Mean (mathematics) Created 2026-10-20 Updated 2026-11-05
A mathematical mean assigns a representative value to several quantities. A real mean of normally lies between their minimum and maximum and returns when every input equals . Different means preserve different operations: the arithmetic mean averages sums, while the geometric mean averages positive multiplicative factors.
For and , the arithmetic mean is and the geometric mean is . If equal distances are traveled at speeds and , the average speed is instead their harmonic mean, . The useful choice depends on which quantity is being added: distances, times, or logarithms of growth factors.
Let be the weighted geometric mean and set . Since the weights sum to one,
Part (i) implies , proving the weighted arithmetic-geometric mean inequality:
Equality holds exactly when for every , or equivalently when all the are equal.
The commands read the 52 cell values into a vector and build the matching two-factor layout. In expand.grid(country, goods) the first factor varies fastest, so the four countries are listed successively within each goods category. This matches row-by-row reading of the displayed table. Converting Goods to a factor is essential: its labels identify 13 categories and are not a quantitative trend. Country supplies the other categorical factor. The summaries suppress coefficient correlations, not coefficient standard errors.
The first fit is the additive two-factor normal linear model
There are free coefficients and residual statistical degrees of freedom. This is two-way analysis of variance without an interaction term. For this complete balanced layout, ordinary least squares gives
These formulas follow from the row- and column-sum normal equations: fitted row and column totals equal their observed totals. The balanced-design coefficient standard errors are for a goods contrast, for a country contrast, and for the reference-cell intercept. These yield the common contrast standard errors and the different intercept standard error in each summary. In particular the intercept is the fitted Software–UK value, not the grand mean or the observed Software–UK entry.
Thus the first intercept predicts million euros for Software in the UK. The Books coefficient adds million euros to the fitted value in any country, relative to Software; every other goods coefficient has this same arithmetic-difference interpretation. The country coefficients subtract , and million euros from the UK prediction for France, Germany and the rest of Europe respectively, regardless of goods. The model therefore assumes common absolute country differences across goods categories. For Software–France its prediction is
which is impossible for positive spending and indicates a limitation of this additive original-scale specification.
The residual standard error is million euros, obtained from ; direct calculation gives . Each coefficient's printed statistic is , with a reference distribution under this normal linear model. The two-sided p-values compare each goods category with Software, or each other country with the UK, after adjustment for the other factor. For example France and the rest of Europe differ significantly from the UK at 5%, but Germany does not. The raw-scale significant goods differences are Books, Clothing, Electronic Goods, Groceries and Leisure Travel; these are individual reference-level comparisons, not an overall test of the goods factor.
The coefficient of determination is . The omnibus F-test compares all 15 nonintercept coefficients with zero:
with reference and . This says at least one factor contrast is nonzero under the assumed error model. There is one observation per cell, so an unrestricted goods-country interaction would saturate the model, leaving zero residual degrees of freedom. Without replication or a specified interaction structure, the residual variation cannot be separated into interaction and independent error.
The second fit applies a logarithmic transformation and fits the same factor design to . The Gaussian log-response model is
Its intercept exponentiates to a fitted Software–UK conditional median of million euros. Goods coefficients now give multiplicative ratios to Software, and country coefficients give multiplicative ratios to the UK. The estimated country ratios are
For example the Books ratio is and the Sports Equipment ratio is , consistently across countries. The model assumes common relative, rather than common absolute, country effects. It also guarantees positive retransformed fitted values. Under normal log errors, the original-scale conditional mean is , whereas is the conditional median and geometric mean. Correcting the retransformation bias here multiplies the naive predictions by approximately .
The log-scale residual standard error is , with the same 36 residual degrees of freedom, corresponding to a typical multiplicative error multiplier . France and the rest remain significant reference-country contrasts; Germany remains nonsignificant at 5%. The log-scale significant reference-goods contrasts are Books, Music, Videos/DVDs, Clothing, Sports Equipment, Electronic Goods, Groceries and Leisure Travel. Event Tickets, Toys, Video Games and Housewares do not individually differ significantly from Software at that level. These p-values remain subject to the usual multiple testing qualification.
The log-scale is computed using variation in the logarithms, so it cannot by itself establish better fit on the original response scale. The change is plausible because spending varies widely in magnitude and relative discrepancies are more meaningful than constant absolute discrepancies, but it should be assessed with regression diagnostics, especially a residual-versus-fitted plot and a normal Q-Q plot. The minimum residual near merits inspection; the command summaries alone do not verify independent normal errors. Moreover these responses are forecasts, so the inferential p-values have a literal sampling interpretation only if an appropriate stochastic model for their uncertainty is justified.
The log-scale analysis of variance adds Goods and then Country. In this balanced design their effect spaces are orthogonal after removing the intercept, so the sequential sums of squares equal the corresponding adjusted factor sums and are unchanged by reversing the two factors:
The residual sum of squares is , giving . Thus
Their tiny upper-tail p-values show strong overall goods and country effects under the log-error model; a globally significant country factor is compatible with the nonsignificant individual Germany contrast. The omnibus fit statistic is on degrees of freedom, whose p-value is approximately . The printed zero reflects numerical formatting, not an exactly zero probability.
If two cell entries were missing, retain their correct factor labels and mark those responses unavailable, rather than shifting the remaining scan values into different cells. Fit the additive models to the 50 observed cells, assuming the missingness mechanism permits those observations to be analyzed without selection bias. All factor levels remain represented, and the row-column incidence graph remains connected: any two goods rows still share an observed country, since only two edges of the complete layout have been removed, and every country still has many observed goods. By identifiability of an incomplete two-factor additive design, the design therefore still has rank . Hence
The coefficients and their covariance must be recomputed by the observed-cell ordinary least squares equations. The balanced row-mean-plus-column-mean formulas and common within-factor standard errors no longer apply. Goods and Country effects are generally no longer orthogonal, so sequential ANOVA sums can depend on order; factor tests adjusted for the other factor should use nested model comparisons. Missing cells do not restore replication or make an unrestricted interaction test possible: that larger model still fits every observed cell exactly. If missingness depends on the unavailable spending values, additional modelling of that mechanism would be needed.
There is no perfect recall. Player 1's later information set joins histories following both of her initial actions, so she has forgotten whether she chose or . It also hides player 2's intervening action, but failure to observe another player's action is imperfect information, whereas forgetting one's own earlier action is the particular violation of perfect recall. Player 2's two nodes form a single information set because player 2 does not observe the initial choice.
For a behavioural strategy, let be player 1's initial probability of and her later probability of . These are separate information sets, so and need not be equal; their random draws are independent. Name player 2's unlabeled branches left and right, and let be the probability of taking the left branch. The only unit-payoff paths are left–left–left and right–right–right. Consequently
Because this is a zero-sum game, the natural meaning of the claimed best behavioural strategy is its maximum guaranteed expected payoff. Player 2 can choose either branch, so the guarantee at is
For nonnegative numbers, their minimum is at most their geometric mean. Also and , hence
Taking makes both possible branch-conditioned success probabilities , so this upper bound is attained against every . Thus the optimal guaranteed behavioural strategy success probability is .
A mixed strategy can instead select the complete plan with probability and the complete plan with probability . The two choices within a sampled plan are correlated. If player 2 chooses left, exactly the sampled plan succeeds; if player 2 chooses right, exactly succeeds. Therefore the success probability is
against every opposing mixed strategy. It is optimal among mixed strategies too: player 2's uniform mixture gives success probability to each of and zero to , so no mixture of the four plans can exceed against that opponent. This establishes mixed and behavioural security levels without perfect recall directly.
Figure 1.
Independent behavioural choices guarantee at most , whereas a mixed plan correlates the choices and guarantees
. Details.
The guarantee qualification matters: if player 2's branch were known to be left, the deterministic plan would succeed with probability one. The number is not an upper bound against every fixed opposing choice separately.
The joint likelihood ratio in favor of the alternative is
It is increasing in the product, so the Neyman-Pearson lemma rejects for large product, equivalently small . Under , the transformation gives
Therefore the are independent exponential random variables of rate two, , and . An exact size- critical region is
Equivalently, reject when the geometric mean satisfies . The boundary distribution is continuous, so no randomization is needed here.
Under , each has expected value and variance . The central limit theorem therefore approximates
For the supplied data this standardized value is
Using the provided approximation , the lower-tail 5% rule rejects when . The observed is in this region, with approximate p-value . Reject at the 5% level in favor of the specified alternative . The unusually large geometric mean is consistent with the direction of that alternative. This is evidence in this comparison, not a proof that the parameter is exactly three.
A Pauling bond-energy excess subtracts a chosen like-atom baseline from the unlike-atom bond-dissociation energy. For an unequal-input arithmetic example, stipulate , and , all with one dissociation convention. The arithmetic mean baseline and excess are
The arithmetic-baseline Pauling scale model therefore gives . The bond-dissociation energy is not itself the excess: subtracting the baseline is essential. All three numbers are stipulated, not experimental values for named chemical elements.
Changing the baseline changes the model. For the same stipulated inputs, a geometric mean baseline would be , yielding an excess of and a difference of approximately . It is a different construction, even though both start from the same bond-dissociation energies. State which baseline is used before comparing electronegativity differences.
The arithmetic-baseline Pauling scale also has a useful inverse: a difference of corresponds to an excess of , and a difference of corresponds to . Doubling the electronegativity difference quadruples the modeled excess. Interchanging the endpoint labels changes neither quantity; the square root alone cannot choose the direction of bond polarization.
In The Nature of the Chemical Bond. IV. The Energy of Single Bonds and the Relative Electronegativity of Atoms (printed pp. 3576–3578, Table II and equation (1)), Linus Pauling organized excess energies for chemical bonds into an approximately additive map of their square roots. He also identified a limitation: his H–F map separation of predicted an excess of , exceeding his table's value by . Those are historical values and a historical model discrepancy, not modern recommended inputs.[ref]
Perplexity 2026-11-05
Perplexity measures how much probability a model assigns to observed outcomes, expressed as the inverse geometric mean of those probabilities. For a sequence evaluated by a neural language model, let be the model's conditional probability of the actually observed token given its allowed context. Then
It is therefore exponentiated mean negative log-likelihood. If cross-entropy is expressed in bits rather than natural-log units, the equivalent expression is .[ref]
A uniform predictor over possible tokens assigns each observed token probability , giving perplexity . For an unequal two-token example with assigned probabilities and ,
The result need not be an integer or an actual vocabulary size. It is an effective branching measure defined by the likelihood calculation.
On a fixed evaluation representation and dataset, lower perplexity means higher geometric mean probability for the realized outcomes. Assigning zero probability to any realized outcome makes the unsmoothed sequence perplexity infinite. Correctness of the most probable token alone does not determine the metric: the probabilities assigned to all realized outcomes matter.
For an empirical example, the 2003 neural-language-model paper reports test perplexity for its MLP1 row, compared with for the five-gram back-off comparison row. These are measurements under the paper's Brown-corpus setup and token-counting conventions.[ref]
Changing tokenization, vocabulary, punctuation or sentence-boundary treatment changes the outcomes and the divisor . Per-word and per-subword scores therefore cannot be compared as if they measured the same event space. Perplexity is also distinct from BLEU, which compares generated text with references rather than directly averaging the probability of observed tokens.
Sandra Prior 2026-11-05
Sandra Prior researches bioassay standards for therapeutic monoclonal antibodies. Her institutional profile describes work linking antibody structure and biological function, including the development of cell-based assays.[ref] Her contribution is to making functional measurements comparable, rather than treating an antibody's chemical amount as a complete description of its biological activity.
Figure 1.
Sandra Prior, identified in the 2020 Immunogenicity and Bioassay Summit speaker biography
. Source.
Prior is first author of a 2017 WHO collaborative-study report evaluating a proposed international standard for the biological activities of rituximab.[ref] The report used comparisons of preparations and response-curve parameters to assess relative biological potency. Its coded duplicates supplied empirical information for study-specific bioassay parallelism criteria; the report distinguished these from routine laboratory validity limits.[ref] These are contributions of the collaborative team, not an attribution of every analytical step to one author.
An original calibration example illustrates the scientific problem. Suppose two laboratories estimate the same test preparation's activity relative to a shared reference as and . Their geometric mean is , whereas their arithmetic mean is . The geometric comparison treats reciprocal departures symmetrically on a logarithmic scale. Agreement of that summary with one does not establish acceptable individual curves or prove that the laboratories share no bias. Assay similarity, experimental controls and uncertainty remain separate questions.
The report also explains why assigned biological units are not simply a conversion from measured antibody mass: first-standard unitage defines a conventional activity scale, while a replacement standard is calibrated against that scale.[ref] This distinction connects Prior's standardization work with reference-standard uncertainty in relative potency and the interpretation of interlaboratory comparisons.
For nonnegative real numbers and positive weights with , the weighted geometric mean does not exceed the weighted arithmetic mean:
For positive , this follows from Jensen's inequality and the concavity of the natural logarithm; a limit extends it to zero inputs. For two terms it gives when .
Weighted geometric mean Created 2026-10-14 Updated 2026-11-05
For positive real numbers and nonnegative weights with , the weighted geometric mean is
It takes a weighted mean of logarithms. A zero weight excludes that input; equal weights give the ordinary geometric mean. If occurs times in a list of total length , its correct weight is . For the list ,
and replacing all four entries by preserves the product .
There is also an optimization interpretation. A logarithmic transformation turns multiplicative discrepancies into additive ones. For a positive representative , the weighted mean minimizes squared deviations identity gives
Thus uniquely minimizes weighted quadratic loss in logarithmic coordinates. The weighted arithmetic mean minimizes squared discrepancies in the original coordinates. These two objectives explain why different choices of what constitutes an error select different summaries.
The strictly concave function and the reversed Jensen's inequality give , where is the weighted arithmetic mean. Apply the same reasoning to the multiplicative inverses to get for the weighted harmonic mean. Thus ; equality holds exactly when all positive-weight inputs agree. For this list, , , and . The weights specify which inputs are represented and must stay the same across the comparison.