Arithmetic-geometric mean iteration Created 2026-10-07 Updated 2026-10-14
Starting from , replace the lower endpoint by the geometric mean and the upper endpoint by the arithmetic mean. The arithmetic-geometric mean inequality gives . The bounded monotone sequence theorem gives limits, and the arithmetic recurrence forces them to coincide. The limiting value is the arithmetic-geometric mean of the initial pair.
BLEU 2026-11-05
BLEU is a reference-based metric for evaluating machine translation. It combines modified n-gram precision at several lengths using a geometric mean, then applies a brevity penalty to discourage very short outputs. Kishore Papineni and colleagues introduced it as a way to make repeated translation evaluation less costly than repeatedly collecting human judgments.[ref]
For each candidate n-gram, count matches only up to its maximum occurrence count in a single reference. Sum these clipped matches across the corpus and divide by the total number of candidate n-grams of that length to obtain . This clipping prevents repeated use of a common word from earning unlimited matches. Corpus aggregation uses counts, rather than simply averaging independently calculated sentence scores.[ref]
BLEU: a Method for Automatic Evaluation of Machine Translation by Kishore Papineni and others. §2.1, footnote 2; §2.1.1, corpus modified precision
With candidate length , effective reference length and weights summing to one,The original baseline uses and . Effective reference length sums the closest reference sentence lengths for the candidates.[ref]
For a synthetic one-pair calculation, take reference
the cat is on the mat and candidate the cat is on mat. The clipped precisions areHere , and the precision product is , givingThis is about on a scale multiplied by one hundred; it is not translation accuracy.The unsmoothed score becomes zero when a required precision is zero. Short sentences make this particularly conspicuous, while legitimate rephrasing can reduce overlap. State tokenization, casing, reference set, smoothing and scale when comparing results. The example illustrates the arithmetic; the paper emphasizes corpus-level use rather than treating each sentence score as an independent human-quality judgment.[ref]
Bose condensation in a three-dimensional harmonic trap Created 2026-10-18 Updated 2026-10-20
Consider noninteracting bosons in one internal quantum state, at thermal equilibrium in the three-dimensional harmonic oscillator potentialHere is the mass, the positive are angular frequencies, and is their geometric mean. Measure the one-particle energy and chemical potential from the ground-state energy. The quantum harmonic oscillator levels then have with nonnegative integer and reduced Planck constant .
When , counting lattice points can be approximated by the volume of a tetrahedron in the positive octant of :This is a density of states, with units of inverse energy. The factor is the unit tetrahedron volume; differentiating produces the factor . Treat the single ground state separately: its macroscopic population would be lost by replacing every level with a smooth density of states.
For , integrate the Bose-Einstein distribution over excited energy eigenstates. As approaches zero from below, their maximum population isHere is the Boltzmann constant and the Riemann zeta function. A short proof uses and . The nonnegative terms permit exchanging summation and integration by Tonelli's theorem, giving . Near zero the integrand is proportional to , so its integral converges. That finite excited-state capacity is the mechanism of Bose-Einstein condensation.
Set . Particle conservation gives, to this continuum accuracy,Above , the macroscopic condensate fraction is zero and the chemical potential adjusts below zero. This cubic power belongs to a three-dimensional harmonic trap; it is not the homogeneous three-dimensional ideal gas exponent. For , and , the model gives . At , , or approximately condensed bosons. These are chosen parameters and calculated values.
Finite gives a smooth crossover rather than a sharp phase transition. The exact ideal gas sum over excited quantum harmonic oscillator levels differs from the continuum integral, especially when is not large relative to every . A sharp trapped-gas limit takes , with fixed and fixed frequency ratios. Interactions also alter the critical temperature, condensate fraction and release profile; a dilute gas can have appreciable interaction energy compared with the trap level spacing. The cubic law is a useful baseline, not a fit guaranteed for either 1995 experiment.
Branch-current mismatch in a resistance ratio measurement is a residual scale error when the sensor and reference resistor carry unequal electric currents. A common supply or common nominal current setting does not establish the shared excitation required by a shared-excitation resistance ratio measurement.
Suppose separate branches carry positive currents and . Ideal potential sensing givesThus the fractional resistance error is exactly . If and , the common scale cancels, whileIn an unbranched series circuit the same electric current passes through both components; this particular error instead concerns separate paths, leakage, or unequal effective sampling times.
For an original example, use , , and . Their potentials are and . The inferred resistance is , high by approximately . Each current is only from its nominal setting, but the opposed errors add to first order.
A useful ideal comparison is current-source swapping in a resistance ratio measurement. In two otherwise unchanged acquisitions, assign source A to the sensor and source B to the reference, then exchange them. If , the two ratios obeyThis exact geometric mean cancellation assumes unchanged source outputs and resistances between acquisitions; switching transients and sensor self-heating in a ratiometric resistance measurement can violate that assumption. The primary circuit guide independently treats current-source mismatch and source interchange in its three-wire lead-cancellation circuit; that topology's correction must follow its own equations rather than be replaced by the two-branch model above.[ref]
A proposed test measures both branch currents through characterized sensing resistors, then changes their assignments while holding the sensor environment fixed. Agreement after swapping tests the mismatch model; agreement from merely changing the common nominal current tests a different cancellation.
Brevity penalty 2026-11-05
The brevity penalty in BLEU reduces the score of a candidate translation corpus that is too short relative to its references. Matching a few reference phrases can give excellent modified n-gram precision while omitting much of the intended translation; the length factor addresses that particular failure.[ref]
BLEU: a Method for Automatic Evaluation of Machine Translation by Kishore Papineni and others. §2.2.2, motivation and sentence brevity penalty
For positive total candidate length and effective reference length ,Equivalently, . The effective reference length sums closest reference sentence lengths across the corpus; the penalty is then applied to the corpus totals.[ref]
For an own example, use reference
the cat is on the mat and candidate the cat is on. Every candidate n-gram of lengths one through four matches the reference, so all four modified precisions are one. Nevertheless, and giveWith the usual four equal weights, the full score in this example is also , because its precision geometric mean is one. Perfect candidate precision has not made an incomplete prefix a perfect translation.At , the factor is one. It remains one for longer candidates; their unmatched or repeated material can instead lower precision. A brevity penalty is therefore a length correction within a combined metric, not a complete measure of translation adequacy.
Let be the Bernoulli distribution with success probability and the Bernoulli distribution with success probability . For ,The two summands have product . Their arithmetic mean is at least their geometric mean, with equality at . ThusThe same argument gives for complementary parameters . At the laws agree and .
With equal prior probabilities, count the successes among observations. The likelihood ratio is , so choose when and when . If is even, either decision or a fair tie randomization gives the same average error at . Under , has the binomial distribution with parameters . Symmetry therefore gives the exact optimal average errorThe last term applies only for even ; it represents half the tie probability. Direct finite sums give , and . Twenty observations reduce the optimal average error to about 0.158%. The finite-sample Chernoff error bound at is , a guarantee rather than the exact error.
Exact repeated-coin error, the finite-sample Chernoff error bound, and the limiting error exponent
. Original finite-binomial calculations under two fixed independent coin laws with equal priors. The first panel uses a logarithmic scale. The normalized negative log error approaches ; finite-sample error need not equal the simple exponential. Details.Reducing the exact error from to illustrates the value of repeated evidence. The example assumes a single fixed coin law throughout the experiment. If repeated sensor readings share an unmodelled common error, increasing their number need not supply this much independent information.
The continuous angular geometric mean of the positive boundary complex modulus isHere the empty product is one. This differs from the angular arithmetic mean . In general by the concavity of the logarithm; equality requires constant boundary complex modulus. For a holomorphic function with no zeros of a function in the closed disk, Jensen's formula gives , even when the boundary complex modulus varies with angle. Thus replacing the mean of a logarithm by the logarithm of an arithmetic mean loses the exact zero identity.
Common-gain cancellation in heterodyne overlap uncertainty is the exact invariance of the normalized beat amplitudewhen the corrected peak beat current and both separately measured arm currents are multiplied by the same factor . The factor occurs once in the numerator and once in the denominator. This is scale invariance: the ratio is a homogeneous function of degree zero. The cancellation holds for every realization of a random common factor, without requiring it to be independent of the currents.
More generally, take and let their residual multiplicative measurement calibration factors be . The observed normalization satisfiesThus the gain-matching condition for heterodyne normalization is ; the three factors need not all be equal. For example, , and leave the ratio unchanged. The denominator depends on the geometric mean of its two factors.
Writing the natural logarithms of the factors as a random vector makes the remaining uncertainty particularly clear. With ,This identity is exact for the logarithmic error. The relative variance of the original ratio is only approximately equal to it when the errors are small, as in the variance of a corrected heterodyne overlap estimate. Keeping the covariance terms follows the general measurement-propagation rule.[ref]
For an original imperfect-sharing model, set . Assume are zero-mean independent random variables. The coefficient of the shared fluctuation is , soTake , , with standard deviations for , for and for each arm's separate error. Then and , giving logarithmic standard uncertainty , approximately relative uncertainty. A large shared fluctuation can leave a small residual, but incomplete matching prevents exact cancellation.
If the log-errors additionally have a normal distribution, then and the normalized readout has a lognormal distribution. Its exact expected value is , not . At , this example gives mean and standard deviation . The lognormal bias of a heterodyne overlap estimate comes from exponentiating the remaining error; the perfectly shared factor still cancels.
The physical calibration paths determine which model applies. NIST's receiver experiment combined two lasers through a beam splitter, polarizer and single-mode optical fiber, while separately correcting electrical loss and impedance mismatch.[ref]
Frequency response metrology for high-speed optical receivers by Paul D. Hale and others. PDF pp. 1–2, Heterodyne measurements and equations 1–4
An electronic gain change confined to the beat-frequency path need not appear in the arm-current measurements. Likewise, adding a current offset is not multiplication by a common factor. Background correction and the sharing of calibration factors must therefore be established before claiming cancellation.Current-source swapping in a resistance ratio measurement exchanges two excitation sources between the sensor and reference resistor, allowing an unequal-current contribution to be identified or canceled. The combining rule depends on the circuit topology. In the separate-branch model of branch-current mismatch in a resistance ratio measurement, the exact combination is a geometric mean.
Let and let the two positive source currents have a fixed ratio . With source A on the sensor and B on the reference, the measured potential ratio is . After exchanging their assignments, . HenceThis assumes that the resistances, current outputs, potential-sensing gains and relevant thermal conditions are unchanged across the two acquisitions. It cancels the multiplicative source ratio, not an additive readout error or an inaccurate assigned reference resistance.
For an original example, let and . The acquisitions give and . Their geometric mean is exactly . Their arithmetic mean is , high by approximately . In general,The arithmetic mean suppresses a small mismatch to second order here, but is not the exact inverse-pair cancellation.
The primary RTD guide uses source interchange in a different three-wire lead-cancellation circuit. Its two readings are averaged according to that circuit's equations; they must not be replaced automatically by the geometric-mean rule above.[ref]
A drift residual after current-source swapping remains if the source ratios in the two acquisitions are and . For unchanged , the same calculation gives . A increase from to therefore makes the combined estimate low by . Sensor self-heating in a ratiometric resistance measurement can also change between source assignments.
A proposed apparatus check switches two characterized current sources between stable sensing resistors, waits for electrical settling, and compares both individual readings and their product. Reversing the order of the two acquisitions helps distinguish a repeatable source mismatch from drift during the sequence. This tests the stated circuit model without presuming that all source-swapping circuits use the same estimator.
For sampling times and , the geometric mean has normally distributed under the stock-numeraire measure in the Black-Scholes model. Its mean is , and its variance is . For descending times the sum is . The drift shift comes from exponential tilting, and must be included when the delivered payment contains a factor .
Drift residual after current-source swapping 2026-11-05
A drift residual after current-source swapping remains when the current ratio changes between the two source assignments. The geometric mean cancellation in current-source swapping in a resistance ratio measurement is exact for a fixed multiplicative mismatch; it does not make a time-varying mismatch constant.
Let the true resistance ratio remain fixed. At the two acquisition times , write . With opposite source assignments,For , the logarithmic error is . A full-interval bound therefore givesFor an original example, a logarithmic source-ratio drift of over a acquisition separation produces . This is a underestimate for that drift direction. Shortening the separation reduces this particular residual, provided both readings still follow electrical settling.
A useful original logarithmic variant of an ABBA measurement sequence uses source-assignment signs at equally spaced times . The four readings obey . Combining them as givesBoth the constant and linear logarithmic drift terms cancel because the signed time sums vanish. Unequal timing, nonlinear drift and a changing sensor resistance can leave residuals. If itself varies, the combination includes its geometric average over the acquisitions rather than automatically recovering a single-time value.
The primary RTD guide establishes source interchange in its three-wire circuit; its topology-specific arithmetic averaging is distinct from the separate-branch logarithmic model above.[ref]
A Basic Guide to RTD Measurements by Joseph Wu. §1.4.2, printed p. 6, Figure 1-6 and equations (9)–(12)
A proposed bench test imposes a slow controlled source-ratio variation, timestamps the readings, and compares a two-reading swap with the balanced four-reading sequence while using stable sensing resistors.The gain-matching condition for heterodyne normalization is the relation between three positive multiplicative calibration factors that leaves a normalized beat unchanged. If the corrected peak beat current is and the separate arm currents are , thenExact cancellation therefore requires , or equivalentlyThe beat-channel gain must equal the geometric mean of the arm-channel gains. The three gains need not individually agree.
An original example uses and . Their geometric mean is 0.99, so gives exact cancellation. Keeping instead multiplies the inferred overlap by . Choosing the arithmetic mean 1.01 for would not solve the problem: it would multiply the result by approximately 1.02020. These are stipulated calibration factors, not receiver specifications.
For time-dependent gains, the condition must hold for the corresponding measurements in each realization if sample-by-sample invariance is claimed. Merely arranging zero mean logarithmic mismatch does not guarantee an unbiased arithmetic overlap: lognormal bias of a heterodyne overlap estimate gives a counterexample under Gaussian log-errors.
A heterodyne RF-to-DC gain transfer check distinguishes electronic amplification in the beat-frequency path from the response used for the arm currents. A change confined to does not cancel. Neither does a generic additive dark-current error satisfy this multiplicative model. Preserve the amplitude convention as well: is a corrected peak current, rather than the unconverted electrical beat power.
The NIST receiver experiment used two lasers, a beam splitter, a polarizer and single-mode collection to establish optical matching, while separately correcting the electrical measurement path.[ref]
Frequency response metrology for high-speed optical receivers by Paul D. Hale and others. PDF pp. 1–2, Heterodyne measurements and Eqs. 1–4
Those different physical operations explain why one common scalar gain is a special case rather than an automatic description of the entire apparatus.One practical test varies a suspected common gain while holding powers, mode matching and collection conditions fixed, and checks whether the normalized result stays constant. Passing that test constrains response to that change; it does not exclude an unchanged systematic mismatch. Report the gain-sharing model and the remaining uncertainty before interpreting normalization as complete calibration cancellation.
Geometric Asian option 2026-10-14
An Asian option whose averaging uses the geometric mean. In the constant-coefficient Black-Scholes model, the logarithm of the continuous geometric average is Gaussian with mean and variance , because . Its forward mean is . The discounted lognormal call formula gives its price. It is strictly cheaper than the terminal European call when and , since both its forward mean and log-variance are no larger, with strictly smaller variance. No unconditional comparison holds for arbitrary negative interest rates: for , sufficiently small positive strikes reverse the inequality.
Geometric mean Created 2026-10-07 Updated 2026-10-13
The geometric mean replaces a list of nonnegative real numbers by one repeated value that preserves their product of numbers. For inputs,For two positive inputs, is their positive square root. The geometric mean of and is three: replacing by retains the same product. In contrast, their arithmetic mean five retains the sum . The geometric mean is zero if any input is zero; the logarithm formula below requires every input to be positive. [ref]
https://dlmf.nist.gov/1.2.iv. NIST Digital Library of Mathematical Functions, §1.2(iv), equation 1.2.18: geometric mean
.Its usefulness becomes immediate for successive growth factors. Suppose a hypothetical positive quantity starts at , is multiplied by , then by . It goes , so the geometric mean factor is : a constant factor of one for both steps gives the same endpoint. The arithmetic mean factor is ; repeating it would give , a different endpoint. A mean of factors summarizes multiplicative change; subtracting one afterwards gives the equivalent change per step. Equal-duration steps are necessary to interpret this as change per common time interval.
For positive inputs, logarithms turn multiplication into addition, soThus the geometric mean is an arithmetic mean in logarithmic coordinates, followed by the exponential function. For positive measurements with a common unit, one can use dimensionless ratios to a fixed reference : . Changing the common unit rescales by the same factor as the inputs. As a check, doubling one input and halving another retains while usually changing the arithmetic mean.
The geometric mean preserves multiplicative endpoints and sits between the harmonic mean and arithmetic mean
. Left: the original toy factors and return a quantity to its starting value; repeating their arithmetic mean would produce a larger endpoint. Right: the dimensionless inputs and keep product one while , , and separate as the inputs become less equal. Lines joining time steps are guides, not a continuous growth model. Details.The arithmetic-geometric mean inequality gives . Apply that same inequality to the multiplicative inverses to obtain , where is the harmonic mean. Consequentlyfor positive inputs, with equality throughout exactly when all inputs agree. For and , the three values are , , and . They preserve, respectively, the sum of inverses, the product, and the sum. No one of these summaries preserves all three for unequal inputs.
Geometric-mean refinement of an asymptotic sequence Created 2026-09-24 Updated 2026-11-05
Let be a positive asymptotic sequence. Define and for , using the positive geometric mean. ThenThe latter limit follows by multiplying the two adjacent ratios of the original asymptotic sequence. However,Thus these are new intermediate scales, not termwise equivalent replacements. For as , the new scales are . Positivity is needed for this real square-root construction.
Growth factor (mathematics) 2026-11-05
For positive quantities measured at successive steps, the growth factor is a dimensionless ratio. It can be below one when the quantity shrinks. Repeated multiplication givesso the common factor reproducing the endpoint after identical steps is their geometric mean. A factor corresponds to a relative change ; factors must be averaged before subtracting one when summarizing compounded change.
For example, increasing by ten percent gives , then decreasing the result by ten percent gives . The factors are and , whose product is . Their geometric mean is , whereas their arithmetic mean is one. Thus equal positive and negative percentage changes need not cancel when applied successively to different bases.
Harmonic mean Created 2026-10-06 Updated 2026-10-14
The harmonic mean of a nonempty finite list of positive real numbers is the multiplicative inverse of the arithmetic mean of their inverses:For two inputs, . Replacing each input by preserves the sum of inverses, making it useful when a fixed amount divided by a rate determines a cost or duration. This positive-input definition avoids division by zero; negative inputs need not retain the usual averaging or ordering properties. [ref]
https://dlmf.nist.gov/1.2.iv. NIST Digital Library of Mathematical Functions, §1.2(iv), equation 1.2.19: harmonic mean
.Travel two legs of equal length at positive speeds , with no stops. Their travel times are and , so the average speed isFor two legs at and , the times are two hours and one hour. The average speed is . The same speeds maintained for one hour each instead cover in two hours, giving the arithmetic mean . A slower equal-distance leg occupies more time, so the data being averaged must include what is held equal. This is total distance traveled divided by elapsed time; a return journey can have zero average velocity while keeping a positive average speed. [ref]
https://undergroundmathematics.org/introducing-calculus/average-speed/averages. University of Cambridge, Underground Mathematics: equal-distance and equal-time speed averages
.The arithmetic-geometric mean inequality applied to the inverses gives ; together with it yields the comparison in geometric mean. For two inputs the arithmetic–harmonic gap has the exact formFor and the gap is five. With one positive input held fixed and another approaching zero, the harmonic mean approaches zero too: a very slow fixed-distance leg dominates the travel time. Unequal leg lengths instead require a weighted harmonic mean.
Jensen's inequality 2026-10-14
Jensen's inequality says that applying a convex function to a weighted mean cannot exceed the same weighted mean of its outputs. For nonnegative weights with and inputs in a convex set,Nonnegative weights are essential: this is a convex combination, rather than an arbitrary affine combination. For two inputs it is precisely the defining inequality for a convex function; repeatedly grouping inputs gives the finite weighted version.
Take , weights and , and inputs and . Their weighted mean is , so the two sides areThe Jensen gap is . The inequality measures the effect of replacing distinct inputs by one representative input before applying a nonlinear function.
The probabilistic form uses an integrable random variable taking values in an interval where is a convex function. If is also an integrable random variable and belongs to that interval, thenThe direction reverses for a concave function. For a strictly convex function, equality in the finite weighted form requires all inputs having positive weight to be equal; in the probabilistic form it requires to be constant almost surely. An affine function gives equality for any admissible distribution. [ref]
https://stanford.edu/~boyd/cvxbook/bv_cvxslides.pdf. Boyd and Vandenberghe's slides, section 3.14, state the expectation form
.For the quadratic function and a random variable with a finite second moment, the difference between the two sides is exactly its variance:Consider two equally likely values and . Their expected value stays at as increases, but their variance is . Thus the expected value of the square grows from to . For the exponential function, the same two-point distribution gives , while . Both are examples of Jensen's inequality, but their Jensen gaps depend differently on the spread of the inputs.
Another useful choice is the concave function on positive inputs. Its version of Jensen's inequality gives . Applying the increasing exponential function shows that the weighted geometric mean cannot exceed the weighted arithmetic mean. For the equally weighted inputs and , these are and . This recovers the arithmetic-geometric mean inequality and illustrates how choosing one function can generate an entire family of bounds.
A short proof idea explains the probabilistic form. If has a subgradient at , the supporting affine function gives . Taking the expected value cancels the linear term because . This argument applies, for example, to finite convex functions on ; more general domains of a function need attention to endpoints and the existence of a subgradient.
Johan Jensen's 1906 paper “Sur les fonctions convexes et les inégalités entre les valeurs moyennes” is an early source for the connection between convex functions and inequalities for means. [ref]
https://zenodo.org/records/2371297. Jensen's original article and publication record, DOI 10.1007/BF02418571
. Kishore Papineni 2026-11-05
Kishore Papineni is a machine translation researcher who coauthored the 2002 paper introducing BLEU. Its original byline places the authors at IBM's research center. The method supplied an automatic way to compare machine output with human reference translations, making repeated evaluation less dependent on commissioning a fresh human assessment for every system change.[ref]
BLEU: a Method for Automatic Evaluation of Machine Translation by Kishore Papineni and others. Page311: original byline, affiliation, Abstract and motivation
One key idea is modified n-gram precision. Repeating a common word cannot earn unlimited matching credit, because the candidate's matched counts are clipped by the counts available in the references. This connects a cheap string-counting procedure with an explicit defense against a trivial way to inflate ordinary overlap.[ref]
BLEU: a Method for Automatic Evaluation of Machine Translation by Kishore Papineni and others. Section2.1: clipped n-gram counts
For an original unigram example, let the candidate be
red red red blue and the reference be red blue green white. Both candidate word types appear in the reference, but clipping permits only one occurrence of red and one of blue, giving . This is only the unigram component: the full BLEU calculation also combines higher-order n-gram precisions through a geometric mean and applies a brevity penalty.The original method aggregates counts over a corpus. It is not simply the arithmetic mean of independently scored sentences, and a short example explaining one component should not be presented as the complete metric. Its usefulness as an automatic comparison also does not turn word-sequence overlap into a direct proof that every sentence conveys the correct meaning.[ref]
BLEU: a Method for Automatic Evaluation of Machine Translation by Kishore Papineni and others. Section2.1.1: corpus aggregation; Sections2.2.2–2.3: brevity penalty and combined score
Logarithmic midpoint 2026-10-18
For positive endpoints , halfway along a logarithmic scale meansHere is the base. Equivalently, using the natural logarithm, . The product of numbers identity for logarithms gives , so the midpoint is the geometric mean . The two endpoints lie at equal multiplicative distances because .
For and , this midpoint is , giving a factor of on each side. The arithmetic mean is halfway on a linear scale, but its two ratios are and . Thus reading halfway between the marks and as on a logarithmic scale gives the wrong value.
Equal ratios give equal logarithmic gaps, and the geometric mean is halfway between two positive values
. Original rulers calculated from logarithms. The lower rulers cover the same endpoints and : the blue geometric mean is centered on the logarithmic scale, while the orange arithmetic mean is centered on the linear scale. Tick labels show original values. Details.





