Singapore School Mathematics Operating Manual · Chapter 32
What we observe is not always the whole mathematical object.
A graph may be sampled at only a few points. A measurement may be recorded only to the nearest unit. A survey may include only part of a population. A time series may be observed every hour even though important changes occur every minute. Two different curves can pass through the same sampled points. A repeated signal can appear slower or simpler when observations are too sparse.
This chapter develops sampling, resolution and aliasing as an operating-manual layer. The goal is not advanced signal processing. It is to teach a school-level discipline: before trusting an observed pattern, ask what the observation process could have hidden.
1. A sample is a partial view
If a population contains 10,000 people and we survey 200, the sample is not the population itself.
The sample provides evidence about the population under assumptions about how it was selected.
A perfectly computed sample mean can still be unrepresentative if the sample is biased.
2. Sampling error is not the same as arithmetic error
A learner may calculate the sample statistic correctly.
The difference between that statistic and the population parameter can still arise because only part of the population was observed.
Correct arithmetic does not eliminate sampling uncertainty.
3. Selection mechanism matters
A voluntary-response survey can overrepresent people with strong opinions.
A sample taken only from one class may not represent the whole school.
A convenience sample may be easy to collect but systematically different from the target population.
Sampling quality depends on how observations enter the data set.
4. Resolution describes how finely a quantity is observed
A thermometer recorded to the nearest degree has coarser resolution than one recorded to the nearest 0.1 degree.
A graph sampled every second has finer time resolution than one sampled every minute.
Higher resolution can reveal variation hidden by coarse measurement.
5. More decimal places do not necessarily mean higher true resolution
If an instrument measures only to the nearest 1 cm, writing 12.000 cm does not create millimetre precision.
Displayed digits and measurement resolution are different things.
6. Rounding is a resolution-limiting operation
Values 7.24 and 7.26 may both become 7.3 to one decimal place under suitable rounding positions.
Different exact states can collapse into the same recorded value.
Inverse reconstruction from the rounded record therefore returns an interval, not one exact hidden value.
7. A table of points does not define the behaviour between them
Suppose a table gives (0,0), (1,1), (2,4), (3,9).
The pattern suggests y=x².
But many other functions can pass through those four points and behave differently elsewhere.
The sampled values support a model; they do not uniquely prove it without additional assumptions.
8. Sparse graph sampling can hide turning points
A curve may rise, fall and rise again between two widely separated sampled x-values.
If only the endpoints are observed, the internal turning points disappear.
A straight segment joining sampled points can therefore create a false impression of monotonic behaviour.
9. Sampling frequency should match the rate of change
If a quantity changes rapidly, observations must be frequent enough to capture the change.
Sampling once a day may be fine for a slowly changing variable but inadequate for a process changing every minute.
The observation schedule is part of the mathematical model.
10. Aliasing means different underlying patterns can produce the same samples
Imagine a repeating signal observed only at times when it happens to return to the same value.
The recorded samples may suggest the signal is constant even though it oscillates between observations.
Two different underlying behaviours have become indistinguishable at the chosen sampling times.
This is the core idea of aliasing.
11. A simple alternating example shows aliasing intuitively
Suppose a sequence alternates 0,1,0,1,0,1,…
If we observe every term, the alternation is obvious.
If we observe only positions 1,3,5,7,…, every recorded value is 0.
The sparse sample falsely suggests a constant sequence.
12. Sampling at the wrong phase can hide periodicity
A periodic process observed once every full period can appear unchanged.
Observed once every half-period, it may appear to alternate.
The same underlying cycle can produce very different sampled sequences depending on timing.
13. Resolution can affect threshold decisions
If a measured quantity is rounded coarsely near a threshold, the true value may lie on either side.
A higher-resolution measurement may be necessary to make a guaranteed decision.
Resolution matters most when the unresolved interval overlaps a critical boundary.
14. Histograms deliberately reduce resolution
Continuous observations are grouped into class intervals.
The exact individual values inside each class are no longer visible.
The histogram reveals distribution shape at class-level resolution, not the original raw data.
15. Wider histogram bins can hide structure
Two clusters may merge into one broad bar if the class interval is too wide.
Narrower bins may reveal more detail but can also produce a noisier appearance.
Representation resolution changes what patterns are visible.
16. Grouped-data means are approximations
When only class intervals and frequencies are known, midpoint values are often used to estimate a mean.
The original within-class positions are hidden.
The resulting mean is therefore an estimate unless every observation is known to equal its class midpoint.
17. Scatter plots are sampled relationships
A scatter plot shows observed pairs, not every possible pair in the population.
A line of best fit summarises the sampled pattern.
It does not imply every future point must lie on the line.
18. Correlation can depend on the observed range
A nonlinear relationship may look approximately linear over a narrow interval.
Over a wider range, curvature may become obvious.
Sampling only one local region can therefore change the apparent relationship.
19. Outliers can be hidden by aggregation
Averages and grouped summaries can conceal extreme individual values.
A mean of 50 does not tell us whether every observation lies near 50 or whether values are widely spread.
Resolution in the data summary matters.
20. Averages can hide oscillation
Values 0,100,0,100 have mean 50.
A constant sequence 50,50,50,50 also has mean 50.
The same summary statistic can represent radically different underlying patterns.
Aggregation is a form of information compression.
21. Sampling and normalisation solve different problems
Normalisation changes scale for comparison.
Sampling chooses which observations are present.
A perfectly normalised biased sample remains biased.
Changing representation cannot recover observations that were never collected.
22. Interpolation assumes something about hidden behaviour
If values are known at x=0 and x=10, estimating at x=5 requires a model for what happens between them.
Linear interpolation assumes straight-line behaviour between the sampled points.
If the true relationship is curved, the interpolation may be wrong.
23. Extrapolation extends the hidden-behaviour assumption farther
Predicting beyond the observed range assumes the pattern continues.
The farther the extrapolation, the more strongly the conclusion depends on model validity.
24. Numerical differentiation is sensitive to sampling spacing
A gradient estimated from two sampled points is an average rate over that interval.
If the curve bends strongly inside the interval, the estimate may differ from the instantaneous derivative.
Closer samples can improve local resolution, subject to measurement noise.
25. Numerical integration also depends on partition resolution
Approximating area under a curve with rectangles becomes more accurate when the partition is refined for a well-behaved function.
Coarse rectangles can miss curvature.
Resolution controls approximation quality.
26. A digital image analogy makes resolution intuitive
A low-resolution image stores fewer sample points.
Fine details may disappear or jagged edges may appear.
A mathematical data set behaves similarly: sparse observations cannot preserve unlimited detail about the underlying object.
27. More samples do not automatically remove bias
Polling 10,000 people from one unrepresentative subgroup can be worse than a smaller carefully selected sample.
Sample size reduces some uncertainty, but selection quality remains fundamental.
28. Random sampling aims to control selection bias
Randomisation gives members of the target population a defined chance of selection.
This supports statistical inference under appropriate conditions.
Random does not mean careless; it means governed by a probability mechanism.
29. Stratification can preserve important subgroups
If a population contains distinct groups, stratified sampling can ensure each is represented.
The design deliberately increases resolution across subgroup structure.
30. Repeated measurements can distinguish noise from signal
If one measurement fluctuates, repeated observations can reveal whether variation is random, periodic or systematic.
One sample point rarely reveals the full mechanism.
31. Missing data can create hidden bias
If certain observations are systematically absent, the visible sample can be distorted.
“No data” should not automatically be interpreted as zero.
Absence of observation and observed absence are different states.
32. Sampling can create false certainty in inverse problems
If only a few outputs are observed, many hidden models may fit them.
The inverse reconstruction can appear unique only because alternative behaviour was never sampled.
This links directly to Inverse Problems and Reconstruction.
33. Resolution and conditioning interact
If a problem is highly sensitive, coarse measurements can produce large uncertainty in the answer.
A stable problem may tolerate coarse resolution; an ill-conditioned problem may require much finer input precision.
34. Sampling should be designed around the question
If the target is average level, one sampling plan may suffice.
If the target is maximum peak, much finer resolution may be required.
If the target is periodicity, sampling times must be chosen to reveal cycles rather than accidentally hide them.
35. A practical sampling audit
Ask:
What is the underlying object? Which observations were actually collected? How were they selected? What resolution was used? What could happen between samples? Could a different underlying pattern produce the same observations? Is aggregation hiding spread, peaks or subgroups? Does the decision require finer resolution?
36. Independent practice
1. A sequence alternates 0,1,0,1,… What is observed if only odd-indexed terms are sampled?
2. Why can three points on a graph fail to prove the whole relationship is linear?
3. A measurement is recorded to nearest 1 cm. Why does writing 12.000 cm not increase resolution?
4. Explain how a wide histogram bin can hide two clusters.
5. Why is a grouped-data mean usually an estimate?
6. Two data sets have the same mean. Must they have the same spread?
7. Why can a very large biased sample remain misleading?
8. What is aliasing in plain language?
9. Why can sampling once per period hide a periodic signal?
10. When is higher measurement resolution especially valuable?
37. Worked answers
1. Every observed term is 0, so the alternation is hidden.
2. Many different functions can pass through the same finite points while differing between or beyond them.
3. The instrument supplied only centimetre-level information; extra written zeros do not create finer measurement evidence.
4. Observations from separate clusters may fall into the same broad interval and merge visually.
5. Midpoints stand in for unknown exact values within each class.
6. No. For example 50,50,50 and 0,50,100 share mean 50 but have different spread.
7. Increasing the number of observations does not fix systematic selection bias.
8. Different underlying patterns look the same because observations are too sparse or timed poorly.
9. Each observation can occur at the same phase, making the recorded value appear constant.
10. Near thresholds, rapid changes, narrow peaks, or sensitive inverse problems where small input uncertainty matters.
38. Batch 08 as one problem-architecture layer
Decomposition and Recomposition controls how a large problem is split and rebuilt.
Dependency Order and Solving Sequence controls when each module should run.
Normalisation and Rescaling controls how quantities are placed on comparable scales.
This chapter controls how much of the underlying mathematical object the observations actually reveal.
Return to the BTT Mathematics Hub for Batch 08.