Entropy and Diversity: The Axiomatic Approach

Tom Leinster

Source-linked AI summary

Entropy and Diversity: The Axiomatic Approach

Tom Leinster

arXiv:2012.02113v4Dec 3, 2020q-bio.PEcs.ITmath.CAmath.CTmath.ITq-bio.QM

TL;DR

How should biological diversity be quantified, and should conservation treat all species as equal? This book uses axiomatic characterizations to study entropy and diversity, showing that a distribution maximizing diversity across all viewpoints exists for every metric space and is unique for generic finite spaces.

Problem
How to quantify biological diversity remains an open question, including whether conservation should treat all species as equal.
Source: Introduction · p. 11, Value · p. 234
Method
The book uses axiomatic characterizations, including power-mean theorems, to derive entropy and abundance-based diversity measures.
Source: Introduction · p. 11
Results
For every metric space, one distribution maximizes diversity across all q viewpoints; for generic finite metric spaces it is unique, and the maximum is q-independent.
Source: Introduction · p. 15
Takeaways & Limitations
The magnitude of a finite metric space can be interpreted as its effective number of completely separate points.
Source: §6.4 · p. 222
The maximum-diversity theorem fails for nonsymmetric similarity matrices, an important restriction.
Source: §6.3 · p. 209

Abstract

from arXiv · show

This book brings new mathematical rigour to the ongoing vigorous debate on how to quantify biological diversity. The question "what is diversity?" has surprising mathematical depth, and breadth too: this book involves parts of mathematics ranging from information theory, functional equations and probability theory to category theory, geometric measure theory and number theory. It applies the power of the axiomatic method to a biological problem of pressing concern, but the new concepts and theorems are also motivated from a purely mathematical perspective. The main narrative thread requires no more than an undergraduate course in analysis. No familiarity with entropy or diversity is assumed.

Interdependence of chapters … 1.3 The q-logarithm

The section uses an axiomatic approach to connect biological diversity and entropy while developing mathematics that also stands independently of biological applications. It focuses on functional equations and their regularity constraints, leading to q-logarithms and the chain-rule basis for characterizing Shannon entropy.

Interdependence of chapters

A dotted line in the chapter-interdependence diagram means one chapter is helpful, but not essential, for another.Source: Interdependence of chapters · p. 6

Note to the reader

The core narrative requires only a first course in rigorous analysis; chapters and sections needing probability, abstract algebra, category theory, geometry, analysis, or statistics can be omitted.No ecology knowledge is needed, and information theory is introduced from the ground up.Source: Note to the reader · p. 7

Introduction

The book develops diversity theory as fertile ground for new mathematics, combining an axiomatic treatment of a biological problem with advances that stand independently of applications.Its origins include research in category theory, information theory, and functional equations.Source: Introduction · pp. 11–12
Further results use probability to solve functional equations, characterize information loss through categorical and operadic ideas, and define entropy for probabilities in Z/pZ.The book also presents entropy as inescapable in category theory, algebra, and topology, apart from its scientific applications.Source: Introduction · p. 12
The Hill numbers Dq form a continuum of diversity viewpoints, and the book proves they are the only measures satisfying specified natural properties.Low q emphasizes rare species, while high q emphasizes the balance of common species; the result also gives a new characterization of Rényi entropies.Source: Introduction · p. 14
Similarity-sensitive diversity measures account for species similarities and extend Rényi entropy from finite sets to metric spaces.For a generic finite metric space, a unique distribution maximizes diversity across all q viewpoints simultaneously, and the maximum is independent of q.Source: Introduction · p. 15
The book extends diversity measures to account for species values and metacommunity structure, and relates relative entropy to local diversity and atypical community components.Logical requirements characterize the value measures; established relative, conditional, and mutual information notions are applied to metacommunities.Source: Introduction · pp. 12, 15
Magnitude for enriched categories unifies size invariants including cardinality and Euler characteristics, and metric-space magnitude can recover geometric quantities such as dimension, volume, and surface area.Source: Introduction · pp. 18–19
Maximum diversity is always the magnitude of some subspace, and in important cases equals the magnitude of the whole metric space.This relationship was used by Meckes to prove a result on Minkowski dimension.Source: Introduction · p. 19
The Hill numbers are the only diversity measures satisfying certain logically fundamental properties for communities modelled by species abundances.The section argues that measures should be justified biologically and behave logically, not merely gain authority from related disciplines.Source: Introduction · pp. 19–20
Characterization theorems determine the Hill numbers, Rényi entropies, q-logarithms, q-logarithmic entropies, value measures, and ℓp norms from short lists of properties.Power-mean characterization theorems are described as the engine of the book; probability theory also enables new characterizations of ℓp norms and power means.Source: Introduction · p. 21
Category theory yields a characterization of Shannon entropy by shifting attention from a probability space to the information lost by a deterministic process.The argument constructs a universal category acted on by the operad of probability distributions on finite sets.Source: Introduction · p. 22
Entropy modulo a prime has a characterization theorem analogous to real Shannon entropy, as does information loss modulo a prime.This extends the book’s axiomatic apparatus from the real field to integers modulo a prime.Source: Introduction · p. 22

Fundamental functional equations

The chapter solves three classical functional equations, with the nontrivial measurable solutions of the two-function equation given by constant multiples of q-logarithms.The ordinary logarithm is one member of this one-parameter family.Source: Fundamental functional equations · p. 26

1.1 Cauchy’s equation

Every measurable additive function from R to R is linear, extending the result from differentiable, continuous, or point-continuous additive functions.The argument proceeds through progressively weaker regularity assumptions, ending with measurability.Source: §1.1 · pp. 27–29
Assuming the axiom of choice, there are additive functions R → R that are not R-linear, while the section identifies measurability as a sufficient regularity condition for linearity.The text notes that without choice, it is consistent that all functions R → R are measurable, whereas choice is also consistent with ZF.Source: §1.1 · p. 30
Measurable solutions of the addition-to-multiplication, multiplication-to-addition, and multiplication-to-multiplication equations have respectively the forms f(x)=e^(cx), f(x)=c log x, and f(x)=x^c.For measurable additive f on [0, ∞), the section also gives f(x)=cx.Source: §1.1 · p. 30

1.2 Logarithmic sequences

A logarithmic sequence need not have the form c log n: its values at the primes can be chosen arbitrarily and determine the sequence uniquely.This contrasts with measurable solutions on the positive reals to f(xy)=f(x)+f(y), which have logarithmic form.Source: §1.2 · p. 33
An increasing logarithmic sequence is exactly of the form f(n)=c log n for all n ≥ 1, with c ≥ 0.This is Theorem 1.2.1, attributed to Erdős [90].Source: §1.2 · p. 33
The unified Erdős–Kátai–Máté theorem characterizes logarithmic sequences as f(n)=c log n under its stated sequence condition, covering increasing sequences and sequences whose consecutive differences tend to zero.The theorem is stated to imply both preceding results; the supplied passage does not show its full hypothesis.Source: §1.2 · p. 35
If an+1 − n/(n + 1) an tends to 0, then an+1 − an tends to 0.Lemma 1.2.4 is presented as useful for verifying the limit condition (1.15).Source: §1.2 · p. 37

1.3 The q-logarithm

The q-logarithms form a continuous one-parameter family of deformations of the natural logarithm, which is the q = 1 case.The text also gives a product formula and quotient formula for q-logarithms, and describes an involution lnq ↔ ln2−q with ln1 as its fixed point.Source: §1.3 · pp. 38–39
For measurable f, existence of a function g satisfying the stated equation is equivalent to f being constant or f = c lnq for some c, q ∈ R.The theorem characterizes the q-logarithms through a single functional equation.Source: §1.3 · p. 40

Shannon entropy

The chapter’s main theoretical goal is to prove that Shannon entropy is essentially the only quantity satisfying the chain rule.The chapter introduces finite-set probability distributions, derives the chain rule, and develops coding-theory and diversity interpretations.Source: Shannon entropy · pp. 42–43

2.1 Probability distributions on finite sets … 2.4 Entropy in terms of diversity

This section develops Shannon entropy and order-1 diversity for finite probability distributions, relating entropy to uniformity and effective species counts. It covers entropy’s bounds and chain rule, coding interpretations and composition, and the bounds, compositional properties, and limitations of order-1 diversity.

2.1 Probability distributions on finite sets

Ecological relative abundances model communities, but species boundaries may be arbitrary (as discussed in Hey), and abundance can mean counts, biomass, or plant-covered area.Source: §2.1 · p. 43
A distribution’s support identifies the elements with positive probability; full support means all elements are present, while the uniform distribution assigns equal probability to each.Source: §2.1 · p. 44
Composition is associative with the one-element distribution as identity, forming an operad; its tensor-product special case represents independent random variables and obeys associativity and identity laws.Source: §2.1 · pp. 47, 49
Composition combines a top-level distribution with conditional distributions; any distribution partitioned into blocks decomposes this way, with block weights uniquely determined, while conditionals on zero-weight blocks remain arbitrary.Source: §2.1 · pp. 47–48

2.2 Definition and properties of Shannon entropy

For distributions on a fixed finite set, Shannon entropy is greatest at the uniform distribution and least at a distribution concentrated on one element; it also represents average information or expected surprise.Source: §2.2 · pp. 50–51
For ecological distributions, exp(H(p)) is presented as a sensible measure of community diversity, and exponentials of other entropies are also meaningful diversity measures.Source: §2.2 · p. 50
Entropy is nonnegative, equals zero only at point masses, is at most log n with equality only for the uniform distribution, and is continuous.Source: §2.2 · pp. 51–52
The function ∂ is a nonlinear derivation, and up to a constant factor it is the only measurable function satisfying the stated product rule.Source: §2.2 · p. 52
The chain rule decomposes the entropy of a composite distribution into the entropy of its first-stage distribution and the weighted entropies of its conditional distributions.For independent events, the information gained from observing both equals the sum of the information gained from observing each.Source: §2.2 · pp. 52, 54

2.3 Entropy in terms of coding

An instantaneous code can assign symbol i the length ⌈log2(1/p_i)⌉, yielding expected word length strictly less than H(2)(p) + 1.When all frequencies are powers of 1/2, an instantaneous code can attain average word length exactly equal to entropy.Source: §2.3 · p. 57
For frequency distribution p, the infimum average word length over instantaneous codes is strictly less than H(2)(p) + 1.Here each symbol i has word length L_i and contributes p_iL_i to the average.Source: §2.3 · p. 59
By encoding symbols in sufficiently large blocks, the average bits per letter can approach H(2)(p) as closely as desired.The argument is valid even when successive symbols are dependent, because the actual block distribution has entropy at most H(p⊗10).Source: §2.3 · p. 59
Three-symbol blocks average 1.117 bits per symbol toward H(2)(p) ≈ 0.971, while blocks of 35 symbols achieve average word length below 1.For p = (0.6, 0.4), larger coding blocks improve on the one-symbol code and approach entropy.Source: §2.3 · p. 60
Ideal codes illustrate the chain rule: encoding French symbols as a base letter plus accent takes the same average bits per symbol as encoding the joint symbol distribution.The average is the entropy of the base-letter distribution plus the weighted average entropy of the accent distributions.Source: §2.3 · pp. 61–62

2.4 Entropy in terms of diversity

Order-1 diversity is the abundance-weighted geometric mean of species rarities; it ranges from 1 for a single present species to n for equal abundances and is an effective species count.A community with nearly all abundance concentrated in one species has diversity barely above 1; equal abundance across n species gives diversity n.Source: §2.4 · pp. 62–63
Diversity depends on relative abundances rather than absolute community size, so uniformly scaling a well-mixed community leaves its diversity unchanged.The text characterizes diversity as an intensive quantity, like density or temperature.Source: §2.4 · p. 63
D falls 90% (10^6 to 10^5) after 90% species loss and 25% (4 to 3) after losing one of four, unlike misleading percentage changes in Shannon entropy.Shannon entropy drops more for the four-species loss, despite that loss being smaller.Source: §2.4 · p. 65
Composite diversity depends only on group sizes and subgroup diversities, not internal distributions; this modularity supports the replication principle that equal, distinct groups multiply diversity.For 16 equal islands of diversity 4, destroying wildlife on half reduces group diversity by 50%.Source: §2.4 · pp. 65–67
For 16 equal islands, Shannon entropy allows claims that either 83% is preserved or 83% is lost when half are destroyed; both conflict with the reasonable 50% loss because entropy fails replication.The example illustrates how choosing a diversity measure can have environmental consequences.Source: §2.4 · p. 67
D ignores variation in species similarities, assigning equal diversity to equally abundant forests even when one contains more varied species.The book presents a remedy in Chapter 6.Source: §2.4 · p. 68

2.5 The chain rule characterizes entropy … 3.4 Relative entropy in measure theory, geometry and statistics

These sections characterize Shannon entropy through continuity and the general chain rule, then develop relative entropy as a coding comparison and diversity measure. They extend it to general measures and connect its geometry and statistical uses to Fisher distance, estimator precision, Fisher information, and the Jeffreys prior.

2.5 The chain rule characterizes entropy

For continuous functions satisfying the general chain rule, Faddeev’s theorem [94] characterizes I as cH without requiring symmetry.The theorem’s proof extends the result from uniform distributions to positive rational distributions, then to all distributions by continuity.Source: §2.5 · pp. 68, 69, 71

Relative entropy

Relative entropy compares distributions on the same finite set; it measures the coding penalty for using a system optimized for a reference distribution and, when exponentiated, relative diversity.Relative diversity reflects how unusual a community with distribution p is relative to a reference distribution r.Source: Relative entropy · p. 72, §3.1 · p. 73

3.1 Definition and properties of relative entropy

Relative entropy is nonnegative and vanishes exactly when p=r; its measurability, permutation invariance, vanishing, and chain rule characterize it up to a constant factor.The section also previews roles for relative entropy in measure theory, geometry, and statistics, including its connection to Fisher information and the Jeffreys prior.Source: §3.1 · pp. 74–76

3.2 Relative entropy in terms of coding

Section 3.2 interprets relative entropy in terms of coding, paralleling the earlier coding interpretation of Shannon entropy.Source: §3.2 · p. 76
Relative entropy is the extra bits per symbol incurred by encoding language p with machine r instead of its ideal machine p, and it is nonnegative, vanishing when p = r.It can also be arbitrarily large and asymmetric: the binary example gives H(2)(p ∥r) ≥ 499, while H(2)(r ∥p) < 1.Source: §3.2 · pp. 77–78
The coding interpretation also explains the relative-entropy chain rule: the extra cost per symbol equals the extra cost per letter plus the mean extra cost per accent.The Swiss–Canadian French example verifies this decomposition by conditioning accent costs on the letter being encoded.Source: §3.2 · p. 79

3.3 Relative entropy in terms of diversity

Relative diversity is the exponential of relative entropy: it is at least 1, equals 1 when p = r, and measures how unusual community p appears from reference community r.If a species occurs in p but is absent from r, relative diversity is infinite.Source: §3.3 · p. 80
Cross diversity measures the average rarity of p’s individuals from r’s viewpoint, while relative diversity measures how much more diverse p appears to r than to itself.For a uniform reference across n species, relative diversity is inversely proportional to p’s intrinsic diversity; it ranges from 1 for a uniform p to n for a single-species p.Source: §3.3 · pp. 80–82
For a subcommunity within a larger community, relative diversity is minimized at 1 when it is representative and reaches 1/w when it is isolated, where w is its share of the whole.The chapter’s relative and cross diversities are nontraditional beta- and gamma-diversities because they describe an individual subcommunity with reference to the larger community [290].Source: §3.3 · pp. 82–84

Measure theory

The direct finite-set formula for Shannon entropy cannot be applied to a probability measure on an arbitrary measurable space because there is no function ‘log ν’; relative entropy, however, generalizes to probability measures ν and µ.The passage introduces this generalization by defining the entropy of ν relative to µ.Source: Measure theory · pp. 84–85
Relative entropy extends to measures through the Radon–Nikodym derivative, allowing an infinite value when absolute continuity fails; with densities, it reduces to the corresponding density expression.Source: Measure theory · p. 85

Geometry

Relative entropy fails the triangle inequality: the example has H(p∥q) + H(q∥r) = 1.190 . . . < 1.757 . . . = H(p∥r), so it only crudely resembles a metric.Source: Geometry · p. 86
On the finite simplex, √H is an infinitesimal metric whose global counterpart is (up to a constant) Fisher distance; the Fisher metric makes it isometric to a radius-2 sphere orthant.Source: Geometry · pp. 86, 90

Statistics

For fixed sample size, the probability of specified observations decreases as their empirical distribution’s entropy increases, from probability 1 at zero entropy to 1/n^n for the uniform distribution.Source: Statistics · p. 92
Maximum likelihood amounts to choosing the parameter that minimizes relative entropy between the empirical distribution and the candidate distribution.Source: Statistics · p. 92

Measure theory, geometry and statistics

For a smooth family of probability densities, the second derivative of relative entropy at the matching parameter is Fisher information, measuring how quickly the distribution changes nearby.Source: Measure theory, geometry and statistics · p. 93, Figure 3.3 · p. 93
The Cramér–Rao bound gives a lower bound on an unbiased estimator’s variance, with low Fisher information implying greater imprecision.When distributions change slowly near the true parameter, nearby values are hard to distinguish; the bound formalizes why unbiased estimates must then have large variance.Source: Measure theory, geometry and statistics · p. 94
The Jeffreys prior is invariant under reparametrization: pushing the prior for standard deviation forward by squaring yields the prior for variance.The prior is normalized, if possible, to integrate to 1.Source: Measure theory, geometry and statistics · pp. 94–95
Uniform priors can fail reparametrization invariance: the example assigns probabilities 1/2 and 1/4 to equivalent standard-deviation and variance events.The probabilities concern standard deviation below 5 and variance below 25, respectively.Source: Measure theory, geometry and statistics · p. 95

3.5 Characterization of relative entropy · 4.1 q-logarithmic entropies · 4.2 Power means

These sections characterize relative entropy, q-logarithmic entropies, and power means through axioms and properties including symmetry, chain rules, naturality, and algebraic structure. They also examine continuity, monotonicity, consistency, and homogeneity, and clarify that q-logarithmic entropy is not itself an effective-number measure of diversity.

3.5 Characterization of relative entropy

Relative entropy is uniquely characterized, up to a scalar multiple, by the four stated properties.The vanishing property is essential: cross entropy satisfies the other listed properties but not vanishing.Source: §3.5 · pp. 95, 98
Measurability and the chain rule force the auxiliary function L to be a scalar multiple of the logarithm, yielding the characterization for distributions with or without full support.The proof first establishes the result for full-support distributions, then extends it to arbitrary distributions.Source: §3.5 · pp. 97–98

Deformations of Shannon entropy

Two one-parameter families deform Shannon entropy, with both recovering it at q = 1; q-logarithmic entropy is uniquely characterizable for q ≠ 1 without regularity assumptions.The same absence of regularity conditions applies to the introduced q-logarithmic relative entropy.Source: Deformations of Shannon entropy · pp. 101–102

4.1 q-logarithmic entropies

The q-logarithmic entropy S_q(p) is expected surprise when the surprise assigned to probability p is lnq(1/p).It vanishes exactly for a point-mass distribution and reaches lnq(n) exactly for the uniform distribution.Source: §4.1 · pp. 103–104
Because S_q(un) = lnq(n) < n, q-logarithmic entropy is not an effective number and should not itself be used as a diversity measure.S_0(p) is one less than the number of species present, S_1(p) is Shannon entropy, and S_2(p) is the probability that two randomly chosen individuals belong to different species.Source: §4.1 · p. 104
q-logarithmic entropies satisfy a simple chain rule, their chief stated advantage over Rényi entropies.The section also notes that S_2(p) is not an effective number and refers to Section 4.3 for a remedy.Source: §4.1 · pp. 105–106
For q ≠ 1, symmetry and the stated q-chain rule characterize ordinary q-logarithmic entropy up to a constant, without regularity conditions.The characterization uses the special chain-rule case stated in the theorem; the passage contrasts this with Shannon entropy, whose characterization requires regularity.Source: §4.1 · p. 106
q-logarithmic relative entropy is characterized up to a constant by its chain rule and permutation invariance, without regularity or the vanishing axiom.The passage contrasts these requirements with the characterization of ordinary relative entropy.Source: §4.1 · p. 109

4.2 Power means

Power means support later treatment of Rényi entropies and diversity measures; their orders 1, 0, and −1 give the arithmetic, geometric, and harmonic means.The section collects their basic properties and terminology.Source: §4.2 · pp. 110–111
Power means are continuous in their order, but means of order ≤ 0 are not continuous in the weights; they increase with order when supported coordinates are positive.The increase is strict unless all coordinates on the support have the same value.Source: §4.2 · pp. 112–114
For every order, power means are natural, which yields symmetry, absence-invariance, and the repetition property.Absence-invariance means zero-weight coordinates can be ignored.Source: §4.2 · pp. 116–117
For every order t ∈ [−∞, ∞], the power mean is homogeneous and increasing on [0, ∞); it is strictly increasing on (0, ∞) for finite t and on [0, ∞) for t > 0.Source: §4.2 · pp. 118–119
For every order, power means satisfy a chain rule that yields modularity: composite means depend only on the outer weights and the component means.The chain rule applies on [0, ∞), and modularity follows as a corollary.Source: §4.2 · pp. 119–120
Power means are multiplicative for every order on [0, ∞).This property follows from the chain rule and homogeneity.Source: §4.2 · pp. 119, 121
A lemma establishes a connection between power means and q-logarithms for q ≥ 0 when the input coordinates on the distribution’s support are positive.The stated condition allows zero-valued coordinates outside the support.Source: §4.2 · p. 121

4.3 R´enyi entropies and Hill numbers · 4.4 Properties of the Hill numbers

These sections define Hill numbers as effective-number versions of Rényi entropies and use diversity profiles to compare diversity across different weightings of rare species. They establish key properties, including continuity, symmetry, monotonicity, bounds, characterization of abundance distributions up to permutation, and chain-rule consequences such as modularity, multiplicativity, and the replication principle.

4.3 R´enyi entropies and Hill numbers

Rényi entropies retain the logarithmic additivity property for all orders, but their characterization theorem is more involved than the one for q-logarithmic entropies.The q-logarithmic and Rényi entropies are related by increasing transformations, so maximizing or minimizing one is equivalent to doing so for the other.Source: §4.3 · pp. 122–123
Hill numbers convert Rényi entropies into effective numbers of diversity, with the order q controlling sensitivity to rare species.Higher q values give rare species less weight; plotting Dq against q displays these viewpoints together as a diversity profile.Source: §4.3 · pp. 124–126
For great apes, D0(p) = 8 records eight species, while diversity at nearly every other q is almost 1 because humans dominate the community.The steep profile drop indicates that seven species are exceptionally rare.Source: §4.3 · p. 127, Figure 4.3 · p. 128
Diversity profiles can reverse community rankings across viewpoints: the bird communities switch from A being more diverse at low q to B at high q.At q = ∞, the comparison reflects dominance by the most common species; profile crossings also occur in experimental data.Source: §4.3 · p. 127, Figure 4.4 · p. 129

4.4 Properties of the Hill numbers

For every distribution, Dq is continuous and decreases with q, strictly unless the distribution is uniform on its support.Although diversity profiles arising from experimental data often appear convex, arbitrary profiles need not be convex.Source: §4.4 · pp. 129–130, Figure 4.5 · p. 130
For q > 0, Dq ranges from 1 at a monopoly to n at the uniform distribution, attaining the same extrema as D1.Equality at the lower bound occurs exactly when one species has probability 1; equality at the upper bound occurs exactly for the uniform distribution.Source: §4.4 · pp. 130–131
For q<0, the uniform distribution does not maximize Hill numbers, so these orders are generally not used as diversity measures and instead measure rarity.They can give high values to communities containing a very rare species.Source: §4.4 · p. 131
For every order q, Dq is continuous in positive probabilities; it is continuous on the simplex for q>0 and discontinuous at boundary points when q≤0.Species richness D0 illustrates the boundary discontinuity: an abundance of 0.0001 differs qualitatively from zero.Source: §4.4 · pp. 131–132
For every order q, Hill numbers are symmetric and absence-invariant: relabeling species or including an absent species does not change diversity.The profile likewise encodes the distribution independently of species ordering.Source: §4.4 · p. 132
A Hill number's diversity profile determines its abundance distribution up to a permutation of species.The converse to symmetry is established: distributions with the same profile must agree up to permutation.Source: §4.4 · p. 132
For every order q, the chain rule expresses a combined distribution's diversity using the group weights and subgroup diversities, with a second form using an escort distribution.The first version uses power means of the subgroup diversities; the alternative weights them by the escort distribution of the group weights.Source: §4.4 · pp. 133–134
For every order q, Hill numbers are modular, multiplicative, and satisfy the replication principle.Modularity means combined diversity depends only on group weights and subgroup diversities; replication gives nD(p) for n disjoint communities with the same distribution p.Source: §4.4 · pp. 136–137

4.5 Characterization of the Hill number of a given order · 5.1 Quasiarithmetic means · 5.2 Unweighted means

This section develops characterizations of positive-order Hill numbers and quasiarithmetic and unweighted means, using decomposability, affine equivalence, homogeneity, and local-to-global structure to connect these means with power means.

4.5 Characterization of the Hill number of a given order

For each q>0, continuous, symmetric effective-number functions satisfying the specified decomposability condition are uniquely Dq.The condition allows the within-group average to use a weighting θ determined by group sizes; the characterization forces that weighting to be the q-escort distribution.Source: §4.5 · pp. 137, 138, 142
The assumed weighting θ is forced to equal the q-escort distribution, completing the route from the decomposability condition to the Hill number.The proof first establishes the result for positive rational distributions, then extends it to all distributions by continuity.Source: §4.5 · pp. 140–141

Means

The chapter focuses on characterizing power means, using homogeneity to usually avoid continuity assumptions and including the non-strictly increasing endpoint means max and min.It culminates in four characterizations of unweighted means and converts them into weighted-mean characterizations.Source: Means · pp. 143–144

5.1 Quasiarithmetic means

A quasiarithmetic mean transports the arithmetic mean along a homeomorphism; all finite-order power means qualify, but the maximum and minimum means do not.The section treats the finite-order power means on (0, ∞) as quasiarithmetic, while M±∞ are not.Source: §5.1 · pp. 145–146
The section frames generator uniqueness, power-mean recognition, and local-to-global quasiarithmeticity as questions whose answers involve affine maps.An affine map on an interval has the form α(x) = ax + b.Source: §5.1 · p. 146
Two quasiarithmetic means are equal for all weighted inputs exactly when their generators differ by an affine map; equality on uniform inputs alone is equivalent.The equivalence also connects equality on all inputs to equality on uniform inputs for every n.Source: §5.1 · pp. 147–148
A quasiarithmetic mean is homogeneous under positive scaling exactly when it is a power mean of some real order.The theorem equates homogeneity on uniform inputs with homogeneity on all weighted inputs.Source: §5.1 · pp. 149–150
If a mean restricts to a quasiarithmetic mean on every nontrivial closed bounded subinterval, it is quasiarithmetic on the entire interval.The proof aligns generators on nested subintervals by affine changes so they extend consistently.Source: §5.1 · pp. 150–151

5.2 Unweighted means

Unweighted power means increase for all orders on [0, ∞); finite orders are strictly increasing on (0, ∞), while orders t≤0 are not strictly increasing on [0, ∞).The extrema M∞ and M−∞ are increasing but not strictly increasing on any nontrivial interval.Source: §5.2 · pp. 153–154
Every power mean is decomposable, and this implies that all quasiarithmetic means are decomposable.Decomposability is presented as the unweighted analogue of the chain rule for weighted means.Source: §5.2 · pp. 154–155
Assuming homogeneity makes the continuity condition in the Kolmogorov–Nagumo theorem unnecessary.Homogeneity is motivated by the focus on power means.Source: §5.2 · p. 156
Repeating each input r times leaves M unchanged for symmetric, consistent, decomposable sequences.The lemma applies for all r, n ≥ 1 and inputs in the interval.Source: §5.2 · p. 156
For a symmetric, consistent, decomposable sequence on [a, b], a rational-indexed function ψ is constructed whose monotonicity follows that of M.The function ψ is intended to be the inverse of a normalized generator ϕ, restricted to the rationals.Source: §5.2 · pp. 156–158
Symmetric, decomposable, homogeneous means on (0, ∞) are multiplicative.The authors first establish that decomposable, homogeneous means on (0, ∞) are automatically consistent.Source: §5.2 · p. 158

5.3 Strictly increasing homogeneous means · 5.4 Increasing homogeneous means · 5.5 Weighted means

This section develops axiomatic characterizations of increasing homogeneous means, including extensions from unweighted to weighted means. It also introduces diversity measures that account for species abundances and similarities, generalize Hill numbers, and connect maximum diversity to matrix magnitude.

5.3 Strictly increasing homogeneous means

A strictly increasing function on [0, 1] ∩ ℚ with the stated matching supremum and infimum conditions extends uniquely to a continuous, strictly increasing function on [0, 1].The conditions ensure that the rational-domain function has no jump discontinuities.Source: §5.3 · p. 159
On (0, ∞), the stated properties characterize M_t for t ∈ ℝ; on [0, ∞), they characterize M_t for t ∈ (0, ∞).These results apply to symmetric, strictly increasing, decomposable, homogeneous unweighted means.Source: §5.3 · pp. 159, 161, 163

5.4 Increasing homogeneous means

For a symmetric, decomposable mean, equality at one pair of coordinate values persists when both values are repeated the same number of times.Specifically, M(x₁, …, xₘ, u) = M(x₁, …, xₘ, v) implies M(x₁, …, xₘ, n * u) = M(x₁, …, xₘ, n * v) for all n ≥ 0.Source: §5.4 · p. 165
If a symmetric, consistent, decomposable mean is increasing but not strictly increasing, some distinct inputs u and v yield equal outputs with a common input x.Under symmetry, consistency, decomposability, increasingness, and homogeneity, an equality M(x, u) = M(x, v) for x > 0 and distinct u, v ≥ x implies M(a, b) = a for some 0 < a < b.Source: §5.4 · pp. 165–166
If such a mean is not strictly increasing, it must be M±∞, the minimum or maximum mean.The proof reduces non-strict increase to a pair on which the mean equals one endpoint, forcing the minimum or maximum.Source: §5.4 · p. 170
A sequence is a power mean Mt for some t ∈[−∞, ∞] exactly when it is symmetric, increasing, decomposable, and homogeneous on (0, ∞).The characterization combines the strictly increasing case with the non-strict case.Source: §5.4 · p. 170
On [0, ∞), the same power-mean characterization requires continuity and that none of the functions be identically zero.Neither condition can be dropped without admitting examples that satisfy the other conditions but are not power means.Source: §5.4 · pp. 170–172

5.5 Weighted means

The section uses elementary property implications and a conversion method to derive four weighted-mean characterization theorems from unweighted results.The resulting theorems are summarized in Table 5.2.Source: §5.5 · p. 172, Table 5.2 · p. 172
For weighted means, consistency together with modularity implies the chain rule, while consistency with the chain rule implies the repetition property.Source: §5.5 · p. 173
A weighted mean with the repetition property and increasingness has the transfer property: shifting weight from a smaller argument to a larger one does not decrease the mean.The condition is xi ≤ xi+1, with a transfer δ satisfying 0 ≤ δ ≤ pi.Source: §5.5 · p. 174
Symmetry and the transfer property imply an approximation property providing distributions with rational coordinates.Source: §5.5 · p. 174
Several weighted properties pass to the unweighted means M(un, −), and consistency with the chain rule makes those unweighted means decomposable.The inherited properties include symmetry, consistency, increasingness, and strict increasingness; homogeneity also passes when the interval is closed under multiplication.Source: §5.5 · p. 175
Four characterization theorems for weighted power means are derived from the corresponding theorems for unweighted means.Theorem 5.5.10 is later used to characterize measures of community value and, in turn, Hill numbers.Source: §5.5 · pp. 177–178, Species similarity and magnitude · p. 180
Symmetry, absence-invariance, consistency, increasingness, modularity, homogeneity, and continuity in the second argument characterize weighted power means M_t for t ∈ [−∞, ∞].Source: §5.5 · p. 178

Species similarity and magnitude

Similarity-sensitive measures combine relative abundances with a species-similarity matrix, generalize Hill numbers, and allow genetic, phylogenetic, or functional similarity.Setting the similarity matrix to the identity recovers the Hill numbers; the framework also yields similarity-sensitive Rényi and q-logarithmic entropies, including on finite metric spaces.Source: Species similarity and magnitude · p. 180
Every similarity matrix has a maximum diversity and a distribution maximizing diversity at every order q, both independent of q.Source: Species similarity and magnitude · p. 180
Maximum diversity is closely related to matrix magnitude, a size-like invariant encompassing quantities such as cardinality, Euler characteristic, volume, surface area, and dimension.The section surveys magnitude and its connections to geometric invariants.Source: Species similarity and magnitude · pp. 180–181

6.1 The importance of species similarity · 6.2 Properties of the similarity-sensitive diversity measures

This section introduces diversity measures that account for species similarities and connects them to species relationships, including graph entropy and non-adjacency. It examines their behavioral properties, including diversity profiles, bounds, continuity, invariances, responses to reclassification and finer taxonomic resolution, and chain-rule, modularity, replication, and entropy properties under specified assumptions.

6.1 The importance of species similarity

Following Leinster and Cobbold, the section introduces diversity measures that combine relative abundances with a species-similarity matrix.Similarity can represent genetic, functional, phylogenetic, or taxonomic relationships, allowing the measures to reflect the scientific problem at hand.Source: §6.1 · pp. 181, 183, 187
A binary similarity matrix corresponds to a reflexive graph, a mathematically useful special case with no claimed ecological relevance.The example sheds light on computational aspects of calculating maximum diversity.Source: §6.1 · p. 183
Similarity-adjusted ordinariness is (Zp)i, its reciprocal extends species specialness, and the resulting order-q measures define diversity for q in [0, ∞].When Z = I, specialness reduces to 1/pi and the measures recover Hill numbers.Source: §6.1 · pp. 184–185
For integer q ≥ 2, diversity can be estimated from sampled group similarities without classifying organisms into species, a potentially useful approach for microbial systems.Repeatedly draw q samples, record their group similarity, and average it to estimate diversity.Source: §6.1 · pp. 186–187
Using Hill numbers amounts to choosing the naive model, which treats distinct species as having nothing in common and always overestimates diversity.The framework makes this ecological assumption explicit rather than eliminating the choice of similarity.Source: §6.1 · p. 188
With naive similarity, the butterfly canopy is at least as diverse as the understorey, but taxonomic similarity makes the understorey more diverse for most q values.The canopy’s concentration in three species of the same genus explains why recognizing their similarity reduces its diversity more.Source: §6.1 · pp. 188–189, Figure 6.1 · p. 189
Similarity-sensitive Rényi and q-logarithmic entropies depend on the metric between symbols and recover their standard finite-set definitions when distinct symbols have infinite distance.The metric can represent how easily symbols are confused or whether one can substitute for another.Source: §6.1 · p. 191
At q=2, q-logarithmic entropy is Rao’s quadratic entropy, the dissimilarity of a random pair; on graphs, it is the probability that randomly chosen vertices are not adjacent.Graph entropy is high when high-probability vertices tend not to be adjacent.Source: §6.1 · pp. 191–192

6.2 Properties of the similarity-sensitive diversity measures

Diversity profiles decrease with q, and are constant exactly when all species present have equal ordinariness; greater perceived similarity means lower perceived diversity.Ordinariness here is (Zp)i.Source: §6.2 · p. 192
For the stated similarity matrices, diversity lies in [1, n]: one-species communities attain 1, while n occurs only for Z=I and the uniform distribution.The measures can be interpreted as effective numbers of completely dissimilar species.Source: §6.2 · p. 193
Diversity is continuous in q and Z, and in p for 0<q<∞; continuity in p can fail at q=0 or q=∞ when similarities are present.At q=0, diversity is species richness; at q=∞, a rare species can affect the maximum ordinariness until it disappears.Source: §6.2 · pp. 194–196
Diversity is unchanged by reordering species, omitting absent species, or merging identical species.Identical species are those with the same similarities to every species.Source: §6.2 · pp. 197–198
Similarity-sensitive measures address the challenge of unstable species boundaries by avoiding an all-or-none treatment of species.Species boundaries can be changeable and somewhat arbitrary, including among well-studied large mammals.Source: §6.2 · p. 198
For order ∞, splitting a species into totally dissimilar descendants raises diversity 100% (1.66... to 3.33...), while assigning them similarity 0.9 changes the profile only slightly.This behavior follows from the identical species property and continuity in Z: treating the descendants as identical leaves the profile unchanged, so near-identity leaves it nearly unchanged.Source: §6.2 · pp. 198–199, Figure 6.2 · p. 199
With coherent similarity assignments, genus-, species-, and subspecies-level surveys are comparable on one scale, and diversity increases with finer resolution without disproportionate jumps.Differences between successive surveys reflect within-genus variation or intraspecific variation, respectively.Source: §6.2 · p. 199
When species on different islands are completely dissimilar, the chain rule expresses total diversity using island sizes and diversities, so the total depends only on q, w, and those diversities.The result assumes distinct, completely dissimilar species across islands.Source: §6.2 · pp. 200–201
For n equal-size islands with the same diversity d, total diversity is nd, establishing the replication principle for the similarity-sensitive measures.The text notes that this principle avoids the problems illustrated by the oil company example.Source: §6.2 · p. 201
For q in [0, ∞), a further chain rule for q-logarithmic entropy has a simple form that generalizes the q-logarithmic entropy chain rule in equation (4.2).It uses the same setup as Proposition 6.2.11.Source: §6.2 · p. 202

6.3 Maximizing diversity · 6.4 Introduction to magnitude

These sections examine maximum diversity for similarity matrices and reflexive graphs, including how maximizers relate to independent sets and matrix conditions. They introduce magnitude as a notion of mathematical size, tracing its development across categories and metric spaces and surveying its behavior and related homology and cohomology theories.

6.3 Maximizing diversity

For symmetric Z, one distribution maximizes diversity at every order q, and its diversity profile is flat at the common maximum Dmax(Z).The maximum value is independent of q.Source: §6.3 · pp. 203–204
To find a distribution maximizing every order, it suffices to maximize any one nonzero order; at q=0, Z=I instead allows any full-support distribution to maximize richness.Source: §6.3 · p. 205

6.4 Introduction to magnitude

For M and M^T admitting weightings, magnitude |M| is the sum of weighting entries, independent of the chosen weighting; for invertible M, it is the sum of M^-1 entries.For symmetric M, magnitude is defined whenever M admits at least one weighting.Source: §6.3 · pp. 205–206

6.3 Maximizing diversity

Dmax(Z) is the largest magnitude |ZB| among submatrices admitting nonnegative weightings, and all maximizers are normalized weightings on subsets attaining that maximum.On a maximizing support, ordinariness (Zp)i is constant for species present; maximizing distributions need not have full support. For the three-species pond example, the unique maximizer is (0.261, 0.261, 0.478).Source: §6.3 · pp. 206–208
Without symmetry, suprema may be unattained or q-dependent and simultaneous maximizers may not exist; nonsymmetric matrices also model phylogenetic diversity.The failure of the theorem is an important restriction because nonsymmetric similarity matrices have practical uses.Source: §6.3 · p. 209
For a reflexive graph G with adjacency matrix Z, maximum diversity Dmax(Z) equals the independence number α(G).Source: §6.3 · p. 210
On a reflexive graph, a maximum-diversity distribution is uniform on a largest independent set, extended by zero; unless there are no edges between distinct vertices, it lacks full support.Different largest independent sets can yield different maximizing distributions.Source: §6.3 · pp. 210–211
Kolmogorov’s ε-entropy log Nε(A) is closely related to the logarithm of maximum diversity as ε approaches zero.The ε-covering number counts the fewest closed ε-balls needed to cover a finite metric space.Source: §6.3 · p. 211
Computing maximum diversity for an arbitrary symmetric similarity matrix is NP-hard, as hard as finding a graph’s independence number.Assuming P ≠ NP, neither maximum diversity nor the support of a maximizing distribution can be computed in polynomial time.Source: §6.3 · p. 212
For ultrametric matrices or matrices sufficiently close to identity, maximum diversity equals magnitude and has a unique full-support maximizing distribution, computable in polynomial time.Ultrametric matrices include those defined from phylogenetic or taxonomic trees.Source: §6.3 · pp. 213–214
Removing a particularly ordinary species can increase diversity because its removal raises the relative abundances of less ordinary species.The example contrasts one oak species with ten pine species, then considers adding and removing an eleventh pine species.Source: §6.3 · p. 214
A full-support maximizing distribution exists iff Z is positive semidefinite with a positive weighting; every maximizer has full support iff Z is positive definite with a positive weighting.The latter condition is equivalent to a unique full-support maximizing distribution and to Dmax(ZB) < Dmax(Z) for every nonempty proper subset B.Source: §6.3 · p. 215

6.4 Introduction to magnitude

Magnitude is presented as a general notion of mathematical size encompassing cardinality, volume, and Euler characteristic.The proposed general notion is the magnitude of an enriched category.Source: §6.4 · pp. 216–217
For a finite category, magnitude is the magnitude of its matrix of Hom-set sizes when it exists; maximum diversity is always the magnitude of a principal submatrix.For some similarity matrices, including those in the cited ultrametric and near-identity cases, maximum diversity equals the magnitude of the whole matrix.Source: §6.3 · p. 214, §6.4 · pp. 216–218
Under suitable hypotheses, categorical magnitude and topological Euler characteristic determine each other.For the illustrated category, both its magnitude and the Euler characteristic of its classifying space are 0.Source: §6.4 · pp. 218–219
Enriched categories unify ordinary categories, linear categories, and metric spaces, allowing magnitude to extend to these settings.For finite-dimensional algebras under finiteness hypotheses, the category of indecomposable projective modules has magnitude given by the Euler form of the direct sum of simple modules; its matrix is the Cartan matrix [67].Source: §6.4 · pp. 220–221
Magnitude of a finite metric space can be interpreted as its effective number of completely separate points.As points in a two-point space move farther apart, magnitude increases toward 2; when all nonzero distances are infinite, it equals the space’s cardinality.Source: §6.4 · pp. 222–223
A three-point long thin triangle has magnitude near 1, about 2, or near 3 as its scale moves from small to moderate to large.More generally, a finite metric space’s magnitude function has finitely many singularities, is increasing for sufficiently large scale, and converges to its cardinality as scale tends to infinity.Source: §6.4 · pp. 223–224
Graph magnitude has rational-function or power-series forms, shares some invariance properties with the Tutte polynomial but is not its specialization, and satisfies multiplicativity and inclusion-exclusion under strict hypotheses.It is proposed as a graph-theoretic analogue of cardinality.Source: §6.4 · p. 225
Magnitude for graphs and enriched categories can be categorified by graded homology theories whose Euler characteristic recovers magnitude.Graph magnitude homology is finer than magnitude because graphs with equal magnitude can have different homology groups [131].Source: §6.4 · pp. 225–226
Magnitude homology is metric rather than topological: for a closed subset of R^n, its first group is trivial if and only if the subset is convex, and all groups are trivial for convex subsets [174] [172].This is presented as a metric analogue of the vanishing of topological homology for contractible spaces.Source: §6.4 · p. 226
For finite metric spaces, the magnitude cohomology ring determines the space uniquely up to isometry [141].Magnitude cohomology has a product formally resembling the ordinary cup product, but it is noncommutative.Source: §6.4 · p. 226

6.5 Magnitude in geometry and analysis · 7.1 Introduction to value · 7.2 Value and relative entropy

This section develops mathematical frameworks for quantifying diversity, from magnitude as a measure of geometric size and dimension to value aggregation as a way to unify established diversity measures. It connects these frameworks to capacity theory, power means, welfare comparisons, the chain rule, and relative entropy and its associated Fisher geometry.

6.5 Magnitude in geometry and analysis

For compact metric spaces of negative type, plausible extensions of finite-space magnitude agree, though some such spaces do not admit weight measures.When a weight measure exists, its total mass equals magnitude; negative type includes Euclidean and ℓ1 subspaces, ultrametric spaces, hyperbolic spaces, and geodesic spheres.Source: §6.5 · pp. 227–228
For convex bodies in R^n with the ℓ1 metric, magnitude is a degree-n polynomial whose i-dimensional coefficients are geometric measures given by ℓ1 intrinsic volumes.These intrinsic volumes sum the i-dimensional volumes of projections onto coordinate subspaces; the line-segment and rectangle examples illustrate the geometric interpretation of magnitude coefficients.Source: §6.5 · pp. 228, 230
For odd-dimensional Euclidean unit balls, magnitude is rational in the radius, while even-dimensional ball magnitudes remain unknown; smooth odd-dimensional sets also have asymptotic coefficients related to intrinsic volumes.For smooth bounded sets, the first three coefficients correspond, up to dimension-dependent constants, to intrinsic volumes; the leading coefficient recovers volume.Source: §6.5 · pp. 230–232
For Euclidean A, maximum diversity equals the Bessel capacity C_(n+1)/2(A), and magnitude differs from this capacity by at most a dimension-dependent factor.Maximum diversity takes a supremum over Borel probability measures, whereas magnitude allows signed measures; consequently Dmax(A) ≤ |A|.Source: §6.5 · p. 232
For compact A ⊆ R^n, the large-scale growth of |tA| recovers its Minkowski dimension, including non-integer dimensions such as log 2/log 3 for the middle-thirds Cantor set.For sets with nonzero volume, |tA| grows like t^n; convex-set magnitude can also be bounded by a polynomial with coefficients proportional to intrinsic volumes.Source: §6.5 · p. 233

Value

A precise account of how a whole’s value aggregates the relative sizes and values of its parts yields diversity and uniquely characterizes the Hill numbers, equivalently the Rényi entropies.The setup assigns relative sizes p1, . . . , pn and values v1, . . . , vn to the parts, then asks for a single aggregate value.Source: Value · p. 234
The family σq includes Hill numbers Dq, similarity-sensitive diversities DZq, phylogenetic diversity measures of Chao, Chiu and Jost, and essentially ℓp norms.Source: Value · p. 235
Under reasonable properties, value measures are exactly the family σq; for communities modeled by relative abundances, reasonable diversity measures are exactly the Hill numbers, characterized as a family rather than for fixed q.Source: Value · pp. 235–236

7.1 Introduction to value

For q > 0, σq(p, v) is at most the sum of part values, with equality exactly when v is proportional to p; fixed total value is maximized by allocating it in proportion to part sizes.Source: §7.1 · p. 238
For q ∈ [0, ∞], σq satisfies the Pigou–Dalton principle, while for q < 0 redistribution from richer to poorer decreases overall welfare.Source: §7.1 · p. 240
With equal values across species, σq gives Hill numbers Dq; assigning species value p_i/(Zp)_i recovers similarity-sensitive diversity DZq.Source: §7.1 · p. 241
For disjoint subcommunities with no shared species and complete dissimilarity between them, σq aggregates their diversities by the chain rule to give whole-community diversity.Source: §7.1 · p. 242
The value measure σq represents the Chao–Chiu–Jost phylogenetic measure and unifies Hill numbers, similarity-sensitive diversity, and diversity across completely dissimilar subcommunities.Branches are valued by their duration as a proportion of evolutionary time, independent of present-day abundances.Source: §7.1 · pp. 245–246
Using branch length L(b) as the value makes σ0 exactly Faith’s phylogenetic diversity, while using relative branch lengths gives a unitless community value.The unitless version equals CCJq; with v(b) = L(b), the value is measured in the same units as branch length.Source: §7.1 · p. 246

7.2 Value and relative entropy

Rènyi and q-logarithmic relative entropies are simple transformations of σq and converge to ordinary relative entropy as q approaches 1.The generalized relative-entropy deformation converges to ordinary relative entropy; choosing λ = log gives Rènyi relative entropy, and λ(x) = x − 1 gives q-logarithmic relative entropy.Source: §7.2 · p. 248
Deforming relative entropy by order q only scales the induced metric by q, leaving the q-analogues of Fisher distance and information proportional to their classical versions.The q-analogue of the Jeffreys prior is exactly the classical notion.Source: §7.2 · p. 249
At q = 1/2, both relative entropies are symmetric in p and r and determine the Fisher distance between the distributions.Order 1/2 is the only parameter value with this symmetry, and the entropies are increasing, invertible transformations of Fisher distance.Source: §7.2 · p. 249

7.3 Characterization of value · 7.4 Total characterization of the Hill numbers · 8.1 Joint entropy, conditional entropy and mutual information

These sections axiomatize value measures and relative-abundance diversity, characterizing the Hill numbers, including a criterion for those of nonnegative order. They then connect information-theoretic quantities to ecological metacommunity measures by introducing joint, conditional, and mutual information for finite-valued variables and describing their bounds, equality conditions, and illustrations.

7.3 Characterization of value

The chain rule requires consistent aggregation whether a whole is evaluated directly or through its parts and subparts.For example, aggregating country values first by continent and then globally must match direct aggregation across all countries.Source: §7.3 · p. 251
Symmetry, absence-invariance, increase, homogeneity, continuity in positive probabilities, an effective number, and the chain rule characterize σ as σq for some q ∈ [−∞, ∞].The theorem focuses on strictly positive values; an analogous characterization on [0, ∞) requires continuity in values as well.Source: §7.3 · pp. 250, 253

7.4 Total characterization of the Hill numbers

Symmetry, absence-invariance, continuity in positive probabilities, normalization, modular-monotonicity, and replication characterize D as a Hill number Dq for some q ∈ [−∞, ∞].The proof constructs a value measure from D and applies the value-measure characterization; the theorem covers all q simultaneously and has simpler hypotheses with more direct ecological explanations than the earlier characterization.Source: §7.4 · pp. 255, 257
Replication and normalization imply that D is an effective number, while modular-monotonicity implies modularity.The proof then establishes that D(p ⊗ r) depends only on D(p) and D(r) when comparing distributions with matching diversity values.Source: §7.4 · p. 257
The proof’s bridge to the value-measure characterization uses multiplicativity and a unique homogeneous extension, yielding a value measure with symmetry, continuity, monotonicity, and the chain rule.The extension is from values in im D to positive inputs; the resulting value measure also has absence-invariance and the effective-number property.Source: §7.4 · pp. 259, 262, 264
Under the stated axioms, every diversity measure on probability distributions is a Hill number Dq for some q in [−∞, ∞].The proof constructs a value measure from D, shows it satisfies the characterization theorem’s hypotheses, and concludes D = Dq.Source: §7.4 · p. 265
The characterization also gives a criterion for q ≥ 0: Dq is maximized at the uniform distribution exactly when Dq(p) ≤ 2 for every binary distribution p.The text gives the equivalent conditions Dq(p) ≤ Dq(un) for all n and p, Dq(p) ≤ 2 for all p in ∆2, and q ∈ [0, ∞].Source: §7.4 · p. 266

Mutual information and metacommunities

Mutual information is motivated as a measure of dependence between finite-valued variables, including cases of complete dependence and independence that covariance and correlation cannot generally cover.The section also notes that covariance and correlation are usually defined for variables in Rn, not arbitrary pairs of finite sets.Source: Mutual information and metacommunities · p. 267
For associated communities, the section raises questions about between-community variation, contributions to overall diversity, and typicality that single-community diversity measures do not answer.Examples include gut microbiomes of healthy and unhealthy adults and aquatic life across salinity zones near a river mouth.Source: Mutual information and metacommunities · p. 267
The chapter exponentiates information-theoretic quantities into ecological measures for metacommunities and reduces them to relative entropy, with diversity measures also expressed in terms of value.Some measures describe individual subcommunities and others the whole metacommunity; the treated diversity measures are the q = 1, Z = I case of Reeve et al. [290], whose framework allows general q and Z.Source: Mutual information and metacommunities · pp. 267–268

8.1 Joint entropy, conditional entropy and mutual information

Section 8.1 introduces classical information-theoretic quantities for pairs of distributions on potentially different finite sets, using the language of random variables.It distinguishes these from relative entropy and cross entropy, which apply to distributions on the same set.Source: §8.1 · p. 268

Joint entropy

Joint entropy is the entropy of the pair (X, Y); it equals H(X) when Y is constant and equals H(X) = H(Y) when X and Y are identical.For independent X and Y, the joint distribution is the product of their distributions.Source: Joint entropy · p. 269

Conditional entropy

Conditional entropy H(X | Y) is the expected entropy of X conditional on Y, is nonnegative, and satisfies H(X | Y) = H(X, Y) − H(Y).It is zero when Y is determined by X.Source: Conditional entropy · pp. 271–272

Mutual information

Mutual information I(X; Y) is the expected information Y gives about X, equals H(X) − H(X | Y) = H(Y) − H(Y | X), is symmetric, and vanishes for independent variables.It is the information that X gives about Y as well.Source: Mutual information · pp. 273–274
A constructed example realizes the entropies in Figure 8.1 as actual cardinalities.Source: Mutual information · p. 276

Extremal cases

Threefold mutual information can be negative when three binary random variables are uniformly distributed over four of the eight possible triples.Source: Extremal cases · p. 276
H(X,Y) lies between max{H(X),H(Y)} and H(X)+H(Y), with equality at the lower bound iff one variable determines the other and at the upper bound iff they are independent.Source: Extremal cases · pp. 277–278
Conditional entropy satisfies 0 ≤ H(X | Y) ≤ H(X), equaling zero iff X is determined by Y and equaling H(X) iff X and Y are independent.Source: Extremal cases · p. 277
Mutual information satisfies 0 ≤ I(X;Y) ≤ min{H(X),H(Y)}; it is zero iff X and Y are independent and reaches its upper bound iff one determines the other.Source: Extremal cases · pp. 277–278
Among distributions on X × Y with fixed marginals, the independent coupling has greatest Shannon entropy, a property not shared by other Rényi or q-logarithmic entropies except trivially at q = 0.Source: Extremal cases · p. 278

8.2 Diversity measures for subcommunities · 8.3 Diversity measures for metacommunities · 8.4 Properties of the metacommunity measures

These sections define and interpret order-1 diversity measures for subcommunities and metacommunities, including alpha-, redundancy, and beta-diversities, and examine their structural properties. They establish independence of alpha from beta and redundancy, identify which measures are invariant to splitting ecologically uniform subcommunities, and extend chain-rule, modularity, and replication properties to metacommunities.

8.2 Diversity measures for subcommunities

For each subcommunity, α_j measures its diversity, β_j its relative diversity against the metacommunity, and γ_j its cross-diversity; equation (8.8) partitions γ_j into components.β_j reflects how atypical the subcommunity is, while γ_j measures species rarity by metacommunity standards.Source: §8.2 · pp. 281–282

8.3 Diversity measures for metacommunities

Metacommunity measures compare the species X and subcommunity Y of a randomly chosen individual, using exponentials of their joint entropy, conditional entropies, and mutual information.Source: §8.3 · p. 283

Metacommunity gamma-diversity

G(P) is whole-metacommunity species diversity, equivalently the size-weighted geometric mean of subcommunity γ_j; it ranges from 1 for one species to S for balanced abundances.Source: Metacommunity gamma-diversity · pp. 283, 285

Metacommunity alpha-diversities

Raw A(P)=D(P) counts effective (species, subcommunity) pairs, overestimating G(P) by at most a factor D(w); it matches G(P) with no species shared and overestimation peaks for identical compositions.Source: Metacommunity alpha-diversities · p. 286
Normalized metacommunity alpha-diversity is the size-weighted geometric mean of local α_j diversities; it equals 1 when each subcommunity has at most one species and G(P) when well-mixed.Source: Metacommunity alpha-diversities · p. 287
Mean subcommunity diversity A(P) is unaffected by species overlap: it equals D(p) for a single subcommunity but is 1 when each subcommunity is exactly one species.The effective number of species–subcommunity pairs is D(p) in both cases.Source: Metacommunity alpha-diversities · p. 288

The redundancy of a metacommunity

Redundancy R(P) is the effective number of subcommunities across which a typical species is spread, reflecting how much diversity persists if subcommunities are lost.It ranges from 1 when no species are shared to D(w) when subcommunity species distributions are identical; it equals N when those subcommunities also have equal sizes.Source: The redundancy of a metacommunity · pp. 289–290

Metacommunity beta-diversity

Beta-diversity B(P) is the exponential of mutual information between species and subcommunity, with B(P) = 1 for well-mixed communities and B(P) = N for equal-sized isolated subcommunities.It partitions gamma-diversity into average within-subcommunity alpha-diversity and between-subcommunity beta-diversity.Source: Metacommunity beta-diversity · pp. 290–292

8.4 Properties of the metacommunity measures

The chapter requires metacommunity measures to have sensible logical and algebraic properties, and states that the measures introduced satisfy this requirement.The following discussion demonstrates particular properties of the alpha-, beta-, and redundancy measures.Source: §8.4 · p. 293

Independence

For each fixed number N of subcommunities, alpha-diversity A is independent of beta-diversity B and of redundancy R, so knowing either measure does not restrict the possible values of A.The alpha–beta result also means that, for any a ≥ 1 and b ∈ [1, N], some metacommunity has A(P) = a and B(P) = b.Source: Independence · pp. 295–296

Identical subcommunities

Splitting a subcommunity into smaller subcommunities of identical composition should leave both effective isolated-subcommunity beta-diversity B and mean subcommunity diversity A unchanged.This invariance would make the decomposition insensitive to arbitrary boundaries between identical subcommunities.Source: Identical subcommunities · p. 297
Splitting a subcommunity into parts with identical species distributions leaves A, B and G unchanged, making these measures independent of ecologically irrelevant boundaries.This invariance applies when a subcommunity is ecologically uniform; finer divisions otherwise can reveal more variation between subcommunities.Source: Identical subcommunities · p. 298
The effective number A of (species, subcommunity) pairs and redundancy R instead increase by a factor of D(w) when a homogeneous subcommunity is split.These measures depend directly on the subcommunity division: A counts species-subcommunity pairs, while R reflects how many subcommunities a typical species spans.Source: Identical subcommunities · pp. 299–300

Chain rule, modularity and replication principle

Metacommunity measures satisfy a chain rule, and each whole-system measure is determined by the corresponding metacommunity measures and their relative sizes.The chain-rule result applies to a multicommunity whose component metacommunities have no species in common.Source: Chain rule, modularity and replication… · pp. 300, 302, 303
The chain-rule formulas distinguish value measures, which aggregate component values, from means, which average them across metacommunities.G, D(w), A for species-subcommunity pairs, and B are value measures; mean subcommunity diversity A and redundancy R are means.Source: Chain rule, modularity and replication… · pp. 303–304
When an identical island is copied with a disjoint species set, G, D(w), A for species-subcommunity pairs, and B double, while mean subcommunity diversity A and R remain unchanged.The mean diversity is unchanged because corresponding subcommunities have the same abundance distributions; redundancy is unchanged because the islands share no species.Source: Chain rule, modularity and replication… · p. 304

8.5 All entropy is relative … 9.2 Large deviations and convex duality

The section relates information measures to relative entropy, extends these ideas to arbitrary q and to probabilistic characterizations of norms and means, and develops variational and convex-duality connections for moment generating functions. It also connects convex duality with large-deviation rates and Cramér’s theorem.

8.5 All entropy is relative

Cross, joint, conditional, and mutual information can all be reduced to relative entropy, illuminating subcommunity and metacommunity diversity measures.Mutual information is itself relative entropy, is nonnegative, and equals zero exactly when the variables are independent.Source: §8.5 · pp. 305–306
Metacommunity beta-diversity measures divergence from a well-mixed state: B(P) is 1 when well-mixed and reaches D(w) when subcommunities share no species.B(P) compares the species–subcommunity distribution with a hypothetical distribution having the same overall species and subcommunity proportions but identical subcommunity composition.Source: §8.5 · p. 307
A subcommunity’s gamma-diversity contribution is high when it is rich in species that are rare across the metacommunity.The value Pij/pi is high when species i is common in subcommunity j but rare in the metacommunity overall.Source: §8.5 · p. 308

8.6 Beyond

For arbitrary q, the section sketches replacing Shannon entropy and its exponential with Rényi entropy and Hill numbers, with q controlling emphasis on rare or common species and smaller or larger subcommunities.The q-analogues of several measures follow by replacing σ with σq; the corresponding B generalization is more subtle and referred to Reeve et al. [290].Source: §8.6 · p. 309
Previously established bounds, extremal cases, and metacommunity properties extend to arbitrary q, while A(P)/G(P) − 1 generalizes the Jaccard index.The Jaccard connection extends to any number of subcommunities and any q emphasis on rare or common species; the cited extension results are shown in Reeve et al. [290].Source: §8.6 · pp. 310–311
For arbitrary q, alpha- and beta-diversity do not generally determine gamma-diversity, and incorporating a species-similarity matrix Z requires more care.The equations connecting these measures are special to q = 1; the section does not develop the Z generalization and refers readers to Reeve et al. [290].Source: §8.6 · pp. 311–312

Probabilistic methods

Probability theory supplies a variational formula for the moment generating function, conceptually the convex conjugate of Cramér’s large deviation theorem, for characterizing ℓp norms and power means.The section says the probabilistic method is applied to characterize ℓp norms and power means in Sections 9.3 and 9.4.Source: Probabilistic methods · p. 314

9.1 Moment generating functions

For iid real random variables, Theorem 9.1.1 gives a variational formula for the moment generating function for λ ≥ 0, allowing infinite values.The section presents the formula as a result also found in Cerf and Petit, but gives a different proof.Source: §9.1 · p. 314

9.2 Large deviations and convex duality

The section places the moment-generating-function formula in context as the convex conjugate of Cramér’s large-deviation theorem.It assumes only basic probability, and its technically sophisticated large-deviations discussion is not logically necessary for what follows.Source: Probabilistic methods · p. 314, §9.2 · p. 317

Cramér’s theorem

Cramér’s theorem establishes an asymptotic rate for P(X_r ≥ x), with exponential decay when x > E(X), addressing rates not given by the law of large numbers or central limit theorem.The discussion concerns deviations larger than those addressed by the central limit theorem.Source: Cram´er’s theorem · p. 318

Convex duality

For finite-valued convex f:R→R, Legendre–Fenchel duality gives f** = f; the text notes extension to finite-dimensional real vector spaces.The vector-space formulation uses a conjugate function on the dual space.Source: Convex duality · p. 322
Willerton’s general categorical construction makes the classical Legendre–Fenchel theorem a special case for functions on finite-dimensional vector spaces valued in [−∞, ∞].The construction uses a dual equivalence between enriched categories induced by the pairing of a vector space and its dual.Source: Convex duality · p. 323

The dual of Cram´er’s theorem

Taking convex conjugates of Cramér’s theorem and applying Legendre–Fenchel yields an expression for log mX and hence mX, with the supremum analysis using x = E(X) for λ ≥ 0.The cumulant generating function log mX is convex; the analysis identifies the relevant supremum at the expectation.Source: The dual of Cram´er’s theorem · p. 324
Theorem 9.1.1 is also a dual form of Cramér’s theorem: (9.18)’s limit as r→∞ is equivalent to its supremum over r≥1, as Cerf and Petit showed.The only difference between the expressions is a limit versus a supremum over r.Source: The dual of Cram´er’s theorem · pp. 324–325

9.3 Multiplicative characterization of the p-norms · 9.4 Multiplicative characterization of the power means · 10.1 Measure-preserving maps

These sections characterize multiplicative systems of norms as p-norms and power-mean sequences through specified structural properties. They then define information loss for finite measure-preserving maps as entropy discarded by a deterministic map and relate it to conditional entropy.

9.3 Multiplicative characterization of the p-norms

Every multiplicative system of norms on the spaces R^n equals the p-norm system for some p.The proof uses a variational formula derived from a moment-generating-function formula and a tensor-power argument.Source: §9.3 · pp. 325, 327, 330
The proof of the p-norm characterization works directly from the moment-generating-function formula, without requiring Cramér’s theorem, the Legendre–Fenchel theorem, or convex conjugates.The authors note that Cerf and Petit showed the formula could be proved without those tools.Source: §9.3 · p. 332

9.4 Multiplicative characterization of the power means

Power means M_t are convex for t in [1, ∞] and are not convex for t in [−∞, 1).Source: §9.4 · p. 333
A sequence of power means is natural, consistent, increasing, multiplicative, and convex if and only if it equals M_t for some t in [1, ∞].This characterization uses multiplicativity and convexity rather than assuming modularity.Source: §9.4 · pp. 332–333
The multiplicative axioms characterize the power means: every qualifying sequence M equals M_t for some t ∈ [1, ∞].The proof constructs a multiplicative system of norms and identifies it with a norm of order s; consistency then gives s = t.Source: §9.4 · pp. 334, 335, 337, 338

Information loss

For measure-preserving maps, reasonable requirements force information loss, up to a constant factor, to equal domain entropy minus codomain entropy.Treating maps rather than probability spaces as the focus gives characterization conditions resembling linearity or homomorphism conditions; this result first appeared in Baez, Fritz and Leinster.Source: Information loss · pp. 339–340

10.1 Measure-preserving maps

A finite map is measure-preserving exactly when its pushforward distribution equals the target distribution, and it factors as a surjection followed by a probability-one subset inclusion.The surjection is a coarse-graining that discards information, while the inclusion adds only events of probability zero.Source: §10.1 · p. 341
Forgetting a fair coin toss loses one bit, while a uniform four-to-one card-to-rank map loses 2 bits by withholding the suit.Source: §10.1 · pp. 343–344
The chain rule equates expected information lost conditional on each letter with the decrease in entropy when accents are forgotten.The two quantities are equal for the example of forgetting accents on French letters.Source: §10.1 · p. 344
Information loss equals the source entropy minus the target entropy, is conditional entropy, and is always nonnegative.For a map f from (Y, s) to (X, p), L(f) = H(s) − H(p) = H(V | U), where U = f(V).Source: §10.1 · pp. 345–346
Conditional entropy is itself information loss: for the projection pr1 from the joint distribution of (U, V) to U, H(V | U) = L(pr1).Source: §10.1 · p. 346

10.2 Characterization of information loss · 11.1 Fermat quotients and the definition of entropy · 11.2 Characterizations of entropy and information loss

These sections characterize information loss and define entropy modulo a prime using Fermat quotients, then develop its chain rule and related characterizations. They show that chain-rule information measures and information loss are scalar multiples of this entropy, with information loss also characterized by linear conditions.

10.2 Characterization of information loss

Information loss is uniquely characterized up to a constant factor by vanishing on isomorphisms, additivity under composition, linearity under parallel mixtures, and continuity.The theorem concludes that any function satisfying these axioms is K = cL for some constant c.Source: §10.2 · pp. 348, 349, 351
The composition condition alone forces K(f) to depend only on the domain and codomain of f, so maps with the same domain and codomain have equal K-values.The text identifies this as an instance of a general categorical fact for functors to a groupoid from a category with a terminal object.Source: §10.2 · p. 351
The q-logarithmic information-loss characterization changes the parallel-mixture rule and requires no continuity or other regularity condition.It is presented as a variant of the main theorem, with its corresponding characterization theorem attributed to Baez, Fritz and Leinster.Source: §10.2 · p. 352

Entropy modulo a prime

The chapter defines entropy and information loss modulo p by characterization theorems analogous to those for real entropy, and relates real entropy to residues modulo p.It also develops an equivalent polynomial-ring approach, more closely connected to polylogarithms.Source: Entropy modulo a prime · pp. 353–354

11.1 Fermat quotients and the definition of entropy

Distributions modulo p are tuples in (Z/pZ)^n whose entries sum to 1; the lack of positive and negative elements poses a problem for transferring the real entropy definition.The text notes that real entropy can be extended to arbitrary real entries summing to 1, without requiring positivity.Source: §11.1 · pp. 354–355
There is no nontrivial homomorphism from (Z/pZ)× to the additive group Z/pZ, so Fermat quotients provide an acceptable substitute for a logarithm modulo p.For n not divisible by p, Fermat’s little theorem implies that p divides n^(p−1)−1, motivating the quotient’s definition.Source: §11.1 · p. 355
Every homomorphism from (Z/p2Z)× to Z/pZ is a scalar multiple of the Fermat quotient, which is characterized by multiplicativity and p2-periodicity.The characterization theorem also identifies functions satisfying these properties as scalar multiples of the Fermat quotient.Source: §11.1 · pp. 356–357
Entropy modulo p is defined using integer representatives of a distribution’s probabilities, and the definition is independent of the representatives chosen.The section presents a characterization theorem for this entropy, analogous to Faddeev’s theorem for real entropy.Source: §11.1 · p. 357
The map ∂: Z → Z/pZ is well-defined modulo p2 and satisfies a Leibniz rule; entropy measures its failure to preserve sums.For a uniform distribution on n elements with p ∤ n, entropy equals the Fermat quotient qp(n).Source: §11.1 · pp. 358–359
Entropy modulo p satisfies the chain rule, and entropy of a tensor product is the sum of the component entropies.The chain rule applies to composed distributions.Source: §11.1 · p. 361

11.2 Characterizations of entropy and information loss

A sequence of information measures satisfies the chain rule if and only if it is a scalar multiple of entropy modulo p; no symmetry condition is needed.This characterizes the entropy defined in the preceding subsection.Source: §11.2 · p. 362
For any information measure satisfying the chain rule, its value on the uniform distribution u_n is p2-periodic in n.This periodicity is established as an intermediate step in the characterization proof.Source: §11.2 · p. 363
I(π)=cH(π) for every distribution π, with c∈Z/pZ.Source: §11.2 · p. 364
The entropy-mod-p characterization also characterizes information loss using only linear conditions, by an argument copied nearly verbatim from the real case.Source: §11.2 · p. 364

11.3 The residues of real entropy · 11.4 Polynomial approach · 12.1 Operads and their algebras

This section develops algebraic and categorical approaches to entropy, from residues and polynomial identities modulo a prime to Shannon entropy as an internal algebra. It also introduces operads and their algebras, connecting them to weighted means, monads, and finitary algebraic theories while noting their limits.

11.3 The residues of real entropy

Mod-p information loss is characterized by vanishing on isomorphisms, additivity under composition, and linearity under mixtures.The theorem states these three properties are equivalent to the specified characterization of K.Source: §11.3 · p. 365
Equal real entropies for rational distributions with denominators prime to p imply equal mod-p entropies, so the residue map is well-defined and additive.The set of real entropy values is closed under addition.Source: §11.3 · pp. 368–369

11.4 Polynomial approach

A polynomial h over Z/pZ represents entropy modulo p on probability distributions, providing a direct definition from their coordinates.The section identifies h as the entropy function on distributions whose coordinates sum to 1.Source: §11.4 · pp. 369–370
The polynomial h is homogeneous of degree p, inducing a polynomial function homogeneous of degree 1, and is presented as the natural polynomial representing mod-p entropy.This homogeneity supplies a further property beyond agreement with entropy on probability distributions.Source: §11.4 · p. 371
The section establishes polynomial identities for h that are stronger than the previously proved functional equations, including one related to the chain rule.The identity equates h applied to all listed variables with h of their grouped sums plus the sum of the corresponding h terms.Source: §11.4 · p. 371
The mod-p entropy polynomial satisfies the chain rule and, in two variables, the cocycle identity.The two-variable identity is h(x, y) − h(x, y + z) + h(x + y, z) − h(y, z) = 0.Source: §11.4 · pp. 372–373
For mod-p entropy, the two-element function satisfies the fundamental equation and uniquely determines the entropy sequence up to a constant factor, without a symmetry axiom in this approach.Assuming symmetry, the fundamental equation is the sole obstacle to extending a function; the chapter’s approach drops symmetry requirements and includes p = 2.Source: §11.4 · pp. 376–377

The categorical origins of entropy

For the simplex operad and the categorical algebra R, the internal algebras are precisely scalar multiples of Shannon entropy.The chapter also describes a free categorical simplex-algebra containing an internal algebra, yielding a concrete characterization of information loss almost exactly like Chapter 10’s.Source: The categorical origins of entropy · pp. 378–379, Figure 12.1 · p. 378

12.1 Operads and their algebras

An operad is a sequence of sets of abstract operations with specified input counts, composition, and identities; associativity and identity laws make composites of operation trees unambiguous.The chapter uses only the definitions and results needed for its goal.Source: §12.1 · pp. 380–381
Operads can be defined in categories with finite products, including Top, and under mild conditions their algebras coincide with those of an induced monad.For Top, the operad’s composition maps are continuous; the stated mild condition is countable coproducts over which products distribute.Source: §12.1 · pp. 383, 386
An operad packages abstract operations with composition and identities, while a P-algebra interprets them as operations on a set satisfying compatibility axioms.The compatibility axioms require the interpretation to respect composition and identity.Source: §12.1 · p. 384
For a weighted mean on an interval, the chain rule supplies the composition axiom for a simplex-operad algebra, while consistency supplies its identity axiom.The identity condition is M(u1, (x)) = x.Source: §12.1 · p. 385
A P-algebra is equivalently a map of operads from P to End(A), and maps preserving the operations form a category of P-algebras.End(A) consists of operations A^n → A.Source: §12.1 · p. 385
Finitary algebraic theories add reindexing maps to operads; bijective reindexings yield symmetric operads, identity-only reindexings yield nonsymmetric operads, and groups cannot be described by an operad.In the group theory example, reindexing turns multiplication into squaring; more generally, most algebraic theories cannot be described by an operad.Source: §12.1 · p. 386

12.2 Categorical algebras and internal algebras · 12.3 Entropy as an internal algebra · 12.4 The universal internal algebra

This section develops categorical and internal algebras and relates internal algebras to strict maps from free categorical algebras. It uses these frameworks to characterize Shannon and q-logarithmic entropy, information loss, and a universal categorical algebra, including its interpretation for the simplex operad.

12.2 Categorical algebras and internal algebras

A categorical P-algebra is an internal category in Alg(P), with its category operations required to be maps of P-algebras.It can also be viewed as an internal category in E on which P acts functorially.Source: §12.2 · p. 387
An internal algebra in a categorical P-algebra A is a lax map of categorical P-algebras from the terminal algebra 1 to A.This definition generalizes internal monoids in monoidal categories.Source: §12.2 · p. 391
In a monoidal category, an internal algebra for the terminal operad is exactly a monoid in that category.A categorical algebra for the terminal operad is a monoidal category.Source: §12.2 · p. 393

12.3 Entropy as an internal algebra

In R with its standard categorical Δ-algebra structure, internal Δ-algebras are exactly γ = cH for c ∈ R.The defining functions are continuous and satisfy operadic coherence; normalization follows from coherence, and Faddeev’s theorem gives the classification.Source: §12.3 · p. 394
In the q-deformed categorical structure on the reals, the internal algebras are precisely the scalar multiples of q-logarithmic entropy.The set-operad result does not require continuity; for q > 0 it also holds for the topological operad of simplices.Source: §12.3 · p. 395
For q > 0, the same characterization holds for the topological operad, whereas the q-deformed structure maps are discontinuous when q ≤ 0.The result for q > 0 follows from the set-operad theorem, whose proof does not require continuity.Source: §12.3 · p. 395

12.4 The universal internal algebra

Up to equivalence, F∆ has finite totally ordered probability spaces as objects and order-preserving measure-preserving maps with extra distributions over zero-probability fibers.The forgetful functor to finite ordered probability spaces is not an equivalence because of this zero-probability complication.Source: §12.4 · p. 400
For any operad P and internal algebra (a, γ) in a categorical P-algebra A, there is a unique strict map from FP to A carrying its canonical internal algebra to (a, γ).This universal property determines FP together with its internal algebra uniquely up to isomorphism.Source: §12.4 · pp. 398, 401
For any operad P and categorical P-algebra A, internal algebras in A correspond canonically to strict maps FP→A.An internal algebra can equivalently be described as a lax map 1→A or a strict map FP→A.Source: §12.4 · p. 402
Strict maps F∆→R are parametrized by c∈R and correspond to scalar multiples of Shannon entropy.Theorem 12.4.9 characterizes such maps by a constant c; the text links them to real scalar multiples of Shannon entropy.Source: §12.4 · pp. 402–403
A similar classification holds for q-logarithmic entropies using the q-deformed categorical ∆-algebra structure on R.Source: §12.4 · p. 404
The symmetric free algebra removes the order-preserving restriction; its only substantial remaining difference from FinProb is extra probability-distribution data on fibres over zero-probability elements.The comparison also notes that skeletality is categorically unimportant.Source: §12.4 · p. 404
Strict maps Fsym∆→R are precisely scalar multiples of information loss, differing from Theorem 10.2.1 only in their treatment of zero probabilities.Discarding the extra data associated with zero-probability elements adapts the result to Theorem 10.2.1; this categorical argument historically led to that theorem.Source: §12.4 · p. 404

Appendix A Proofs of background facts … Index

This section gathers deferred proofs and mathematical facts used throughout the book, including results on entropy, diversity, convexity, couplings, affine functions, and polynomial representations over finite fields. It also provides reference tables and an index to topics spanning diversity, entropy, geometry, and related mathematical frameworks.

A.1 Forms of the chain rule for entropy

The general chain rule for symmetric entropy functions is equivalent to the special cases used by previous authors.The appendix proves the equivalence among four forms; it also notes a similar assertion for q-logarithmic entropies.Source: §A.1 · p. 405

A.2 The expected number of species in a random sample

The Hurlbert, Smith and Grassle diversity index for a random sample with replacement is expressed in terms of the Hill numbers Dq(p).The sample statistic is the expected number of different species represented in a sample of m individuals.Source: §A.2 · p. 408

A.3 The diversity profile determines the distribution

Two finite distributions have the same diversity profile exactly when one is a permutation of the other.The lemma states the equivalent condition that their Hill numbers agree on a set of orders unbounded above.Source: §A.3 · pp. 409–410

A.4 Affine functions

A function on real intervals is affine if and only if it preserves midpoints, equivalently if it satisfies the stated general affine-combination condition.The lemma gives these conditions as equivalent to α(x) = ax + b.Source: §A.4 · p. 411

A.5 Diversity of integer orders

The appendix states a computation of diversity DZ_q(p) for integer orders q ≥ 2, using notation defined earlier in the book.The supplied passage does not include the computation itself.Source: §A.5 · p. 412

A.6 The maximum entropy of a coupling

Except at q = 0 and q = 1, the product of the marginals does not maximize Rényi entropy among couplings: some coupling has Hq(P) > Hq(p) + Hq(r).The proof constructs such a coupling for each q ∈ (0, 1) ∪ (1, ∞], while the stated inequality holds for q = 0.Source: §A.6 · pp. 413–414
For q > 1, including q = ∞, there are couplings whose entropy exceeds the sum of the marginal entropies for sufficiently large N.The proof constructs such examples using a matrix P and matching marginals p and r.Source: §A.6 · p. 415

A.7 Convex duality

The Legendre–Fenchel theorem states that every convex function f on R equals its biconjugate, f**.The proof establishes the needed inequality by convexity.Source: §A.7 · p. 416

A.8 Cumulant generating functions are convex

Hölder’s inequality proves that the cumulant generating function of a real random variable is convex without assuming twice differentiability.The cited Hölder statement allows one or more integrals to be infinite.Source: §A.8 · p. 417

A.9 Functions on a finite field

Every function from K^n to a finite field K with q elements has a unique representing polynomial of degree less than q in each variable.Evaluating that polynomial at each input recovers the function.Source: §A.9 · p. 418

Weighted means

The weighted-means table summarizes conditions including absence, chain rule, consistency, convexity, modularity, multiplicativity, and quasiarithmetic form.It gives abbreviated definitions and references to their full definitions in the text.Source: Appendix B · p. 419, Weighted means · p. 419

Unweighted means

The unweighted-means table expresses conditions using repeated inputs and records dependencies on multiplicities and associated means.The supplied table text is fragmentary.Source: Unweighted means · p. 420

Diversity measures

For diversity measures on finite probability distributions without species similarity, the table lists absence-invariance, modularity, and modular-monotonicity conditions.The entries also point to the full definitions in the text.Source: Diversity measures · pp. 420–421

Value measures; Standard notation; Index

The appendices also list the effective-number condition for value measures, standard notation, and indexed topics for reference.The effective-number condition is σ(u_n, (1, ..., 1)) = n.Source: Value measures · p. 421, Standard notation · p. 441, Index · pp. 443, 445

Index

Maximum diversity entries point to computation, species elimination, geometry and analysis, comparison with magnitude, and a theorem.The index also lists maximum diversity for matrices and nonsymmetric similarity.Source: Index · p. 448
Metacommunity diversity entries include order q, relative diversity, gamma-diversity, modularity, and replication principles.The index gives page references 293–299 for these topics.Source: Index · p. 448
Modularity entries cover diversity measures, Hill numbers, metacommunity diversities, similarity-sensitive diversity, and weighted and unweighted means.The index separately identifies a modularity principle for metacommunity diversities.Source: Index · p. 448
Replication-principle entries cover diversity of order 1, Hill numbers, and metacommunity diversities.The index gives references to pages 56, 127, and 294.Source: Index · p. 451