Why Measurement Precision Matters
Tables of values can suggest what a function approaches near an input. But a recorded value is not always the exact value: it may have been rounded, or it may include measurement error. If the uncertainty is large compared with the changes in the readings, a seemingly clear pattern may not determine a reliable numerical estimate.
The key is to interpret each reported value as information about a range of possible values. Then ask whether those ranges are narrow enough to support the estimate. Even a very consistent pattern in finitely many measurements is evidence, not a proof of a limit: a limit concerns function values at every sufficiently nearby input, not just the inputs that happened to be measured.
What a Rounded Reading Tells You
Suppose a value is rounded to the nearest multiple of \(q\), where \(q>0\) is the reporting increment. A displayed reading \(r\) means that the unrounded value is within \(q/2\) of \(r\). Thus its possible values lie in the interval from \(r-q/2\) to \(r+q/2\). Whether an endpoint is included depends on the rule used for ties; using a closed interval is a safe way to record the possible range.
For example, a reading of \(2.37\) to the nearest hundredth has \(q=0.01\), so \(q/2=0.005\). Its rounding interval is \([2.365,2.375]\). The displayed digits do not say that the true value is exactly \(2.370\).
Measurements may also have a stated error bound. If a reading \(r\) is known to differ from the true value by at most \(e\), its possible values lie in \([r-e,r+e]\). This is the same interval reasoning, with the uncertainty specified directly rather than arising from rounding.
Reading a Trend Without Overclaiming
To assess a numerical limit estimate, first note how the inputs approach the target: from one side or from both sides, and how close they actually are. Then record the uncertainty in each output. A pattern is more persuasive when readings at progressively nearer inputs remain compatible with a common candidate limit and when the uncertainty is small relative to the observed changes.
This is still an assessment of evidence, not a proof. In the epsilon-delta definition, a limit \(L\) at \(a\) requires that for every \(\varepsilon>0\), there is a \(\delta>0\) such that \(0<|x-a|<\delta\) implies \(|f(x)-L|<\varepsilon\). A finite set of measurements does not check every input in any whole neighborhood of \(a\). Rounded data can support or weaken an estimate, but cannot alone establish that universal condition.
Worked Example: A Trend Consistent With a Limit of 5
Worked Example: A Trend Consistent With a Limit of 5
Suppose the function is sampled near \(x=1\). Each output is reported to the nearest hundredth. The readings are \(4.90\) at \(x=0.90\), \(4.99\) at \(x=0.99\), \(5.01\) at \(x=1.01\), and \(5.10\) at \(x=1.10\). Assess the estimate \(L=5\).
For each output, \(q=0.01\), so the uncertainty is \(q/2=0.005\). The readings correspond to these intervals:
The two closest inputs are \(0.99\) and \(1.01\), each at distance \(0.01\) from \(1\). Their output intervals lie within \(0.015\) of \(5\): the first interval runs from \(4.985\) to \(4.995\), and the second from \(5.005\) to \(5.015\). Thus both nearby readings are consistent with outputs close to \(5\), on opposite sides of the target input.
The pattern also has an exact-function witness: \(f(x)=4+x\) gives \(f(0.90)=4.90\), \(f(0.99)=4.99\), \(f(1.01)=5.01\), and \(f(1.10)=5.10\). Moreover, \(f(x)-5=x-1\), so \(|f(x)-5|=|x-1|\), which tends to zero as \(x\) approaches \(1\). This verifies that the readings fit a function with limit \(5\). It does not show that the measured function must be this function, or prove its limit from the finite readings alone.
Check: The calculated outputs agree with all four displayed readings, and each lies in its corresponding rounding interval. The estimate \(5\) is well supported by this pattern, but remains an estimate unless further information about the function is available.
When the Available Precision Is Too Coarse
A reading can be compatible with several substantially different true values. If the uncertainty intervals for different candidate values overlap the same reported readings, those readings cannot distinguish the candidates. In particular, several rounded zeros do not necessarily mean that the underlying values are zero or that a limit is zero.
Worked Example: Rounded Zeros Do Not Identify the Limit
Suppose measurements at \(x=-0.10,-0.01,0.01,0.10\) are all reported as \(0\) to the nearest whole unit. The increment is \(q=1\), so each reading has the interval \([-0.5,0.5]\).
Consider two possible functions, \(f(x)=0.4\) and \(g(x)=-0.4\). At each of the four inputs, \(f(x)=0.4\), and \(g(x)=-0.4\). Both outputs lie in \([-0.5,0.5]\), so both functions produce the displayed reading \(0\) at every sampled input.
But their limits at \(0\) differ. For every \(x\), \(f(x)=0.4\), so \(\lim_{x\to 0}f(x)=0.4\). Likewise, \(g(x)=-0.4\) for every \(x\), so \(\lim_{x\to 0}g(x)=-0.4\). The rounded measurements are identical, although the limits differ by \(0.8\).
Check: Both constants are within \(0.5\) of the reported zero, and their difference is \(0.4-(-0.4)=0.8\). Therefore the reported values do not determine which limit is present. Finer precision or additional information is needed.
Worked Example: Noise Can Hide the Direction of Change
Worked Example: Noise Can Hide the Direction of Change
Suppose the target input is \(0\). Measurements at \(x=0.01\) and \(x=0.001\) are \(1.03\) and \(1.01\), respectively. Each measurement has an error of at most \(0.02\). Decide whether these readings establish that the output decreases as \(x\) gets closer to zero.
The possible true output at \(x=0.01\) lies in \([1.03-0.02,1.03+0.02]=[1.01,1.05]\). At \(x=0.001\), it lies in \([1.01-0.02,1.01+0.02]=[0.99,1.03]\).
Let \(y_1\) be the true output at \(0.01\) and \(y_2\) the true output at \(0.001\). The change is \(y_2-y_1\). Its smallest possible value is \(0.99-1.05=-0.06\), and its largest is \(1.03-1.01=0.02\). Thus \(y_2-y_1\) could be negative, zero, or positive. The displayed readings decrease by \(1.01-1.03=-0.02\), but the error bounds are large enough to conceal the true direction of change.
Check: The intervals overlap on \([1.01,1.03]\), and the possible change interval \([-0.06,0.02]\) contains zero. Therefore the readings do not establish a decrease. They also do not determine a limit; they provide only two uncertain measurements.
Why the Interval Test Works
The interval method follows directly from what rounding and error bounds mean. If a reported value \(r\) is within \(e\) of the unknown value \(y\), then \(|y-r|\le e\). By the definition of absolute value, this is equivalent to \(-e\le y-r\le e\). Adding \(r\) throughout gives \(r-e\le y\le r+e\). Every value outside this interval would violate the stated error bound.
This gives a practical test of whether an estimate is resolved by the measurements. To check a candidate \(L\), compare each nearby output interval with the distance from that interval to \(L\). If the entire interval is close to \(L\), that reading supports the estimate at the given precision. If it extends far from \(L\), the reading is not precise enough to confirm closeness. If intervals accommodate very different candidate values, the measurements cannot distinguish those candidates.
There is a related uncertainty in the input if \(x\) itself has been rounded. A recorded input of \(1.0\) to the nearest tenth represents an actual input in the interval \([0.95,1.05]\), subject to the tie rule. It does not specify the exact distance between the true input and \(1\). When the question concerns behavior extremely close to a target, uncertainty in the input can therefore matter as much as uncertainty in the output.
A Common Mistake
A frequent error is to treat the displayed digits as exact and to report their difference as the exact change. For example, readings of \(1.2\) and \(1.2\) to the nearest tenth do not prove that the two true values are equal. Each has uncertainty \(0.1/2=0.05\), so each lies in \([1.15,1.25]\). The true values could be \(1.16\) and \(1.24\), a difference of \(0.08\), even though both readings display as \(1.2\).
The correction is to keep the possible-value intervals in view. State what the readings support at their stated precision; do not turn rounded measurements into exact values or turn a finite pattern into a proof of a limit.
Using Measurement Evidence Carefully
When judging a numerical limit estimate from rounded or noisy measurements, keep three questions separate:
- Are the inputs close to the target? Check their distances from the target and account for any rounding of the inputs.
- How uncertain are the outputs? Convert the rounding increment or error bound into an interval for each true value.
- What conclusion is justified? Say whether the intervals are consistent with the estimate, whether the precision distinguishes competing values, and whether the evidence is only suggestive.
A stable observed trend can make a numerical estimate reasonable, especially when measurements close to the target are precise enough to cluster near it. But measurements always have limited scope and resolution. A defensible conclusion describes both the estimate and the uncertainty that remains.
Check Your Understanding
For each question, use intervals to distinguish what the reported measurements show from what they leave unresolved.
- A value is reported as \(3.6\) to the nearest tenth. What interval of true values is consistent with that reading?
- Two readings are both \(2.0\) to the nearest tenth. Does this establish that the underlying values are equal? Explain using their intervals.
- Readings near a target input appear to approach \(4\), but their uncertainty intervals extend from \(3\) to \(5\). Is \(4\) determined to high precision by these readings? Why or why not?
- Why can a finite collection of exact function values support a limit estimate without proving the limit?
- A recorded input is \(0.0\) to the nearest tenth. What is the approximate range of possible true inputs, and why might that matter when studying a limit at zero?
