Pseudomedian¶

The median of the distribution of the average of two independent draws from one real-valued distribution.

Version: v1 · 2026-10-25 · History
Domain-specific #: 14431
Domain group: Formal Sciences
Origin domain: Mathematics
Subdomains: Statistics, Nonparametric Inference → Mathematics

Core Idea¶

The pseudomedian of a real-valued probability law $F$ is a median of the law of $(X+Y)/2$, where $X$ and $Y$ are independent draws from $F$.^[1] The first step changes the distribution by pair averaging; the second takes its ordinary median. If the resulting law has several medians, the definition gives a set unless a selection convention is declared.^[1]

A symmetric law and its pair-average law share their center of symmetry as a median; their full median sets need not coincide.^[1] When both medians are unique, as for a normal law, the two reported centers agree. For a skewed law they can differ. For a rate-one exponential law, the original median is $\ln 2\approx0.693$, while the pseudomedian is approximately $0.839$.^[1] The latter number follows from the pair-average distribution, not from taking a sample's middle observation.^[1] These numbers are a specific calculation, not a universal claim about the ordering of median, mean, and pseudomedian for skewed laws.

Structural Signature¶

Sig role-phrases: ordered real law — two independent same-law draws — arithmetic pair average — median of induced law — optional finite-sample Walsh estimate.

Ordered real law. $F$ supplies scalar values whose probability mass can be cut in half. Without an ordered scalar output space, the bare scalar median is undefined. A selection convention is needed only when one number must be reported from a nonunique median set.^[1]
Independent same-law draws. $X$ and $Y$ are drawn independently from one $F$. Changing the coupling or using two different populations changes the induced law and therefore the target.^[1]
Arithmetic pair average. The transformation $A=(X+Y)/2$ is constitutive. Taking the median directly on $F$ instead gives the ordinary median; taking a maximum or another summary gives another functional.^[1]
Induced median. A value $m$ qualifies when $P(A\le m)\ge1/2$ and $P(A\ge m)\ge1/2$.^[1] Weak inequalities preserve the possibility of ties and nonunique medians.
Optional estimator. For observations $x_1,\dots,x_n$, the one-sample Hodges–Lehmann point estimate is the median of the $n(n+1)/2$ Walsh averages $(x_i+x_j)/2$ for $i\le j$. It is a finite-sample computation associated with the signed-rank procedure; it is not itself the population definition.^[1]

What It Is Not¶

It is not the median of $F$ as a functional. Symmetry gives the original and pair-average laws a shared center that is a median of each, though their entire median sets need not agree; unique medians make the familiar equality precise. For skewed laws, the pseudomedian and the median can differ.^[1]

It is not the mean. A law can have a median of its pair-average distribution even when an ordinary finite mean of $F$ is unavailable. Existence of a median does not require a moment calculation.^[1]

It is not the two-sample Hodges–Lehmann location-difference estimate, which uses cross-group differences rather than Walsh averages within one sample.^[2]^[1]

It is not Pratt, Cooper, and Kabir's pseudomedian filter. Their signal operator combines local max-min and min-max envelopes; that formula-bearing homonym appears in the frozen source and remains a separate review branch.^[3] Royce and colleagues later use the word “filter” for a sliding-window Walsh-average median, which is the sample operation discussed here, not proof that the two signal filters are the same.^[3]^[4]

Scope of Application¶

In one-sample nonparametric location inference, the Walsh-average median is the point estimate associated with the Wilcoxon signed-rank procedure. Under a symmetric continuous law, its center agrees with the ordinary median. Geyer's worked example computes all Walsh averages and reports a point estimate of $9.625$.^[1] He also cautions that, without symmetry, the procedure targets the pseudomedian rather than the original median, and that the latter target may not be the desired center for a particular scientific question.^[1]

In paired comparisons, one forms one difference per observed pair and applies the one-sample Walsh calculation to those differences.^[5] Geyer’s paired $y-x$ worked example reports a Walsh-average median of $-0.46$.^[5] The candidate population law is the law of a within-pair difference, not the marginal law of either measured variable. The point calculation must not be mistaken for an unconditional guarantee of exact signed-rank inference: symmetry, continuity, and treatment of ties matter to the test.^[1]^[5]

In tiling microarray analysis, Royce, Carriero, and Gerstein calculate local medians of pairwise averages of observed probe intensities in sliding genomic windows, including an ENCODE ChIP-chip dataset.^[4] Their reported computation is an application of the finite-sample operation to signals; the paper does not, by that fact alone, establish an iid population law for each local window or claim a unique population pseudomedian of such a law.^[4]

Clarity¶

The order of operations matters. For $X,Y\sim F$ independently, first form the law of $A=(X+Y)/2$ and then take a median of $A$. Computing $\operatorname{median}(F)$ first and averaging that number with itself simply returns the ordinary median. In general, median and pair averaging do not commute.^[1]

For $F=\mathrm{Exp}(1)$, $P(A\le a)=1-(1+2a)e^{-2a}$ for $a\ge0$, by the $\mathrm{Gamma}(2,1)$ convolution of $X+Y$.^[1] Solving $(1+2a)e^{-2a}=1/2$ gives $a\approx0.8391735$, while $\operatorname{median}(F)=\ln2\approx0.6931472$.^[1] This is an author calculation from the defining transformation, agreeing with the frozen source's rounded comparison; it does not establish a general skewness inequality.^[1]

Manages Complexity¶

The functional compresses a whole distribution into a location summary while retaining a precise statement about how pairs contribute. In a sample, $n(n+1)/2$ Walsh averages become one reported point; Geyer's example shows how that point can sit beside an associated interval and test.^[1]

The compression does not retain the distribution's shape, dependence structure, or reason for choosing this center. A report of a pseudomedian should say which population $F$ is being described, how observations were sampled, whether it reports the population target or a finite-sample statistic, and how nonunique medians are resolved. In a sliding window, it should also specify the window and signal construction.^[1]^[4]

Abstract Reasoning¶

Begin with the estimand. Specify one ordered real law $F$ and independent $X,Y\sim F$. Derive or characterize the induced pair-average law. Then solve the two weak half-mass inequalities for that law. If a unique number is needed, state the selection rule where several medians satisfy them.^[1]

For observed data, form the Walsh averages with $i\le j$, take their median under an explicit even-count convention, and only then interpret that number against a population or experimental model. The one-sample statistic is computable even when the signed-rank test's exact null assumptions are not met. A question about the ordinary median of a skewed population needs a different estimand and possibly a different procedure.^[1]^[5]

Knowledge Transfer¶

The same pair averages then median sample computation appears in paired-difference inference and genomic probe smoothing. The first setting supplies observed paired differences; the second supplies observed probe intensities selected by genomic window. Both instantiate the finite calculation, but they do not inherit one another's sampling assumptions or interpretation.^[5]^[4]

At the population level, the transformation applies to any real-valued probability law. Its literal transfer requires a law, independent same-law pair draws, and a median on their average. Calling another local nonlinear operator “pseudomedian” does not transfer that mechanism.^[1]^[3]

Examples¶

Symmetric normal law and its boundary. Let $X,Y$ be independent $N(\mu,\sigma^2)$ draws. Their average is $N(\mu,\sigma^2/2)$, whose unique median is $\mu$.^[1] Thus the pseudomedian and ordinary median coincide. Mapped back: $F$ is the normal law; independence builds the pair-average law; the median step returns its center.^[1] Symmetry explains the coincidence, not the entire identity.^[1]

Skewed exponential law. Let $X,Y$ be independent rate-one exponential draws. Their average has CDF $1-(1+2a)e^{-2a}$ for $a\ge0$.^[1] Its median is about $0.839$, compared with $\ln2\approx0.693$ for $F$.^[1] Mapped back: one law and independent draws enter; averaging changes the law; taking the new median produces a distinct center.^[1] The formula and numbers are a derivation for this law, not a source-reported empirical estimate.^[1]

Paired-data point estimate. In Geyer’s worked paired data, form one difference within each observed pair, then take all self-inclusive pairwise arithmetic averages of the differences.^[5] The median of these Walsh averages is $-0.46$.^[5] Mapped back: the observed differences are the finite carrier and candidate population-law sample; their pair averages supply the transformed values; their median is the reported point estimate.^[5] The computation alone does not justify a signed-rank interval without its separate assumptions.^[1]^[5]

ENCODE ChIP-chip smoothing. Royce and colleagues place measured probe intensities in a 250-nt genomic window, form all $n(n+1)/2$ pairwise averages, and take their median as a smoothed value at a feature position.^[4] They repeat this computation as the window slides and compare algorithm runtimes on real ENCODE ChIP-chip data.^[4] Mapped back: the observed intensities are the finite carrier; the Walsh pair averages and their median are the same sample operation.^[4] The paper does not equate this output with Pratt's max-min/min-max filter or identify it with a specified local population-law pseudomedian.^[4]

Structural Tensions¶

Population target versus convenient sample center. The finite Walsh calculation is well defined from observed numbers. Interpreting it as a scientifically useful population center requires a choice of estimand and sampling model. Symmetry aligns it with the familiar median; asymmetry separates those targets. The exact computation buys a stable named statistic while a casual “median of the population” description can point to the wrong parameter. Diagnostic: Is the scientific question about the median of $F$ or of its pair-average law?^[1]

Induced-law fidelity versus a convenient local statistic. The population definition admits a median set, while an analysis table or sliding probe window needs one number. Selecting an endpoint or midpoint can make reporting reproducible but hides a nonunique induced-law median; narrowing or widening a genomic window changes which observed values enter the Walsh calculation. Diagnostic: Was the reported value selected from a population median set, or computed from a declared finite window and tie convention?^[4]

Structural–Framed Character¶

The pseudomedian has a structural mathematical core: a specified transformation of a scalar probability law followed by a half-mass location operation. It is not an evaluative label; a larger pseudomedian is not intrinsically better. Its practice dependence enters when analysts decide whether this rather than the ordinary median is the relevant estimand, how observations or windows are sampled, and how a nonunique output is reported. Geyer's signed-rank discussion and Royce's genomic windows provide literal, differing uses of the calculation, not a license to treat their inferential assumptions as interchangeable.^[1]^[4]

No institution confers the mathematical definition; it is fixed by the iid pair transformation and the induced median. Statistical teaching, software, and research papers shape how the point estimate is named and used, so their labels cannot silently change the estimand. The term's vocabulary travels within statistics and data analysis when the pair-average-median operation is literally imported, as in local probe smoothing. The Pratt signal filter shares the name but has a different operative formula; recognizing similar vocabulary or a shared smoothing purpose is not importing this operation. Its character: a precise distributional location functional with a corresponding sample computation whose interpretation depends on the data-generating setting.^[3]^[4]

Structural Core vs. Domain Accent¶

The core is one real law → two independent same-law draws → arithmetic average → median of the induced law. A finite Walsh-average median operationalizes a corresponding sample calculation. Paired observed differences and microarray probe signals fill the observed-value role differently; neither changes the mathematical composition.^[1]^[5]^[4]

The dependence on real-valued probability laws, independence, and the induced median operation keeps the named identity in statistical mathematics. A broader “transform then summarize by order” skeleton might be a future-Prime question if demonstrated across genuinely different carriers and mechanisms; no such Prime is claimed or assigned here. Removing the statistical roles from this entry leaves generic aggregation, which fails the Prime bar for this named pseudomedian. The median step supplies the actual half-mass operation, but pseudomedian is not merely an alias for Median: it changes the input law first, as the exponential case demonstrates. The recorded composition relation, with Median inside the child operation, remains under separate official DAG review.^[1]

This entry includes Median.

Median is the staged composition/part_of parent pending official graph audit: the induced pair-average law is processed by the ordinary median operation. This is not a strict subsumption claim that every pseudomedian is the median of its original law. Order is indirectly required because the induced law is real-valued and its median needs order. Robustness describes some reasons to use rank-based location summaries, but a general robustness guarantee is not part of this definition. The typed edge and best rejected alternatives remain subject to the official current-catalog graph audit.^[1]

Relationships to Other Abstractions¶

Current abstraction Pseudomedian Domain-specific

Parents (1) — outgoing typed relations

Pseudomedian includes Median Domain-specific

The ordinary median is an identity-bearing constituent operation applied to the induced pair-average law.

Hierarchy paths (3) — routes to 3 parentless roots

Pseudomedian → includes Median → presupposes Order → presupposes Comparison

Show alternative paths (2)

Neighborhood in Abstraction Space¶

Pseudomedian sits in a sparse region of the domain-specific corpus (86^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (2769 abstractions)

Nearest neighbors

Monotone Likelihood Ratio Property — 0.85
Hellinger Distance — 0.81
Median — 0.81
Linear order — 0.80
Jensen's Inequality — 0.80

Computed from structural-signature embeddings · 2026-10-25

Not to Be Confused With¶

Median of $F$: computed before any pair-average distribution is formed; it can differ on skewed laws.^[1]
Sample median: the middle of the original observations, rather than the median of their Walsh averages.^[1]
Two-sample location-shift estimate: uses cross-group differences and answers a distinct question.^[2]
Pratt's pseudomedian filter: a max-min/min-max local signal operator in the frozen source, still an independently routed source branch.^[3]
Royce's sliding-window pseudomedian filter: the finite Walsh-average operation applied to local probe intensities, not the Pratt operator.^[4]

References¶

[1] Charles J. Geyer, Stat 5102 Notes: Nonparametric Tests and Confidence Intervals (2007), §§2.1–2.4, especially §2.3 Walsh averages and §2.4 pseudomedian definition and caution, https://www.stat.umn.edu/geyer/s06/5102/notes/rank.pdf. University instructor notes; mathematical definition, signed-rank assumptions and worked data consulted directly. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u ↩v ↩w ↩x ↩y ↩z ↩27 ↩28 ↩29 ↩30 ↩31 ↩32 ↩33 ↩34 ↩35 ↩36 ↩37 ↩38 ↩39 ↩40

[2] R Core Team, “Wilcoxon Rank Sum and Signed Rank Tests,” R stats manual, two-sample estimator note, https://stat.ethz.ch/R-manual/R-devel/library/stats/html/wilcox.test.html. Official method documentation distinguishes a cross-group difference median from the one-sample signed-rank estimate. registry ↩a ↩b

[3] William K. Pratt, Thomas J. Cooper and Ihtisham Kabir, “Pseudomedian filter,” Proceedings of SPIE 534 (1985): 34–43, https://doi.org/10.1117/12.946562. Bibliographic/abstract-level identity checked; the frozen Wikipedia revision, lines 14–18, gives the separate max-min/min-max formula. The full SPIE paper was not available for this audit. registry ↩a ↩b ↩c ↩d ↩e

[4] Thomas E. Royce, Nicholas J. Carriero and Mark B. Gerstein, “An efficient pseudomedian filter for tiling microarrays,” BMC Bioinformatics 8 (2007): 186, https://doi.org/10.1186/1471-2105-8-186; original author-hosted full-text PDF https://archive.gersteinlab.org/papers/e-print/pseudomedian/preprint.pdf, Algorithm, Figure 2, and Testing. Supports the $n(n+1)/2$ sliding-window Walsh calculation and the ENCODE ChIP-chip application, not a population-target assertion. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n

[5] Charles J. Geyer, “Wilcoxon Signed Rank Test,” Statistics 5601 examples (2007), paired-difference setup and Associated Point Estimate, https://www.stat.umn.edu/geyer/f07/5601/examp/signrank.html. University worked example; the point-estimate calculation and its test assumptions must be distinguished. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j