Pseudomedian¶
The median of the distribution of the average of two independent draws from one real-valued distribution.
Core Idea¶
The pseudomedian of a real-valued probability distribution \(F\) is a median of the distribution of \((X+Y)/2\), where \(X\) and \(Y\) are independent draws from that same \(F\).[1] In other words, average pairs first, then take the middle of the resulting distribution. If several values qualify as a median, the definition gives a median set until a one-number convention is chosen.[1]
For a symmetric distribution, its symmetry center is a median both before and after pair averaging.[1] That does not mean the two full median sets always coincide. For a rate-one exponential distribution, which is skewed, the ordinary median is about $0.693$ and the pseudomedian is about $0.839$.[1] This particular difference follows from the pair-average law, not a rule that one center always lies between two others.[1]
Scope of Application¶
In nonparametric statistics, the one-sample Hodges–Lehmann point estimate takes the median of all \(n(n+1)/2\) Walsh averages from a sample, including each observation paired with itself.[1] It is a sample statistic corresponding to the population pseudomedian, not the population definition itself. Geyer calculates $9.625$ in a worked one-sample example[1] and cautions that, without symmetry, this statistic targets the pseudomedian rather than the ordinary median.[1]
For paired observations, calculate one difference per pair and take the Walsh-average median of those differences.[2] Geyer's worked paired example yields \(-0.46\).[2] The point calculation remains distinct from the assumptions for an exact signed-rank test.[2][1]
The same finite calculation appears in tiling microarray signal smoothing. Royce and colleagues take pairwise averages of probe intensities within sliding genomic windows, then their median; they test the computation on ENCODE ChIP-chip data.[3] This is a local sample calculation, not proof of a specified population law for every window.[3]
Clarity¶
Order matters: \(\operatorname{median}(\operatorname{law}((X+Y)/2))\) is not generally \(\operatorname{median}(F)\). Taking the original median first and averaging it with itself would just return that original median. The exponential example shows that this shortcut can give a different answer.[1]
For independent rate-one exponential draws, the pair-average CDF is \(1-(1+2a)e^{-2a}\) for \(a\ge0\).[1] Solving its half-probability equation gives about $0.839$, compared with \(\ln 2\approx0.693\) for the original law.[1] This is an author calculation from the cited definition and the frozen source's rounded values.[1]
Manages Complexity¶
The pseudomedian gives one location summary after a specified pair transformation. For a sample, a table of Walsh averages becomes one center; for a microarray, many probe intensities in a window become one smoothed value.[1][3]
That compact number omits distribution shape and the reason for choosing this particular center. A useful report says whether it is a population quantity or a sample statistic, what observations or window supplied the values, and what was done if several medians were possible.
Abstract Reasoning¶
To find a population pseudomedian, choose one real-valued law \(F\), draw two values independently from it, describe the distribution of their arithmetic average, then find a median of that new distribution. To find the sample version, form self-inclusive Walsh averages and take their median. Keep the point calculation separate from claims about a test or an interval.[1][2]
If the scientific question is about the ordinary median of a skewed distribution, check whether the pseudomedian is actually the desired target before interpreting the number as that median.[1]
Knowledge Transfer¶
The finite operation transfers from paired differences to local genomic probe signals because both supply observed scalar values, pairwise arithmetic averages, and a median. Their scientific meanings and sampling assumptions do not transfer with the arithmetic. A similarly named max-min/min-max signal filter is a different construction.[2][3][4]
Examples¶
-
Symmetric normal law. Independent draws from \(N(\mu,\sigma^2)\) average to a normal law with center \(\mu\) and variance \(\sigma^2/2\).[1] Its unique median is \(\mu\).[1] Mapped back: the normal law supplies the draws; their average changes the spread; the median step returns the shared symmetry center.[1]
-
Skewed exponential law. For independent rate-one exponential draws, averaging produces a law whose median is about $0.839\(; the original law's median is about \$0.693\). Mapped back: the same-law independent pair is transformed first, and the median is taken only on its resulting law.[1]
-
Paired differences. Geyer's worked paired example takes each \(y-x\) difference, forms self-inclusive Walsh averages, and reports their median as \(-0.46\). Mapped back: the observed differences are the sample values; their pair averages and median give the point estimate.[2]
-
Microarray windows. Royce and colleagues use probe intensities in a 250-nt genomic window, take all self-inclusive pair averages and their median, and repeat as the window slides across ENCODE ChIP-chip data.[3] Mapped back: intensities fill the sample-value role; pair averaging and the median supply the local smoothing output.[3] This does not assert an iid population target for each window.[3]
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
Neighborhood in Abstraction Space¶
Pseudomedian sits in a sparse region of the domain-specific corpus (86th 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¶
-
Ordinary median: it cuts the original distribution, while pseudomedian cuts the pair-average distribution.[1]
-
Two-sample location-difference estimate: it uses cross-group differences, not the one-sample Walsh averages here.[5]
-
Pratt's pseudomedian filter: it is a distinct local max-min/min-max signal operator in the frozen source. The source branch is preserved as a separate research lead and is not covered by this statistical entry.[4]
References¶
[1] Charles J. Geyer, Stat 5102 Notes: Nonparametric Tests and Confidence Intervals (2007), §§2.1–2.4, https://www.stat.umn.edu/geyer/s06/5102/notes/rank.pdf. University instructor notes; mathematical definition, Walsh averages, symmetric inference and the worked one-sample example directly checked. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u
[2] Charles J. Geyer, “Wilcoxon Signed Rank Test,” Statistics 5601 examples (2007), Associated Point Estimate and tied-data discussion, https://www.stat.umn.edu/geyer/f07/5601/examp/signrank.html. University worked paired-difference example, with reported Walsh-average median \(-0.46\). registry ↩a ↩b ↩c ↩d ↩e ↩f
[3] 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 https://archive.gersteinlab.org/papers/e-print/pseudomedian/preprint.pdf, Algorithm and Testing. Supports finite-sample sliding-window Walsh smoothing and ENCODE ChIP-chip testing, not a population target. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g
[4] 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. Only bibliographic/abstract metadata checked; the frozen Wikipedia revision gives the separate signal formula. The full original article was unavailable for this audit. registry ↩a ↩b
[5] 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. registry ↩