Excess Growth Rate in Theory & Practice

• Excess growth rate is defined as the benchmark-relative logarithmic growth differential observed in diverse fields like finance, ecology, and cell-population dynamics.
• In portfolio theory, it is derived via Jensen's inequality and linked to the Kullback–Leibler divergence, highlighting a nonnegative premium when returns vary.
• In ecological and biological contexts, its sign and magnitude capture transient deviations from intrinsic growth laws, often reflecting the impact of variability and feedback mechanisms.

“Excess growth rate” denotes several benchmark-relative growth functionals rather than a single universally standardized quantity. In portfolio theory, it is a logarithmic functional measuring the gap between the log return of a rebalanced portfolio and the weighted average of constituent log returns; in equity return modeling, it is the continuously compounded log return in excess of the risk-free rate; in ecology and cell-population dynamics, it denotes the difference between realized growth and an intrinsic, asymptotic, or mean single-cell benchmark; and in growth diagnostics it can denote acceleration beyond a constant-rate exponential law (Campbell et al., 29 Oct 2025, Alswaidan et al., 10 Mar 2026, Deveau et al., 2013, Hein et al., 2022, Hüsler et al., 2011). This suggests a family resemblance centered on baseline-subtracted logarithmic growth, with the precise baseline determined by the surrounding theory.

1. Conceptual Scope

The term appears in multiple technical literatures with distinct formal definitions and distinct sign properties. Some versions are nonnegative by construction, such as the portfolio-theoretic Jensen gap; others are signed, such as excess return above the risk-free rate or transient demographic deviation from intrinsic growth; still others are strictly positive under stochastic amplification mechanisms, such as the population-level growth premium generated by variability and memory in single-cell growth (Campbell et al., 29 Oct 2025, Alswaidan et al., 10 Mar 2026, Deveau et al., 2013, Hein et al., 2022, Rochman et al., 2017, Hüsler et al., 2011).

Domain	Formal quantity	Baseline
Portfolio theory	$$\Gamma(\pi,R)=\log(\sum_i \pi_i R_i)-\sum_i \pi_i\log R_i$$	Weighted average of constituent log returns
Equity time series	$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$	Continuously compounded risk-free rate
Ecology	$\epsilon(t)=\frac{d}{dt}\ln N(t)-r$	Intrinsic Lotka–Euler growth rate
Stochastic cell populations	$$\Delta=\lambda_p-\langle\lambda\rangle=\sigma_\lambda^2\tau_c$$	Mean single-cell growth rate
Cell-cycle variability	$\Delta\Lambda=\Lambda-\ln 2/\langle\tau\rangle$	Naive doubling-time benchmark
FTS growth diagnostics	any $r(t)$ above fixed-$r$ exponential growth	Constant-rate exponential law

A common source of confusion is the assumption that “excess growth rate” always means a positive premium. In the supplied literature, positivity is domain-specific. The portfolio functional is nonnegative by Jensen’s inequality, the cell-population excess is positive in the models considered, but the ecological quantity $\epsilon(t)$ may be positive or negative, and the equity quantity $G_{i,j}$ can vary with daily returns and the risk-free benchmark.

2. Portfolio-Theoretic Excess Growth Rate

In portfolio theory, the excess growth rate is a fundamental logarithmic functional defined for $\pi\in\Delta_n$ and deterministic gross returns $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$0 by

$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$1

Equivalently, for $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$2,

$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$3

In the probabilistic setting, for a nonnegative random gross-return vector $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$4 with law $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$5 and $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$6 almost surely,

$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$7

The deterministic functional is nonnegative by Jensen’s inequality and vanishes if and only if $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$8 is constant on $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$9 (Campbell et al., 29 Oct 2025).

The paper establishes several equivalent interpretations. Defining the perturbation $\epsilon(t)=\frac{d}{dt}\ln N(t)-r$0 as the normalization of $\epsilon(t)=\frac{d}{dt}\ln N(t)-r$1, one has

$\epsilon(t)=\frac{d}{dt}\ln N(t)-r$2

where $\epsilon(t)=\frac{d}{dt}\ln N(t)-r$3 is the Kullback–Leibler divergence. The same quantity is also exactly the Jensen gap for $\epsilon(t)=\frac{d}{dt}\ln N(t)-r$4: $\epsilon(t)=\frac{d}{dt}\ln N(t)-r$5 Within information geometry, for the special exponentially concave generator $\epsilon(t)=\frac{d}{dt}\ln N(t)-r$6, the induced logarithmic divergence coincides with the excess growth rate.

Three axiomatic characterization theorems identify the excess growth rate, up to multiplicative constant, via relative entropy, the Jensen-gap structure, and perturbation-invariant logarithmic divergence. The axioms include joint measurability, permutation invariance, support invariance, vanishing on constant returns, and a general chain rule in the relative-entropy characterization; homogeneity and affine behavior on level sets in the Jensen-gap characterization; and perturbation invariance in the logarithmic-divergence characterization. These results make the functional structurally rigid rather than merely convenient.

Optimization properties further distinguish the functional from the growth-optimal criterion. For deterministic maximization of $\epsilon(t)=\frac{d}{dt}\ln N(t)-r$7, the KKT conditions imply that any optimizer has support on exactly two assets corresponding to the extreme values of $\epsilon(t)=\frac{d}{dt}\ln N(t)-r$8. For expected excess-growth maximization, $\epsilon(t)=\frac{d}{dt}\ln N(t)-r$9 is concave, and the optimizer satisfies a KKT system involving $$\Delta=\lambda_p-\langle\lambda\rangle=\sigma_\lambda^2\tau_c$$0 and the mean log-return vector $$\Delta=\lambda_p-\langle\lambda\rangle=\sigma_\lambda^2\tau_c$$1. If $$\Delta=\lambda_p-\langle\lambda\rangle=\sigma_\lambda^2\tau_c$$2 is a constant vector, the maximizer of expected excess growth coincides with the growth-optimal, or Kelly, portfolio.

The same work places the functional at the intersection of several other theories. It is linked to Helmholtz free energy through

$$\Delta=\lambda_p-\langle\lambda\rangle=\sigma_\lambda^2\tau_c$$3

admits a Gibbs-principle variational representation,

$$\Delta=\lambda_p-\langle\lambda\rangle=\sigma_\lambda^2\tau_c$$4

appears in Campbell’s average code length as the gap between generalized and Shannon mean code lengths, and serves as the large-deviation rate function for scaled-Dirichlet laws. In this literature, excess growth rate is therefore not merely a financial heuristic but a canonical logarithmic divergence.

3. Excess Growth Rate in Equity Return Modeling

In the equity time-series setting, the daily excess growth rate of asset $$\Delta=\lambda_p-\langle\lambda\rangle=\sigma_\lambda^2\tau_c$$5 at day $\Delta=\lambda_p-\langle\lambda\rangle=\sigma_\lambda^2\tau_c$6 is defined by

$$\Delta=\lambda_p-\langle\lambda\rangle=\sigma_\lambda^2\tau_c$$7

It is a continuously compounded, time-additive log return minus the risk-free benchmark, has units of year$$\Delta=\lambda_p-\langle\lambda\rangle=\sigma_\lambda^2\tau_c$$8, and isolates the premium over the safe rate (Alswaidan et al., 10 Mar 2026).

The paper “Hybrid Hidden Markov Model for Modeling Equity Excess Growth Rate Dynamics: A Discrete-State Approach with Jump-Diffusion” (Alswaidan et al., 10 Mar 2026) models the process by first discretizing the empirical series into Laplace quantile-defined states. A Laplace distribution $$\Delta=\lambda_p-\langle\lambda\rangle=\sigma_\lambda^2\tau_c$$9 is fit by maximum likelihood, and quantile cutoffs

$\Delta\Lambda=\Lambda-\ln 2/\langle\tau\rangle$0

with $\Delta\Lambda=\Lambda-\ln 2/\langle\tau\rangle$1 and $\Delta\Lambda=\Lambda-\ln 2/\langle\tau\rangle$2, define the hidden states. Each observation is assigned to state $\Delta\Lambda=\Lambda-\ln 2/\langle\tau\rangle$3 when $\Delta\Lambda=\Lambda-\ln 2/\langle\tau\rangle$4. This partition is intended to preserve heavy-tail structure without enforcing equal bin counts.

On this discretization, the model

$\Delta\Lambda=\Lambda-\ln 2/\langle\tau\rangle$5

combines regime switching with a Poisson-driven jump-duration mechanism. The hidden regimes are $\Delta\Lambda=\Lambda-\ln 2/\langle\tau\rangle$6, the observation space is $\Delta\Lambda=\Lambda-\ln 2/\langle\tau\rangle$7, and the regime transition matrix is $\Delta\Lambda=\Lambda-\ln 2/\langle\tau\rangle$8. Conditional on $\Delta\Lambda=\Lambda-\ln 2/\langle\tau\rangle$9, emissions follow

$r(t)$0

with $r(t)$1 taken as the empirical mean and standard deviation of training observations in state $r(t)$2. At each step, with probability $r(t)$3, the chain enters a tail episode of length $r(t)$4, during which the state is forced into a negative-tail or positive-tail set, with a small bias $r(t)$5 toward negative tails. Otherwise the chain evolves according to the usual Markov rule

$r(t)$6

Parameter estimation is performed by direct transition counting rather than Baum–Welch EM. After discretization into states $r(t)$7, counts

$r(t)$8

are normalized row-wise to obtain $r(t)$9. The stated advantages are triviality, determinism, and freedom from local maxima or initialization issues.

Synthetic-data quality is evaluated by the two-sample Kolmogorov–Smirnov statistic, the two-sample Anderson–Darling statistic, Wasserstein-1 distance, Hellinger distance on binned histograms, and

$r$0

With 100 quantile states and Student-$r$1 emissions, the framework yields in-sample excess-growth kurtosis close to the observed $r$2, KS and AD pass rates of $r$3 and $r$4, and out-of-sample rates above $r$5. The pure HMM without jumps sits at the i.i.d. floor for volatility clustering, with $r$6, while the hybrid HMM with jumps reduces this to $r$7 in-sample and $r$8 out-of-sample. No single benchmark dominates all quality dimensions: GARCH$r$9 attains the best $\epsilon(t)$0 at approximately $\epsilon(t)$1 but has a $\epsilon(t)$2 KS pass rate, whereas the standard HMM without jumps attains $\epsilon(t)$3 KS but cannot generate persistent high-volatility regimes.

The framework is presented as a tool for stress testing, risk model validation, and scenario design. Its discrete states provide an interpretable “market mood” index, and a Single-Index Model extension

$\epsilon(t)$4

propagates a simulated SPY factor path to a 424-asset universe while preserving cross-sectional correlation structure.

4. Intrinsic and Transient Excess in Ecological Demography

Ecological theory uses several formally different notions of growth rate: the Malthusian law $\epsilon(t)$5, the discrete-time multiplier $\epsilon(t)$6 with $\epsilon(t)$7, the finite-window census rate

$\epsilon(t)$8

the birth–death relation $\epsilon(t)$9, the Lotka–Euler equation

$G_{i,j}$0

and the reproductive-value weighting

$G_{i,j}$1

Deveau, Karsten, and Teismann emphasize that field estimates of “the” population growth rate can oscillate, depend on the fitting window, and vary across models, especially in stage-structured populations with strong transients (Deveau et al., 2013).

Their alternative point of view is that growth rate is model-dependent, but its identity is preserved by reproductive-value weighting. In the continuous-age PDE

$G_{i,j}$2

the adjoint eigenfunction $G_{i,j}$3 defines a weighted total $G_{i,j}$4 that grows exactly exponentially at the Lotka–Euler rate $G_{i,j}$5, regardless of initial age distribution. Raw totals

$G_{i,j}$6

need not do so.

Within that framework, an instantaneous excess growth rate can be written as

$G_{i,j}$7

or in discrete form

$G_{i,j}$8

Here $G_{i,j}$9 is the intrinsic Lotka–Euler rate and $\pi\in\Delta_n$0 is the realized per-capita growth of the raw head count. Positive $\pi\in\Delta_n$1 corresponds to growth above the intrinsic benchmark, for example because the current age structure contains a temporary surplus of fertile adults; negative $\pi\in\Delta_n$2 corresponds to lagging growth, for example because the population is dominated by juveniles yet to mature.

The same framework turns robustness of $\pi\in\Delta_n$3 into a model-selection criterion. A good model is said to yield a narrow distribution of estimated $\pi\in\Delta_n$4 across replicates and relative insensitivity to the chosen time window. In the mite-data comparison summarized in the supplied material, the demographic models that compute $\pi\in\Delta_n$5 through Euler–Lotka were the most stable and window-independent. In this literature, excess growth rate is therefore best understood as a transient structural residual around an intrinsic demographic law, not as a universal population constant.

5. Stochastic Population Growth, Variability, and Growth-Rate Gain

In proliferating cell populations, excess growth rate often denotes the population-level premium induced by variability, temporal correlation, or lineage selection. Hein and Jafarpour consider single-cell volume growth

$\pi\in\Delta_n$6

where $\pi\in\Delta_n$7 is a stationary Gaussian process with mean $\pi\in\Delta_n$8, variance $\pi\in\Delta_n$9, and exponential autocorrelation $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$00. For the equivalent Ornstein–Uhlenbeck process, the asymptotic population growth rate is

$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$01

The excess is therefore

$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$02

which is always positive in this model and increases with growth-rate variability and its correlation time (Hein et al., 2022).

A central result of that analysis is asymptotic decoupling: in balanced growth, the population growth rate depends only on the single-cell growth-rate process, while the population cell-size distribution depends only on division and cell-size regulation. The same excess $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$03 also sets the timescale for relaxation to steady state through damped generational oscillations and traveling waves in the cell-size distribution. The supplied details state that the damping rate satisfies

$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$04

A related but distinct formulation arises in age-structured cell-cycle models. Starting from the Euler–Lotka equation

$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$05

where $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$06 is the cell-cycle-duration density, a cumulant expansion yields, for small $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$07,

$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$08

Relative to the naive rate $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$09, the excess or gain is

$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$10

When mother–daughter cycle times have correlation $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$11, the leading correction becomes

$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$12

Positive $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$13 amplifies the gain, while sufficiently negative $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$14 can suppress it (Rochman et al., 2017).

The interpretation emphasized in that work is that variability alone can increase population growth because fast-dividing lineages produce more descendants and are therefore overrepresented in a growing ensemble. The authors tie this to asymmetric division and to the failure of “ergodicity II,” namely the claim that one lineage observed over many generations reproduces the statistics of the entire expanding ensemble. Across both cell-population literatures, excess growth rate is thus a stochastic selection premium generated by convexity, inheritance, and biased lineage sampling.

6. Excess Relative to Benchmark Exponentials

A further usage treats excess growth rate as acceleration above a fixed exponential benchmark. For standard exponential growth,

$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$15

the growth rate is constant. For finite-time-singular power-law growth,

$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$16

the instantaneous rate is

$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$17

which diverges as $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$18. The supplied summary explicitly states that an “excess” growth rate is any $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$19 that rises above the fixed-$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$20 exponential benchmark (Hüsler et al., 2011).

In the empirical calibration reported there, human population displays a transition from an earlier super-exponential regime to a late-period exponential fit

$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$21

with $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$22 for 1970–2010. Atmospheric $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$23 content, by contrast, is reported to be at least exponentially increasing and more likely super-exponential. For the combined 1850–2009 record, the FTS fit

$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$24

outperforms the best exponential fit in residual error, and monthly Mauna-Loa window scans yield $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$25, often greater than $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$26, for recent intervals, implying that the inferred $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$27 was still accelerating as of 2009. The interpretation given is that the demographic driver alone cannot account for the accelerating carbon trajectory; the excess arises from coupled positive feedbacks among population, capital, technology, output, and emissions.

Cosmology uses a related benchmark-relative notion for linear structure formation. Hirano and Komiya define the matter growth rate by

$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$28

and show that in Galileon gravity the effective Newton constant $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$29 exceeds $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$30 on cosmological scales, enhancing growth relative to $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$31CDM. Writing

$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$32

the supplied semi-analytic expression is

$$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$33

At fixed $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$34, this produces a growth rate in excess of the $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$35CDM prediction, leading growth-only fits to prefer $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$36 and $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$37 at $$G_{i,j}=\frac{1}{\Delta t}\ln(P_{i,j}/P_{i,j-1})-r_f$$38 confidence (Hirano et al., 2010).

Across these literatures, a common misconception is that “excess growth rate” refers to one invariant scalar concept. The record assembled here indicates otherwise. In some fields it is a divergence, in others a premium over a safe benchmark, a transient residual around intrinsic growth, a stochastic population-level gain, or a diagnostic of super-exponential acceleration. The unifying structure is comparative rather than ontological: excess growth rate measures growth after subtraction, normalization, or benchmarking against a theoretically privileged reference.