Harmonic Mean

Summarize a nonempty finite list of positive values by taking the reciprocal of the arithmetic mean of their reciprocals.

Version

v1 · 2026-10-17 · History

Domain-specific #

14178

Domain group

Formal Sciences

Origin domain

Mathematics

Subdomains

Theory of Means, Descriptive Statistics → Mathematics

Core Idea

The harmonic mean summarizes positive values by averaging their reciprocals and then inverting that average. For a nonempty finite list \(x_1,\ldots,x_n\) of strictly positive finite real numbers,

\[ H_n(x_1,\ldots,x_n) =\left(\frac1n\sum_{i=1}^{n}\frac1{x_i}\right)^{-1} =\frac{n}{\sum_{i=1}^{n}1/x_i}. \]

NIST's Dataplot documentation gives this formula. The positive-input restriction states the mathematical scope used here; it is not a claim about which other inputs its software accepts.^[1]

The operation has three linked stages: reciprocal coordinates, an equal arithmetic average, and inverse return. The normalization by \(n\) makes it a mean. For equal inputs \(x\), the result remains \(x\); simply inverting the reciprocal sum would instead give \(x/n\). Both formulas can be useful, but they identify different operations.

Choosing this mean for an application requires a further argument. Equal-distance travel provides one: slower legs consume more time, and the total-time calculation yields harmonic averaging. Equal-time travel with the same speeds yields arithmetic averaging. A harmonic mean can be computed in either case; the physical interpretation distinguishes which calculation answers the question.^[2]

Structural Signature

Sig role-phrases: positive values; cardinality and multiplicity; reciprocal transform; additive reciprocal pool; equal normalization; inverse return.

• Admissible input values. Each \(x_i\) is a strictly positive finite real number on a compatible numeric scale. This makes its reciprocal available and keeps the reciprocal average positive.
• Cardinality and multiplicity. The number \(n\geq1\) counts every admitted observation once. Repeated equal values remain repeated observations; deleting them changes their weight.
• Reciprocal transform. Each value becomes \(1/x_i\). For speeds, these coordinates have units of time per distance.
• Additive reciprocal pool. The transformed values contribute to \(S=\sum_i1/x_i\). Replacing this pool with a product, maximum or selected subset changes the computation.
• Equal mean normalization. Divide \(S\) by \(n\), assigning each reciprocal weight \(1/n\). Omitting this stage produces the unnormalized reciprocal-additive result.
• Inverse return. Return \(1/(S/n)\), restoring the input scale and, where applicable, its units. Stopping at the average reciprocal leaves a different quantity.^[1]

Equal distance, equal expenditure, a preferred scoring interpretation or an official reporting institution is not a seventh universal role. Such conditions can justify particular uses of the mathematical operator.

What It Is Not

• Every average of rates. Equal-time and equal-distance speed problems have different weighting structures. Rate units alone do not select the harmonic mean.^[2]
• An inverse sum without normalization. For inputs \(6,3\), the harmonic mean is \(4\), whereas \(1/(1/6+1/3)=2\).
• A literal formula admitting zero here. A positive-weight zero input makes \(1/x_i\) undefined. A convention extending a mean to zero, or software handling of that case, must be stated separately.
• A positive-mean guarantee for signed inputs. Reciprocals of \(1,-1\) cancel, leaving no finite result under the displayed formula. Other signed cases also lose the positive-domain guarantees.
• A weighted mean by default. Unequal weights define an extension with additional data.
• A claim of official adoption. ABS discusses a harmonic price-relative formula as an option mentioned in the literature. That discussion does not establish its operational use in an ABS index.^[3]

Scope of Application

The literal scope is finite positive-input averaging under the specified rule. Its applications need their own account of what the values represent and why reciprocal averaging is appropriate.

• Equal-distance effective speed. Speeds across equal-length legs occupy the value roles. Inverse speeds determine the time required per unit distance, so their equal average and inverse give total distance divided by total time. Gavin supplies a complete worked construction and an equal-time contrast.^[2]
• Equal-weight price-relative aggregation. The values are matched-product ratios \(p_i^t/p_i^0\) between specified periods. ABS gives the harmonic aggregation formula for these relatives; its example price observations permit a transparent numerical substitution. This is a mathematical index construction, with no claim that the derived number was published as an adopted official index.^[3]

These are different substantive quantities filling the same mathematical roles. The speed calculation does not explain why a price-index designer should choose one index formula over another.

Clarity

The abstraction separates the coordinate being averaged from the units of the reported answer. Original values and reciprocals encode different gaps. Averaging speeds directly treats their numeric values equally; averaging their reciprocals treats times per equal unit distance equally. Inverting at the end does not undo that choice.

It also separates mean normalization from reciprocal addition. Forgetting \(n\) is detectable without knowing an application: equal inputs should return that same value. For \((x,x)\), a result \(x/2\) belongs to the unnormalized rule, while the harmonic mean is \(x\).

Finally, distinguish a source's stated formula from a calculation made using it. In the price example below, the formula and inputs are supplied by ABS, while the displayed harmonic numerical result is calculated from those inputs.

Manages Complexity

A list of arbitrary length produces one representative value through the reciprocal sum and its count. For a singleton, \(H_1(x)=x\): the output preserves the sole observation. For multi-observation inputs, different lists can have the same harmonic mean, and permuting the observations leaves it unchanged; the scalar therefore does not generally recover the original observations or their order.

For later pooling, group counts and reciprocal sums can replace retention of the full lists; the mean alone generally does not suffice. For positive groups \(A,B\) of sizes \(n,m\), the defining formula gives

\[ H(A\cup B)= \frac{n+m}{n/H(A)+m/H(B)}. \]

This is an algebraic consequence: \(n/H(A)\) and \(m/H(B)\) recover the two reciprocal sums. Retaining the group counts therefore permits exact pooling without retaining every original value. An unweighted average of group summaries generally discards information required for that task.

Abstract Reasoning

Several useful checks follow directly from the definition.

Internality. If \(a=\min_i x_i\) and \(b=\max_i x_i\), then [ \frac1b\leq\frac1n\sum_i\frac1{x_i}\leq\frac1a. ] Inversion gives \(a\leq H\leq b\). A result outside the positive input range signals an error or a different convention.

Symmetry, scaling and monotonicity. Permuting observations leaves the reciprocal sum unchanged. Multiplying every input by \(\lambda>0\) multiplies \(H\) by \(\lambda\). Increasing one positive input decreases its reciprocal contribution and increases the mean. Equal inputs yield that common value.

Arithmetic comparison. Let \(A=n^{-1}\sum_i x_i\). Then [ \frac{A}{H}=\frac{(\sum_i x_i)(\sum_i1/x_i)}{n^2}\geq1, ] because the numerator minus \(n^2\) is [ \sum_{i<j}\frac{(x_i-x_j)^2}{x_ix_j}\geq0. ] Thus \(H\leq A\), with equality exactly when all values coincide. These are deductions from the formula, rather than separate experimental findings.

Pooling check. Suppose \(A=(1)\) and \(B=(3,3)\). Their means are \(1\) and \(3\), whose binary harmonic mean is \(3/2\). The original pooled list has harmonic mean \(9/5\). Carrying the group sizes resolves the discrepancy through the pooling formula.

Knowledge Transfer

The full computation transfers literally from the speed construction to the price-relative construction: identify positive values, preserve their multiplicities, take reciprocals, average with the stated weights and invert. Positive-domain and normalization checks remain useful in either setting.

The application argument must be rebuilt. In travel, equal distances make total time proportional to the reciprocal-speed sum. For price relatives, the methodology specifies an equal-weight index formula. A conservation argument for a journey supplies no justification for adopting an economic weighting convention.

Unequal weights can be handled by a separately declared extension: [ H_w=\frac{\sum_i w_i}{\sum_i w_i/x_i}, \qquad w_i\geq0,\quad 0<\sum_i w_i<\infty. ] Here the inputs are positive finite reals and weights are finite nonnegative reals. Zero-weight terms can be omitted. This formula follows by normalizing a weighted reciprocal average and inverting; it is not the definition of the unweighted entry. The shorter expression \(1/\sum_i w_i/x_i\) assumes normalized weights.

Examples

Canonical

Gavin considers four consecutive \(100\)-km legs at \(100,110,90,120\) km/h. The total distance is \(400\) km and the elapsed time is [ \frac{100}{100}+\frac{100}{110}+\frac{100}{90}+\frac{100}{120} \approx3.853535 \text{hours}. ] Consequently, [ \bar v= \frac{400}{3.853535\ldots} =\frac{4}{1/100+1/110+1/90+1/120} \approx103.8 \text{km/h}. ] The source works both the total-time calculation and harmonic calculation. Its next example assigns the same speeds to equal one-hour intervals and obtains \(105\) km/h by arithmetic averaging.^[2]

Mapped back: positive speeds → four counted legs → inverse speeds in hours per kilometre → their additive pool → division by four → inversion to effective kilometres per hour. Equal leg distances justify the effective-speed interpretation. This is a worked formal journey, not observed travel data.

Applied / practice

ABS defines a harmonic mean of price relatives by [ I_H=\left(\frac1n\sum_{i=1}^{n}\frac{p_i0}{p_i^t}\right). ] Its example table supplies period-\(0\) to period-\(1\) price pairs \((10,12),(12,13),(15,17)\). Applying the harmonic formula to those inputs gives [ I_H= \frac{3}{10/12+12/13+15/17} =\frac{3978}{3499} \approx1.13689626. ] Expressed on base \(100\), this is approximately \(113.6896\). This numerical harmonic value is a calculation from the source's formula and example prices, rather than a harmonic result printed by ABS. The methodology discusses harmonic averaging as another possible formula; this example establishes no official operational adoption.^[3]

Mapped back: three positive matched-product price relatives → cardinality three → inverse relatives → their sum → equal normalization → inverse aggregate relative. Multiplying by \(100\) changes the reporting scale. These dimensionless values require no claim that every price relative represents an equal-distance or fixed-numerator rate.

Structural Tensions

T1: One-number summary versus valid regrouping. Reporting only each group's harmonic mean is compact, but it generally prevents correct pooling across unequal group sizes. The same group summaries can arise from different counts, and the pooled result depends on those counts. Carrying each count alongside its mean costs additional summary information while preserving exact regrouping; retaining only the scalar sacrifices that capability. The \((1)\) and \((3,3)\) example shows the cost through the difference between \(3/2\) and \(9/5\). Diagnostic: Before combining these harmonic summaries, are their group counts equal, or have counts or reciprocal totals been retained explicitly?

This tension follows from the computation and the information retained in a summary. Positivity and zero handling remain admissibility boundaries rather than additional opposing pressures.

Structural–Framed Character

Harmonic Mean is predominantly structural as a mathematical computation, with application-dependent choices about the quantities and weights supplied to it.

• Evaluative weight. The formula does not judge an answer desirable, a small value bad, or one application policy optimal. Those evaluations belong to the purpose of a particular summary.
• Human-practice dependence. Its arithmetic and domain conditions can be specified without a person or institution. Deciding which observations to admit and which average answers a practical question involves human practice.
• Institutional origin. Documentation by NIST and ABS supplies evidence of the named rule and its uses. Neither institution is a constitutive part of the operator.
• Vocabulary travel. Positive list, reciprocal, cardinality, normalized average and inverse retain their literal meanings across the two applications.
• Import versus recognition. Applying the mean imports an exact mathematical rule into a problem under independently defended quantity and weighting assumptions. The sources do not establish one physical causal mechanism shared by speeds and price indices.

Its character: strongly structural within mathematical averaging, with contextual choices about inputs and interpretation; the named computation remains a bounded domain-specific identity.

Structural Core vs. Domain Accent

The broader Quasi-arithmetic Mean family strictly includes the harmonic member: on positive inputs the continuous strictly decreasing generator \(f(x)=1/x\) gives exactly this equal-weight reciprocal average and inverse return for every admitted \(n\). Other generators, such as the identity, yield different means. This genus relation does not establish every broader summarization description: the singleton case \(H_1(x)=x\) preserves its input. Aggregation applies to the multi-input branch when observations are collapsed into a summary that loses detail.

The autonomous identity fixes positive real inputs, multiplicity, reciprocal coordinates, equal normalization and inverse return. It supports its own rate-selection, normalization and pooling diagnostics. Removing these commitments leaves another averaging operation or a broad summarization pattern.

Speeds and price relatives are literal applications of the same formal operator. Their different interpretations remain essential: the formula does not determine which distances, products, periods or weights a problem should use. Harmonic Mean therefore serves here as a named mathematical computation with independently justified applications. It supplies no universal rule that all rates or ratios ought to be averaged harmonically. Aggregation offers a broader comparison for genuine multi-item compression, rather than a description of every admitted harmonic-mean computation.

Instantiates / Related Primes

This entry is a kind of Quasi-arithmetic mean.

Quasi-arithmetic Mean is the broader mathematical family of which this mean is a reciprocal-generator member for every admitted positive list, including a singleton. Aggregation concerns deliberately combining multiple items into a lower-granularity summary while discarding detail. Multi-observation harmonic summaries can exhibit that pattern, subject to their purpose and retained information. The admitted singleton computation \(H_1(x)=x\) performs no such collapse, so the Aggregation relation holds only for the multi-input branch.

Function (Mapping) concerns a specified input-output assignment. The harmonic mean is such an assignment on its declared positive domain; that fact does not distinguish it from the many other functions used as summaries.

Reciprocal Additivity (Optic Equation) is related through reciprocal coordinates but has a different result. Its rule is \(1/c=\sum_i1/x_i\), yielding \(c=H/n\). For \((6,3)\), the values \(2\) and \(4\) distinguish reciprocal combination from normalized reciprocal averaging.

Relationships to Other Abstractions

Current abstraction Harmonic Mean Domain-specific

Parents (1) — more general patterns this builds on

• Harmonic Mean is a kind of Quasi-arithmetic mean Domain-specific

  Every positive equal-weight harmonic mean is the quasi-arithmetic mean generated by the reciprocal function; the broader family also contains nonreciprocal generators.

Hierarchy path (1) — routes to 1 parentless root

• Harmonic Mean → Quasi-arithmetic mean → Aggregation → Transformation → Function (Mapping)

Neighborhood in Abstraction Space

Harmonic Mean sits in a sparse region of the domain-specific corpus (75^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Number-Theoretic Properties & Tests (22 abstractions)

Nearest neighbors

• Large Set (Combinatorics) — 0.84
• Least Common Multiple — 0.84
• AM–GM Inequality — 0.83
• Average Order of an Arithmetic Function — 0.83
• Lifting Scheme — 0.82

Computed from structural-signature embeddings · 2026-10-17

Not to Be Confused With

• Quasi-arithmetic Mean. A broader transformed-average family. On positive inputs, \(f(x)=1/x\) gives the harmonic member; the identity generator instead gives the arithmetic mean. For \((1,4)\), their results are \(1.6\) and \(2.5\).
• Arithmetic mean. Averages original values. It solves the equal-time speed contrast and is not interchangeable with equal-distance harmonic averaging.^[2]
• Unnormalized reciprocal sum. Inverting the sum gives \(H/n\), changing the equal-input result and the operation's interpretation.
• Weighted harmonic mean. Requires declared weights and their normalization. Group harmonic summaries need count weights when the underlying group sizes differ.
• Mean of a Function. A normalized integral does not by itself specify the additional reciprocal transform and inverse return.
• Arithmetic–Geometric Mean. A common limit of coupled arithmetic and geometric iterations, rather than this finite reciprocal average.
• Other price-index formulas. The harmonic mean of individual price relatives differs from arithmetic averaging of those relatives and from taking the relative of arithmetic mean prices. ABS treats these as separate index constructions.^[3]

References

[1] National Institute of Standards and Technology. “Harmonic Mean.” Dataplot Reference Manual, last updated November 16, 2015. https://www.itl.nist.gov/div898/software/dataplot/refman2/auxillar/harmmean.htm. Official full HTML text, Description displayed formula and command documentation. The positive-input scope here is a mathematical restriction, not an asserted software rejection rule. registry ↩a ↩b

[2] Henri P. Gavin (2005). "Arithmetic Mean, Harmonic Mean and Geometric Mean." Duke University, January 26, 2005, PDF p. 2, Harmonic Mean section, equal-distance worked example and following equal-time contrast. https://people.duke.edu/~hpgavin/Risk/mean-example.pdf. Author-hosted two-page teaching note, full PDF text consulted. The displayed formula and arithmetic include normalization; the opening verbal shorthand omits it. No observed-travel dataset or visual figure verification is claimed. registry ↩a ↩b ↩c ↩d ↩e

[3] Australian Bureau of Statistics (2022). “Price index theory.” Producer and International Trade Price Indexes: Concepts, Sources and Methods, 2022 edition, released April 29, 2022, “Unweighted or equal-weight indexes,” footnote 7 and Table 3.2. https://www.abs.gov.au/statistics/detailed-methodology-information/concepts-sources-methods/producer-and-international-trade-price-indexes-concepts-sources-and-methods/2022/chapter-3-technical-methodology/price-index-theory. Official full HTML text consulted. The harmonic formula and example prices are source material; the harmonic numerical substitution above is an explicit calculation, not an ABS printed or adopted index. registry ↩a ↩b ↩c ↩d