Upper-Tail Quantiles and Exceedance Thresholds

From an Upper-Tail Probability to a Threshold

A threshold is often chosen by specifying how rare an observation above it should be. For example, a safety limit might be set so that only a small fraction of outcomes exceed it. The key question is what “exceed” means: \(X>t\) and \(X\ge t\) can have different probabilities when \(X\) has an atom at \(t\). Quantiles provide a precise way to choose a threshold while keeping that distinction visible.

Let \(F(x)=P(X\le x)\) be the CDF of a real-valued random variable \(X\), and let its survival function be \(S(x)=P(X>x)=1-F(x)\). For a target upper-tail probability \(\alpha\), where \(0<\alpha<1\), the corresponding CDF probability is \(1-\alpha\). We use the generalized-inverse quantile convention: \(Q(p)=\inf\{x:F(x)\ge p\}\), for \(0<p<1\).

Definition: The upper-tail quantile at exceedance level \(\alpha\) is \(t_\alpha=Q(1-\alpha)\). Equivalently, \(t_\alpha=\inf\{x:S(x)\le\alpha\}\). It is the first threshold at which the strict survival probability is no greater than \(\alpha\). It need not satisfy \(P(X>t_\alpha)=\alpha\), particularly when the distribution has a jump.

The equivalence follows directly from \(S(x)=1-F(x)\): the condition \(F(x)\ge 1-\alpha\) holds exactly when \(S(x)\le\alpha\). Thus the quantile convention specifies which point to select even if the CDF is flat over an interval or jumps over the target probability.

What Probability Does the Quantile Guarantee?

At a continuous distribution’s quantile, the target tail probability is often attained exactly. At a discrete distribution’s quantile, the CDF can jump across the target level, so the strict exceedance probability may be smaller than requested. The general statement is a pair of bounds that covers both cases.

Theorem: Let \(0<\alpha<1\) and \(t_\alpha=Q(1-\alpha)\). Then \(P(X>t_\alpha)\le\alpha\le P(X\ge t_\alpha)\). Equivalently, the chosen threshold has strict exceedance probability at most \(\alpha\), while the probability of reaching or exceeding it is at least \(\alpha\).

Proof. Put \(p=1-\alpha\) and \(q=Q(p)=\inf\{x:F(x)\ge p\}\). Since \(0<p<1\) and a CDF tends to 0 and 1 at the two ends of the real line, \(q\) is finite. Right-continuity of \(F\) gives \(F(q)\ge p\): points in the set \(\{x:F(x)\ge p\}\) can be found arbitrarily close to \(q\) from the right, and right-continuity carries the inequality to \(q\). For every \(x<q\), the definition of the infimum implies \(F(x)<p\). Taking the limit as \(x\) increases to \(q\) yields \(F(q-)\le p\), where \(F(q-)=\lim_{x\uparrow q}F(x)=P(X<q)\). Therefore,

P(X>q)=1-F(q)\le 1-p=\alpha, \qquad P(X\ge q)=1-F(q-)\ge 1-p=\alpha.

This proves both bounds. The gap between them is \(P(X=q)\), the probability mass at the threshold. If there is no atom at \(q\), the two probabilities agree, and the bounds force both to equal \(\alpha\). \(\square\)

A useful exact case is a continuous CDF. If \(F\) is continuous, then \(F(q-)=F(q)\). The theorem gives \(F(q)\le p\) and \(F(q)\ge p\), so \(F(q)=p\), and consequently \(P(X>q)=\alpha\). If, in addition, \(F\) is strictly increasing around the target level, this is the unique solution to \(F(t)=1-\alpha\), or equivalently \(S(t)=\alpha\). Without strict increase, the probability can still be exact even if more than one threshold solves the equation.

Worked Threshold Calculations

Worked Example: An Exponential Upper-Tail Threshold

Let \(X\) have an exponential distribution with rate \(\lambda=2\), so \(S(x)=e^{-2x}\) for \(x\ge0\). Find the threshold with strict exceedance probability \(0.05\).

Set \(S(t)=0.05\). Since \(e^{-2t}=0.05\), taking logarithms gives \(-2t=\ln(0.05)\), and hence \(t=-\ln(0.05)/2=\ln(20)/2\). Numerically, \(\ln(20)\approx2.9957\), so \(t\approx1.4979\).

Checking the result, \(P(X>t)=e^{-2t}=e^{-\ln(20)}=1/20=0.05\). The exponential CDF is continuous and strictly increasing on its support, so this is the unique threshold for the requested tail probability.

Worked Example: A Uniform Threshold on a Finite Interval

Let \(X\) be uniform on \([0,10]\). Its CDF on this interval is \(F(x)=x/10\), and its survival function is \(S(x)=(10-x)/10\). Find the threshold for an upper-tail probability of \(0.1\).

Solving \(S(t)=0.1\) gives \((10-t)/10=0.1\), so \(10-t=1\) and \(t=9\). Equivalently, \(F(t)=1-0.1=0.9\), and \(F(9)=9/10=0.9\).

The distribution has no atom at 9. In fact, \(P(X>9)\) is the length of the interval \((9,10]\) divided by the total interval length: \(1/10=0.1\). Also \(P(X\ge9)=0.1\), since \(P(X=9)=0\).

Worked Example: A Discrete Threshold and an Unattained Tail Probability

Suppose \(X\) has the following probability mass function: \(P(X=0)=0.20\), \(P(X=1)=0.35\), \(P(X=2)=0.25\), and \(P(X=4)=0.20\). The probabilities sum to \(1\). Find the upper-tail quantile for \(\alpha=0.30\), then compare strict and inclusive exceedance.

Here \(p=1-\alpha=0.70\). The cumulative probabilities are \(F(0)=0.20\), \(F(1)=0.55\), \(F(2)=0.80\), and \(F(4)=1\). The first value at which the CDF reaches or exceeds \(0.70\) is 2, so \(t_{0.30}=Q(0.70)=2\).

The strict exceedance probability is \(P(X>2)=P(X=4)=0.20\), which is below \(0.30\). The inclusive probability is \(P(X\ge2)=P(X=2)+P(X=4)=0.25+0.20=0.45\), which is above \(0.30\). The theorem’s bounds hold: \(0.20\le0.30\le0.45\).

No threshold in this distribution gives strict exceedance probability exactly \(0.30\). For example, \(P(X>1)=0.25+0.20=0.45\), whereas \(P(X>2)=0.20\); moving the threshold across the atom at 2 makes the tail probability jump over \(0.30\). The quantile still gives the well-defined first threshold whose strict tail is no greater than the target.

Strict and Inclusive Exceedance

In continuous models, a single point has probability zero, so \(P(X>t)=P(X\ge t)\). In discrete or mixed distributions, that shortcut can be wrong. The exact relationship is \(P(X\ge t)=P(X>t)+P(X=t)\). Consequently, a phrase such as “the probability of exceeding the limit” should be translated into the event actually intended before calculating a threshold.

For the quantile \(t_\alpha\), the defining guarantee concerns strict exceedance: \(P(X>t_\alpha)\le\alpha\). If an application instead requires a threshold with \(P(X\ge t)\le\alpha\), the same quantile need not satisfy that requirement. Its inclusive probability is at least \(\alpha\), and can be substantially larger when there is a large atom at the threshold. The requested event—not just the numerical value of the tail level—determines which inequality is relevant.

The quantile also provides a practical inversion rule. If a continuous CDF is available in closed form, solve \(F(t)=1-\alpha\) or \(S(t)=\alpha\). For a discrete law, accumulate probabilities until the CDF first reaches \(1-\alpha\); do not interpolate through a jump unless a separate convention explicitly calls for it.

Translate the tail level.
For a target strict exceedance probability \(\alpha\), use the CDF level \(p=1-\alpha\).

Choose the quantile threshold.
Set \(t_\alpha=Q(p)=\inf\{x:F(x)\ge p\}\), or equivalently take the first \(x\) for which \(S(x)\le\alpha\).

Check the event and the result.
For a continuous law, verify whether \(S(t_\alpha)=\alpha\). For a law with atoms, calculate \(P(X>t_\alpha)\) and \(P(X\ge t_\alpha)\) separately.

Common Mistake

A common error is to assert that a \(1-\alpha\) quantile always has exactly \(\alpha\) probability above it. That conclusion silently assumes there is no probability mass at the threshold. In the discrete example, the threshold is 2, but \(P(X>2)=0.20\), not \(0.30\). The correct general claim is \(P(X>t_\alpha)\le\alpha\le P(X\ge t_\alpha)\); equality follows when the distribution is continuous at the quantile. A second error is to replace \(>\) by \(\ge\) without checking for an atom, even though the difference is precisely \(P(X=t_\alpha)\).

Key takeaway: An upper-tail level \(\alpha\) corresponds to the quantile \(Q(1-\alpha)\). It is the first threshold whose strict survival probability is at most \(\alpha\). Continuous distributions attain the target exactly; discrete distributions may jump over it, so strict and inclusive exceedance probabilities must be distinguished.

Check Your Understanding

Try these questions using the generalized-inverse quantile convention.

For an upper-tail level of \(0.025\), which CDF quantile level should be used?
State the two probability bounds that hold at \(t_\alpha=Q(1-\alpha)\), and explain which event uses the strict inequality.
If a distribution has a continuous CDF at \(t_\alpha\), why must \(P(X>t_\alpha)=\alpha\)?
Suppose \(P(X=1)=0.6\) and \(P(X=3)=0.4\). Find \(Q(0.5)\), then calculate \(P(X>Q(0.5))\) and \(P(X\ge Q(0.5))\).
Why can a discrete distribution fail to have any threshold with strict exceedance probability exactly equal to a specified \(\alpha\)?