Pi

Pi, written π, is the constant obtained by dividing the circumference of any circle by its diameter. Its value begins 3.14159 and continues without end or repetition. Because it is irrational and transcendental, it cannot be written as a fraction or as the root of any polynomial with rational coefficients. Its importance comes from the fact that it links circles to analysis, trigonometry, probability and physics, so it appears in formulas that have nothing obvious to do with circles, such as the normalization of the normal distribution or the Basel sum over power series reciprocals.

The entry follows the constant from its earliest approximations in Babylonian, Egyptian, Greek, Chinese and Indian mathematics, through the infinite series and products of the seventeenth to nineteenth centuries and the proofs of its irrationality and transcendence, to the modern race to compute trillions of digits with supercomputers and algorithms built on complex-analytic identities. It then treats the open question of whether the digits behave like a random sequence, the constant's roles in calculus, geometry, probability and physics, the notational history of the symbol, its handling in floating-point arithmetic, and its place in popular culture. The entry closes with unresolved problems.

Definition and basic value

In Euclidean plane geometry, π is defined as the ratio of a circle's circumference $C$ to its diameter $d$, so $\pi = C/d$. The definition only makes sense if this ratio is the same for every circle, and similarity guarantees that. Any two circles are similar: one can be mapped onto the other by a translation, a rotation, and a uniform scaling. Scaling multiplies every length by the same factor, so circumference and diameter change together and their quotient stays fixed. Euclid's Elements (Book XII) proves that the areas of circles stand in the same proportion as the squares of their diameters, but it does not give the ratio of circumference to diameter a numerical value. Because the ratio is dimensionless, π has the same value in any system of length units.

The decimal expansion of π begins $\pi \approx 3.14159265358979323846264338327950288419716939937510$, with the digits shown to fifty places after the decimal point. Because π is an irrational number, this expansion never terminates and never repeats. For practical work, two fractions are often used as approximations. The fraction $22/7 \approx 3.142857$ differs from π by about $1.26 \times 10^{-3}$, so it is accurate only to about two decimal places. The fraction $355/113 \approx 3.14159292$ differs from π by about $2.7 \times 10^{-7}$, and it agrees with π to six decimal places. Both are rational numbers, and the fact that such short fractions come so close is a standard illustration of how good rational approximations to an irrational number can be.

The constant also enters through area and through angle measure. A circle of radius $r$ has area $\pi r^2$. For a unit circle, with $r = 1$, this gives an area of exactly $\pi$. A circle of radius $r$ has circumference $2\pi r$, so the unit circle has circumference $2\pi$, and half of that, the length of a semicircle, is $\pi$. This is why π is often described as the distance travelled along a unit circle in a half turn.

The same relation fixes the link between π and angles. In trigonometry, an angle measured in radians is the length of the arc it cuts out on a unit circle. A full revolution therefore corresponds to an arc of length $2\pi$, so a full turn equals $2\pi$ radians, a half turn equals $\pi$ radians, and a right angle equals $\pi/2$ radians. Since $360^\circ$ corresponds to $2\pi$ radians, one radian is $180^\circ/\pi$, or about $57.3^\circ$. This is the reason the constant appears so often when angular quantities are expressed in radians, a point that the later sections on analysis and physics develop further.

Irrationality and transcendence

Irrationality

The first rigorous proof that π is irrational came from Johann Heinrich Lambert. Sources differ on the date: one records the argument in a 1761 memoir, while MacTutor dates his presentation of the result to the Berlin Academy in 1768. Lambert showed the result in 1761 in his work Mémoire sur quelques propriétés remarquables des quantités transcendantes circulaires et logarithmiques. In a paper presented to the Berlin Academy in 1768 Lambert showed that, if x is a nonzero rational number, then neither e^x nor tan x can be rational.

His method was to expand a ratio of trigonometric power series into a continued fraction. Starting from the sine and cosine series, he applied the Euclidean algorithm to obtain an expansion of tan r in which each partial denominator is a simple linear or quadratic expression in r². He then showed that the tails of this expansion, under the sign conditions it satisfies, can never terminate, so tan x is irrational whenever x is a nonzero rational number. The conclusion for π follows from a single value: tan(π/4) = 1 is rational, so π/4 cannot be rational, and therefore π is not rational either. Later authors gave simpler routes to the same result. Ivan Niven and Charles Hermite's approach led to shorter proofs, and Miklós Laczkovich showed that Lambert's continued fraction for tan x can be derived from the hypergeometric function, which gave a more streamlined version of the argument.

Transcendence

Irrationality says only that π is not a root of a linear equation with integer coefficients. The stronger claim, that π is not a root of any non-zero polynomial with integer coefficients, is what makes π a transcendental number. The proof of this came from a chain of results. In 1873 Charles Hermite proved that e is transcendental, but nearly all proofs of the transcendence of π used e's transcendence, so Hermite's result was the necessary first step. Ferdinand von Lindemann adapted this approach and, in 1882, proved that π is transcendental.

The most important practical consequence concerns geometry. A square with the same area as a unit circle has side √π, and a length can be constructed with compass and straightedge only if it is a root of an algebraic equation over the rationals. Because π is transcendental, the classical problem of squaring the circle has no solution with these tools; this follows from Euler's identity e^{iπ} = −1, which is equivalent to the impossibility of squaring the circle with only ruler and compass, a problem that had remained open for more than two thousand years.

The argument rests on the Lindemann–Weierstrass theorem. In 1885 Karl Weierstrass succeeded in proving a much more general result than Hermite's and Lindemann's special cases. In its usual form the theorem says that if α₁, …, αₙ are distinct algebraic numbers, then the exponentials e^{α₁}, …, e^{αₙ} are linearly independent over the algebraic numbers. Suppose π were algebraic. Then iπ would also be algebraic, since if π is a root of a polynomial p, then iπ is a root of the polynomial p(ix)p(−ix), hence also algebraic. Applying the theorem to the distinct exponents iπ and 0 would then forbid the algebraic relation e^{iπ} + 1 = 0, which is false. Hence π is transcendental.

Open questions about related constants

Knowing that π is transcendental does not settle the status of combinations of π with other famous constants. At least one of πe and π+e (and probably both) are transcendental, but transcendence has not been proven for either number on its own. The sum π + e is described in the literature as a longstanding problem in number theory. Some special cases are settled: the Gelfond theorem, for example, shows that e^π is transcendental, because e^π = (−1)^(−i).

Irrationality measures give a finer view of how well π can be approximated by rationals. The irrationality measure μ(x) is the smallest exponent such that |x − p/q| > q^(−μ−ε) holds for all sufficiently large q and all ε > 0. It is not hard to see that the irrationality measure of e is 2, but the exact irrationality measure of π is unknown. The first upper bound, of 42, was proved in 1953 by Kurt Mahler. Vladislav Salikhov's bound of 7.606308 was announced in 2008. In 2019 Doron Zeilberger and Wadim Zudilin, adapting Salikhov's method, established that μ(π) is at most 7.103205334137, and a 2026 arXiv preprint reports a slightly lower bound, μ(π) < 7.101862832357. Lower bounds for μ(π) are much weaker: the best one known is only 2, which every irrational number satisfies. Whether the true value is 2 remains open.

Ancient approximations

The oldest known numerical values for the circle ratio come from Mesopotamia and Egypt. They were practical rules for measuring fields, granaries and basins, not statements about an abstract constant. Their dating is uncertain, so the figures below should be read as approximate.

Babylonian computations mostly rested on a crude figure. In some texts the circumference was thought to be 3 times the diameter, but one text from Susa uses essentially 25/8 = 3.125 instead. The Susa text reaches this value by comparing the perimeter of an inscribed hexagon with the circle's circumference, and this number is probably one of the most disputed constants in Babylonian mathematics. The reasoning is simple once stated. A regular hexagon inscribed in a circle has perimeter six radii, so the naive comparison always returns 3. A ratio of 24/25 between hexagon perimeter and circumference, as the tablet implies, corrects this to 3/π = 24/25, hence π = 25/8. Dating the tablet is also disputed. Cuneiform tablets from c. 1900–1600 BCE give a value for pi derived from a hexagon. Most of the surviving Babylonian mathematical tablets are later, dating from 1800 to 1600 BC.

Egyptian evidence is more explicit. The Rhind papyrus gives a procedure for the area of a circular field: subtract one-ninth of the diameter and square the remainder, yielding 256/81, or about 3.16045. Writing the diameter as 2r, the rule gives (8/9 · 2r)² = (256/81)r², so the implied value is π ≈ 256/81 ≈ 3.1605. The fractional term 256/81 approximates the value of π as being 3.1605…, an error of less than one percent. The problem is circle areas are approximated by treating the circle as a square with side length equal to eight-ninths of the diameter. The same value appears in the Moscow Mathematical Papyrus, where it defines a unit of volume. The hekat unit was defined in the Moscow Mathematical Papyrus by approximating π to around 3.16. The papyrus dates also disagree: one source puts it at c. 1650 bce, another at dated to 1700 BCE, and a popular account suggests 2000 BC for the Egyptian value. Egyptian practice was not uniform. Earlier in RMP 38 Ahmes informally corrected granary losses by a Pi estimate of 22/7.

Biblical references are weaker evidence. The description of Solomon's bronze basin in 1 Kings 7:23 gives a diameter of ten cubits and a circumference of thirty cubits, which implies π = 3. A cubit generally extended from the elbow to the tip of the middle finger and could vary slightly between cultures, and large objects in ancient settings were often measured in whole units for ease of recordkeeping. The verse therefore reflects rough measurement, not a computed constant. Later readings that find a precise value hidden in the text, such as a gematria-based ratio proposed by Matityahu Munk in 1962, are modern interpretations rather than evidence of ancient knowledge. The coded value is 111/106, a highly accurate approximation of Pi/3. Indian material is similarly difficult to date. The Shulba Sutras, texts on altar construction, are usually placed in the first millennium BCE, and one popular summary treats π as (9785/5568)² ≃ 3.088, though such secondary figures should be checked against scholarly editions.

Early methods were limited by their tools. Measuring a real basin, rope or wheel gives only a few reliable digits, and the hexagon trick returns 3 by construction unless the polygon has more sides. Greek mathematics changed the question. The first theorems relating to circles are attributed to Thales around 650 BC. Book III of Euclid's Elements deals with properties of circles and problems of inscribing and escribing polygons. Greek geometers also took up the problem of finding a square with the same area as a given circle. The decisive step was to replace measured values with proven limits. Archimedes was the first mathematician to bound pi in this way, by calculating its upper and lower limits. His polygon method, described in the Archimedes entry, is the start of the geometric tradition, and it is discussed in the next section.

Classical geometric methods

Before infinite series were known, the most reliable way to pin down π was geometric. A circle lies between the perimeters of regular polygons drawn outside and inside it, and the bracket tightens as the number of sides grows. Greek, Chinese, Persian and European mathematicians used this idea to move the value from a few percent of accuracy to many digits, and the same polygon framework shaped the search for better bounds for about two thousand years.

Archimedes

Around the third century BCE, Archimedes of Syracuse worked with regular polygons circumscribed about a circle and inscribed in it. He started with hexagons and repeatedly doubled the number of sides, reaching 12, 24, 48 and finally 96. For each polygon, the Pythagorean theorem gives the side length from the previous stage, and the perimeter then gives a bound: the inscribed perimeter falls short of the circumference and the circumscribed one exceeds it. The result was 223/71 < π < 22/7, or roughly 3.1408 < π < 3.1429. The upper bound 22/7 later became known as "Archimedes' constant", although he did not present it as a value of π itself.

His calculations were demanding because every doubling required square roots that had to be bounded by hand. Modern analyses show that the exact perimeters of the 96-gons give a slightly narrower interval than the one Archimedes stated, because his rational bounds on the square roots loosened the result. Each doubling reduces the error by about a factor of four, which is why the method is steady but slow per digit.

Liu Hui and the method of exhaustion

In the third century, during the Three Kingdoms period, Liu Hui wrote a commentary on the Nine Chapters on the Mathematical Art that included a polygon method for the circle. He argued that the area of an inscribed polygon approaches the area of the circle as the sides multiply, a reasoning that is sometimes called the method of exhaustion. Liu Hui first noted that the inscribed hexagon's perimeter ratio of three implies that π must exceed 3, then used a 96-gon to obtain bounds and proposed 157/50 as a usable value while admitting it was slightly small.

His later and more important contribution was a faster procedure that worked with areas rather than perimeters, which allowed him to avoid repeated square roots. He also bounded the excess area with rectangles, which made his upper estimate sharper than Archimedes'. With a 3072-gon he reached 3.14159, and some sources credit him with 3.1416 from a 96-gon using his quick method. The attribution of these figures varies between sources, so the exact sequence of his steps should be read with care.

Zu Chongzhi

In the fifth century, during the Liu Song and Southern Qi dynasties, Zu Chongzhi (429–500) used the polygon approach to a much higher standard. According to the History of the Sui Dynasty, he established that π lies between 3.1415926 and 3.1415927, a bound that was not surpassed for nearly nine centuries. The calculation is generally believed to have used an inscribed 24,576-gon and many hundreds of square roots worked to about nine decimal places.

He also gave two rational approximations: 22/7, which he called the "approximate ratio", and 355/113, which he called the "close ratio" (milü). The fraction 355/113 is correct to six decimal places and is the best rational approximation of π with a denominator of four digits or fewer. The attribution of his methods remains uncertain. His treatise Zhui Shu is lost, so historians cannot be certain how he moved from the decimal bounds to 355/113. Some scholars suggest a fraction-interpolation technique attributed to his contemporary He Chengtian, while others doubt that this reconstruction is secure. The later commentator Zhao Youqin reportedly used a 16,384-gon to check the interval, which shows how the Chinese tradition continued to build on the polygon method.

Al-Kashi and European polygon computations

In the fifteenth century, the Persian mathematician Al-Kashi, working at Samarkand, computed 2π to 16 sexagesimal places, roughly 28 decimal digits, in his Treatise on the Circumference (about 1424). He achieved this with a polygon of about 3 × 2²⁸ sides and by carefully managing the arithmetic. This was the most precise value of its time and far beyond the reach of Archimedes' method.

European efforts followed in the late sixteenth and early seventeenth centuries. Ludolph van Ceulen spent years on polygon calculations and reached about 35 digits, a figure later published posthumously and often recalled as engraved on his tombstone. Christoph Grienberger reached roughly 38 or 39 digits in 1630 using polygons with an enormous number of sides; sources differ on the exact count. The polygon record was overtaken in 1699, when infinite series produced many more digits.

Limits of the polygon approach

The polygon method has a structural weakness: each doubling of the sides cuts the error by roughly a factor of four, so each additional correct digit costs much more work than the last. Archimedes stopped at 96 sides, and records such as Al-Kashi's required enormous computation. Liu Hui's area technique and the later use of infinite series show that the same problem was attacked in different ways, but only the series approach, developed from the fourteenth century onward, made high-precision computation practical.

Infinite series and products

The earliest infinite expressions for π came from India. The Kerala school of astronomy and mathematics, founded by Madhava of Sangamagrama (c. 1340–c. 1425), used the power series of the arctangent to obtain π. Setting the argument to 1 in the series for arctan gives $\pi/4 = 1 - \tfrac{1}{3} + \tfrac{1}{5} - \tfrac{1}{7} + \cdots$. Madhava's own works are lost, and his results are known through later Kerala writers such as Nilakantha Somayaji and the Malayalam Yuktibhāṣā (c. 1530). Sources differ on how many digits he obtained. The figure most consistently attributed to the tradition is 3.14159265359 (eleven decimals). The same series reached Europe in the seventeenth century, where it is usually called the Gregory–Leibniz series. Its terms alternate and shrink slowly, so its error after $n$ terms is roughly $1/(2n)$. Ten correct digits would therefore need on the order of billions of terms, which makes it unsuitable for high-precision work. The Kerala tradition addressed this with correction terms, and the Gregory–Leibniz series is the basic building block of that approach.

François Viète's 1593 formula is generally regarded as the first infinite product in European mathematics and the first exact analytic expression for π. It reads $\frac{2}{\pi} = \frac{\sqrt{2}}{2}\cdot\frac{\sqrt{2+\sqrt{2}}}{2}\cdot\frac{\sqrt{2+\sqrt{2+\sqrt{2}}}}{2}\cdots$. Viète, a French lawyer and amateur mathematician, derived it from the repeated use of the half-angle (double-angle) formula for cosine, which is equivalent to a telescoping product of the areas or perimeters of nested polygons. Its convergence is only linear, so it is a historical landmark rather than a practical method. Viète himself computed π to nine decimals with the polygon method of Archimedes instead, using a 393,216-sided polygon.

John Wallis published his product in Arithmetica Infinitorum in 1655. It states that $\frac{\pi}{2} = \frac{2}{1}\cdot\frac{2}{3}\cdot\frac{4}{3}\cdot\frac{4}{5}\cdots$. Wallis reached it by interpolation over a table of ratios between a square and its circumscribed circle, a procedure that he himself could not justify rigorously. Modern textbooks instead prove the product from the integrals $\int_0^{\pi/2}\sin^n x\,dx$, using the fact that the ratio of consecutive values approaches 1. Wallis's product converges faster than Viète's but is still too slow for computing many digits. Its importance lies in showing that π can be written as a limit of rational factors, a point that prepared the ground for the development of calculus and limit arguments in the following century.

Arctangent series became practical for computation through Machin's formula of 1706. John Machin showed that $\frac{\pi}{4} = 4\arctan\frac{1}{5} - \arctan\frac{1}{239}$. Because both arctangent arguments are small, their series converge rapidly. Machin used this identity to compute 100 digits of π by hand. The same idea, with other pairs of arctangent terms, was repeated by later calculators in the eighteenth and nineteenth centuries. William Shanks published 707 digits in 1873 by this method, and an error discovered in the twentieth century shows how hard long hand calculations were to check.

These series illustrate a general principle that runs through the history of π computation: the closer the argument of the series is to zero, the faster the series converges. The Gregory–Leibniz series, the Viète and Wallis products, and Machin-type formulas all rest on the same trigonometric and power-series foundations. Their differences in speed explain why they were replaced, in the twentieth century, by formulas and algorithms that were designed specifically for numerical analysis on machines.

Ramanujan-type and rapidly converging formulas

Srinivasa Ramanujan set down seventeen series for $1/\pi$ in 1914, and the sums he wrote down became a model for later work on digit computation. The survey literature covers the methods of proof for these formulas and generalizations that were found later, some not yet proven. His best-known example is $$\frac{1}{\pi} = \frac{2\sqrt{2}}{9801}\sum_{n=0}^{\infty}\frac{(4n)!\,(1103+26390n)}{(n!)^4\,396^{4n}}.$$ The square roots in such formulas sit in the constant prefactor, not inside the summands. Each term of this series adds about 8 digits of accuracy, so it is fast by the standards of its era but far slower than what came later. Ramanujan wrote few details of his proofs, and the first rigorous deductions were made by the Borwein brothers. Later work generalized the pattern into families such as the Ramanujan–Sato series, and recent constructions report much higher yields, with one arXiv preprint estimating about 153 digits per term for a new series of this type.

The Chudnovsky formula

The most important descendant of Ramanujan's work is the Chudnovsky formula, which the brothers David and Gregory Chudnovsky built from Ramanujan's ideas on modular equations and hypergeometric series. The Chudnovsky brothers introduced a formula for π based on Ramanujan's earlier work on modular equations and hypergeometric series. It is usually written as $$\frac{1}{\pi} = 12\sum_{n=0}^{\infty}\frac{(-1)^n\,(6n)!}{(3n)!\,(n!)^3}\cdot\frac{13591409+545140134\,n}{640320^{3n+3/2}}.$$ Each term yields a little over fourteen correct decimal digits; the figure is commonly given as on average 14.1816 decimal digits of π per iteration, and it is being used in most world record computations since 1989. The dating of the formula varies between sources: one gives 1987 as the year it was introduced, while another notes that it was published by the Chudnovsky brothers in 1988.

Fourteen digits per term does not by itself make the computation cheap. One analysis notes that the Chudnovsky series, taken alone, still has O(p²) computational cost due to the need to evaluate a large number of terms, each involving high-precision arithmetic. For this reason serious implementations use the Chudnovsky binary splitting method, which groups the terms recursively so that most of the work is done with large integer multiplications.

The Bailey–Borwein–Plouffe formula

The Bailey–Borwein–Plouffe (BBP) formula, discovered in 1995 by Simon Plouffe and named after the authors of the article in which it was published, David H. Bailey, Peter Borwein, and Plouffe, is a different kind of result. It reads $$\pi=\sum_{k=0}^{\infty}\frac{1}{16^{k}}\left(\frac{4}{8k+1}-\frac{2}{8k+4}-\frac{1}{8k+5}-\frac{1}{8k+6}\right).$$ Its importance lies in digit extraction rather than speed of convergence: it gives rise to a spigot algorithm for computing the nth base-16 (hexadecimal) digit of π without computing the preceding digits. This is also why the formula came as a surprise, since it had been widely believed that computing the nth digit of π is just as hard as computing the first n digits. It does not extend directly to decimal digits: this does not compute the nth decimal digit of π (i.e., in base 10). The method grew out of an observation about $\ln 2$, where Borwein and Plouffe found that binary digits could be computed beginning at an arbitrary starting position. Fabrice Bellard later found a formula that is about 43% faster than the Bailey–Borwein–Plouffe formula.

The Gauss–Legendre and Brent–Salamin algorithms

A different route to π comes from the arithmetic–geometric mean (AGM), not from series. The Gauss–Legendre algorithm, which Richard Brent and Eugene Salamin independently published in 1976, starts from $a_0=1$, $b_0=1/\sqrt{2}$, $t_0=1/4$ and $p_0=1$, and iterates $$a_{n+1}=\frac{a_n+b_n}{2},\quad b_{n+1}=\sqrt{a_nb_n},\quad t_{n+1}=t_n-p_n(a_n-a_{n+1})^2,\quad p_{n+1}=2p_n.$$ After each step, $\pi$ is approximated by $(a_{n+1}+b_{n+1})^2/(4t_{n+1})$. The method converges quadratically, meaning that each step roughly doubles the number of correct digits, so only a few dozen iterations are needed even for very large precision. The price is that every step needs a full-precision square root and multiplication, which makes the algorithm costly per digit compared with the Chudnovsky series. For that reason it is more often used as an independent check than as the main engine of record computations.

Digit computation and records

From hand calculation to electronic machines

The record for digits of $\pi$ moved onto machines in 1949. Von Neumann and colleagues used ENIAC to compute 2,037 digits, a calculation that took 70 hours. The run was partly a statistical experiment: John von Neumann was interested in whether the digits could give a measure of how random their distribution is. The hardware was large but had little memory. The machine filled a room and held over 17,000 vacuum tubes, yet its entire internal storage was 200 decimal digits. The following year, George Reitwiesner published an ENIAC determination of $\pi$ and $e$ to more than 2000 decimal places.

Digit counts then grew with each generation of hardware. The million-digit mark was passed in 1973. Through the 1980s and 1990s the frontier moved to the billions and then the trillions. The 1989 work of David and Gregory Chudnovsky pushed past the billion-digit mark on a homemade supercomputer, using a series that converges far faster than the older formulas. Yasumasa Kanada's group in Tokyo then held the record for much of the following two decades, reaching a trillion digits in 2002 with heavily optimised code on vector and parallel machines. By the late 2000s, the record had reached the low trillions. As of January 2010, the record was almost 2.7 trillion digits. Later that year, Shigeru Kondo and Alexander Yee reached five trillion digits, and in 2011 Kondo completed a run of about ten trillion, a computation that took 371 days.

Cloud and cluster era

Commercial and corporate hardware took over the record in the late 2010s. In 2019, Google calculated 31.4 trillion digits of $\pi$. In 2022, Google Cloud pushed the figure to 100 trillion digits, running y-cruncher across a massive fleet of cloud instances and moving tens of petabytes of I/O. Both runs relied on the y-cruncher program together with the Chudnovsky formula, which makes the arithmetic fast and leaves storage and data movement as the main constraints.

Independent groups then took the lead. StorageReview published its own 105 trillion-digit and 202 trillion-digit results in 2024. In May 2025, Linus Media Group and Kioxia announced 300 trillion digits. The computation ran on a storage cluster of 2.2 petabytes of Kioxia PCIe NVMe solid-state drives. The run lasted nearly seven and a half months. Guinness World Records officially verified the 300 trillion digits. Sources differ on the earlier Guinness benchmark, with one citing 62 trillion digits and another 64 trillion digits.

The 314-trillion-digit single-server run

The most recent record was set by StorageReview. It used a single 2U Dell PowerEdge R7725 server with dual AMD EPYC 192-core CPUs and forty 61.44 TB Micron 6550 Ion SSDs. The storage layer is where the authors say the record was actually won. Thirty-four of the 40 drives were allocated to y-cruncher, which gave the program about 2.1 petabytes of space for the 314-trillion-digit run. The remaining six drives formed a RAID 10 volume that stored the output. The calculation ran continuously for 110 days. Other reports give about 102 days or "nearly four months", so the exact duration varies by source. StorageReview reported sustained bandwidth of around 280 GB/s to handle the large stream of intermediate results.

These figures show why modern records are limited by data movement rather than arithmetic speed. Intermediate results can be far larger than the final digit string, and the program must write, read and check them repeatedly over months. Reliability over the run is therefore part of the achievement, and the storage bandwidth, memory layout and system stability all affect the outcome.

Verification

A computed expansion is only useful if it can be trusted, so each record is checked by a second, independent method. The usual approach is to recompute the digits with a different formula or algorithm and compare the results digit by digit. Because the two methods share few failure modes, agreement makes it very unlikely that a hardware fault or software bug has gone unnoticed. Techniques such as the Bailey–Borwein–Plouffe formula, which can extract individual hexadecimal digits without computing the earlier ones, are useful for spot-checks of this kind. Most verification is done this way rather than by trusting one run.

Normality and statistical properties of the digits

The normality conjecture

Pi is conjectured to be a normal number. In base ten, that would mean each digit occurs about one time in ten, each two-digit pair about one time in a hundred, and so on, so that a string of $k$ digits has limiting frequency $10^{-k}$. A number is said to be normal in a given base if all possible sequences of digits of a given length appear with equal frequency. The property is demanding. It requires uniformity at every block length, not only among single digits.

Borel showed in 1909 that almost every real number is normal in every base, a result that follows from ergodic theory applied to multiplication by the base modulo 1. These numbers were introduced in 1909 by É. Borel in an article where he proved that Lebesgue almost every number is normal in any base. That result is measure-theoretic and says nothing about any specific constant. For π, the question remains open. It is not difficult to construct explicit normal numbers, but there is no known example of a "natural" transcendental constant that is normal in every base. Even more basic facts are unknown: it is not even known if a given digit appears infinitely often in the base 10 expansion for π. Gowers observed that mathematicians widely believe each digit takes up about ten percent of the expansion in the long run, yet no proof or informal argument supports this belief.

Empirical tests

Because a proof is missing, digit records serve as the main evidence. New record computations of π are often used to perform empirical consistency checks for the normality of π. A 2016 study by Trueb examined the first 22.4 trillion decimal digits and counted the frequencies of all sequences of length one, two and three in base 10 and base 16. All evaluated frequencies are found to be consistent with the hypothesis of π being a normal number in these bases. The same computation reported that the largest deviation from the expected binomial variance was only about 1.33 standard deviations. Earlier analyses of the first six billion decimal places showed that each digit from 0 through 9 appears approximately six hundred million times, and such counts are typically assessed with chi-square goodness-of-fit tests.

These tests have clear limits. A finite prefix can only support or contradict finite-sample uniformity; it cannot establish a property defined by an infinite limit. Passing a battery of tests is also compatible with a hidden structure that the tests were not designed to detect. Consistency with randomness is evidence, not proof. Some researchers have proposed probabilistic models in which the digits of constants like π behave like independent uniform draws, and such models imply normality, but they remain conjectures.

Other bases and hexadecimal digit extraction

Normality is base-dependent. An example that demonstrates how this property depends on the base is the number 1/3 which is simply normal in base 2, but obviously not simply normal in the usual base 10. A number that is normal in every integer base is called absolutely normal, and empirical data strongly seems to suggest that constants like π, e, log 2 and the golden mean are absolutely normal, yet no proof nor disproof of normality in any base has ever been given for any of these numbers.

This is where hexadecimal digits matter. The Bailey–Borwein–Plouffe formula, which is a BBP-type formula, extracts hexadecimal digits from the numerical constant π beginning at any digit without having to compute any of the preceding digits. This makes it possible to sample blocks of base-16 digits far out in the expansion without computing everything before them, so base-16 statistics can be examined at positions that would otherwise be unreachable. Researchers have also proposed a route to proof: an approach to prove normality in base b for any transcendental constant which admits a BBP formula in base b is proposed. The authors of that line of work note that π itself is not covered by the known results, so the approach has not yet resolved the question for π.

Notable patterns: the Feynman point

The best-known local pattern is the Feynman point. The Feynman point is the sequence of six 9s which begins at the 762nd decimal place of π. The name comes from a supposed remark by the physicist Richard Feynman, who was said to want to memorize the digits up to that point so he could end with "and so on." The attribution is doubtful: it is not clear when, or even if, Feynman ever made such a statement, as it is neither mentioned in his memoirs nor is it known to his biographer James Gleick. The earliest known mention of this idea occurs in Douglas Hofstadter's 1985 book Metamagical Themas.

The pattern is interesting but not evidence against normality. For a normal number, the probability of a specific sequence of six digits occurring this early in the decimal representation is about 0.08%. Yet long random-like sequences contain many such runs, and some will appear early by chance. Another run of six 9s begins at position 193,034, so the Feynman point is not unique. Ordinary strings also recur freely: in the first million digits of π, the sequence '123456' occurs several times. The joke that the nines continue forever would make π rational, which it is not. Feynman points are therefore cultural curiosities rather than statistical findings.

Pi in calculus and complex analysis

Pi appears in analysis in places where no circle is drawn. The best-known case is the Basel problem, which asks for the value of the sum of reciprocal squares. Earlier mathematicians, including members of the Bernoulli family, could not find it. Leonhard Euler's 1735 solution showed that the sum equals $\pi^2/6$:

Euler's argument treated $\sin x / x$ as a polynomial with roots at $\pm n\pi$ and expanded it as an infinite product. Comparing the coefficient of $x^2$ with the one from the Taylor series of $\sin x$ gives the sum. The method is not fully rigorous by modern standards, but it was later justified. It generalizes: the same reasoning gives $\zeta(2k)$ as a rational multiple of $\pi^{2k}$, for example $\zeta(4) = \pi^4/90$ and $\zeta(6) = \pi^6/945$, where $\zeta$ is the Riemann zeta function. No comparable closed form in powers of π is known for $\zeta(3)$, $\zeta(5)$, and other odd arguments.

The Gaussian integral gives a second route by which $\sqrt{\pi}$ enters calculus:

The integral has no elementary antiderivative, so the standard evaluation squares it and converts the double integral to polar coordinates. The angular integral runs over a full turn and contributes $2\pi$, and the radial integral contributes $1/2$, so the square of the integral is $\pi$. The same value appears as $\Gamma(1/2)$ in the gamma function, and it fixes the normalizing constant of the normal distribution (discussed in the probability section).

Euler's identity connects π with the exponential function. For real $x$, the power series of $e^{ix}$ splits into those of $\cos x$ and $\sin x$, giving $e^{ix} = \cos x + i\sin x$. Setting $x = \pi$ yields

This is a direct consequence of the power series definitions and the exponential function's behaviour under trigonometric periodicity, so it is better read as a compact statement about how $\pi$ sets the period of the complex exponential than as a surprising coincidence.

Complex analysis gives a method that produces π from integrals of rational functions. For a function $f$ that is analytic inside and on a closed contour, the residue theorem states that the contour integral equals $2\pi i$ times the sum of the residues at the enclosed poles. The factor $2\pi i$ is a full turn in the complex plane. For example, $1/(1+z^2)$ has a simple pole at $z = i$ with residue $1/(2i)$, so

Closing the contour in the upper half-plane and letting its large arc vanish gives this value. Related definite integrals, such as $\int_0^\infty \frac{\sin x}{x}\,dx = \frac{\pi}{2}$, are commonly computed with contour or Fourier techniques, and these calculations are the usual route by which π appears in analysis without any geometric input.

Pi in geometry and trigonometry

Circles, disks and spheres

The circumference of a circle of radius $r$ is $C = 2\pi r$, and its area is $A = \pi r^2$. The same constant appears in three dimensions: a sphere of radius $r$ has surface area $4\pi r^2$ and enclosed volume $\tfrac{4}{3}\pi r^3$. The two circle formulas are linked. If a disk is cut into many thin sectors and rearranged into an approximate parallelogram, its height is $r$ and its base is half the circumference, so the area is $\tfrac12 \cdot 2\pi r \cdot r = \pi r^2$. The factor π in the area therefore comes from the same circular geometry that produces the factor 2π in the circumference. The sphere volume formula can be derived by summing the areas of circular cross-sections, which again brings in π.

Radian measure and periodicity

Angles are most naturally measured in radians. An arc of length $s$ on a circle of radius $r$ subtends the angle $\theta = s/r$, so a full turn is $2\pi$ radians, a half turn is $\pi$, and a right angle is $\pi/2$. Because the unit is defined through the circle itself, the constant is part of the definition rather than a conversion factor. This matters in analysis. With angles in radians, the derivative of $\sin x$ is exactly $\cos x$, and the Taylor series $\sin x = x - \tfrac{x^3}{3!} + \tfrac{x^5}{5!} - \cdots$ holds with no extra constants. In degrees, the same derivative would carry a factor of $\pi/180$.

Sine and cosine are periodic with period $2\pi$: $\sin(x+2\pi)=\sin x$ and $\cos(x+2\pi)=\cos x$. The number 2π is the fundamental period of both functions, which is why it governs the frequency definitions used in Fourier series and in signal processing.

The isoperimetric property

Among all closed plane curves of a given length, the circle encloses the largest area. Equivalently, among all shapes with a given area, the circle has the smallest perimeter. In inequality form, for a region with area $A$ and perimeter $L$, one has $4\pi A \le L^2$, with equality only for a circle. The question is ancient; the legend of Dido's problem is a famous example, and Greek mathematicians understood the result for regular polygons and circles. A rigorous general proof was not given until the nineteenth century, when Jakob Steiner produced a synthetic argument in the early 1840s. Here π appears as the constant that links area and perimeter in the extremal case.

Curvature and non-Euclidean geometry

The role of π depends on the geometry of the surface. On a sphere of radius 1, a circle of geodesic radius $r$ has circumference $2\pi\sin r$ and area $2\pi(1-\cos r)$. In the hyperbolic plane, where curvature is $-1$, the circumference is $2\pi\sinh r$ and the area is $2\pi(\cosh r-1)$. For small $r$, both expressions approach the Euclidean values, so the constant 2π is a local feature of the plane, not a global law of every space.

The Gauss–Bonnet theorem makes this connection exact. For a closed, two-dimensional Riemannian manifold, the integral of Gaussian curvature over the surface equals $2\pi$ times its Euler characteristic, $\iint_M K\,dA = 2\pi\chi(M)$. Thus the total curvature of a sphere is $4\pi$ however it is dented, because the Euler characteristic of a sphere is 2. For a geodesic triangle, the interior angles sum to $\pi + \iint K\,dA$. In the flat plane the sum is $\pi$; on a sphere the excess is positive, and on a hyperbolic surface it is negative. Here π measures how curvature and topology are balanced on a surface.

Pi in probability and statistics

The constant appears in probability most visibly in the normal distribution. Its density is $$\varphi(x)=\frac{1}{\sqrt{2\pi}}\,e^{-x^{2}/2},$$ and the prefactor $1/\sqrt{2\pi}$ is what makes the total area under the curve equal one. That area reduces to the Gaussian integral $\int_{-\infty}^{\infty}e^{-x^{2}}\,dx=\sqrt{\pi}$, and a rescaling of the variable turns $\sqrt{\pi}$ into $\sqrt{2\pi}$. Because the normal law arises as the limit of sums of many independent effects, as described by the central limit theorem, this constant carries into a large family of models. Other distributions show π directly or through special functions. The Cauchy density $1/(\pi(1+x^{2}))$ has π in its prefactor, and the normalizing constants of the chi-squared and Student $t$ distributions are written with the gamma function, whose value at one-half is $\sqrt{\pi}$.

The earliest well-known link between chance and π is Buffon's needle problem, posed by Georges-Louis Leclerc, Comte de Buffon, in 1777. If you drop a needle of length L on a plane containing parallel lines spaced a distance D from each other, the expected number of intersections of the needle with the lines is 2L/(πD). The restriction that the needle be no longer than the spacing is what makes this a probability of crossing, and the formula still holds if the needle is bent into any reasonable plane curve. Since π appears in the answer, an experimenter who counts crossings can invert the formula and obtain an estimate of π. Buffon's solution remained dormant for 35 years, until another great French analysis by Laplace, who recognized the same method as a way to estimate π. Laplace also remarked that the experiment would give π "... although with large error". The best-known physical run was made by Mario Lazzarini in 1901, who tossed a needle 3,408 times, and obtained the well-known approximation 355/113 for π, accurate to six decimal places. Later analysis suggests the set-up was tuned to that target. Lazzarini chose needles whose length was 5/6 of the width of the strips, and the count 3,408 is 213 × 16, consistent with a stopping strategy aimed at 355/113. Statisticians treat the result as a textbook case of confirmation bias. Hand experiments are in any case inefficient, since the probability of obtaining a correct answer to five decimal places in 3,400 throws would be less than 1.5 percent.

The Monte Carlo method applies the same principle with a computer. Points are sampled uniformly at random in a square, and the fraction that lands inside the inscribed circle estimates the ratio of the areas, which is π/4. The probability that a point is inside the quarter circle equals the area of the quarter circle divided by the area of the unit square, which is π/4. Multiplying the observed fraction by four gives the estimate, and it is an unbiased estimator of π/4. The weakness is speed of convergence. The standard deviation of the estimate falls only like the inverse square root of the sample size, roughly $1.64/\sqrt{N}$ for N points. Each additional correct digit therefore costs about a hundredfold more samples, and ten reliable digits would require on the order of $3\times10^{20}$ points. Small runs illustrate the noise: a single Monte Carlo calculation with 500 trials might suggest that Pi is 3.04, or that it is 3.20, depending on your luck. For high-precision work, this is why the infinite series and product formulas discussed elsewhere in this entry are far more practical, while the Monte Carlo approach remains valuable as a teaching device and as a general tool for problems without closed forms.

Stirling's approximation gives the asymptotic size of the factorial. The usual form is $$n!\sim\sqrt{2\pi n}\,\left(\frac{n}{e}\right)^{n},$$ where the sign ∼ means that the two quantities are asymptotic, that is, their ratio tends to 1 as n tends to infinity. The relative error is roughly $1/(12n)$, so the formula is already accurate for moderate n. The $\sqrt{2\pi}$ factor is not arbitrary. The factor of √(2π) falls out exactly from the Gaussian integral that also normalizes the normal distribution, so π enters the factorial asymptotics through the same mechanism.

In statistical mechanics and combinatorics, the factorial is usually needed in logarithmic form, because multiplicities of microstates are multinomial coefficients. For large N, the commonly quoted simplification $\ln N!\approx N\ln N-N$ yields the standard entropy expressions for an ideal gas. The generally called Stirling approximation is actually an approximation of Stirling's formula, and the term that is dropped is the one that takes fluctuations into account. Keeping the full formula therefore matters in some analyses. In combinatorics, the same asymptotics give binomial coefficients, for instance $\binom{2n}{n}\sim 4^{n}/\sqrt{\pi n}$, so π appears in the sizes of central probabilities in simple coin-tossing models.

Pi in physics and engineering

Physical theories seldom mention circles directly, yet π appears in many of their equations. The usual reason is geometric. A field or flux that spreads out from a point, line, or plane fills a sphere or cylinder, and the surface area of those shapes carries a factor of 4π or 2π. Coulomb's law is the standard case. In SI form the force between two point charges is $F = \frac{1}{4\pi\varepsilon_0}\frac{q_1 q_2}{r^2}$. The 4π comes from the area $4\pi r^2$ of a sphere over which the field lines spread.

The choice of unit system decides where these factors sit. In rationalized systems, including SI and Heaviside–Lorentz units, Maxwell's equations in differential form contain no explicit 4π, and the factor appears in the inverse-square force laws instead. In unrationalized Gaussian units, Coulomb's law has no 1/(4π), but Gauss's law carries a 4π in front of the charge density, as $\nabla\cdot\mathbf{E} = 4\pi\rho$. Heaviside–Lorentz units are rationalized in the same way, which partly explains their use in quantum field theory, since the Lagrangian then has no 4π factors. The historical background is awkward. Nineteenth-century electrical units based on Coulomb's and Ampère's laws produced many factors of 4π even in problems with no spherical or circular symmetry. From about 1892, Oliver Heaviside argued that the 4π belonged in the denominator of the force law, and his proposal led to the rationalized convention. In SI, the vacuum permeability μ₀ was defined as exactly $4\pi \times 10^{-7}$ N/A² until the 2019 redefinition of the ampere, after which μ₀ became a measured quantity.

In quantum mechanics, the factor 2π enters through the change from frequency to angular frequency. Planck's constant h relates energy to frequency, $E = h\nu$, while the angular frequency $\omega = 2\pi\nu$ gives the relation $E = \hbar\omega$, where $\hbar = h/2\pi$. Using ħ removes the repeated 2π that would otherwise appear in formulas for energy, momentum, and angular momentum. The Schrödinger equation, $i\hbar\,\partial_t\psi = \hat H\psi$, is written this way, as are the commutation relations and the uncertainty principle. Angular momentum is quantized in multiples of ħ, so the 2π is built into the basic quantum of action in its reduced form.

Fourier analysis shows the same pattern. A sinusoid of frequency f is written as $\sin(2\pi f t)$ or, in angular form, as $\sin(\omega t)$, and the choice of convention changes where the constants of the transform appear. Some texts define the transform with $e^{-i\omega t}$ and a normalizing factor, while others use $e^{-2\pi i f t}$ with none. The physics is identical, but the constants move. In signal processing and engineering, the same distinction matters for circuit impedances, where the angular frequency ω appears in expressions such as $1/(j\omega C)$, and for wave phenomena, where the wavenumber is $k = 2\pi/\lambda$. In these fields π enters both through the periodicity of the sinusoid and through the definition of the unit of frequency.

Physicists disagree about what this pattern means. One view holds that π in a law reflects real geometry, such as the total solid angle of $4\pi$ steradians around a point source or the spherical flux that Gauss's law counts. On this view the appearance of π is informative. A second view treats the factors as artifacts of how units were chosen, and holds that a well-designed system should absorb them. The debate has practical consequences. Engineers usually prefer SI because it keeps permittivity and permeability as explicit quantities, while many theoretical texts prefer Gaussian or Heaviside–Lorentz conventions. The gravitational case shows that π is not confined to electromagnetism. In general relativity the field equations carry $8\pi G$, and the factor 8π is fixed by matching the Newtonian limit, where the gravitational potential obeys an equation with $4\pi G\rho$. Whether such constants reveal the structure of nature or the conventions of description remains an open interpretive question.

Notation and the tau debate

Origin of the symbol

In 1706 the mathematician William Jones printed the first known use of π for the ratio of circumference to diameter, in his Synopsis palmariorum matheseos. He is listed in MacTutor as William Jones (1675-1749). The choice of letter is usually explained as an abbreviation of perimetron, the Greek word for perimeter, though this motive is inferred from the choice rather than documented in Jones's text.

Jones's notation drew little attention at first. Its spread depended on Leonhard Euler. Sources give different dates for his adoption. MacTutor reports that in 1736 Euler wrote 1 : π in his Mechanica, and he popularized the symbol in his 1748 Introductio in Analysin Infinitorum. Other accounts put his adoption in 1737. The idea that Euler chose the letter as the initial of his own name is considered unlikely, since he was notably modest about credit.

Earlier and regional names

Before a single symbol became standard, the ratio was described in words or as fractional bounds. Archimedes' tradition of giving the upper and lower limits as fractions persisted into the seventeenth century. Ludolph van Ceulen's tombstone, for example, still gives its bounds as decimals placed in the Archimedean layout.

The number also carried personal names. In Germany π was once called the Ludolphine number. In the Netherlands it is often still referred to by that name. The name honours Ludolph van Ceulen, who published a 20-decimal value in his 1596 book Van den Circkel and later extended it to 35 decimals. He requested that the two limiting numbers be carved on his gravestone.

The tau proposal

The case for replacing π with τ = 2π has two stages. Bob Palais's 2001 article "π Is Wrong!" argued that the circumference-to-radius ratio is the more natural circle constant. Michael Hartl then popularized the idea through his 2010 Tau Manifesto. Other people had suggested the same symbol independently before or around that time.

Hartl gave two reasons for τ. τ is the number of radians in one turn, and both τ and turn begin with a /t/ sound. Second, τ visually resembles π, whose association with the circle constant is unavoidable. Hartl also argued that the factor 2π appears so often in mathematics that writing it as a single constant would clarify many formulas, and that unnecessary factors of two arising from π tend to cancel awkwardly when divided by even numbers. He called the resulting results absurd.

Responses

Proponents concede that use of τ in place of 2π does not change any of the underlying mathematics; their claim concerns clarity. Critics have pointed to cases where the change would make formulas harder to read. For example, torque involves circular motion, which must involve a circle constant, so those formulas would get hairier if each 2π were replaced with τ. Hartl replied that people underestimate how good physicists, engineers and mathematicians are at handling that kind of notational ambiguity.

Hartl also blamed the academic community for inertia, saying that a new symbol needs a name, popularization and typesetting support, and that mathematicians have had over 300 years to fix this π problem. Physics has partly adopted the alternative in practice: the reduced Planck constant, written ħ, is h divided by 2π, and this usage is now standard. Even so, π remains the dominant notation in textbooks and research. Tau advocates have continued to promote the idea; they call for an annual Tau Day on June 28 to raise awareness of the concept.

Computing practice and approximations

Storing π in binary arithmetic

Almost no program stores π exactly. Most languages use the 64-bit double format defined by IEEE 754, which has a 52-bit significand plus an implicit leading bit. That format gives roughly sixteen decimal digits of precision (almost 16 digits of precision). The double closest to π is 3.141592653589793115997963468544185161590576171875, which matches π to about sixteen digits (approximates pi accurately to about 16 decimal digits). No finite string of binary digits can represent π, so every program works with a rounded value (This is the case in any base, so if we wanted to store π in base-2, we'd need an infinite number of bits in memory.).

Standard libraries mostly share this value. Python's math.pi is the fastest no-dependency answer for most scripts, and NumPy's constant is identical to it. MATLAB's pi returns the floating-point number nearest to the value of π in IEEE double-precision. Some APIs state the constant with a few more printed digits, but the stored value is still the same 64-bit number. For example, the .NET Double.Pi field is written as 3.1415926535897931.

Why sin(π) is not zero

Floating-point rounding is visible in trigonometry. In Python, math.sin(math.pi) returns about 1.22 × 10⁻¹⁶ rather than 0. The stored constant lies slightly below π, and near π the sine changes at a rate close to −1, so the output is roughly the gap between the two numbers. MATLAB documents the same behaviour, noting that the result is not exactly zero in floating point arithmetic. Some functions avoid this problem. MATLAB's sinpi function computes sin(Xpi) without explicitly computing Xpi, and for integers, sinpi(n) is exactly zero.

Different languages and hardware may still disagree in the last bits of trigonometric results. Libraries use different algorithms and rounding strategies, so exact bit-level reproducibility across platforms is not guaranteed, even when the input constant is identical.

Error propagation from rounding

Errors from a truncated constant scale linearly with the quantities they multiply. Replacing π with 3.14 introduces an absolute error of about 0.0016, so for a circle of radius 1000 m the computed circumference is about 3.2 m too short. Replacing π with the double value changes the result by roughly 10⁻¹⁶ times the radius, which is far below any physical measurement. Errors of this size matter mainly when subtracting nearly equal quantities, where cancellation can amplify small rounding differences; this is a central concern of numerical stability analysis in numerical-analysis.

How many digits are needed

For most engineering work, double precision is more than enough. Popular accounts often cite the Jet Propulsion Laboratory, which uses only 15 digits of pi for its highest-accuracy calculations for interplanetary navigation. A frequently repeated example concerns a circle as large as the Voyager 1 distance, where the circumference error is less than 1.5 inches. Other retellings use a 30-billion-mile circle and state an error of less than half an inch (about one centimeter). These figures differ among secondary sources, so they are best read as illustrations of scale rather than as official specifications.

The same reasoning shows where extra digits stop helping. The limiting factor in navigation is usually the accuracy of physical measurements, not the constant. For a circle as large as the observable universe, one source notes that about 40 digits of Pi would give an error smaller than a single hydrogen atom.

Random sampling, integration and special functions

Scientific software uses π in many routine places. Random-angle generators typically map a uniform variable on $[0,1)$ to an angle by multiplying by $2\pi$, and the rounding of the stored constant perturbs the resulting distribution only at roughly the 10⁻¹⁶ level, which is negligible for simulation. Monte Carlo methods estimate π by sampling points in a square and counting the fraction that falls inside the inscribed quarter circle. The standard error falls roughly as $1/\sqrt{N}$, so about 10⁸ samples yield an error on the order of 10⁻⁴, far less efficient than the series discussed elsewhere in this article (see Monte Carlo method).

Numerical integration and special-function libraries also depend on π. Quadrature tables, normalization constants for normal distributions, and values such as $\Gamma(1/2)=\sqrt{\pi}$ for the gamma function are normally stored in double precision, so their accuracy is bounded by the same 16-digit limit. Library authors therefore usually treat the stored constant as exact for practical purposes while still testing their routines against higher-precision references.

Cultural history and public recognition

Pi's public profile rests largely on one day. In March 1988, physicist Larry Shaw of the Exploratorium, a science museum in San Francisco, linked the date 3/14 with the first digits of π during a staff retreat. The first observance was small. Staff walked around a circular plaque, the "Pi Shrine," which bears the constant's first hundred digits, and then ate fruit pies. The event was set for 1:59 p.m., the digits that follow 3.14. Later celebrations at the museum kept these elements: a procession, pie, and "Happy Birthday" sung to Albert Einstein, whose birthday is also March 14. Shaw was later nicknamed the "Prince of Pi."

The holiday spread from one museum to the wider public. The U.S. House of Representatives backed a formal designation in 2009, and the observance moved into classrooms, where teachers use it for lessons, and into bakeries and grocery stores, which often discount pies that day. Pi Day also became a setting for memorization. Competitions in reciting digits, a pursuit sometimes called piphilology, are now a standard part of the celebrations. Some years later, UNESCO adopted March 14 as the International Day of Mathematics, which shows how far a joke from one museum's staff retreat has travelled.

The public also knows π through a legislative episode. In 1897, Indiana House Bill 246, introduced by Representative Taylor I. Record at the urging of the physician Edwin J. Goodwin, claimed to offer a "new mathematical truth" for squaring the circle. Its wording implied that the ratio of diameter to circumference was 5/4 to 4, that is, π = 3.2. The House passed the bill 67 to 0 on February 5, 1897. A Purdue mathematics professor, Clarence A. Waldo, who happened to be at the statehouse, helped stop it, and the Senate postponed it indefinitely in February. Accounts differ on the details: some describe senators laughing before tabling the bill, and the text itself is so contradictory that one mathematician, David Singmaster, counted several different values for π within it. The episode is usually told as a comedy, but it is a documented example of trying to set a mathematical fact by statute.

Pi also appears in fiction, though often in titles rather than as the constant itself. Carl Sagan's novel Contact (1985) and its 1997 film adaptation center on a message hidden in the digits of π, a plot device that plays on the public belief that the digits contain meaning. Darren Aronofsky's film Pi (1998) follows a mathematician obsessed with finding patterns in numbers. Yann Martel's novel Life of Pi (2001) takes its title from a character's nickname, not from the constant. The appeal of the digits lies in their apparent randomness combined with their endless, exact, and easily stated character. Many people who would never compute a circle's area know that the digits never end.

Approximation days show a different kind of public appeal. Pi Approximation Day falls on July 22, written 22/7 in the day-month format. The fraction 22/7 is accurate to two decimal places and dates back to Archimedes, so the day celebrates a simple rational stand-in rather than the constant itself. A rival observance, Tau Day, falls on June 28 (6/28) and honors 2π, reflecting the proposal to replace π with τ = 2π. Taken together, these observances show that the public has long preferred a memorable number, a date, or a plate of pie over the abstract ratio.

Open problems and future directions

Whether π is normal

The most famous open question is whether π is a normal number. Normality in base ten would mean that every block of digits of a given length occurs with the same long-run frequency. No one has been able to establish that π is b-normal for any integer b, much less for all bases simultaneously. The empirical picture is strongly favorable. In the first six billion decimal places of pi, each of the digits from 0 through 9 shows up about six hundred million times. A 2016 check of 22.4 trillion decimal digits examined every block up to length three in bases ten and sixteen, and all evaluated frequencies were found to be consistent with the hypothesis of π being a normal number in these bases.

These results are not proofs, and the gap is structural. Finite data can only falsify normality in a limited sense, and testing any finite number of digits will never suffice for a proof. A weaker property, disjunctivity, asks only that every finite string appear somewhere in the expansion, and we don't know whether pi is disjunctive either. Consequently, the open problem is not only to find the right statistics but to find a proof technique that goes beyond counting. Researchers have argued that more powerful techniques must be brought to bear on the question of normality before real progress is possible.

Irrationality measure and transcendence of combinations

Known results are much weaker for sums and products of known constants. The irrationality measure μ(π) quantifies how well π can be approximated by rationals. The Zeilberger–Zudilin bound of 2020 gave 7.103205334137... improving on the value 7.606308 previously found by Salikhov (2008). A 2026 preprint claims a further small reduction to 7.101862832357 (the arXiv listing reports this as a refinement of the same method rather than a new structural result). Such bounds are steadily improving, but they remain far from the value many experts suspect. It is widely suspected that μ(π) = 2, and progress has shown that μ(π) ≤ 7.104..., with this bound steadily decreasing over time. A proof of μ(π) ≤ 5/2 would have a concrete consequence: the convergence of the Flint-Hills series would be settled.

Transcendence is a different matter. The Lindemann–Weierstrass theorem settles many algebraic combinations of exponentials, but it does not decide whether $\pi + e$ or $\pi e$ is transcendental. Both numbers are believed to be transcendental, and it is known that at least one of them must be, but neither has been individually proven. The broader question of which sums and products of familiar constants are transcendental remains largely open, and few general tools exist to attack it.

The limits of digit records

Digit records continue to advance, and they raise questions about what limits remain. In May 2025, a joint team of Kioxia and Linus Media Group announced that they had successfully calculated 300 trillion digits. StorageReview then reported 314 trillion digits, using a single 2U Dell PowerEdge R7725 server equipped with dual AMD EPYC 192-Core CPUs and forty 61.44 TB Micron 6550 Ion SSDs. The run ran continuously for approximately 110 days before completion.

The bottleneck is increasingly storage rather than arithmetic. The raw output of the run, including checkpoints, extended beyond 2PB of data. Future records will therefore depend on sustained storage bandwidth, power, cooling and months-long reliability as much as on new algorithms. A further question is whether faster series or hardware will keep extending the frontier indefinitely. The practical demand is slight: according to NASA's Jet Propulsion Laboratory, only about 37 decimal places are needed to calculate the circumference of the observable Universe to within the width of a hydrogen atom. The larger digit counts therefore serve benchmarking, hardware testing and statistical checks rather than any physical requirement.

Structure in the digit sequence

Whether the digits contain detectable structure is also unresolved. Graphical methods have been used to display digits as random walks, but large-scale statistical analyses on the digits of π have so far produced no clear verdict on normality. Machine-learning approaches have been tried as well: a 2018 study noted that it is not known if π, or π' (in a binary disguise), is normal or not, although the first 30 million digits of π are quite uniform.

Open questions include whether compression algorithms can distinguish π's digits from random sequences of the same length, and whether spectral analysis of the digit sequence reveals any regularity. Answering such questions would show whether the digits are random in a meaningful algorithmic sense, a claim that current evidence does not establish.

References

1. faculty of science and engineering Irrationality proofs of e and πfse.studenttheses.ub.rug.nl
2. Johann Heinrich Lambert (1728 - 1777) - Biography - MacTutor History of Mathematicsmathshistory.st-andrews.ac.uk
3. Transcendence of e and π Taha Rhaouti April 2023 1 Introduction 1.1 e and πjonathanlove.info
4. info prev up next book cdrom email homearchive.lib.msu.edu
5. Some transcendence results from a harmless irrationality theoremar5iv.labs.arxiv.org
6. The Irrationality Measure of Pi is at most 7.103205334137...arxiv.org
7. The irrationality measure of {\pi} is at most 7.101862832357arxiv.org
8. Proof that pi is irrationalen.wikipedia.org
9. Some transcendence results from a harmless irrationality theoremarxiv.org
10. Proof that π is irrational - HandWikihandwiki.org
11. www.arxiv.orgarxiv.org
12. Linear Differential Equations as a Data-Structurearxiv.org
13. graphsearch.epfl.chgraphsearch.epfl.ch
14. THE IRRATIONALITY MEASURE OF π IS AT MOST 7.103205334137 . . .sites.math.rutgers.edu
15. A New World Record for the Irrationality Measure of Pivimeo.com
16. Irrationality Measure -- from Wolfram MathWorldmathworld.wolfram.com
17. Two hypergeometric tales and a new irrationality measure of $\zeta(2)$arxiv.org
18. Open Diophantine Problemsarxiv.org
19. Mesopotamian Slidespublic.csusm.edu
20. Pythagorean Theorem in Elamite Mathematicsarxiv.org
21. Pibritannica.com
22. Babylonian mathematicsen.wikipedia.org
23. Units of Measurement: Pi in the Ancient Egyptian Rhind Mathematical Papyrus Brewminate: A Bold Blend of News and Ideasbrewminate.com
24. Rhind Mathematical Papyrus — Grokipediagrokipedia.com
25. Hekaten.wikipedia.org
26. Circleen.wikipedia.org
27. Pi Before Pimentalfloss.com
28. Rhind Mathematical Papyrusplanetmath.org
29. Why does the Bible miscount the value of pi in 1 Kings 7:23?biblehub.com
30. A Reconstruction, Assessment, Error Analysis and Simulation of a Method for Measuring {\pi} That Could Have Been Used 3000 Years Agoarxiv.org
31. www.slideshare.netslideshare.net
32. Circle - MacTutor History of Mathematicsmathshistory.st-andrews.ac.uk
33. Origins: 3.14159265… - Biblical Archaeology Societybiblicalarchaeology.org
34. Approximations of π — Grokipediagrokipedia.com
35. How Archimedes showed that π is approximately equal to 22/7arxiv.org
36. Archimedes and Pics.miami.edu
37. Archimedes' calculations of square rootsarxiv.org
38. Approximations of pien.wikipedia.org
39. Ramanujan in Computing Technologyarxiv.org
40. Liu Hui’s 𝜋 algorithm vs Archimedes’ 𝜋 algorithm - Maths from the Pastmaths-from-the-past.org
41. Liu Hui（a famous mathematician of the Wei and Jin Periods）_Baiduwikibaike.baidu.com
42. Liu Huien.wikipedia.org
43. Liu Hui%27s %CF%80 algorithmen.wikipedia.org
44. Method of exhaustionen.wikipedia.org
45. Zhao Youqinen.wikipedia.org
46. Zu Chongzhien.wikipedia.org
47. Zu Chongzhi (429 - 501) - Biography - MacTutor History of Mathematicsmathshistory.st-andrews.ac.uk
48. Mil%C3%BCen.wikipedia.org
49. Zu Chongzhi - New World Encyclopedianewworldencyclopedia.org
50. Madhava seriesen.wikipedia.org
51. On M\=adhava and his correction terms for the M\=adhava-Leibniz series for $\pi$arxiv.org
52. Kerala School - Ancient Indian Mathematics - Ancient Indian Mathematicsoboe.com
53. Madhava - The Founder of The Kerala Schoolstoryofmathematics.com
54. Madhava series - Bharatpediaen.bharatpedia.org
55. Madhava series — Grokipediagrokipedia.com
56. Madhava of Sangamagrama: Infinite Series for π, Sin & Cos — 300 Years Before Newtondsremo.com
57. Viète's formulaen.wikipedia.org
58. Oldest series for pijohndcook.com
59. On the periodic continued radicals of 2 and generalization for Vieta productarxiv.org
60. Wallis producten.wikipedia.org
61. New derivation of pi links quantum physics and pure mathphys.org
62. An elementary proof of Wallis’ product formula for pi Johan W¨astlunddiva-portal.org
63. Wallis-Ramanujan-Schur-Feynmanarxiv.org
64. arXiv:0712.1332v2 [math.NT] 19 May 2008 Ramanujan-type formulae for 1/π:arxiv.org
65. Practical Implementation of π Algorithms.hvks.com
66. (PDF) A method for proving Ramanujan series for 1/πresearchgate.net
67. Ramanujan–Sato series — Grokipediagrokipedia.com
68. An extension of the Chudnovsky algorithmarxiv.org
69. A detailed proof of the Chudnovsky formula with means of basic complex analysis -- Ein ausf\"uhrlicher Beweis der Chudnovsky-Formel mit elementarer Funktionentheoriearxiv.org
70. Chudnovsky algorithmen.wikipedia.org
71. Bailey%E2%80%93Borwein%E2%80%93Plouffe formulaen.wikipedia.org
72. The BBP Algorithm for Pi David H. Bailey∗ September 17, 2006 1. Introductiondavidhbailey.com
73. Bellard%27s formulaen.wikipedia.org
74. [2104.12412] On the elegance of Ramanujan's series for $π$arxiv.org
75. (PDF) The BBP Algorithm for Piresearchgate.net
76. History of the formulas and algorithms for piarxiv.org
77. A174277 - OEISoeis.org
78. The Eniac and Piwittenberg.edu
79. The first 4,000,000 digits of Pi, visualized in a single imagegizmodo.com
80. Pi Day: Breakthrough 'Obliterates' The World Record For Calculating Pi : ScienceAlertsciencealert.com
81. StorageReview Sets New Pi Record: 314 Trillion Digits on a Dell PowerEdge R7725 - StorageReview.comstoragereview.com
82. Pi has been calculated to trillions of digits ‪—‬ is that completely irrational?livescience.com
83. KIOXIA and Linus Media Group set world record for pi calculationeurope.kioxia.com
84. Kioxia-powered storage cluster helps push PI calculation to a whopping 300 trillion digitstechradar.com
85. A single server just crushed cloud computing with 314 trillion digits of pi and far less powertechradar.com
86. Pi Day Special: From Archimedes to Modern Computing - Kronecker Walliskroneckerwallis.com
87. Kioxia and Linus Media Group Set World Record for Pi Calculationapac.kioxia.com
88. 300 trillion digit Pi value calculated in new world record by KIOXIA and Linus Tech Tipstweaktown.com
89. Pi calculated to 314,000,000,000,000 digits breaks world record - GIGAZINEgigazine.net
90. Can $\pi$ generate itself? A Monte Carlo analysis of 314 trillion digitsarxiv.org
91. An Analysis of Digit Summation Patterns in the Expansion of Pi: An Inquiry into Order and Randomnesszenodo.org
92. On the genesis of BBP formulasarxiv.org
93. Digit Statistics of the First 22.4 Trillion Decimal Digits of Piarxiv.org
94. Hausdorff's forgotten proof that almost all numbers are normalarxiv.org
95. faculteit Wiskunde en Natuurwetenschappen BBP-numbers and normalityfse.studenttheses.ub.rug.nl
96. Pi Fractions for Generating Uniformly Distributed Sampling Points in Global Search and Optimization Algorithmsarxiv.org
97. Feynman point - Wikipedia, the free encyclopediataggedwiki.zubiaga.org
98. Six nines in pien.wikipedia.org
99. What does feynman point mean?definitions.net
100. Does π contain all possible number combinations? - GeeksforGeeksgeeksforgeeks.org
101. Normality of pi tested with 22.4 trillion decimal digitspi2e.ch
102. Chance and the Continuum Hypothesisarxiv.org
103. Gauss%E2%80%93Bonnet theoremen.wikipedia.org
104. Gauss-Bonnet Formula -- from Wolfram MathWorldmathworld.wolfram.com
105. Gauss-Bonnet Theorem - an overviewsciencedirect.com
106. Isoperimetric inequalityen.wikipedia.org
107. Isoperimetric inequality — Grokipediagrokipedia.com
108. The Isoperimetric Inequality: Proofs by Convex and ...scholar.rose-hulman.edu
109. How many zeros of a random polynomial are real?arxiv.org
110. Chapter 2 THE BUFFON’S NEEDLE PROBLEM: FIRST MONTE CARLO SIMULATIONmragheb.com
111. Three Lectures on Probability and Statisticsarxiv.org
112. Buffon%27s needle problemen.wikipedia.org
113. Monte Carlo estimates of pi and an important statistical lesson - The DO Loopblogs.sas.com
114. The wrapping hull and a unified framework for volume estimationarxiv.org
115. Introduction to Monte Carlo Methods in Chemistrypeople.chem.ucsb.edu
116. Stirling's approximationen.wikipedia.org
117. Stirling's Approximation: How Big Is n Factorial, Really?world-of-mathematics.com
118. (PDF) Plea for the use of the exact Stirling formula in statistical mechanicsresearchgate.net
119. Buffon's needle problemen.wikipedia.org
120. Estimating Pi Using Monte Carlo Methodsmedium.com
121. Stirling's Approximation Calculator - CalculatorHubcalculatorhub.com
122. Heaviside%E2%80%93Lorentz unitsen.wikipedia.org
123. Gaussian unitsen.wikipedia.org
124. Babel of Units. The Evolution of Units Systems in Classical Electromagnetismarxiv.org
125. 1.5: Coulomb's Law - Physics LibreTextsphys.libretexts.org
126. The vacuum impedance and unit systemsarxiv.org
127. From Qubit to Qubit: A Graduate Course in Quantum Mechanicsarxiv.org
128. Reduced Planck's Constantfiveable.me
129. Planck relationen.wikipedia.org
130. An Ahistorical Approach to Elementary Physicsarxiv.org
131. Ferromagnetic Phase Transitions in Neutron Starsarxiv.org
132. Earliest Uses of Symbols for Constants - MacTutor History of Mathematicsmathshistory.st-andrews.ac.uk
133. Digits of Piams.org
134. Ludolph Van Ceulen (1540 - 1610) - Biography - MacTutor History of Mathematicsmathshistory.st-andrews.ac.uk
135. Ludolph van Ceulen - The Linda Halllindahall.org
136. Ludolph van Ceulenen.wikipedia.org
137. Tau (mathematics) — Grokipediagrokipedia.com
138. Tau (mathematics)en.wikipedia.org
139. THE TAU MANIFESTO Michael Hartl Tau Day, 2010 updated Tau Day, 2023tauday.com
140. The Tao of Tauscientificamerican.com
141. The Tau Manifesto Michael Hartl Tau Day, 2010 updated Half Tau Day, 2013 1blogimages.bloggen.be
142. No, really, pi is wrong: The Tau Manifestotauday.com
143. A Brief History of Pi (π)exploratorium.edu
144. Michael Hartlen.wikipedia.org
145. The Tau Manifesto Michael Hartl Tau Day, 2010 1 The circle constanthexnet.org
146. Embedded Wednesdays: Floating Point Numbers — Embeddedembedded.fm
147. Pi and e In Binary - Exploring Binaryexploringbinary.com
148. Why $\sin(\pi)\neq0$ in python — Keith Pedersenhepguy.com
149. How to Write Pi in Python: math.pi, numpy.pi, scipy, and Moredocs.kanaries.net
150. pi - Ratio of circle's circumference to its diameter - MATLABmathworks.com
151. Double.Pi Fieldlearn.microsoft.com
152. lost-contact.mit.edulost-contact.mit.edu
153. cs400-web.cs.wisc.educs400-web.cs.wisc.edu
154. even after 22 trillion digits were still no closer to the end of pifivethirtyeight.com
155. The Pi Paradox: Why NASA and Your Computer Only Care About the First 15 Digitsoddlyverified.com
156. Your pi, my piideapharma.substack.com
157. How Many Decimal Places is Pi Accurate in Single, Double, and Extended-Precision Floating-Point Numbers? — codegenes.netcodegenes.net
158. www.ibm.comibm.com
159. Floating-Point Numbers - MATLAB & Simulinkmathworks.com
160. Pi Dayen.wikipedia.org
161. Larry Shaw (physicist)en.wikipedia.org
162. Guide to Celebrating Pi (π) Dayexploratorium.edu
163. What is Pi Day and why do we celebrate the holiday? - ABC Newsabcnews.com
164. Pi Day: San Francisco Exploratorium honors holiday it madesfexaminer.com
165. Indiana pi billen.wikipedia.org
166. Pi Saved - LIX (Polytechnique)lix.polytechnique.fr
167. The Eccentric Crank Who Tried To Legislate The Value Of Pigizmodo.com
168. The Indiana Pi Bill and the Attempt to Redefine Mathematics Through Legislative Fiat - Commonplace Fun Factscommonplacefacts.com
169. π (Pi) Approximation Day: July 22risingkashmir.com
170. Pi Day is upon us again and we still do not know if Pi is normalcarmamaths.org
171. Are The Digits Of Pi Random? A Berkeley Lab Researcher May Hold The Keysciencedaily.com
172. Does Pi Contain All of Shakespeare?scientificamerican.com
173. Bounds on Irrationality Measures and the Flint-Hills Seriesarxiv.org
174. Backblaze Now Serving 314 Trillion Digits of Pibackblaze.com
175. Learning from Pseudo-Randomness With an Artificial Neural Network - Does God Play Pseudo-Dice?arxiv.org
176. The irrationality measure of π is at most 7.103205334137...msp.org