Exponential Versus Polynomial Tail Decay

Comparing the Rate of Tail Decay

A tail probability answers a direct question: how likely is an observation to exceed a large threshold? Two distributions can both assign positive probability to arbitrarily large values while assigning very different probabilities to those values. Comparing their rates of tail decay helps describe that difference and interpret plots on logarithmic scales.

Here, “tail” means the survival probability \(S(x)=P(X>x)\), for a random variable \(X\) and a threshold \(x\). The focus is on tails that, for large \(x\), behave like an exponential function or a negative power of \(x\). A polynomial tail can be called heavy relative to an exponential tail in this comparison; the terminology does not mean that every use of “heavy-tailed” has the same definition.

Definition: An exponential-decay tail has \(S(x)\sim A e^{-cx}\) as \(x\to\infty\), where \(A>0\) and \(c>0\). A polynomial-decay tail has \(S(x)\sim Bx^{-\alpha}\) as \(x\to\infty\), where \(B>0\) and \(\alpha>0\). The notation \(g(x)\sim h(x)\) means \(g(x)/h(x)\to 1\). The constants affect the tail probabilities at particular thresholds, while the decay forms describe their behavior for large thresholds.

These forms can be seen in standard models. If \(X\) has an exponential distribution with rate \(c>0\), then \(S(x)=e^{-cx}\) for \(x\geq 0\). If \(X\) has a Pareto distribution with lower endpoint 1 and parameter \(\alpha>0\), then \(S(x)=x^{-\alpha}\) for \(x\geq 1\). Both are valid survival functions, but their tails decrease at different rates.

What Logarithmic Plots Reveal

Taking logarithms turns the two decay forms into different straight-line patterns. With natural logarithms, an exponential tail satisfies \(\log S(x)\sim \log A-cx\). A plot of \(\log S(x)\) against \(x\), often called a semilog plot, is therefore approximately linear with slope \(-c\). A polynomial tail instead satisfies \(\log S(x)\sim \log B-\alpha\log x\). A plot of \(\log S(x)\) against \(\log x\), called a log-log plot, is approximately linear with slope \(-\alpha\).

Worked Example: Identifying the Plot Scale

Compare an exponential survival function \(S_E(x)=e^{-0.4x}\) with a Pareto survival function \(S_P(x)=x^{-1.5}\), for \(x\geq 1\). Calculate their natural-log survival probabilities at \(x=1,10,100,\) and \(1000\).

For the exponential model, \(\log S_E(x)=-0.4x\). For the Pareto model, \(\log S_P(x)=-1.5\log x\). The values are:

\(x\)	\(\log S_E(x)=-0.4x\)	\(\log S_P(x)=-1.5\log x\)
1	-0.4	0
10	-4	approximately -3.454
100	-40	approximately -6.908
1000	-400	approximately -10.362

On a plot of log survival against \(x\), the exponential values follow a straight line: increasing \(x\) by 10 changes \(\log S_E(x)\) by \(-4\). The Pareto values do not follow a straight line on that scale. On a plot of log survival against \(\log x\), the Pareto values follow a line of slope \(-1.5\), since \(\log S_P(x)=-1.5\log x\). The axes matter: a straight line on one of these plots does not imply a straight line on the other.

In real data, survival probabilities at large thresholds are estimated from observations, and the most extreme estimates may be imprecise. A near-linear pattern on an appropriate log scale is a clue about tail behavior, not a proof that a particular formula holds. The relevant range, sample size, and possibility of other tail forms all matter.

Why Polynomial Tails Eventually Dominate

For any fixed positive \(c\) and \(\alpha\), an exponential factor \(e^{-cx}\) eventually becomes smaller than a power \(x^{-\alpha}\). The key comparison is between \(cx\), which grows in proportion to \(x\), and \(\alpha\log x\), which grows much more slowly.

Theorem: Suppose two positive survival functions satisfy \(S_E(x)\sim A e^{-cx}\) and \(S_P(x)\sim Bx^{-\alpha}\), where \(A,B,c,\alpha>0\). Then \[ \frac{S_P(x)}{S_E(x)}\longrightarrow\infty \qquad\text{as }x\to\infty. \] In particular, for all sufficiently large \(x\), \(S_P(x)>S_E(x)\).

Proof. By the two asymptotic equivalences,

\frac{S_P(x)}{S_E(x)} \sim \frac{B}{A}x^{-\alpha}e^{cx} = \frac{B}{A}\exp\!\left(cx-\alpha\log x\right).

For \(x\geq 1\), write \(\log x=\int_1^x t^{-1}\,dt\). Since \(t^{-1}\leq t^{-1/2}\) for \(t\geq1\),

0\leq\frac{\log x}{x} \leq\frac{1}{x}\int_1^x t^{-1/2}\,dt =\frac{2(\sqrt{x}-1)}{x} \leq\frac{2}{\sqrt{x}} \longrightarrow 0.

It follows that \(\alpha\log x/x\to0\), so \((cx-\alpha\log x)/x\to c>0\). Consequently \(cx-\alpha\log x\to\infty\), and the displayed ratio tends to infinity. Since the ratio is eventually greater than 1, the polynomial survival probability is eventually larger. \(\square\)

The result establishes an eventual ordering, not a universal ordering at every threshold. The constants and decay parameters can make either tail larger over a substantial finite range. The phrase “eventually dominates” is therefore important: it refers to sufficiently large thresholds, not necessarily to the range visible in a small sample or a short plot.

Worked Example: A Finite Threshold Versus the Eventual Ordering

Take an exponential model with \(S_E(x)=e^{-0.5x}\) and a Pareto model with \(S_P(x)=x^{-2}\) for \(x\geq1\). Compare their probabilities of exceeding 10 and 100.

At \(x=10\),

S_E(10)=e^{-5}\approx0.006738, \qquad S_P(10)=10^{-2}=0.01.

The polynomial tail is already larger at this threshold. At \(x=100\),

S_E(100)=e^{-50}\approx1.929\times10^{-22}, \qquad S_P(100)=100^{-2}=0.0001.

The difference is now much greater. For this particular pair of models, the ratio of the Pareto tail to the exponential tail is \(x^{-2}e^{0.5x}\), which tends to infinity by the theorem. The calculation at \(x=10\) is only a finite-threshold comparison; the theorem explains the eventual behavior for all positive choices of these parameters.

Threshold Probabilities in Practice

A useful comparison can be made by fixing a meaningful threshold and calculating \(P(X>x)\) under each model. The numerical values show what the models imply for that threshold, while the decay forms help explain how the comparison changes as the threshold grows. Parameters should be included: an exponential rate \(c\) has units inverse to those of \(x\), and both the threshold and parameter determine the probability.

Worked Example: Comparing Exceedance Probabilities

Suppose a positive measurement is modeled either by an exponential survival function \(S_E(x)=e^{-0.2x}\) or by a Pareto survival function \(S_P(x)=x^{-1.5}\) for \(x\geq1\). Calculate the probability of exceeding 20 and 100 under each model.

At \(x=20\),

S_E(20)=e^{-0.2(20)}=e^{-4}\approx0.01832, \qquad S_P(20)=20^{-1.5} =\frac{1}{20\sqrt{20}} \approx0.01118.

At this threshold, the exponential model assigns the larger exceedance probability. At \(x=100\),

S_E(100)=e^{-20}\approx2.061\times10^{-9}, \qquad S_P(100)=100^{-1.5}=0.001.

Now the polynomial model assigns a much larger probability. There is no contradiction: one comparison concerns a finite threshold, while the asymptotic result concerns the ordering for sufficiently large thresholds. In an application, the tail form and the numerical threshold probability answer related but distinct questions.

The examples use survival functions. A density is not interchangeable with a survival probability: the survival function is the probability remaining beyond a threshold, obtained by accumulating probability to the right. In particular, if a density has a polynomial form \(f(x)\) proportional to \(x^{-\beta}\) with \(\beta>1\), its integrated survival tail is proportional to \(x^{-(\beta-1)}\), not \(x^{-\beta}\). Always check whether a stated power law describes a density or a survival tail before interpreting its exponent or plotting it.

Common Mistakes

Assuming exponential decay is always smaller. The eventual ordering does not guarantee the ordering at every threshold. Compare the actual survival probabilities when a specific threshold matters.
Using the wrong log scale. Exponential decay is approximately linear on a plot of \(\log S(x)\) against \(x\); polynomial decay is approximately linear on a plot of \(\log S(x)\) against \(\log x\).
Confusing a density with a survival tail. The exponent for a polynomial density need not equal the exponent for its survival probability.
Reading a straight line as proof of a tail model. A log-scale plot can suggest a pattern, but finite observations do not establish the form of the tail for arbitrarily large thresholds.

Key takeaway: Exponential tails decrease like \(e^{-cx}\), while polynomial tails decrease like \(x^{-\alpha}\). Logarithmic plots reveal different linear patterns, and polynomial tails eventually exceed exponential tails, even though their probabilities need not be larger at every finite threshold.

Check Your Understanding

Use the definitions and comparisons above to answer these questions.

For an exponential survival tail, which quantity should be on the horizontal axis to obtain an approximately straight line when the vertical axis is \(\log S(x)\)?
For a polynomial survival tail \(S(x)\sim Bx^{-\alpha}\), what is the slope of the line when \(\log S(x)\) is plotted against \(\log x\)?
Why does \(S_P(x)/S_E(x)\to\infty\) not imply that the polynomial tail is larger at every threshold?
Explain why a polynomial density proportional to \(x^{-\beta}\), for \(\beta>1\), has a survival exponent \(\beta-1\).
What should be checked before interpreting a straight-line pattern on a logarithmic plot as evidence about a tail?