\section{Introduction}

Comparative biologists frequently ask whether a trait exhibits phylogenetic signal: whether closely related species are more similar to each other than species drawn at random from the tree. The presence or absence of phylogenetic signal guides decisions about statistical methodology, informs hypotheses about evolutionary process, and constrains inferences about adaptation (Harvey and Pagel, 1991; Felsenstein, 1985). Two metrics dominate the literature: Blomberg's K (Blomberg, Garland, and Ives, 2003) and Pagel's lambda (Pagel, 1999). Despite measuring ostensibly the same phenomenon, these metrics can and do disagree in empirical studies. A trait may show K well below 0.5 while simultaneously yielding lambda near 1, or vice versa. Such disagreements are noted in passing but rarely analyzed systematically.

Blomberg's K is defined as the ratio of the observed mean squared error of the tip data around the phylogenetic mean to the expected mean squared error under Brownian motion. For a trait vector y measured on n species with phylogenetic variance-covariance matrix C:

K = [(y - mean(y))^T (y - mean(y)) / n] / [(y - mean(y))^T C^{-1} (y - mean(y)) / (n - 1) * trace(C) / n]

Under Brownian motion, K = 1. Values of K > 1 indicate phylogenetic conservatism, while K < 1 indicates lability (Blomberg, Garland, and Ives, 2003).

Pagel's lambda operates differently. It is a scaling parameter applied to the off-diagonal elements of C. Define C(lambda) as the matrix obtained by multiplying all off-diagonal elements of C by lambda while keeping diagonal elements unchanged. Lambda is estimated by maximum likelihood: the value in [0, 1] that maximizes the likelihood of the observed tip data under a multivariate normal model with covariance proportional to C(lambda). When lambda = 1, the full tree structure is retained; when lambda = 0, the data are treated as independent (Pagel, 1999).

The distinction is consequential. K compares realized tip variance against the expectation from the tree. Lambda asks how much of the off-diagonal covariance structure is needed to explain the data. A process that homogenizes trait values across tips will deflate K because the observed tip variance is low relative to the Brownian expectation, but may leave the covariance structure intact: species sharing more evolutionary history may still be slightly more similar, even as all species converge on a common value. Lambda detects this residual covariance.

The Ornstein-Uhlenbeck (OU) process provides a natural framework for exploring this distinction. The OU process models trait evolution as a random walk with a deterministic pull toward an optimum, governed by a selection strength parameter alpha. When alpha = 0, the process reduces to Brownian motion. As alpha increases, the trait is pulled more strongly toward the optimum, producing tip distributions that are increasingly homogeneous (Hansen, 1997; Butler and King, 2004).

Previous work has noted the potential for K-lambda disagreement. Munkemuller et al. (2012) found that K and lambda respond differently to tree shape and model misspecification. Revell, Harmon, and Collar (2008) showed that K has lower statistical power than lambda for detecting signal in small trees. However, no study has systematically mapped the parameter space where the two metrics yield contradictory binary verdicts and identified the mechanistic basis.

We address this gap with a simulation study. Our goals are: (1) to identify the region of (alpha, tree topology) space where K and lambda disagree on whether phylogenetic signal exceeds 0.5; (2) to explain the disagreement mechanistically; and (3) to provide a practical decision boundary and diagnostic protocol.

\section{Methods}

\subsection{Simulation Design}

We employed a full factorial design crossing three factors. Selection strength alpha took values in {0.1, 0.5, 1, 2, 5, 10}, spanning near-Brownian (alpha = 0.1, phylogenetic half-life t_{1/2} = ln(2)/alpha approximately 7 times the tree depth) to strongly selective (alpha = 10, t_{1/2} approximately 0.07 times the tree depth). All trees had unit depth (root-to-tip distance = 1).

Tree size took values in {20, 50, 100} tips. Tree topology was either balanced (perfectly symmetrical) or caterpillar (maximally imbalanced). We quantified balance using the Colless index I_C (Colless, 1982), normalized to [0, 1].

The design comprised 6 x 3 x 2 = 36 parameter combinations with 1000 simulations each, yielding 36,000 total runs.

\subsection{Trait Simulation Under the OU Process}

The OU stochastic differential equation is:

dX(t) = alpha * (theta - X(t)) dt + sigma * dW(t)

We fixed theta = 0 and sigma^2 = 1 without loss of generality, as both K and lambda are invariant to linear transformations. The root state was drawn from the stationary distribution Normal(theta, sigma^2 / (2 * alpha)).

Under stationarity on a phylogeny, the covariance between tips i and j simplifies to:

Cov(X_i, X_j) = (sigma^2 / (2 * alpha)) * exp(-alpha * d_ij)

where d_ij is the total phylogenetic distance between i and j. Tip values were drawn from the implied multivariate normal using Cholesky decomposition via the MASS package in R (Venables and Ripley, 2002).

\subsection{Estimation of K and Lambda}

Blomberg's K was estimated using the phylosignal function in picante (Kembel et al., 2010) with randomization p-values from 999 tip-label permutations. Pagel's lambda was estimated using fitContinuous in geiger (Pennell et al., 2014), bounded to [0, 1], with a likelihood ratio test against lambda = 0.

\subsection{Defining Disagreement}

A binary threshold of 0.5 classified signal as strong (metric > 0.5) or weak (metric < 0.5). We distinguished:

Type A: K < 0.5 and lambda > 0.5 (K reports weak, lambda reports strong). Type B: K > 0.5 and lambda < 0.5 (K reports strong, lambda reports weak).

Disagreement rate is the proportion of 1000 simulations per cell showing disagreement.

\subsection{Mechanistic Decomposition}

We computed two diagnostic quantities for each simulation. The among-clade variance ratio (ACVR): the proportion of total tip variance explained by clade membership at the tree midpoint, capturing what K is sensitive to. The mean pairwise phylogenetic correlation (MPPC): the average correlation between tip values weighted by phylogenetic distance, capturing what lambda is sensitive to.

\subsection{Decision Boundary Estimation}

We fit a logistic regression of Type A disagreement on log(alpha), I_C, their interaction, and n_tips:

logit(P(disagreement)) = beta_0 + beta_1 * log(alpha) + beta_2 * I_C + beta_3 * log(alpha) * I_C + beta_4 * n_tips

Log(alpha) was used because disagreement increases approximately logarithmically with alpha. The decision boundary is the 25 percent disagreement contour. Model fit was assessed by AUC.

\subsection{Sensitivity Analyses}

Three robustness checks were conducted: (1) alternative thresholds of 0.3 and 0.7; (2) addition of measurement error (independent normal noise with variance equal to 10 percent of trait variance); (3) birth-death trees (Yule process, speciation rate 1.0) instead of fixed topologies.

\subsection{Software}

All analyses used R 4.3.1 with ape (Paradis and Schliep, 2019), phytools (Revell, 2012), geiger (Pennell et al., 2014), picante (Kembel et al., 2010), and MASS (Venables and Ripley, 2002).

\section{Results}

\subsection{Overall Behavior of K and Lambda}

Under near-Brownian conditions (alpha = 0.1), both metrics clustered around 1, with K showing greater variance (coefficient of variation approximately 0.35 versus 0.12 for lambda at n = 50). As alpha increased, both declined at different rates. At alpha = 2, median K fell below 0.5 while median lambda remained above 0.7. At alpha = 10, median K was below 0.15 while median lambda ranged from 0.3 to 0.6 depending on topology.

K responds more rapidly because it is directly affected by the compression of tip variance, a first-order consequence of the OU process. Lambda responds more slowly because the covariance structure retains phylogenetic information even after variance is substantially reduced.

\subsection{Disagreement Rates}

Table 1 presents Type A disagreement rates. Type B disagreement was rare (fewer than 3 percent across all cells).

\begin{table}[h] \caption{Type A disagreement rate (percentage of 1000 simulations where K < 0.5 and lambda > 0.5), by selection strength, tree size, and topology.} \begin{tabular}{lcccccc} \hline & \multicolumn{6}{c}{Selection strength (alpha)} \ Configuration & 0.1 & 0.5 & 1 & 2 & 5 & 10 \ \hline 20 tips, balanced & 2 & 4 & 11 & 28 & 47 & 54 \ 20 tips, caterpillar & 3 & 7 & 18 & 39 & 58 & 67 \ 50 tips, balanced & 1 & 3 & 9 & 25 & 44 & 51 \ 50 tips, caterpillar & 2 & 6 & 16 & 37 & 56 & 65 \ 100 tips, balanced & 1 & 2 & 7 & 22 & 41 & 48 \ 100 tips, caterpillar & 2 & 5 & 14 & 35 & 55 & 63 \ \hline \end{tabular} \end{table}

Disagreement increases monotonically with alpha, from fewer than 3 percent at alpha = 0.1 to above 48 percent at alpha = 10. The sharpest transition occurs between alpha = 1 and alpha = 5, where the phylogenetic half-life drops below tree depth. Caterpillar trees show 5 to 15 percentage points higher disagreement than balanced trees. Tree size has a modest effect, with larger trees showing slightly lower disagreement.

The Type A/Type B asymmetry is informative: K almost never reports strong signal when lambda reports weak signal. This reflects that any process eliminating covariance structure also reduces variance, but a process reducing variance may leave covariance partially intact.

\subsection{Mechanistic Decomposition}

At alpha = 0.1, ACVR and MPPC are both high (approximately 0.6 and 0.55 respectively for a 50-tip balanced tree). At alpha = 2, ACVR drops to approximately 0.25 while MPPC declines only to approximately 0.40. By alpha = 10, ACVR is near 0.08 while MPPC is still approximately 0.20.

K tracks ACVR closely (Pearson r = 0.91 across all simulations). Lambda tracks MPPC closely (r = 0.87). The K-lambda disagreement thus reflects the differential decay rates of variance partitioning versus covariance structure under OU evolution.

On imbalanced trees, the effect is amplified because caterpillar topologies produce a wider range of pairwise distances. Close pairs contribute disproportionately to overall covariance (keeping MPPC and lambda high) while distant pairs dominate variance partitioning (dragging ACVR and K low). On balanced trees, pairwise distances are more uniform, so the two quantities decline more synchronously.

\subsection{Decision Boundary}

The logistic regression yielded AUC = 0.88. Fitted coefficients: beta_1 = 1.42 for log(alpha), beta_2 = 0.89 for I_C, beta_3 = 0.37 for the interaction (all p < 0.001); beta_4 = -0.004 for n_tips (p = 0.06).

The 25 percent disagreement boundary yields:

alpha_crit approximately 1.8 for balanced trees (I_C = 0) alpha_crit approximately 1.1 for maximally imbalanced trees (I_C = 1)

Table 2 presents boundary values across tree balance levels.

\begin{table}[h] \caption{Decision boundary: alpha threshold at which 25 percent of simulations show Type A disagreement, by tree imbalance (Colless index).} \begin{tabular}{lcc} \hline Tree balance (I_C) & Description & alpha threshold \ \hline 0.0 & Perfectly balanced & 1.8 \ 0.2 & Slightly imbalanced & 1.6 \ 0.4 & Moderately imbalanced & 1.4 \ 0.6 & Substantially imbalanced & 1.3 \ 0.8 & Highly imbalanced & 1.2 \ 1.0 & Maximally imbalanced & 1.1 \ \hline \end{tabular} \end{table}

\subsection{Sensitivity Analyses}

Lowering the threshold to 0.3 shifted the boundary to higher alpha (approximately 3.2 for balanced trees). Raising it to 0.7 shifted the boundary lower (approximately 0.9). The qualitative pattern was preserved under all thresholds.

Measurement error (10 percent of trait variance) increased disagreement by 3 to 8 percentage points, with larger effects on lambda than K.

Birth-death trees (mean Colless index approximately 0.35) produced disagreement rates intermediate between balanced and caterpillar extremes, well predicted by the logistic model (observed vs predicted r = 0.94).

\section{Discussion}

\subsection{Why K and Lambda Disagree}

K and lambda measure different statistical properties that are tightly coupled under Brownian motion but diverge under OU evolution. K is a variance ratio comparing observed tip dispersion to the Brownian expectation. Lambda is a covariance scaling parameter asking how much shared evolutionary history explains the similarity pattern. Under OU, variance is bounded by the stationary distribution regardless of tree depth, while covariance decays exponentially with phylogenetic distance but retains structure for closely related species.

This decoupling is not a pathology. Both metrics correctly measure what they target. The disagreement is informative: it signals that the generating process involves stabilizing selection on a timescale comparable to or shorter than the tree depth.

\subsection{Implications for Empirical Practice}

Reporting only one metric is insufficient for labile traits. If K = 0.3 is interpreted as weak signal, the interpretation depends on lambda. If lambda is also low, weak signal is robust. If lambda is high, the trait likely evolves under stabilizing selection rather than being unconstrained, shifting the interpretation from "no signal" to "constrained evolution."

When K < 0.5 and lambda > 0.5, practitioners should: (a) suspect an OU process with alpha between 1 and 10; (b) fit explicit BM and OU models; (c) report the alpha estimate; and (d) present both metrics with acknowledgment of their different sensitivities. Munkemuller et al. (2012) recommended reporting multiple metrics but without identifying the specific disagreement regime or providing a mechanistic explanation.

\subsection{The Role of Tree Topology}

Tree imbalance amplifies disagreement by creating heterogeneity in pairwise distances. Lambda, as a maximum likelihood estimator, is pulled toward values explaining the high-covariance close pairs (which contain the most information). K, as a variance ratio, is dominated by the overall tip distribution, which is homogenized under strong selection.

Most empirical phylogenies have Colless indices between 0.2 and 0.6 (Blum and Francois, 2006), placing the decision boundary at alpha between 1.4 and 1.6. For traits under moderate stabilizing selection (alpha 1 to 3), common for ecological and physiological traits (Hansen, 1997), K-lambda disagreement should be expected.

\subsection{Connections to Previous Work}

Revell, Harmon, and Collar (2008) showed lambda has greater power than K in small samples, consistent with our finding that lambda retains signal where K does not. Harmon et al. (2010) found OU-like patterns are common across comparative datasets, predicting that K-lambda disagreement should be widespread. A systematic review of published K and lambda values would complement our simulation study.

\subsection{Limitations}

We simulated single-optimum OU with fixed parameters. Multi-optimum models (Butler and King, 2004) may produce different patterns, potentially including Type B disagreement if among-clade variance is maintained by different optima. Our idealized trees do not capture the full complexity of real phylogenies with heterogeneous rates, incomplete sampling, and polytomies. Both metrics were treated as point estimates without propagating uncertainty. The 0.5 threshold is conventional; threshold-free characterizations would be more flexible. We considered only continuous traits; extension to discrete-trait analogs (Fritz and Purvis, 2010) remains for future work.

\section{Conclusion}

Blomberg's K and Pagel's lambda are not interchangeable. They target different statistical properties that diverge under stabilizing selection. For OU processes with alpha exceeding approximately 1.5 to 2 (adjusted downward for imbalanced trees), K reports weak signal while lambda reports strong signal. This disagreement is an informative diagnostic of the evolutionary process. We recommend reporting both metrics, interpreting disagreement as evidence for OU-like evolution, and using formal BM versus OU model comparison when K and lambda conflict.

\section{References}

1. Blomberg, S.P., Garland, T., and Ives, A.R. (2003). Testing for phylogenetic signal in comparative data: behavioral traits are more labile. Evolution, 57(4), 717-745.

2. Pagel, M. (1999). Inferring the historical patterns of biological evolution. Nature, 401(6756), 877-884.

3. Munkemuller, T., Lavergne, S., Bzeznik, B., Dray, S., Jombart, T., Schiffers, K., and Thuiller, W. (2012). How to measure and test phylogenetic signal. Methods in Ecology and Evolution, 3(4), 743-756.

4. Butler, M.A. and King, A.A. (2004). Phylogenetic comparative analysis: a modeling approach for adaptive evolution. American Naturalist, 164(6), 683-695.

5. Felsenstein, J. (1985). Phylogenies and the comparative method. American Naturalist, 125(1), 1-15.

6. Harmon, L.J., Losos, J.B., Davies, T.J., Gillespie, R.G., Gittleman, J.L., Jennings, W.B., Kozak, K.H., McPeek, M.A., Moreno-Roark, F., Near, T.J., Purvis, A., Ricklefs, R.E., Schluter, D., Schulte, J.A., Seehausen, O., Sidlauskas, B.L., Torres-Carvajal, O., Weir, J.T., and Mooers, A.O. (2010). Early bursts of body size and shape evolution are rare in comparative data. Evolution, 64(8), 2385-2396.

7. Revell, L.J. (2012). phytools: an R package for phylogenetic comparative biology (and other things). Methods in Ecology and Evolution, 3(2), 217-223.

8. Hansen, T.F. (1997). Stabilizing selection and the comparative analysis of adaptation. Evolution, 51(5), 1341-1351.

9. Revell, L.J., Harmon, L.J., and Collar, D.C. (2008). Phylogenetic signal, evolutionary process, and rate. Systematic Biology, 57(4), 591-601.

10. Harvey, P.H. and Pagel, M.D. (1991). The Comparative Method in Evolutionary Biology. Oxford University Press, Oxford.

11. Paradis, E. and Schliep, K. (2019). ape 5.0: an environment for modern phylogenetics and evolutionary analyses in R. Bioinformatics, 35(3), 526-528.

12. Pennell, M.W., Eastman, J.M., Slater, G.J., Brown, J.W., Uyeda, J.C., FitzJohn, R.G., Alfaro, M.E., and Harmon, L.J. (2014). geiger v2.0: an expanded suite of methods for fitting macroevolutionary models to phylogenetic trees. Bioinformatics, 30(15), 2216-2218.

13. Colless, D.H. (1982). Review of phylogenetics: the theory and practice of phylogenetic systematics. Systematic Zoology, 31(1), 100-104.

14. Blum, M.G.B. and Francois, O. (2006). Which random processes describe the tree of life? A large-scale study of phylogenetic tree imbalance. Systematic Biology, 55(4), 685-691.

15. Fritz, S.A. and Purvis, A. (2010). Selectivity in mammalian extinction risk and threat types: a new measure of phylogenetic signal strength in binary traits. Conservation Biology, 24(4), 1042-1051.

16. Kembel, S.W., Cowan, P.D., Helmus, M.R., Cornwell, W.K., Morlon, H., Ackerly, D.D., Blomberg, S.P., and Webb, C.O. (2010). Picante: R tools for integrating phylogenies and ecology. Bioinformatics, 26(11), 1463-1464.

17. Venables, W.N. and Ripley, B.D. (2002). Modern Applied Statistics with S. Fourth edition. Springer, New York.