Log-Log Survival Plot¶

Visual tail diagnostic — instantiates Tail-Dominance Modeling and Control

Plots the survival function on log-log axes so a heavy, slowly-decaying tail shows up as a near-straight line — a fast visual test of whether thin-tailed reasoning is even allowed.

Before you invest in any tail model, you need to know whether the tail is even heavy. The Log-Log Survival Plot answers that at a glance: plot the survival function — the probability of exceeding x — on logarithmic axes, and a power-law tail linearizes into a near-straight line whose slope reveals how heavy it is, while a thin (exponential-ish) tail curves sharply down. Its defining move is to turn "is this distribution dangerous?" into a shape you can see: a straight, slowly-falling line is a veto on averages and finite-variance reasoning and a green light for tail methods. It is a qualitative gate and prompt, not a fitted model or a certified number — it tells you which world you are in and hands the precise measurement to others.

Example¶

A seismic-hazard analyst holds a catalog of earthquake energies and must decide whether to summarize seismic release with a thin-tailed distribution or treat it as heavy-tailed. Plotting the survival function of seismic moment on log-log axes produces a strikingly straight line across several orders of magnitude — the visual signature of the Gutenberg–Richter law, in which earthquake frequency falls off as an approximate power of magnitude.

That single picture vetoes any plan to describe seismic energy by an average or a finite variance: with a tail this heavy, the largest event dominates the total and the "typical" quake is not the story. The workflow is gated into extreme-value territory, and the slope of the line gives a first, rough read on how heavy the tail is — a read to be pinned down later by formal estimation, not trusted from the eye alone.

How it works¶

The trick is the axis transform: on log-log axes a power law becomes a straight line whose slope is its exponent, because taking logs of a power relationship yields a linear one. Rendering the survival function rather than a binned histogram is deliberate — the survival form is far more robust in the sparse tail, where histograms turn to noise. You read three things: is the tail roughly linear (heavy) or curving down (thin); over how many decades does that linearity hold; and what is the approximate slope? The linear-vs-curved verdict is the gate; the slope is the shape hint.

Tuning parameters¶

Survival vs density rendering — plotting the complementary CDF instead of a binned PDF avoids the binning-dependent noise that wrecks a histogram's tail; the survival form is the robust default.
Lower cutoff (x-min) — where the inspected tail region begins. Power-law behaviour usually holds only above some point, so including the body bends the line and hides the tail.
Reference overlays — drawing candidate slopes or an exponential reference line so "straight" is judged against a standard rather than by wishful eyeballing.
Binning (for any density overlay) — logarithmic vs linear bins; log bins tame tail noise but coarsen detail. (A survival plot sidesteps most of this.)

When it helps, and when it misleads¶

Its strength is speed and legibility: one cheap picture can veto an entire thin-tailed modeling plan, and it communicates "the tail is the story here" to a non-specialist better than any statistic.

Its danger is that eyeballing linearity is treacherous. Distributions that are not power laws — lognormal, stretched exponential — can look nearly straight over a limited range, so the plot can manufacture a power law that is not there; and the far tail, where the claim matters most, is where the plot is noisiest. The classic misuse is declaring "it's a power law!" from a hand-drawn straight line and skipping the formal fit — or choosing the lower cutoff to maximize apparent straightness. The discipline is to use the plot as a gate and a prompt, never a verdict: confirm heaviness with formal tail-index estimation and goodness-of-fit, and keep the actionable claim modest — "heavy-tailed, so averages are unsafe," not "exactly a power law."^[1]

How it implements the components¶

tail_shape_diagnostic — the linearity and slope of the log-log survival curve are a direct visual read of the tail's heaviness and rough shape.
distributional_assumption_gate — a straight, slowly-decaying line vetoes mean- and variance-based reasoning and gates the workflow into tail-aware methods.

It does not produce a numeric tail index with confidence bounds (the quantitative side of tail shape — Tail-Index Estimation), fit an extrapolation model above a threshold (tail_boundary_definition, model_scope_boundary — Extreme-Value Threshold Model), or track shape drift over time (tail_drift_monitor — Expected Shortfall Dashboard).

Instantiates: Tail-Dominance Modeling and Control — it is the entry diagnostic that decides whether the whole archetype even applies.
Sibling mechanisms: Tail-Index Estimation · Extreme-Value Threshold Model · Cumulative Contribution Curve · Expected Shortfall Dashboard · Exposure Cap Policy · Heavy-Tail Simulation Scenario Set · Rare-Event or Importance Sampling · Reserve Buffer Policy · Robust Tail Statistic Review · Stress Test and Reverse Stress Test · Tail Incident Review

Notes¶

"Heavy-tailed" and "power-law" are not the same claim. The plot reliably supports the first — the actionable one, do not trust averages — but only weakly the second, a specific parametric form. Treating a roughly straight line as proof of an exact power law is the field's classic error, and it drives overconfident extrapolation; let the plot open the investigation, not close it.

References¶

[1] A widely-cited methodological caution (associated with the Clauset–Shalizi–Newman work on fitting power laws to empirical data): approximate linearity on a log-log plot is necessary but not sufficient evidence of a power law, since other heavy- and even medium-tailed distributions mimic it over a finite range. Formal estimation and goodness-of-fit are needed before asserting the parametric form. ↩