E-measure: A Cross-Disciplinary Overview

• E-measure is a term describing discipline-specific measure-like objects that quantify evidence, alignment, or bounds depending on the field.
• In statistics and risk backtesting, E-measures use closure rules and antitonicity to ensure coherent evidence accumulation and validate risk forecasts.
• In computer vision, transcendence theory, and measure theory, E-measures offer enhanced evaluation of binary maps, explicit transcendence bounds, and integral representations for co-occurrence.

Searching arXiv for papers on “E-measure” to ground the article in the current literature. E-measure is a cross-disciplinary term rather than a single standardized construct. In recent statistical literature, it denotes a measure-like generalization of the E-value on a hypothesis class; in sequential risk validation, it refers to betting-style quantities built from e-values, e-statistics, and e-processes; in computer vision, it denotes the Enhanced-alignment measure for binary foreground map evaluation; in transcendence theory, it denotes a transcendence measure for numbers such as $e^{1/n}$; and in measure-theoretic work on co-occurrence, an associated E-measure is induced from E-integrals by applying them to indicator functions (Koning, 22 Apr 2026, Wang et al., 2022, Fan et al., 2018, Dujella et al., 2023, Wang et al., 2022).

1. Terminological scope

The literature represented here uses the same label for several mathematically unrelated objects. The term therefore requires domain-specific disambiguation rather than a single universal definition.

Domain	Meaning of “E-measure”	Representative source
Statistical evidence	A map $\mathbf{e}:\mathcal{H}\to[0,\infty]$ satisfying an infimum closure rule over unions of hypotheses	(Koning, 22 Apr 2026)
Risk backtesting	A family of e-values, backtest e-statistics, and e-processes for VaR and ES forecasts	(Wang et al., 2022)
Computer vision	The Enhanced-alignment measure for binary foreground map evaluation	(Fan et al., 2018)
Transcendence theory	Any explicit upper bound for $\omega(k,H)$, yielding lower bounds on $|P(\xi)|$	(Dujella et al., 2023)
Measure-theoretic co-occurrence	A set function induced from an E-integral by $A\mapsto E_{[\cdots]}(\mathbf{1}_A)$	(Wang et al., 2022)

A recurrent source of confusion is that the 2022 backtesting paper explicitly uses “E‑measure” for betting-style evidence objects and states that this is not the information-retrieval F1-type measure, whereas the 2018 computer-vision paper uses “E-measure” precisely for a foreground-map evaluation score (Wang et al., 2022, Fan et al., 2018). The term is therefore best understood as a family resemblance across disciplines rather than a single invariant concept.

2. E-measure as a measure-like generalization of the E-value

In "The E-measure" (Koning, 22 Apr 2026), the object is defined on a hypothesis space $(\mathcal{P},\mathcal{H})$, where a hypothesis is a subset $H\subseteq\mathcal{P}$ and $\mathcal{H}$ is closed under arbitrary unions. The starting point is an E-function $\mathbf{e}:\mathcal{H}\to[0,\infty]$ satisfying $\mathbf{e}(\emptyset)=\infty$, followed by the antitonicity requirement for E-capacities,

$\mathbf{e}:\mathcal{H}\to[0,\infty]$0

The defining axiom of an E-measure is the closure rule

$\mathbf{e}:\mathcal{H}\to[0,\infty]$1

for every $\mathbf{e}:\mathcal{H}\to[0,\infty]$2. This makes evidence behave compatibly with logical implication: more specific hypotheses cannot have less evidence against them, and unions are evaluated by infimum rather than by addition (Koning, 22 Apr 2026).

This formulation places E-measures in deliberate contrast with classical measures. Classical measures are monotone and additive on disjoint sets, whereas E-measures are antitone and combine by infimum on arbitrary unions. On the $\mathbf{e}:\mathcal{H}\to[0,\infty]$3-value scale $\mathbf{e}:\mathcal{H}\to[0,\infty]$4, the closure rule becomes

$\mathbf{e}:\mathcal{H}\to[0,\infty]$5

which the paper identifies as a maxitive measure in the sense of Shilkret (Koning, 22 Apr 2026).

A central simplification occurs for intersection-closed hypothesis classes. If $\mathbf{e}:\mathcal{H}\to[0,\infty]$6 is closed under arbitrary intersections and contains $\mathbf{e}:\mathcal{H}\to[0,\infty]$7, then every $\mathbf{e}:\mathcal{H}\to[0,\infty]$8 has a least hypothesis

$\mathbf{e}:\mathcal{H}\to[0,\infty]$9

For an E-capacity $\omega(k,H)$0, its closure satisfies

$\omega(k,H)$1

Thus the values on least hypotheses form an E-density that determines the entire E-measure. The closure operator is also minimal: it is the smallest E-measure dominating the original E-function (Koning, 22 Apr 2026).

The data-dependent version is an E-kernel $\omega(k,H)$2, which is an E-measure in $\omega(k,H)$3 for each $\omega(k,H)$4 and a valid E-variable in $\omega(k,H)$5 for each $\omega(k,H)$6. Under intersection-closure and a countability condition on distinct least hypotheses, Theorem 4.4 states that closure preserves validity and that any non-dominated, hypothesis-wise valid E-capacity kernel must already be an E-measure kernel (Koning, 22 Apr 2026). The same framework yields familywise evidence control and false evidence rate control without multiplicity correction when the hypothesis class is intersection-closed, and it supports a frequentist update from E-prior to closed E-posterior by pointwise multiplication followed by closure (Koning, 22 Apr 2026).

The paper further abstracts “hypothesis” away from subsets of $\omega(k,H)$7 to subsets of other spaces, leading to predictive E-measures. In that setting, a predictive E-kernel gives evidence against claims $\omega(k,H)$8, and under intersection-closure predictive validity reduces to validity of the single E-variable attached to the least hypothesis $\omega(k,H)$9 for each outcome $|P(\xi)|$0 (Koning, 22 Apr 2026).

3. Betting-style E-measures in risk backtesting

In "E-backtesting" (Wang et al., 2022), the relevant objects are e-values, e-statistics, and e-processes for model-free, non-asymptotic, anytime-valid backtesting of Expected Shortfall and Value-at-Risk forecasts. For a hypothesis $|P(\xi)|$1, an e-variable is a nonnegative random variable $|P(\xi)|$2 with

$|P(\xi)|$3

and a realized value is an e-value. In sequential form, an e-process is a nonnegative adapted process $|P(\xi)|$4 such that

$|P(\xi)|$5

for all stopping times $|P(\xi)|$6, equivalently a nonnegative supermartingale with $|P(\xi)|$7. Ville’s inequality gives

$|P(\xi)|$8

so rejection can occur at any time without violating level-$|P(\xi)|$9 validity (Wang et al., 2022).

To connect these objects to risk forecasts, the paper defines point e-statistics, one-sided e-statistics, and backtest e-statistics for functionals $A\mapsto E_{[\cdots]}(\mathbf{1}_A)$0. A backtest e-statistic is calibrated so that the expected e-value is at most $A\mapsto E_{[\cdots]}(\mathbf{1}_A)$1 under correct or conservative forecasts and strictly greater than $A\mapsto E_{[\cdots]}(\mathbf{1}_A)$2 under underestimation. Monotonicity means that larger risk forecasts produce smaller e-values, thereby rewarding prudence (Wang et al., 2022).

For $A\mapsto E_{[\cdots]}(\mathbf{1}_A)$3, the canonical example is

$A\mapsto E_{[\cdots]}(\mathbf{1}_A)$4

which is a monotone backtest e-statistic. For $A\mapsto E_{[\cdots]}(\mathbf{1}_A)$5, the paper introduces

$A\mapsto E_{[\cdots]}(\mathbf{1}_A)$6

and Theorem 2.4 shows that this is a monotone backtest e-statistic for $A\mapsto E_{[\cdots]}(\mathbf{1}_A)$7. The construction uses the Rockafellar–Uryasev convex dual representation of ES, so that the normalized excess-tail loss has expectation $A\mapsto E_{[\cdots]}(\mathbf{1}_A)$8 under correct forecasts and exceeds $A\mapsto E_{[\cdots]}(\mathbf{1}_A)$9 when ES is underestimated (Wang et al., 2022).

Evidence is accumulated over time by a betting transform. If $(\mathcal{P},\mathcal{H})$0 is the per-period e-variable and $(\mathcal{P},\mathcal{H})$1 is predictable, then

$(\mathcal{P},\mathcal{H})$2

is an e-process. The paper studies several choices of betting fraction: GRO, GREE, GREL, and the mixture GREM, and measures their performance through e-power $(\mathcal{P},\mathcal{H})$3. Under the stated iid or convergence conditions, GREE and GREL are asymptotically optimal in the sense that their normalized log-growth matches that of the oracle GRO, while GREM is asymptotically optimal whenever either GREE or GREL is (Wang et al., 2022).

The paper also establishes characterization results. For VaR, any reasonable one-sided e-statistic is bounded above by a mixture of $(\mathcal{P},\mathcal{H})$4 and $(\mathcal{P},\mathcal{H})$5; for ES, under the stated monotonicity and continuity conditions, any such e-statistic is a mixture of $(\mathcal{P},\mathcal{H})$6 and $(\mathcal{P},\mathcal{H})$7. This makes the proposed constructions essentially canonical within the model-free framework (Wang et al., 2022). In that sense, the paper’s “E-measure” is a sequential evidence process aligned with regulatory concern about one-sided underestimation of risk.

4. Enhanced-alignment E-measure in binary foreground evaluation

In "Enhanced-alignment Measure for Binary Foreground Map Evaluation" (Fan et al., 2018), E-measure is a scalar score for comparing a binary foreground map $(\mathcal{P},\mathcal{H})$8 with a binary ground-truth map $(\mathcal{P},\mathcal{H})$9. The stated motivation is that existing binary foreground map measures treat pixel-level match or image-level information independently, whereas the proposed measure combines local pixel values with the image-level mean value in one term, jointly capturing image-level statistics and local pixel matching information (Fan et al., 2018).

The construction begins by mean-centering both binary maps: $H\subseteq\mathcal{P}$0 where $H\subseteq\mathcal{P}$1 is the image-level mean value and $H\subseteq\mathcal{P}$2 is an all-ones matrix. Writing the centered maps as $H\subseteq\mathcal{P}$3 and $H\subseteq\mathcal{P}$4, the alignment matrix is

$H\subseteq\mathcal{P}$5

This quantity lies in $H\subseteq\mathcal{P}$6, is nonnegative when the centered values have the same sign, and couples local agreement to global foreground proportions through the means $H\subseteq\mathcal{P}$7 and $H\subseteq\mathcal{P}$8 (Fan et al., 2018).

To amplify positive alignment and suppress negative alignment, the paper applies the convex map

$H\subseteq\mathcal{P}$9

defining an enhanced alignment matrix $\mathcal{H}$0. The final E-measure is the spatial average

$\mathcal{H}$1

so that $\mathcal{H}$2 (Fan et al., 2018).

The empirical evaluation uses four popular datasets and five meta-measures: ranking models for applications, demoting generic maps, demoting random Gaussian noise maps, ground-truth switch, and human judgments. The paper reports large improvements in almost all meta-measures and, for application ranking, an improvement ranging from $\mathcal{H}$3 to $\mathcal{H}$4 compared with other popular measures (Fan et al., 2018). It also states that E-measure and S-measure are the only tested measures that never prefer random noise over state-of-the-art maps on the reported datasets, while E-measure achieves the best alignment with human judgments in the FMDatabase experiment (Fan et al., 2018).

The stated limitations are specific rather than general. On PASCAL-S, which contains more structurally complex images, S-measure can outperform E-measure on structural meta-measures, and a reported failure case shows a generic map ranked above a map from RFCN because the metric exploits shape and alignment but not semantic information (Fan et al., 2018). The scope of the method is therefore binary foreground-map evaluation rather than soft saliency-map assessment.

5. E-measure in transcendence theory

In "Transcendence measure of $\mathcal{H}$5" (Dujella et al., 2023), an E-measure is a transcendence measure. For a transcendental number $\mathcal{H}$6 and a polynomial

$\mathcal{H}$7

the paper defines $\mathcal{H}$8 as the infimum of real numbers $\mathcal{H}$9 such that

$\mathbf{e}:\mathcal{H}\to[0,\infty]$0

for all nonzero integer coefficient vectors with $\mathbf{e}:\mathcal{H}\to[0,\infty]$1. Any function greater than or equal to $\mathbf{e}:\mathcal{H}\to[0,\infty]$2 is a transcendence measure of $\mathbf{e}:\mathcal{H}\to[0,\infty]$3 (Dujella et al., 2023).

Specializing to $\mathbf{e}:\mathcal{H}\to[0,\infty]$4, the main theorem states that for $\mathbf{e}:\mathcal{H}\to[0,\infty]$5,

$\mathbf{e}:\mathcal{H}\to[0,\infty]$6

where $\mathbf{e}:\mathcal{H}\to[0,\infty]$7, and that one may take

$\mathbf{e}:\mathcal{H}\to[0,\infty]$8

for $\mathbf{e}:\mathcal{H}\to[0,\infty]$9, with separate explicit constants for $\mathbf{e}(\emptyset)=\infty$0, under the condition

$\mathbf{e}(\emptyset)=\infty$1

As $\mathbf{e}(\emptyset)=\infty$2 with fixed $\mathbf{e}(\emptyset)=\infty$3, the correction term tends to $\mathbf{e}(\emptyset)=\infty$4, so the exponent approaches $\mathbf{e}(\emptyset)=\infty$5 from above (Dujella et al., 2023).

The paper compares this bound with Mahler’s 1975 result specialized to the same setting. The stated conclusion is that the new bound is better than Mahler’s bound, because the correction term in Mahler’s exponent has the asymptotic form $\mathbf{e}(\emptyset)=\infty$6, whereas the new exponent differs from $\mathbf{e}(\emptyset)=\infty$7 by about $\mathbf{e}(\emptyset)=\infty$8 (Dujella et al., 2023). The method is an explicit Hermite–Mahler style auxiliary-function construction based on simultaneous Padé-type approximations, determinant arguments, and integral estimates, specialized to the sequence $\mathbf{e}(\emptyset)=\infty$9 (Dujella et al., 2023).

In this domain, “E-measure” is tied to the constant $\mathbf{e}:\mathcal{H}\to[0,\infty]$00 rather than to evidence measures. The term refers to quantitative lower bounds for nonzero polynomial values at transcendental numbers related to $\mathbf{e}:\mathcal{H}\to[0,\infty]$01, and the paper identifies its contribution as the first explicit measure tailored for roots of $\mathbf{e}:\mathcal{H}\to[0,\infty]$02 (Dujella et al., 2023).

6. E-integrals and associated E-measures for complex co-occurrence

In "Measure-Theoretic Probability of Complex Co-occurrence and E-Integral" (Wang et al., 2022), the primitive object is the E-integral rather than an E-measure in the sense of hypothesis testing. The setting consists of a probability space $\mathbf{e}:\mathcal{H}\to[0,\infty]$03, measurable spaces $\mathbf{e}:\mathcal{H}\to[0,\infty]$04, and random objects $\mathbf{e}:\mathcal{H}\to[0,\infty]$05. For finite index sets, the paper defines co-occurrence probabilities such as

$\mathbf{e}:\mathcal{H}\to[0,\infty]$06

and conditional probabilities of co-occurrence by normalizing joint co-occurrence when the conditioning event has positive probability (Wang et al., 2022).

The E-integral is then defined as an integral of a measurable function with respect to a co-occurrence measure or a conditional co-occurrence kernel. In the simplest case,

$\mathbf{e}:\mathcal{H}\to[0,\infty]$07

where

$\mathbf{e}:\mathcal{H}\to[0,\infty]$08

More general forms integrate with respect to measures such as $\mathbf{e}:\mathcal{H}\to[0,\infty]$09, $\mathbf{e}:\mathcal{H}\to[0,\infty]$10, or conditional kernels $\mathbf{e}:\mathcal{H}\to[0,\infty]$11 (Wang et al., 2022).

An associated E-measure arises by applying the E-integral to indicator functions: $\mathbf{e}:\mathcal{H}\to[0,\infty]$12 For example,

$\mathbf{e}:\mathcal{H}\to[0,\infty]$13

In this sense, the paper treats expectation-like functionals as primary and recovers measures from them through indicator functions, explicitly aligning the discussion with the expectation functional approach associated with Whittle and Pollard (Wang et al., 2022).

The paper establishes linearity, monotonicity, absolute-value inequalities, monotone convergence theorems, tower-type identities, and factorization properties under independence for these E-integrals (Wang et al., 2022). Its stated motivation is sparse high-dimensional co-occurrence data, where joint densities may fail to exist or may be analytically inconvenient; the E-integral framework remains measure-theoretic and kernel-based, so it can represent expectations under complex co-occurrence and conditioning structures without relying on smooth density models (Wang et al., 2022).

The resulting terminology differs sharply from the evidence-theoretic E-measure of (Koning, 22 Apr 2026). Here, the “E” is attached to an expectation-like integral operator over co-occurrence measures, and the induced E-measure is simply the set function generated from that operator.

7. Conceptual distinctions and recurring themes

The surveyed uses share notation but not ontology. In (Koning, 22 Apr 2026), an E-measure is a logically coherent evidence assignment on a hypothesis class. In (Wang et al., 2022), the operative objects are e-values, backtest e-statistics, and e-processes whose sequential growth quantifies evidence against under-conservative risk forecasts. In (Fan et al., 2018), E-measure is a deterministic image-comparison functional. In (Dujella et al., 2023), it is a transcendence measure for $\mathbf{e}:\mathcal{H}\to[0,\infty]$14. In (Wang et al., 2022), it is a measure induced from E-integrals in a co-occurrence framework.

A common misconception is therefore to treat “E-measure” as if it named one established metric across fields. The recent literature does not support that reading. The strongest unifying statement supported by these sources is narrower: each usage introduces a structured scalar- or measure-valued object that aggregates information under a discipline-specific notion of coherence—logical closure in hypothesis classes, supermartingale validity in sequential testing, global-local alignment in binary masks, explicit lower bounds in transcendence theory, or integral representation for co-occurrence measures (Koning, 22 Apr 2026, Wang et al., 2022, Fan et al., 2018, Dujella et al., 2023, Wang et al., 2022). This suggests that the term functions primarily as a local technical label whose meaning must be fixed by context.