Finding a Tipping Threshold by Bracketing

Why Bracket a Tipping Threshold?

A sensitivity analysis evaluates a decision rule at specified values of a sensitivity parameter. Sometimes the decision changes between two evaluated scenarios, but neither scenario is exactly at the boundary. Bracketing turns that observation into a numerical statement: the tipping threshold lies between a value where one decision holds and a value where the other holds, subject to stated conditions. Further evaluations can narrow that interval.

The earlier material on defining and tracing a one-parameter tipping point establishes the decision rule and the direction in which the sensitivity parameter is varied. Here the task is narrower: locate the threshold numerically from computed scenarios, and report how precisely it has been located. The parameter might be a delta adjustment, but the bracketing calculation applies to any one-dimensional sensitivity parameter.

Definition: Let \(\delta\) be a scalar sensitivity parameter, and let \(M(\delta)\) be a signed margin relative to a specified decision boundary. The tipping threshold is a value \(\delta^*\) satisfying \(M(\delta^*)=0\). For example, the sign convention may define \(M(\delta)>0\) as retaining the decision of interest and \(M(\delta)<0\) as not retaining it. The margin, its zero, and its sign interpretation must all correspond to the same prespecified decision rule.

The margin is a convenient numerical representation of the decision boundary; its units and scale depend on the rule. It could be a contrast minus a decision limit, for example. A zero of \(M\) marks the boundary, but the sign of \(M\) alone does not state how clinically important an effect is. That interpretation comes from the decision rule and the outcome scale.

What a Valid Bracket Establishes

Choose two sensitivity values \(a<b\) and evaluate the margin at both. If the evaluations are reliable and have opposite signs, the endpoints bracket a sign change. To infer that an exact zero lies between them, the margin must also be continuous on \([a,b]\). Continuity means that sufficiently small changes in \(\delta\) produce sufficiently small changes in \(M(\delta)\). It rules out a jump from a positive value to a negative value without taking the value zero.

M(a)>0,\quad M(b)<0,\quad M \text{ continuous on } [a,b] \quad\Longrightarrow\quad \text{at least one } \delta^*\in(a,b)\text{ has }M(\delta^*)=0.

This existence conclusion is the intermediate value theorem. It does not establish that there is only one zero. To guarantee uniqueness, require, for example, that \(M\) is strictly monotone throughout the bracket. A strictly decreasing margin can cross zero at most once; a strictly increasing margin can do so at most once. More generally, uniqueness follows if the zero set within the bracket is known to consist of a single point. Merely assuming a non-strict monotone pattern is not enough: a function could equal zero across an interval.

Key distinction: Continuity and opposite endpoint signs guarantee at least one zero. Strict monotonicity on the bracket (or another explicit uniqueness condition) guarantees at most one. Without both existence and uniqueness, a bracket need not identify a single tipping threshold.

In practice, a computed analysis often provides evaluated scenarios rather than a proof of continuity or strict monotonicity. Treat those properties as modeling or numerical assumptions to assess, not as facts established by the endpoint signs. If a decision statistic jumps at a cutoff, or if the computed margin is noisy, the results may locate a change between scenarios without establishing an exact zero of a continuous function.

Interpolation and Its Limits

Once opposite signs are found, linear interpolation gives a quick estimate of where a zero would fall if the margin changed linearly between the endpoints. The straight line through \((a,M(a))\) and \((b,M(b))\) reaches zero at:

\widetilde{\delta} = a+\frac{M(a)}{M(a)-M(b)}(b-a), \qquad M(a)>0,\quad M(b)<0.

The fraction is between zero and one, so the estimate falls inside the bracket. This is the zero of the secant line, not necessarily the zero of the actual margin. It equals the actual threshold when the margin is linear over the bracket; otherwise it is an approximation. Without further information about curvature or a justified error bound, the interpolation calculation alone does not establish how close the estimate is to the true threshold.

Bisection provides a different kind of information. Evaluate the margin at the midpoint \(m=(a+b)/2\). If its sign matches \(M(a)\), retain \([m,b]\); if it matches \(M(b)\), retain \([a,m]\). Under continuity and reliable evaluations, the retained interval still contains at least one zero, and its width is half the previous width. Bisection narrows a bracket without assuming linearity. With a unique zero, it narrows the interval containing that one threshold.

Worked Example: Interpolating Within a Computed Bracket

Suppose hypothetical sensitivity calculations use \(\delta\) measured in standardized sensitivity units. The prespecified signed margin is positive when the decision is retained and negative otherwise. The computed margins are \(M(1.5)=0.90\) margin units and \(M(2.5)=-0.30\) margin units. Assume, for this calculation, that the margin is continuous and strictly decreasing on this interval.

The endpoint signs are opposite, so continuity guarantees a zero between 1.5 and 2.5; strict decrease guarantees that there is only one zero there. Linear interpolation estimates its location as:

\widetilde{\delta} =1.5+\frac{0.90}{0.90-(-0.30)}(2.5-1.5) =1.5+\frac{0.90}{1.20}(1.0) =2.25.

Thus the secant-line estimate is \(\delta=2.25\) standardized sensitivity units. The evaluated scenarios themselves support the bracket \([1.5,2.5]\), whose width is 1.0 unit. Unless linearity or a suitable error bound is justified, 2.25 is an interpolation estimate, not evidence that the actual threshold is known to the nearest hundredth. A careful report gives both the estimate and the bracket, and explains the basis for any additional precision.

Refining the Bracket by Bisection

Bisection is useful when evaluations are available at chosen parameter values and the main objective is a defensible interval. It requires a bracket with reliable opposite signs and continuity for the zero-containment conclusion. For a single threshold, also require uniqueness, such as strict monotonicity throughout the bracket. At each step, evaluate the midpoint and keep the half-interval whose endpoints still have opposite signs.

Worked Example: Narrowing a Threshold and Reporting Its Precision

Consider a second hypothetical analysis, with \(\delta\) in standardized sensitivity units and \(M(\delta)\) in standardized margin units. Suppose \(M\) is continuous and strictly decreasing over the region considered, and these computed scenarios are obtained: \(M(1)=3.0\), \(M(5)=-5.0\), \(M(3)=-1.0\), \(M(2)=1.0\), \(M(2.5)=-0.20\), and \(M(2.25)=0.30\).

Start with the opposite-sign endpoints \([1,5]\). The midpoint is 3, where the margin is negative, so retain \([1,3]\). Its midpoint is 2, where the margin is positive, so retain \([2,3]\). At the next midpoint, 2.5, the margin is negative; retain \([2,2.5]\). The midpoint of that interval is 2.25, where the margin is positive, leaving \([2.25,2.5]\). Every retained interval has opposite-sign endpoint margins, and the assumptions ensure it contains the unique zero.

The final bracket has width \(2.5-2.25=0.25\) standardized sensitivity units. Linear interpolation within that bracket gives:

\widetilde{\delta} =2.25+\frac{0.30}{0.30-(-0.20)}(2.5-2.25) =2.25+\frac{0.30}{0.50}(0.25) =2.40.

The result can be reported as a unique threshold bracketed between 2.25 and 2.50 units, with a secant interpolation estimate of 2.40 units. The bracket width, rather than the number of decimal places displayed in 2.40, describes the guaranteed numerical resolution from these evaluations. More midpoint evaluations could narrow the bracket if greater precision were needed.

After \(k\) bisection steps, a starting interval of width \(b-a\) has width \((b-a)/2^k\), provided each step produces a valid sign-based half-interval. This gives a direct stopping rule: choose a width tolerance in the units of \(\delta\), then stop when the bracket is no wider than that tolerance. The tolerance should reflect the resolution needed for interpretation, not merely the number of digits a computer can print.

Why Interpolation Can Mislead

Worked Example: A Curved Margin With a Poor Secant Estimate

For a third hypothetical example, let \(\delta\) be dimensionless and let the margin, in standardized margin units, be \(M(\delta)=1-\delta^2\) on the interval \([0,2]\). This function is continuous and strictly decreasing there: for \(0\leq x<y\leq2\), \(x^2<y^2\), so \(1-x^2>1-y^2\). The endpoints give \(M(0)=1\) and \(M(2)=-3\), establishing a bracket for one zero.

Linear interpolation gives:

\widetilde{\delta} =0+\frac{1}{1-(-3)}(2-0) =\frac{1}{4}(2)=0.5.

But direct evaluation shows \(M(1)=1-1^2=0\), so the actual threshold is 1, not 0.5. The secant estimate is substantially displaced because the margin is curved on this wide interval. Bisection would evaluate the midpoint 1 and immediately identify the zero in this example. In a general analysis, the lesson is not that interpolation is unusable; it is that interpolation is an estimate whose accuracy depends on the shape of the margin.

Diagnostics, Alternatives, and Limitations

Before reporting a bracket, inspect how the evaluated margins behave across the relevant sensitivity range. A useful sequence of checks is:

Verify the target.
Confirm that every scenario uses the same signed margin and decision boundary. A change in the rule or sign convention can create an apparent crossing that has no consistent interpretation.

Verify the endpoint evaluations.
Check the parameter values, units, computed margins, and signs. If repeated or more precise calculations change an endpoint sign, the original bracket was not stable.

Inspect the path between endpoints.
Evaluate intermediate scenarios where appropriate. A reversal in sign pattern can indicate nonmonotonicity, and a jump or unstable computation can undermine the continuity assumption.

State what precision is supported.
Report the final bracket, its width, the interpolation estimate if used, and the assumptions that justify interpreting the bracket as containing one threshold.

A coarse grid scan is helpful for finding candidate intervals, but it may miss a sign change between grid values. Bisection systematically refines a known bracket and does not need a linearity assumption. Interpolation can propose a more targeted evaluation and can be efficient when the margin is nearly linear, but the secant estimate should be checked by evaluating the margin there. An analytic solution, when available and based on the actual margin rather than an approximation, is preferable for locating an exact root.

Several limitations deserve explicit attention. First, opposite signs do not by themselves establish continuity or uniqueness. A discontinuous decision calculation can switch sides without ever taking the value zero. Second, if the margin is nonmonotone, the bracket may contain more than one zero; a single interpolated estimate then does not identify which zero is being described. Third, Monte Carlo or other numerical noise can make signs unstable near the boundary. In that case, stabilize or increase the precision of the scenario computations before treating a sign as decisive. The resulting bracket concerns the computed sensitivity analysis; it is not, by itself, an uncertainty interval for a threshold in a population.

A common mistake is to say, “The endpoint signs differ, so the threshold is exactly the interpolated value, and monotonicity proves it is unique.” The correction is twofold: opposite signs plus continuity establish at least one zero, while strict monotonicity (or an explicit single-zero condition) is needed for uniqueness. Linear interpolation supplies a secant estimate, not an exact root in general. Report the bracket and its width so readers can distinguish what the evaluations guarantee from what the approximation suggests.

Key takeaway: A defensible numerical threshold report separates three claims: a zero exists in a bracket under continuity and opposite endpoint signs; it is unique under an additional condition such as strict monotonicity; and its reported numerical precision is supported by the final bracket, not simply by the digits in an interpolation estimate.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

A continuous margin has \(M(4)=2\) and \(M(8)=-1\). What does this establish, and what additional condition would establish a unique threshold in \([4,8]\)?
A bisection procedure starts with a bracket of width 3.2 sensitivity units. What is its width after four valid bisection steps?
For \(M(a)=0.6\), \(M(b)=-0.4\), \(a=2\), and \(b=5\), calculate the linear interpolation estimate. What can and cannot be concluded from it?
Why does a non-strict monotone margin with opposite endpoint signs not, by itself, guarantee a unique zero?

Solutions. (1) The intermediate value theorem establishes at least one zero in \((4,8)\). Strict monotonicity on the bracket, or another condition ensuring the zero set there is a single point, establishes uniqueness. (2) The width is \(3.2/2^4=3.2/16=0.20\) sensitivity units. (3) The estimate is \(2+\{0.6/[0.6-(-0.4)]\}(5-2)=2+(0.6/1.0)(3)=3.8\). This is the zero of the secant line. It is not necessarily the zero of the actual margin, and its precision is not established without further assumptions or evaluations. (4) A non-strict monotone function can be zero throughout an interval, so the bracket can contain multiple zero values. Strict monotonicity rules out that possibility.