A ratio can have a perfectly ordinary point estimate and a confidence set that reaches without limit. This can happen when the denominator is not well distinguished from zero. Forcing the result into a neat finite interval can hide the very uncertainty the analysis is meant to describe.
The short answer is to test candidate ratios through a difference. For a proposed ratio r, examine whether the estimated numerator minus r times the estimated denominator is compatible with zero under a specified statistical model. Inverting that check produces an inequality in r. Depending on the evidence, its solution can be a bounded interval, two separated unbounded parts, or the whole real line.
We will derive that result using small, original examples. The inputs are invented, dimensionless estimates, not measurements of a named system. We use a deliberately explicit jointly normal model with known covariance so that the arithmetic and its coverage interpretation can be checked separately. Actual data require a suitable method and justified uncertainty model.
The Uncertainty Quantification subject guide develops uncertainty budgets, dependence and nonlinear transformations. This branch asks a more specific question: what shape should uncertainty about a ratio have when division by a near-zero quantity is genuinely unresolved?
eduKateSG · Uncertainty Quantification
Find your next learning step
Choose the route that matches the question you are trying to answer, or follow the chapters in order. Each worked example states its assumptions before asking you to apply the idea in a changed setting.
ROUTE 3 · CHAPTERS 9–12
Test covariance and approximation
Compare joint information and local approximations.
ROUTE 5 · CHAPTERS 17–20
Check and connect
Diagnose errors, preserve evidence and follow further reading.
Full chapter contents · Uncertainty Quantification subject guide
Chapter contents
Define the model · Chapters 1–4
1. Define the ratio of parameters, not a mean of ratios
2. State the statistical model that gives the check meaning
Read the confidence-set shape · Chapters 5–8
5. Work a bounded example completely
6. Read the leading coefficient as a denominator check
Test covariance and approximation · Chapters 9–12
9. Work a case in which no finite ratio is excluded
10. Handle the boundary without dividing by zero
Change the context · Chapters 13–16
13. Distinguish a confidence set from division of hard bounds
14. Change units without changing the underlying conclusion
15. Keep a zero observed denominator separate from a zero parameter
16. Solve a new bounded case without the familiar coefficients
Check and connect · Chapters 17–20
17. Diagnose the first incorrect step in an explanation
18. Write an evidence record another reader can reconstruct
CHAPTER 1 OF 20 · Define the model
1. Define the ratio of parameters, not a mean of ratios
Let μX and μY be two fixed but unknown quantities. The target is R = μY/μX, with μX assumed nonzero so the target is defined. Let x and y be estimates of those quantities. The familiar point estimate is y/x when the observed x is nonzero.
This target is not automatically the expected value of a random quotient Y/X. Nor is it the average of individual ratios from paired observations. Those are different quantities, and their existence or behaviour can require different assumptions. Name the target before selecting an interval method.
For a concrete teaching picture, imagine two estimated signed changes relative to reference conditions. A signed change can be positive, negative or near zero. Asking how one change compares with the other can therefore create a ratio whose denominator is not safely separated from zero. The mathematical issue is visible without attaching the example to a medical, commercial or engineering decision.
Write the statement card as “the ratio of these two specified parameters, in this order.” Reversing numerator and denominator changes the target and the confidence set. It is not just a different label for the same number. In the worked cases below, x always estimates the denominator and y always estimates the numerator. That convention keeps the signs and variance terms traceable as the conditions change.
CHAPTER 2 OF 20 · Define the model
2. State the statistical model that gives the check meaning
Our teaching model assumes the estimator pair (x, y), across repetitions, is jointly normal with mean (μX, μY). Its known covariance matrix has entries vX for the variance of x, vY for the variance of y, and c for their covariance. Assume the matrix is positive definite, avoiding zero-variance degeneracies.
These are variances of the estimators being used, not automatically the variances of individual raw measurements. If x and y are sample means, their uncertainty inputs must describe those means. Substituting a raw-data standard deviation without the required conversion would change the procedure.
Choose a positive critical value k according to the intended normal-model coverage. We use k = 2 for simple arithmetic. In this exact known-covariance normal model, the corresponding two-sided coverage is about 95.45%, not exactly 95%. A desired 95% normal procedure would use its appropriate critical value instead.
The model is intentionally restrictive. With estimated covariance, unknown scale or a different sampling distribution, the reference distribution and critical value need their own justification. A familiar multiplier cannot create coverage merely by being inserted into an equation. Separating this statistical contract from the algebra lets us see precisely which parts are conditional on normality and which parts are ordinary manipulation of a supplied inequality.
CHAPTER 3 OF 20 · Define the model
3. Replace the quotient question with a difference
If a proposed value r equals the true ratio R, then μY − rμX = 0. That suggests inspecting the estimated difference y − rx. Its variance under the stated covariance model is vY − 2rc + r²vX.
The covariance sign comes from the minus sign in the difference. Positive covariance can reduce the variance of some such differences. It is not enough to list two marginal standard errors and quietly assume their dependence has no effect.
For a fixed candidate r, standardise the difference by its standard deviation. Under the model, if r is the true ratio, Z(r) = (y − rx)/√(vY − 2rc + r²vX) has a standard normal distribution. The check retains r when |Z(r)| ≤ k.
Notice the direction of reasoning. We are not treating r as a random value drawn from a posterior distribution. We are constructing a set of parameter values not rejected by a specified family of checks. The confidence property concerns repeated use of that procedure under its assumptions.
This difference-based route is useful because it avoids beginning with a local approximation to y/x. The denominator uncertainty remains present in the variance of the candidate-specific difference. As r becomes large, the r²vX term also becomes large. That behaviour is central to understanding why the final accepted set need not have finite endpoints.
CHAPTER 4 OF 20 · Define the model
4. Turn the candidate check into a quadratic inequality
The variance in our positive-definite model is positive for every finite r. Squaring the check therefore gives (y − rx)² ≤ k²(vY − 2rc + r²vX). Expand both sides and move everything to the left.
The result is Ar² + Br + C ≤ 0, where A = x² − k²vX, B = −2xy + 2k²c, and C = y² − k²vY. These coefficient letters describe the inequality; they are not new observations.
This is a confidence-set inversion of the kind associated with Fieller’s method. The primary paper A Geometric Approach to Confidence Sets for Ratios explains the relation between linear-combination checks and ratio confidence sets, including unbounded cases. Our known-covariance normal model and numerical cases below are explicitly supplied teaching constructions.
Do not solve the equality alone and stop. Its roots are boundaries, while the inequality decides which regions belong to the set. An upward-opening quadratic can be nonpositive between its roots. A downward-opening quadratic can be nonpositive outside them. Reporting the same middle interval in both cases loses the essential distinction.
Before applying the quadratic formula, record A’s sign and the discriminant B² − 4AC. Then test one value in each relevant region. A short sign check is more dependable than assuming every confidence result must be the interval between two printed numbers.
CHAPTER 5 OF 20 · Read the confidence-set shape
5. Work a bounded example completely
Let x = 4 and y = 8. Set vX = vY = 1, c = 0 and k = 2. The point ratio is two. The inequality is (8 − 4r)² ≤ 4(1 + r²), which simplifies to 12r² − 64r + 60 ≤ 0.
Divide by four to obtain 3r² − 16r + 15 ≤ 0. Its roots are (8 − √19)/3 and (8 + √19)/3, approximately 1.214 and 4.120. Since the leading coefficient is positive, the accepted set lies between the roots, including the endpoints.
Check r = 2 directly. The estimated difference is zero, so the candidate passes. Check r = 0: the squared difference is sixty-four, greater than the allowance four, so zero fails. Check r = 5: the squared difference is 144 while the allowance is 104, so five fails too.
Those checks make the interval’s orientation visible. The set is not symmetric around the point estimate two. Its longer upper side reflects the nonlinear ratio relationship rather than an error in the root calculation.
A useful written result names both the set and the procedure: under the stipulated known-covariance normal model and k = 2 check, the ratio confidence set is approximately [1.214, 4.120]. Without that model statement, the numerical endpoints would look more generally justified than they are.
CHAPTER 6 OF 20 · Read the confidence-set shape
6. Read the leading coefficient as a denominator check
The coefficient A = x² − k²vX compares the squared denominator estimate with its uncertainty allowance. It is positive exactly when |x| > k√vX. In our bounded example, four is larger than two, so the corresponding marginal denominator interval x ± k√vX excludes zero.
That observation explains why the denominator is important, but do not replace the entire calculation with a slogan. The numerator and covariance still determine the roots and the shape within the relevant case. Knowing that A is positive does not supply the interval endpoints.
Change x from four to one while leaving its variance at one and k at two. Now A is negative. The denominator’s uncertainty scale is large relative to its estimate, and the marginal interval reaches across zero. Large positive or negative ratios may remain compatible with the candidate checks.
An unbounded set does not prove that the true denominator is zero. Our target still assumes a nonzero μX. It says that the available evidence and chosen procedure do not exclude denominator values close enough to zero to rule out arbitrarily large ratio magnitudes.
This is an important distinction between a parameter being undefined and a confidence set being unbounded. The first concerns the target’s actual mathematical definition. The second concerns what this evidence can resolve about a defined target. Replacing either statement with “division failed” loses useful information about the reason for the result.
CHAPTER 7 OF 20 · Read the confidence-set shape
7. Work the two-part unbounded case
Use x = 1, y = 3, vX = vY = 1, c = 0 and k = 2. The candidate inequality becomes (3 − r)² ≤ 4(1 + r²), or −3r² − 6r + 5 ≤ 0.
The roots are −1 − (2/3)√6 and −1 + (2/3)√6, approximately −2.633 and 0.633. The quadratic opens downward, so the accepted set is r ≤ −2.633 or r ≥ 0.633. It consists of two separated unbounded parts.
Check the gap. At r = 0, the squared difference is nine while the allowance is four, so zero fails. At r = 3, the difference is zero and the candidate passes. At r = 100, the squared difference is 9,409 while the allowance is 40,004, so that large candidate also passes. Large negative candidates pass as well.
| Candidate ratio r | Squared difference | Allowance | Result |
|---|---|---|---|
| 0 | 9 | 4 | Excluded |
| 3 | 0 | 40 | Retained |
| 100 | 9,409 | 40,004 | Retained |
The two roots do not mean that the interval between them is accepted. That middle region is precisely the excluded gap. A graphing or spreadsheet routine that automatically fills between the roots would misrepresent the confidence set while displaying the correct numerical boundaries.
CHAPTER 8 OF 20 · Read the confidence-set shape
8. Interpret the negative and positive parts honestly
The point estimate in the last example is positive three, yet the confidence set includes sufficiently negative ratios. That is possible because the denominator’s sign is not securely resolved by the stipulated evidence. A positive numerator divided by a small negative denominator can produce a large negative ratio.
Do not remove the negative part merely because it looks surprising. If the parameter is allowed to be negative under the model, the confidence set must preserve that possibility. The sign uncertainty is part of the result, not a cosmetic defect.
Now suppose independent subject knowledge established beforehand that the target ratio must be positive. Intersecting the confidence set with that parameter domain leaves r ≥ approximately 0.633. The upper end remains unbounded. A domain restriction can remove inadmissible values without creating a finite upper limit that the evidence does not support.
The restriction must come from the actual parameter definition or justified prior knowledge about its domain. It should not be invented after seeing an inconvenient result. A signed change, for example, cannot be declared positive merely to simplify its quotient.
This chapter tests interpretation rather than another quadratic calculation. Ask a learner to distinguish three sentences: the point ratio is positive; the confidence procedure does not resolve its sign; the scientifically permitted domain is positive. Those sentences concern different information. Keeping them separate prevents a central estimate from quietly becoming an unsupported constraint on the entire uncertainty statement.
CHAPTER 9 OF 20 · Test covariance and approximation
9. Work a case in which no finite ratio is excluded
Let x = 1 and y = 1, with the same independent unit variances and k = 2. The inequality is (1 − r)² ≤ 4(1 + r²), or −3r² − 2r − 3 ≤ 0.
The discriminant is four minus thirty-six, which is negative. Because the quadratic opens downward, it stays below zero for all real r. The confidence set is the whole real line.
This result does not say that all ratios are equally plausible, nor that the ratio has a uniform probability distribution over the real numbers. It says that this particular inversion does not exclude any finite candidate at the chosen threshold under the supplied model.
The point estimate is still one. Reporting “one with a small-looking standard error” would therefore omit the main limitation the exact candidate checks reveal. A central estimate can exist while the available information provides very little resolution about the ratio.
For an independent check, test r = 0, r = 1 and r = 1,000 in the original inequality. All pass. Completing the square gives another explanation: −3r² − 2r − 3 = −3(r + 1/3)² − 8/3, always negative. This algebraic check removes any ambiguity about how a negative discriminant should be interpreted when the leading coefficient is also negative.
CHAPTER 10 OF 20 · Test covariance and approximation
10. Handle the boundary without dividing by zero
Set x = 2 and y = 3, retaining vX = vY = 1, c = 0 and k = 2. Now A = 2² − 4 = 0. The quadratic term disappears, leaving −12r + 5 ≤ 0.
The solution is r ≥ 5/12, approximately 0.417. It is a one-sided unbounded set at this exact boundary. A general quadratic-root expression that divides by 2A would be undefined here, even though the original inequality has a simple meaningful solution.
This is a good lesson in solving the equation actually present. Check for a zero leading coefficient before using a formula that assumes it is nonzero. If the remaining linear coefficient also vanished, the constant term would need separate examination.
In a continuous statistical model, observing an exact algebraic boundary may have probability zero. It still matters when checking a computational implementation, rounded teaching data or symbolic derivation. A method should not convert an interpretable limiting case into an unexplained software failure.
Use the original inequality as a final check. At r = 5/12, both sides agree. At r = 0, the candidate fails; at r = 1, it passes. The accepted direction is therefore clear without relying on a graph. State that this is the exact stipulated boundary case rather than a claim about how often such a result occurs in practice.
CHAPTER 11 OF 20 · Test covariance and approximation
11. Put covariance back into the calculation
Return to x = 4 and y = 8 with unit variances, but now set covariance c = 0.5. This is a valid positive-definite covariance matrix. The variance of y − rx becomes 1 − r + r², and the candidate inequality simplifies to 12r² − 60r + 60 ≤ 0.
The roots are (5 − √5)/2 and (5 + √5)/2, approximately 1.382 and 3.618. The accepted set remains bounded but differs from the independent case’s approximately [1.214, 4.120]. Two unchanged marginal variances did not preserve the ratio confidence set.
The reason is visible in the candidate difference. For positive r, positive covariance reduces its variance relative to independence. This does not justify a general statement that positive correlation always reduces uncertainty for every output. Its effect depends on the signs and structure of the model.
A practical uncertainty record must therefore include how covariance was obtained or justified. If it is unknown, a convenient choice of zero is an additional assumption. A sensitivity exercise can compare plausible covariance models, but its results should be labelled as such rather than treated as measured dependence.
This example also helps diagnose a common coding error: a covariance term entered with the wrong sign can produce plausible-looking finite endpoints. Reconstruct the variance of y − rx first, then expand the inequality. The original candidate check provides an independent way to verify the final coefficients.
CHAPTER 12 OF 20 · Test covariance and approximation
12. Compare with a local symmetric approximation
For the ratio function y/x near the bounded example’s centre, the derivatives with respect to y and x are 1/x and −y/x². With x = 4, y = 8 and independent unit estimator variances, first-order propagation gives approximate variance 1/16 + 64/256 = 5/16.
Its standard deviation estimate is √5/4, approximately 0.559. Multiplying by two and placing the allowance symmetrically around the point ratio two gives approximately [0.882, 3.118]. That differs from the candidate-inversion interval [1.214, 4.120]. Neither its symmetry nor its extra lower-side reach establishes that it is more informative.
The local approximation treats the ratio through its derivatives near the estimated point. Candidate inversion asks a different question for every r while retaining the denominator’s uncertainty in the difference variance. The distinction becomes particularly consequential near a denominator that is not separated from zero.
Do not turn this comparison into a universal claim that one interval must contain the other. Our bounded example already shows that their endpoints can cross. The method, assumptions and coverage properties must be evaluated on their actual terms.
A finite symmetric formula can be useful in an appropriate regime. The mistake is assuming that its finite output proves the ratio has a well-resolved finite confidence interval, regardless of the denominator evidence. A numerical format cannot supply information that the underlying model lacks.
CHAPTER 13 OF 20 · Change the context
13. Distinguish a confidence set from division of hard bounds
Suppose a different problem supplies guaranteed hard bounds X between two and six and Y between six and ten, with X strictly positive. Ordinary interval arithmetic gives Y/X between one and five. That is a deterministic consequence of those bounds and their allowed joint combinations.
It is not the same procedure as our normal-model ratio confidence set. Taking separate marginal confidence endpoints and dividing them as if they were simultaneous hard bounds silently changes their interpretation. Their joint coverage and dependence would need examination before any probabilistic claim about the resulting rectangle could be made.
The geometric interpretation of Fieller’s method is helpful, but its contour should also retain its stated calibration. A contour built to support projected one-dimensional candidate checks is not automatically a two-dimensional joint confidence region with the same percentage. Degrees of freedom and coverage calibration belong to the statistical construction, not merely to the shape drawn on the page.
For a learner, keep two worksheets separate. One begins with hard input bounds and computes all possible quotients under those bounds. The other begins with estimator distributions and inverts a calibrated candidate check. Both use algebra. Their guarantees differ because the starting information differs. If the explanation swaps the words bound and confidence without changing the assumptions, pause and repair that interpretation before doing more arithmetic.
CHAPTER 14 OF 20 · Change the context
14. Change units without changing the underlying conclusion
Imagine the numerator is re-expressed in units one hundred times smaller, so its numerical estimate y becomes one hundred times larger. Its variance becomes ten thousand times larger, and its covariance with x becomes one hundred times larger. The numerical ratio and every accepted ratio value should scale by one hundred.
Changing only y while leaving its uncertainty inputs untouched would describe a different model. This is why a unit conversion must travel through the covariance matrix as well as the central estimates. The confidence set is a relationship among all of those quantities.
For an easier common-scale check, multiply both x and y by ten and multiply every covariance-matrix entry by one hundred. The ratio target is unchanged. Every coefficient A, B and C is multiplied by the same positive factor one hundred, leaving the solution set of the inequality unchanged.
This gives a strong implementation check with no new statistical assumptions. A program that returns a different dimensionless confidence set after a common unit rescaling has likely mishandled a variance, covariance or coefficient. The mathematical structure supplies the expected invariance.
State units even when an exercise uses dimensionless numbers. Real ratios can have units, such as one measured change per unit of another. The sign, scale and domain of the target must remain consistent from the measurement record to the final confidence-set statement.
CHAPTER 15 OF 20 · Change the context
15. Keep a zero observed denominator separate from a zero parameter
If the observed denominator estimate x equals zero, the plug-in point estimate y/x is undefined. That does not automatically prevent the candidate-difference procedure from being evaluated. Its inputs are y − rx and the variance expression, neither of which requires dividing by the observed x.
For a teaching case, set x = 0, y = 3, independent unit variances and k = 2. The inequality becomes 9 ≤ 4(1 + r²), giving r² ≥ 5/4. Thus the accepted finite candidates satisfy r ≤ −√5/2 or r ≥ √5/2, approximately outside [−1.118, 1.118]. The point quotient is absent while the confidence set remains describable.
This does not define the target when the true parameter μX equals zero. We still assume μX is nonzero. An observed zero is an estimator outcome; a true zero is a property of the parameter and can make the chosen ratio target undefined. Those statements must not be merged.
The example is useful because it exposes a hidden software dependency. A routine that begins by calculating y/x may stop before reaching a perfectly evaluable candidate inequality. A method should respect the mathematical structure of the intended inference rather than require an intermediate point estimate that is not essential to its definition.
CHAPTER 16 OF 20 · Change the context
16. Solve a new bounded case without the familiar coefficients
Take x = 5, y = 5, vX = vY = 1, c = 0 and k = 2. Form the inequality yourself before reading the answer. Expanding (5 − 5r)² ≤ 4(1 + r²) gives 21r² − 50r + 21 ≤ 0.
The discriminant is 2,500 − 1,764 = 736, or sixteen times forty-six. The roots are (25 − 2√46)/21 and (25 + 2√46)/21, approximately 0.545 and 1.836. Since the leading coefficient is positive, the accepted set lies between them.
The point ratio one passes because the candidate difference is zero. The product of the two roots is C/A = 1, which provides an extra arithmetic check in this deliberately symmetric x–y setup. Do not expect that root-product property for arbitrary examples.
Now increase only vX to nine. The leading coefficient becomes twenty-five minus thirty-six, or −11. Before solving the new inequality, predict that the denominator-separation condition has changed. The old bounded interval is no longer a justified answer.
This is a transfer task, not a speed test. A good solution identifies the target, forms the candidate variance, expands the coefficients, solves the correct inequality and checks a candidate. Recalling the earlier approximate endpoints would demonstrate memory of one case, not control of the method.
CHAPTER 17 OF 20 · Check and connect
17. Diagnose the first incorrect step in an explanation
A learner reports the middle interval between the two roots in the unbounded example. Ask which way the quadratic opens and test r = 0. The coefficient arithmetic may be correct; the missing idea is the sign of the inequality across regions.
Another learner obtains covariance coefficient −2rc in the variance but changes it to the wrong sign after moving terms. Ask them to compare the expanded polynomial with the original squared inequality at a chosen r. Repair the expansion rather than repeat the entire confidence-interval introduction.
A third learner calls k = 2 exactly 95% because that is a familiar classroom phrase. Return to the reference distribution and critical value. The multiplier’s numerical convenience does not establish its coverage. With estimated variances, even the distribution used to choose the multiplier needs care.
Finally, a learner interprets the whole real line as all values having equal probability. The arithmetic can be entirely correct while that sentence is wrong. The set records which candidates survive this check, not a probability density over candidates.
These are different errors. Correcting the exact unstable step is more effective than assigning a larger collection of nearly identical quadratics. Follow each repair with a changed case that requires the learner to identify the same distinction independently.
CHAPTER 18 OF 20 · Check and connect
18. Write an evidence record another reader can reconstruct
A useful result names the ratio’s numerator and denominator parameters, their domains, the estimator pair and its covariance matrix. It states the distributional assumptions, how the critical value was chosen and whether the solution is bounded, separated or completely unbounded.
Retain the original candidate inequality alongside any rounded endpoints. That makes it possible to check a proposed value without trusting a transcription of the quadratic formula. For an unbounded set, write the two inequalities explicitly rather than show only two numbers separated by a dash.
| Evidence-record field | What it preserves |
|---|---|
| Target and domain | Which ratio is defined and which values are admissible |
| Estimates and covariance | The supplied joint uncertainty information |
| Reference distribution and k | The procedure’s coverage interpretation |
| Candidate inequality | A reproducible acceptance check |
| Set shape and limits | What the evidence does and does not exclude |
A practical next question follows from the limitation. If denominator uncertainty dominates the loss of resolution, ask which justified observation or model improvement would address it. More numerator precision alone may not separate the denominator from zero. This is a question for an appropriate study design, not an instruction to repeat measurements indiscriminately.
The uncertainty statement should remain useful even when it cannot offer a finite range. Explaining that the available evidence does not support a finite upper bound is a substantive result. Replacing it with an arbitrary numerical cap makes the report look tidier while making its meaning worse.
CHAPTER 19 OF 20 · Check and connect
19. Questions about unusual confidence-set shapes
Does an unbounded result mean the calculation broke?
Not necessarily. The original candidate inequalities demonstrate valid unbounded solutions under the stated model. Check the assumptions and arithmetic, then preserve the set’s shape. A software error and a mathematically unbounded result are different diagnoses.
Does it prove the true denominator is zero?
No. The target here assumes a nonzero denominator parameter. The evidence may still fail to distinguish it from values arbitrarily close to zero strongly enough to bound the ratio.
Can a positive point ratio justify deleting negative candidates?
No. A point estimate does not establish a parameter-domain restriction. A justified positive domain can be applied explicitly, but it may still leave an unbounded positive set.
Are these ordinary error bars around a quotient?
They are sets obtained by inverting specified candidate checks. Their shape need not be symmetric, connected or finite. A plotting format should accommodate the result rather than determine it.
Can the same k be used for every dataset?
Only with the statistical justification appropriate to the actual procedure. Our known-covariance normal model is explicit. Unknown variance, dependence estimation and other sampling conditions require suitable treatment rather than an automatic multiplier.
CHAPTER 20 OF 20 · Check and connect
20. Continue through inference, modelling and measurement
For the wider uncertainty framework, return to Uncertainty Quantification and How Uncertainty Quantification Works. For the interpretation of repeated-procedure coverage, use How Confidence Intervals Work.
For the mathematical foundations, follow Statistics, Probability and Mathematical Modelling. The first helps define inference, the second dependence and distributions, and the third the relationship between the target and its inputs.
The primary technical source is von Luxburg and Franz, A Geometric Approach to Confidence Sets for Ratios, reviewed on 22 October 2026. Its broader treatment includes settings beyond our known-covariance exercise. The supplied estimates, coefficient calculations, unit checks, boundary case and diagnostic tasks here are original teaching examples. They do not establish that an actual dataset satisfies the required model.
Continue through Learning That Transfers into Life and How X Works. A useful finish is to solve one new inequality and explain its set shape in words. The understanding is secure when the explanation remains accurate even when the result is less tidy than two finite error bars.
Contents · Previous chapter · Return to Uncertainty Quantification