimated error. |
| Why Mathematics? | Savitzky–Golay Filters, Local Polynomial Fits, Smoothing and Numerical Derivatives | Connect smoothing with numerical derivatives. |
| Why Mathematics? | Welford’s Algorithm, Running Means and Online Variance | Follow stable running-mean and variance calculations. |
| Why Mathematics? | Freivalds’ Algorithm, Random Vectors and Matrix-Product Verification | Examine matrix-product verification. |
| Why Mathematics? | CORDIC Rotations, Shift-and-Add Arithmetic, Sine and Cosine | Understand CORDIC rotations and shift-and-add computation. |
| Why Mathematics? | Vincenty Ellipsoidal Geodesics, Distance and Bearing Iteration | See iteration limits in ellipsoidal geodesics. |
| Why Mathematics? | RANSAC, Random Samples, Consensus Sets and Robust Model Fitting | Explore robust fitting and the role of model assumptions. |
Simulation and application checks
| Read next | The question this article develops |
|---|---|
| Why Mathematics? | Finite-Element Analysis, Mesh Convergence and Stress Concentrations | Distinguish mesh convergence from physical validation. |
| How Mathematics Improves The World | Testing a Structure Before Reality Has To | Follow mathematical modelling into structural questions. |
| How Mathematics Improves The World | Designing Airflow Before Building the Aircraft | Connect airflow models with computational approximation. |
| How Mathematics Improves The World | Predicting a Flood Before the River Reaches the Street | Explore the modelling chain in flood prediction. |
| How Mathematics Improves The World | Getting Clean Water to Every Tap at the Right Pressure | Connect equations and water-network behaviour. |
Professional and educational connections
| Read next | The question this article develops |
|---|---|
| The Gold Standard of Computing | Connect algorithms with computing systems. |
| The Gold Standard of Software | Preserve implementation quality and understandable status. |
| The Gold Standard of Optimisation | Follow numerical choices into optimisation. |
| The Gold Standard of Statistics | Distinguish data variation from numerical error. |
| The Gold Standard of Education | Learning That Transfers into Life | Check independent use after the task changes. |
| How X Works | How Things, Systems, Subjects and the World Work | eduKateSG | Explore the wider subject hub. |
Sources and further reference
Institutional references checked on 21 October 2026. The numerical exercises and fictional cases in this guide are original teaching examples, not measurements of actual products, people or sites.
MIT OpenCourseWare: Introduction to Numerical Analysis — Formal coverage of roots, interpolation, quadrature, equations and linear algebra.
MIT OpenCourseWare: Introduction to Numerical Methods — Accuracy, floating-point arithmetic, conditioning, stability and backward error.
NIST Digital Library of Mathematical Functions: Interpolation — Interpolation methods and formal remainder relationships.
NIST Digital Library of Mathematical Functions: Differentiation — Numerical derivative formulas and their mathematical context.
NIST Digital Library of Mathematical Functions: Quadrature — Numerical integration methods and error-related structure.
MIT: Euler’s Numerical Method — An institutional continuation into numerical differential equations.
Return to The Gold Standard of Education: learning that transfers into life or How X Works. Choose one relationship, explain it in your own words and test it after a relevant condition changes.