Arithmetic–Geometric Mean¶

The arithmetic–geometric mean is the common limit of coupled arithmetic-mean and geometric-mean iterations on two positive numbers.

Version: v1 · 2026-10-03 · History
Domain-specific #: 12989
Domain group: Formal Sciences
Origin domain: Mathematics
Subdomains: Elliptic Integrals, Numerical Analysis → Mathematics
Aliases: AGM, AGM Method

Core Idea¶

For two positive real numbers $a_0$ and $g_0$, repeatedly replace one by the arithmetic mean and the other by the geometric mean of the current pair:

\[a_{n+1}=\frac{a_n+g_n}{2},\qquad g_{n+1}=\sqrt{a_ng_n}.\]

Both sequences approach the same value $M(a_0,g_0)$, the arithmetic–geometric mean (AGM). The result is neither the first arithmetic mean nor the first geometric mean; those are successive bounds or approximations. The input order does not matter, because both update rules are symmetric. NIST's Digital Library of Mathematical Functions (DLMF) defines the AGM this way and gives a quadratic recurrence for the gap between the iterates.^[1]

The construction matters numerically because that gap contracts very rapidly near the common limit. It also connects to complete elliptic integrals. A formula can turn evaluation of such an integral into evaluation of an AGM, and additional recurrence variables can turn related transformations into a fast approximation of π. The last step needs care: the AGM limit alone is not a formula for π. The Gauss–Legendre/Brent–Salamin algorithm contains a correction sequence as well as the two AGM iterates.^[1]^[2]

Structural Signature¶

Sig role-phrases:

Positive input pair: $a_0,g_0>0$ supplies a well-defined positive square root and an ordered real comparison. Complex AGM variants need branch rules and are outside this entry's claim.
Arithmetic update: $a_{n+1}=(a_n+g_n)/2$ draws one coordinate toward the center of the current pair.
Geometric update: $g_{n+1}=\sqrt{a_ng_n}$ draws the other by multiplication and a positive square root. Both updates use the same current pair; updating one then using that new value in the other is a different algorithm.
Common limit: the paired updates shrink their separation toward $M(a_0,g_0)$. A finite iterate is a numerical approximation, not the definition's terminal value.
Reduction identity: for an external problem, an independently justified identity relates the wanted object to AGM values. The identity is an application layer, not part of AGM membership.^[1]

In compressed form: positive pair → simultaneous arithmetic and geometric updates → common limit, optionally followed by a proved relation to a target quantity.

What It Is Not¶

Not the elementary arithmetic or geometric mean. Those are single operations. AGM names the infinite coupled construction's limit.
Not a generic average of more than two numbers. Multivariate generalizations exist, but the named classical AGM here is a two-input construction.
Not the whole Gauss–Legendre π algorithm. That algorithm starts the same two recurrences at $1$ and $1/\sqrt2$ but tracks a separate correction term to recover π.^[2]
Not automatic complex arithmetic. Over positive reals, $\sqrt{a_ng_n}$ has a chosen positive value and inequalities constrain the iterates. Complex square roots are multivalued; the simple real squeeze proof cannot be copied without additional conventions.

Scope of Application¶

The core mathematical scope is positive real inputs. Since arithmetic mean is at least geometric mean, after one update the arithmetic coordinate is no smaller than the geometric coordinate. Both remain positive. The upper coordinate decreases and the lower increases; they therefore squeeze toward a common finite value. DLMF gives a sharper relation: if $c_n=\sqrt{a_n^2-g_n^2}$, then $c_{n+1}=(a_n-g_n)/2=c_n^2/(4a_{n+1})$. Squaring of the small separation explains local quadratic convergence.^[1]

The elliptic-integral application is precise, not a loose resemblance. For $0\le k<1$, take $a_0=1$ and $g_0=\sqrt{1-k^2}$. Then the complete elliptic integral of the first kind satisfies $K(k)=\pi/[2M(1,\sqrt{1-k^2})]$. DLMF also identifies AGM-based computation of complete elliptic integrals. The relation is real and parameter-bounded here; no claim is made about every incomplete or complex elliptic integral.^[1]^[3]

Brent's original high-precision work uses AGM and elliptic transformations for elementary functions and constants, including an algorithm for π. This is a second setting in the sense of algorithmic role: the AGM iteration becomes a fast computational component inside a larger formula. It is not an independent physical setting, and the transfer does not license calling every rapidly convergent algorithm “AGM.”^[2]

Clarity¶

The two coordinates serve different update rules even though the limit is symmetric in the initial inputs. Writing $M(a,g)$ distinguishes the limit from $A(a,g)=(a+g)/2$ and $G(a,g)=\sqrt{ag}$. For example, starting at $1$ and $1/\sqrt2$, the first arithmetic update is $(1+1/\sqrt2)/2$, whereas the first geometric update is $2^{-1/4}$. Neither is the final AGM. A subsequent update must use those two first-step values together.^[1]

“Quadratic convergence” describes asymptotic reduction of error, not a promise that any finite-precision machine doubles correct digits at every step. Rounding in multiplication and square root can arrest progress. An implementation must increase arithmetic precision and stop according to an error test appropriate to the target formula, not count iterations from a slogan.^[1]^[2]

Manages Complexity¶

An elliptic integral initially looks like a continuous accumulation over an interval. The AGM identity moves a complete first-kind case to two scalar state variables and repeated square roots. That replaces difficult direct quadrature with a rapidly shrinking gap. It does not abolish analytical work: the reduction identity and parameter conversion must be established first. If the wrong complementary modulus is supplied, the fast iteration will accurately compute the wrong integral.^[1]^[3]

Likewise, the π algorithm needs an auxiliary correction sequence. This exposes an important complexity boundary: a compact computational kernel may be reusable without being the whole application. Omitting the correction because the AGM converges quickly destroys the claimed π computation.^[2]

Abstract Reasoning¶

Assume $a_0\ge g_0>0$. After each simultaneous update, $g_{n+1}\le a_{n+1}$, while $g_n\le g_{n+1}$ and $a_{n+1}\le a_n$. The sequences are bounded and monotone. If their limits were distinct, applying the same update equations at the limits would require two distinct positive numbers to equal both their arithmetic and geometric means, impossible by strict AM–GM. Thus the limits coincide. DLMF's $c_{n+1}=c_n^2/(4a_{n+1})$ shows why the gap then falls quadratically.^[1]

This reasoning separates correctness of the limit from utility of a reduction. The squeeze works for any positive pair. The elliptic identity works only after matching $k$ to $\sqrt{1-k^2}$. The π method requires still more structure. If a supposed AGM application lacks a proved reduction, one can compute a beautiful common limit without having solved the external problem.

Knowledge Transfer¶

The literal transfer is between mathematical tasks whose formulas genuinely reduce to an AGM: complete elliptic integrals and high-precision elementary-function/constant algorithms. What travels is the coupled recurrence and its fast convergence, with new identities supplying task-specific output extraction.^[3]^[2]

In optimization or organizational change, an “arithmetic–geometric mean process” may be a metaphor for balancing two kinds of averaging. Unless the state is two positive numbers updated by these equations, it is not this mathematical construction. A broader prime about iteration can describe many refinement loops; it cannot supply the AGM's special limit or its elliptic identities by itself.

Examples¶

Complete elliptic integral at $k=1/\sqrt2$¶

The first-kind complete integral $K(k)$ is $\int_0^{\pi/2}(1-k^2\sin^2\theta)^{-1/2}\,d\theta$. For $k=1/\sqrt2$, its complementary parameter is $\sqrt{1-k^2}=1/\sqrt2$. Set $a_0=1$, $g_0=1/\sqrt2$, then iterate the two mean rules. The first pair is $a_1=(1+1/\sqrt2)/2\approx0.8535533906$ and $g_1=2^{-1/4}\approx0.8408964153$. Once the common limit is approximated, DLMF's identity yields $K(1/\sqrt2)=\pi/[2M(1,1/\sqrt2)]$. The same starting pair, without a correction sequence, is not a numerical π algorithm.^[1]

Mapped back: the two positive starting values are the input pair; the displayed first-step values instantiate the arithmetic and geometric updates; their common limit is $M$; the $K$-identity is the independently justified reduction layer.

Brent–Salamin computation of π¶

Brent's original 1976 paper, §5, gives an AGM-based π algorithm with $a_0=1$, $g_0=1/\sqrt2$, correction $t_0=1/4$, and weight $p_0=1$. One simultaneous round gives $a_1=(a_0+g_0)/2\approx0.8535533906$ and $g_1=\sqrt{a_0g_0}\approx0.8408964153$; the correction becomes $t_1=t_0-p_0(a_0-a_1)^2\approx0.2285533906$, and $p_1=2p_0=2$. Brent's improved output formula then gives $\pi_1=(a_1+g_1)^2/(4t_1)\approx3.1405792505$. This first approximation is close to, but not equal to, π. If $t_1$ is omitted, the remaining operation still computes an AGM but not π by this algorithm.^[2]

Mapped back: the positive pair and two executed mean updates are the AGM kernel; the common limit supplies rapid convergence; the explicitly computed $t_1$ and $\pi_1$ are application-specific reduction and output, not hidden rules in the AGM definition.

Structural Tensions¶

Further steps versus precision cost. The gap falls quadratically, so another round may deliver many additional correct digits. Yet each high-precision multiplication and square root consumes time, and after the target accuracy is reached or rounding dominates, another round may produce no usable gain. Stopping early risks truncation error; continuing indefinitely spends precision work without improving the answer. Diagnostic: under the intended output formula, is the remaining AGM gap (plus rounding error) below the requested tolerance?^[1]^[2]

The cited AGM recurrence and squeeze result are stated for positive real inputs. Complex inputs require a separately specified branch convention and proof; one cannot silently transfer the positive-real argument. That is a scope boundary for this entry, not a second competing objective within its admitted recurrence.^[1]

Structural–Framed Character¶

The AGM lies near the structural end: given positive inputs, the equations and common limit are mathematical facts independent of a user's purpose. Evaluative weight appears only in calling it fast or useful relative to arithmetic cost and a numerical target. Human practice determines which target to reduce and what precision is worth paying for, but not whether the recurrence has its limit. The historical names “arithmetic” and “geometric” are standard mathematical vocabulary; they travel as definitions across fields without importing an institution's authority. Recognizing the same coupled equations in a new numerical problem is literal; calling any compromise or hybrid average an AGM merely imports the label. Its character: a structural mathematical limit generated by a specific coupled recurrence, with task-framed value in numerical computation.^[1]

Structural Core vs. Domain Accent¶

The skeletal relation is feedback through repeated state update, with the current pair feeding the next pair. The live prime Iteration is an independently challenged prerequisite under composition/presupposes, not a taxonomic genus of the limiting value. The domain mechanism is narrower and indispensable: one arithmetic and one geometric update on positive reals, a shared limit, and elliptic identities that sometimes extract a target. The named AGM is not a prime because those equations and positive-real conditions do not survive transfer to unrelated organizational, biological or engineering iterations. Conversely, stripping the equations down to “repeated improvement” loses the AGM. No broader new prime is needed unless an independently recurring coupled-mean skeleton is demonstrated beyond this mathematics.

This entry presupposes Iteration.

Iteration is the strict prerequisite under composition/presupposes: the AGM value requires repeated coupled state updates to the common limit, but the value is not itself a procedure. Convergence is related to the limit and rate, but not asserted as an additional strict parent. An elliptic integral or π algorithm is an application, not a parent.

Relationships to Other Abstractions¶

Current abstraction Arithmetic–Geometric Mean Domain-specific

Parents (1) — more general patterns this builds on

Arithmetic–Geometric Mean presupposes Iteration Prime

The AGM value presupposes coupled iteration.

Hierarchy path (1) — routes to 1 parentless root

Arithmetic–Geometric Mean → Iteration

Neighborhood in Abstraction Space¶

Arithmetic–Geometric Mean sits in a sparse region of the domain-specific corpus (84^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Number-Theoretic Properties & Tests (20 abstractions)

Nearest neighbors

Least Common Multiple — 0.83
AM–GM Inequality — 0.82
Average Order of an Arithmetic Function — 0.82
QR Decomposition — 0.81
Hensel's Lemma — 0.81

Computed from structural-signature embeddings · 2026-10-08

Not to Be Confused With¶

Single-step arithmetic and geometric means: ingredients of the recurrence, not its limiting value.
Gauss–Legendre/Brent–Salamin π algorithm: uses AGM iterates plus an independent correction state.^[2]
An arbitrary fast fixed-point iteration: quadratic convergence alone does not identify AGM.

References¶

[1] NIST Digital Library of Mathematical Functions, §19.8(i), Gauss's arithmetic–geometric mean, especially equations 19.8.1–19.8.4. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m

[2] Richard P. Brent, “Fast multiple-precision evaluation of elementary functions,” JACM 23 (1976), 242–251, original full paper, §5 algorithm for π and improved return formula. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i

[3] NIST Digital Library of Mathematical Functions, §19.36, methods of computation, complete elliptic integrals. registry ↩a ↩b ↩c