Cauchy Distribution¶

A real-line probability law with a rational density, heavy tails, undefined mean and variance, and an iid sample average that retains the original law.

Version: v1 · 2026-10-31 · History
Domain-specific #: 14470
Domain group: Formal Sciences
Origin domain: Mathematics
Subdomain: Probability Theory → Mathematics
Aliases: Lorentz distribution

Core Idea¶

The Cauchy distribution is a probability law on the whole real line with location $\mu$ and positive scale $\gamma$. Its density is $f(y)=\gamma/[\pi((y-\mu)^2+\gamma^2)]$. Changing $\mu$ shifts its center; changing $\gamma$ changes its width. The density integrates to one over the real line, making it a law for a real-valued random outcome, not merely a curve with the same shape.^[1]^[2]

Its tails decay like $1/y^2$ at large absolute values. Consequently the usual expectation and variance are undefined; $\mu$ is the median and mode, not an expected value.^[1]^[2] For independent identically distributed Cauchy draws with a common location and scale, their arithmetic average has the same Cauchy law as one draw. Increasing the sample size therefore does not make that mean concentrate around a finite population mean. NIST contrasts a simulated Cauchy sample mean of 3.70 with a median of −0.016 to show why the median can be a more useful location summary here.^[1]^[2]

The name marks an exact probability object. A Lorentzian physical intensity may resemble its density, but an approximately proportional line profile alone does not assert a normalized probability law for a random wavelength.^[3]

How would you explain it like I'm…

The Wild-Spinner Pattern

Imagine a magic spinner that picks numbers. Most of the time it lands near the middle, but every now and then it lands super, super far away. Those far-away ones happen often enough that averaging lots of spins never settles down. So to find the middle, it's better to line up the results and pick the one in the middle.

Averages That Never Settle

The Cauchy distribution is a rule in math for how likely different numbers are when something random happens. Its picture is a hill with a center and a width you can change. But the sides of the hill fade away slowly, so really huge numbers keep popping up now and then. Because of that, if you take the average of many results, it jumps around just as much as one single result — it never settles down to a fixed answer. The middle-most result, called the median, does a much better job of finding the center.

Heavy-Tailed Law Without a Mean

The Cauchy distribution is a probability distribution over all real numbers, with a location μ (its center) and a positive scale γ (its width). Its density is f(y) = γ / [π((y−μ)² + γ²)], which integrates to 1, so it truly describes a random number, not just a bell-like curve. Its tails shrink only like 1/y², so slowly that the mean and variance are undefined; μ is the median and the most likely value, not an average. Surprisingly, the average of many independent Cauchy draws with the same μ and γ has exactly the same distribution as a single draw, so bigger samples don't make the mean more precise. That's why the median is a better way to estimate the center here. A physics line shape (a Lorentzian) can look like this curve without being a probability distribution.

The Cauchy distribution is a continuous probability law on the real line with location parameter μ and scale parameter γ > 0, density f(y) = γ / [π((y−μ)² + γ²)]. Shifting μ moves the center; changing γ changes the width; the density integrates to one, so it is a genuine law for a real-valued random variable. Its tails decay like 1/y², which makes the expectation and variance undefined; μ is the median and mode rather than an expected value. For i.i.d. Cauchy draws with common location and scale, the sample mean has the same Cauchy law as a single observation, so the mean does not concentrate as the sample size grows — the usual law-of-large-numbers intuition fails. NIST illustrates this with a simulated sample whose mean was 3.70 while its median was −0.016, showing why the median is the more useful location summary. The name refers to the exact normalized probability object; a Lorentzian intensity profile in physics may resemble the density, but proportional shape alone does not assert a probability law for a random wavelength.

Structural Signature¶

Real-valued outcome. The modeled random coordinate ranges over the full real line. Restricting the support to one side changes the object: the reflected Brownian first-hit density discussed below lives only on $y>0$.^[4]
Location and positive scale. The pair $(\mu,\gamma)$ identifies a member of the family; $\mu$ is its center and $\gamma>0$ sets its width. The Brownian first-hit case has $\mu=0$ and $\gamma=x$, where $x$ is the positive starting distance from the axis.^[1]^[2]^[4]
Normalized rational density. The precise function $\gamma/[\pi((y-\mu)^2+\gamma^2)]$ allocates probability over that outcome space. Merely proportional or approximate Lorentzian intensity lacks this full probability-law assertion.^[1]^[2]^[3]
Heavy-tail and moment readout. The density entails slow tails and undefined mean and variance. This is a consequence of the law, not an extra parameter or a claim that every observed finite sample lacks a computable arithmetic mean.^[1]^[2]
Stability under iid averaging. A sample average of iid draws from the same Cauchy law retains that law. This characteristic readout is not a requirement that a single variable be generated by averaging, nor a claim about arbitrary independent variables with different laws.^[1]^[2]^[5]

The first three roles specify the object. The last two tell us what follows from that specification and how to recognize the inferential danger of using an ordinary sample mean.

What It Is Not¶

Not a distribution with mean $\mu$. The location parameter is a median and mode; the expectation and variance do not exist in the conventional finite sense.^[1]^[2]
Not proof that no statistic can estimate location. NIST uses the contrast between its simulated sample mean and median precisely because the median remains informative when the mean behaves poorly.^[1]^[2]
Not every heavy-tailed law. Heavy tails alone do not identify the rational Cauchy density, its full-line support, or its exact stability property.
Not a reflected half-Cauchy law. Arista and O’Connell first give a full-line Brownian hitting density $h'(x,y)=x/[\pi(x^2+y^2)]$. Their subsequently reflected quadrant process has $h(x,y)=2x/[\pi(x^2+y^2)]$ only for $y>0$; its support and normalization differ.^[4]
Not every Lorentzian-looking signal. NIST says pressure-broadened spectral intensity is often approximately Lorentzian. Proportional intensity does not by itself provide a random-wavelength sample space, exact normalization, or a claim that measured wavelengths follow this law.^[3]

Scope of Application¶

The law applies wherever a genuine real-valued random outcome has the exact Cauchy density. In statistical modeling, one may posit that law for generated observations and use its tail and location properties to decide which summaries are meaningful. NIST’s 10,000-draw simulation is an explicit case; its extreme observations make the arithmetic mean quite unlike the median in that realization.^[1]^[2]

It also appears as an exact mathematical hitting law. For ordinary planar Brownian motion started distance $x>0$ from the $y$-axis, the first $y$-axis hit coordinate ranges over the whole real line and has the Cauchy density with scale $x$.^[4] The same family can arise as the ratio of independent standard normal variables or through a uniform-angle ray construction.^[5] These are different constructions of one law, not evidence that every unrelated heavy-tailed process is Cauchy.

Clarity¶

Calling a variable “Cauchy” commits to more than saying its histogram looks peaked with long tails. It fixes a normalized density, an outcome space, and the roles of its location and scale. That lets a reader distinguish a sample mean, which can be computed for any finite list of observed numbers, from a population expectation, which this law does not possess. NIST’s simulated sample provides both a numerical sample mean and a very different median without implying that the theoretical mean is 3.70.^[1]^[2]

The exact formula also separates the two Brownian cases: the full-line $h'$ matches the named distribution; the positive-half-line $h$ has a doubled density because it is a different law on a restricted support.^[4]

Manages Complexity¶

The two parameters and a single formula replace a potentially unbounded list of event probabilities. From them one can determine the center, width, tail behavior and whether a proposed mean-variance summary is even defined.^[1]^[2] In the Brownian hitting problem, the starting distance $x$ becomes the scale parameter, so changes to the geometry can be read directly as changes to the hit-coordinate law.^[4]

NIST describes pressure-broadened intensity as often approximately Lorentzian.^[3] Knowing that a curve has roughly this shape is less information than knowing that a random outcome follows the exact normalized full-line Cauchy law. Shape resemblance alone cannot carry every consequence of the law.

Abstract Reasoning¶

NIST gives the exact full-line density and its moment and averaging properties.^[1]^[2] Arista and O’Connell show that restricting the Brownian construction to a reflected positive quadrant gives a different half-line density.^[4] To test a proposed Cauchy identification, ask what random outcome is modeled, check its support and normalization, then compare its density with the Cauchy formula. Only a matching law licenses the same moment and iid-averaging conclusions.

For data analysis, the inferential move is especially sharp: do not assume averaging more iid observations will make the sample mean settle at $\mu$. Under the Cauchy law the mean’s sampling distribution remains Cauchy, whereas the median is a meaningful location readout.^[1]^[2] This is a statement about that estimator under that law, not a prohibition on learning location by every method.

Knowledge Transfer¶

Within probability and statistics, the literal law moves between simulation, sampling-theory calculations and geometric constructions. The two examples below instantiate the same full-line density even though one begins with generated numbers and the other with a Brownian first-hit coordinate.^[1]^[2]^[4] A normal-ratio construction gives another route to the same formal object.^[5]

NIST describes a pressure-broadened physical intensity profile that is often approximately Lorentzian.^[3] Outside an explicit random-outcome model, such a profile is a useful comparison but its shape alone does not transport the Cauchy distribution’s probability and estimator claims. The more general possibility of allocating probability over outcomes belongs to the broader Probability Distribution and Probability concepts; the Cauchy name carries this particular density and its consequences.

Examples¶

NIST’s simulated Cauchy sample. The handbook introduces histograms from 10,000 generated draws per distribution, including a Cauchy sample. It reports that sample’s mean as 3.70 and median as −0.016; values outside the displayed $[-10,10]$ range extend to roughly −29,000 and 89,000.^[1]^[2] Mapped back: each generated value is a real-valued outcome; the generating family has location and positive scale, though NIST does not specify their exact values in this passage; NIST identifies the normalized rational density in its distribution section; extreme draws reveal the heavy-tail and moment readout without defining a population mean; and the handbook states the iid averaging stability of the Cauchy sample mean. These reported numbers describe one simulation, not every Cauchy sample.

Brownian first hit on a line. Arista and O’Connell state that the $y$ coordinate of the first hit of the $y$-axis by ordinary planar Brownian motion started at horizontal distance $x>0$ has $h'(x,y)=x/[\pi(x^2+y^2)]$ for $y\in\mathbb R$.^[4] Mapped back: the hit coordinate is the real-valued outcome; location and positive scale are $0$ and $x$; $h'$ is exactly the normalized rational density. The heavy-tail and moment readout and iid averaging stability follow mathematically from that identified law; the paper does not report repeated hit experiments or sample means. Its later reflected process instead has the half-line density $h(x,y)$ and is not this full-line example.^[4]

Structural Tensions¶

The available sources do not establish a genuine opposed-pressure trade-off intrinsic to this distribution. They establish an estimator asymmetry: an arithmetic mean can be computed for a finite sample, but this law has no population expectation and the iid sample mean retains the original broad distribution. NIST presents the median as a meaningful location summary for its simulated case.^[1]^[2] The practical diagnostic is whether a proposed summary has the concentration needed for the claim being made; the law does not make one estimator universally best. No additional structural tension is asserted from these sources.

Structural–Framed Character¶

NIST identifies the exact density in a generated Cauchy sample,^[1]^[2] while Arista and O’Connell identify it in a Brownian first-hit coordinate.^[4] That invariance across constructions places the entry toward the structural end of the structural–framed spectrum. Choosing the law for actual data has evaluative weight because a wrong family changes tail and estimator conclusions, but the choice does not alter the law’s identity.

The named object does not depend on a particular human practice or institution for its operation. Its historical mathematical vocabulary and publication settings explain how people discuss it, not how the density is constituted. NIST describes physical intensity as often approximately Lorentzian.^[3] Calling that profile “Cauchy” does not by itself establish a full-line probability law for a random outcome. The portable skeleton is broader probability allocation over outcomes, represented in the catalog by Probability Distribution and ultimately Probability; neither makes this specific rational density cross-domain on its own. Its character: a highly structural, domain-bound formal law whose exact consequences travel only with its probabilistic preconditions.

Structural Core vs. Domain Accent¶

The structural core is a normalized allocation of probability across a measurable set of possible outcomes. Probability Distribution captures that broader measure-level relation; the Cauchy law adds the real-line rational density, location-scale parameters, undefined moments, and iid averaging stability.^[1]^[2] These conditions distinguish it from other distributions such as the normal or from a half-line fold.^[4]

That specificity is why the named Cauchy law remains a domain-specific abstraction. A nonprobabilistic curve with a similar shape does not inherit its expectation or stability statements. The broader allocation idea may reach other domains through Probability or Probability Distribution, but evidence for such reach should not be assigned to this named law by analogy. The test for a more general abstraction is whether a source-backed nonprobabilistic case preserves the full organizing relation and usable inferences without simply borrowing this density’s name. The present cases do not supply one.

Mathematical comparison: A Cauchy law is a probability distribution distinguished by its exact density and consequences. A probability distribution need not be Cauchy. Probability describes the still broader allocation of weight over possibilities.

Related concepts: Random Variable names the measurable quantity that may have this law, rather than a class of distribution laws. NIST reports approximately Lorentzian physical line shapes.^[3] A Lorentz oscillator model or similar profile does not automatically supply the probability law. The distinction is whether a full-line random outcome carries the exact normalized density.

Neighborhood in Abstraction Space¶

Cauchy Distribution sits in a sparse region of the domain-specific corpus (95^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (2835 abstractions)

Nearest neighbors

Local Time (Mathematics) — 0.79
Wigner Surmise — 0.78
Carleman's equation — 0.77
Gamma Distribution — 0.77
Reflection principle (Wiener process) — 0.77

Computed from structural-signature embeddings · 2026-10-31

Not to Be Confused With¶

A generic Lorentzian function: ask whether there is a random variable with this exact normalized full-line density. An approximate pressure-broadened intensity profile alone does not answer yes.^[3]
The folded or half-Cauchy distribution: ask whether negative outcomes remain possible under the asserted law. The reflected quadrant density in Arista and O’Connell lives on $y>0$ and has a factor of two.^[4]
An arbitrary heavy-tailed distribution: ask whether the exact rational density and its parameterization hold; qualitative long tails are insufficient.
A Cauchy random variable versus its distribution: the variable is the random map or outcome, while this entry describes its induced law. Different constructions, including simulation and Brownian first hitting, can share that law.^[1]^[2]^[4]

References¶

[1] National Institute of Standards and Technology, “Cauchy Distribution,” NIST/SEMATECH e-Handbook of Statistical Methods, §1.3.6.6.3, https://itl.nist.gov/div898/handbook/eda/section3/eda3663.htm (density and distribution properties). registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t

[2] National Institute of Standards and Technology, “Measures of Location,” NIST/SEMATECH e-Handbook of Statistical Methods, §1.3.5.1, https://itl.nist.gov/div898/handbook/eda/section3/eda351.htm (simulated-sample and location discussion). registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t

[3] National Institute of Standards and Technology, “Spectral Line Shapes, Widths, and Shifts,” in Atomic Spectroscopy, https://physics.nist.gov/Pubs/AtSpec/node20.html (official HTML inspected; Pressure Broadening passage says intensity profiles are often approximately Lorentzian and gives a proportional profile, not an exact random-wavelength Cauchy probability assertion). registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h

[4] Jonas Arista and Neil O’Connell, “Loop-Erased Walks and Random Matrices,” Journal of Statistical Physics 177 (2019), 528–567, DOI 10.1007/s10955-019-02378-1, https://link.springer.com/article/10.1007/s10955-019-02378-1 (original mathematical research, publisher full HTML inspected; §4.3 gives full-line ordinary Brownian $h'$ immediately before Equation (4.7), then reflected positive-quadrant $h$ in Equation (4.7)). registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n

[5] Giovanni Puccetti, Pietro Rigo, Bin Wang, and Ruodu Wang, “Centers of probability measures without the mean,” author-hosted full manuscript (2018), https://sas.uwaterloo.ca/~wang/papers/2018Puccetti-Rigo-Wang-Wang-JTP.pdf (printed p.2 / one-based PDF p.3 gives standard-normal-ratio, stability, and ray constructions; printed p.5 / one-based PDF p.6 gives standard Cauchy density). These are mathematical background facts in the article, not reported experiments. registry ↩a ↩b ↩c