Processes underlying complex patterns of song trait evolution in a Setophaga hybrid zone

Abstract During secondary contact between two species when hybrids are less fit than parents, mating signals are expected to diverge, while aggressive signals are expected to converge. If a single signal trait is used in both mating and aggression, then the dynamics between these two forces could influence the evolutionary trajectory of that trait. We studied such a situation in an avian hybrid zone between two Setophaga species, where birdsong is used in both mate attraction and territory defense. We hypothesized that song modules of the two species will show separate and distinct geographic patterns due to the influence of selective pressures for effective territorial aggression and for effective mate attraction. We conducted geographic cline analyses and playback experiments across this hybrid zone. We found an unexpected geographic pattern of asymmetric introgression of song rhythm, which may be explained by results of the playback experiments that suggest that differences in song rhythm serve a greater role in mate attraction than in territory defense. In contrast, differences in syllable morphology show little evidence of importance in mate attraction or territorial defense. Song features converge in the hybrid zone, yet patterns of trait change suggest that the song production modules may vary in their modes of development and inheritance. Syringeal motor gesturing, which gives rise to syllable morphology, shows a nonclinal mosaic pattern, suggesting that this trait may be predominantly learned. In contrast, respiratory patterning, which forms song rhythm, shows a clinal geographic transition, suggesting that this trait could be more innate. The results indicate that opposing forces act independently on song via distinct modules of the song production mechanism, driving complex patterns of song trait evolution.

populations and thus increases speciation rates compared with when divergence in mating signals is absent (Pfennig & Pfennig, 2009;West-Eberhard, 1983). Given their clear evolutionary importance and often charismatic nature, mating signals predominate biological discussion of signaling, but signals with other functions are also used in abundance.
Aggressive signals are found widely in animals and aid in securing access to resources, including food or mates (Grether et al., 2017).
While often more species-general than mating signals, aggressive signals are also expected to diverge during the speciation process via adaptive (Endler, 1992;Morton, 1975) or neutral (Irwin et al., 2008) processes. However, unlike mating signals, aggressive signals are expected to converge upon secondary contact if genetic or ecological divergence between species is incomplete, a condition that encourages interspecific competition or territoriality (convergent agonistic character displacement; Grether et al., 2017). This phenomenon presumably arises out of selection for effective aggressive communication across species boundaries while reducing the potential costs of direct physical encounters (Cody, 1969;Grether et al., 2017;Tobias & Seddon, 2009). Convergence in aggressive signals should decrease loss of fitness due to hybridization (Grether et al., 2017), since mates and mating territories will be defended against competing members of both species.
Birdsong can be used as a mating signal and as an aggressive signal (Marler 1960, Collins, 2004Searcy & Nowicki, 2005). Thus, in cases of secondary contact between two closely related species, we expect disparate selective pressures to act on song. In such a situation, selection for the facilitation of aggressive signaling should result in song convergence between populations, while selection for effective mate signaling should result in song divergence between populations. Changes to song upon secondary contact therefore present opportunities for dissecting these competing selective pressures.
It has been hypothesized that in cases of dual-purpose signals, the mating function should be the primary driver of evolutionary change (Okamoto & Grether, 2013). This suggests that the songs of two species should diverge upon secondary contact when hybrids are less fit than parents. However, empirical studies have shown mixed results, with some finding upon secondary contact song convergence (de Kort et al., 2002, Secondi et al., 2003, Haavie et al., 2004, Tobias & Seddon, 2009, Kenyon et al., 2011, Laiolo, 2012 and others upon divergence (Haavie et al., 2004;Kirschel et al., 2009) or stasis (Halfwerk et al., 2016). These conflicting results are intriguing: Why would there be such disparate outcomes in seemingly similar situations?
One possibility is that the varied patterns of song trait evolution in response to secondary contact are, in part, the result of the action of the competing selective pressures described above. As was expressed in the original theoretical formulation (Okamoto & Grether, 2013), the selection imparted by the mating function of a dual-use signal should be greater than that imparted by the aggressive function when the signal is truly simple. If, however, the signal is complex and multivariate (i.e., arising through contributions from multiple physiological mechanisms or expressed through multiple signaling modalities), then other possibilities emerge. For example, divergence in the social function of birdsong could arise, if two distinct features of song evolve divergently in response to specific, singular selection pressures. This possibility contrasts with a scenario where potentially opposing pressures exert either stabilizing or directional selection on song as a whole. Since birdsong is a complex trait, perhaps the discrepant results from past studies of song evolution in zones of secondary contact could be explained in this way.
Complicating the situation of potentially opposing trajectories of song in response to secondary contact is the fact that for all investigated oscine species, some aspects of species-specific song develop through learning (e.g., Grant & Grant, 1996;Marler, 1970;Thorpe, 1958). It is thought that learning increases rates of trait change through a higher copy-error rate compared with the absence of learning, and song learning has been implicated in the rapid diversification of vocal learning songbirds (Lachlan & Servedio, 2004).
However, it remains difficult to form hypotheses of precisely how vocal learning influences song evolution, because details of the learning process vary (e.g., species differences in tutor-tutee relationships, timing of the sensitive period) and substantial interspecific variation exists in which features of song must be learned (Love et al., 2019), Disentangling culturally transmitted from genetically inherited song traits would allow a more accurate assessment of song trait evolution but has proved to be difficult. A promising approach involves the distinction between three main production "modules" that must be coordinated into an integrated system of control during song production. This distinction of modules is based on the three main neuromotor mechanisms for song production: respiratory pattern, syringeal motor gesture, and central program for syllable sequence. Each of these modules contributes specific song features (song rhythm, syllable morphology, and song syntax). Critically, these modules show evidence of differential reliance on learning for their development. Both comparative and experimental studies suggest that, in many species, the rhythm of song (the coarse patterns of sound and silence, which are produced primarily by the respiratory system) is likely to be less subject to modification through learning than other production modules (syllable acoustic morphology and syntax; Marler & Sherman, 1983, Ali et al., 2013, Araki et al., 2016, Love et al., 2019, also see Lipkind et al., 2017. Consequently, distinguishing between production modules in the analyses of song across contact zones may allow for more detailed evaluation of the direction, strength, and targets of selection relative to an approach that assumes uniform learning of all aspects of song. Hermit warblers (Setophaga occidentalis) and Townsend's warblers (Setophaga townsendi) are sister species that are estimated to have diverged in glacial refugia during the middle or late Pleistocene and reached secondary contact approximately 5,000 years ago (Lovette & Bermingham, 1999;Rohwer et al., 2001). Despite their distinct plumage phenotypes, individuals of both species are interspecifically territorial and hybridize in three small areas in the mountains of the northwestern United States where their ranges overlap (Rohwer & Wood, 1998). Importantly, hybrids between the two species appear to have reduced fitness, as they produce fewer eggs per clutch than either parent species (Pearson, 1997;Pearson & Rohwer, 1998). Therefore, we should expect reinforcing selection to act on mating signals and decrease fitness loss due to hybridization. Townsend's males appear to be more aggressive than Hermit males (Pearson, 2000;Pearson & Rohwer, 2000), and Townsend's females produce larger average clutch sizes than Hermit females.
Both factors have been used as an explanation for the supposed movement of the hybrid zone from north to south, with Townsend's replacing Hermit warblers (Krosby & Rohwer, 2010;Pearson, 2000;Pearson & Rohwer, 2000). However, recent genomic evidence suggests that this hybrid zone has not, in fact, been moving as was previously thought . While previous researchers have stated that both species "sing the same songs" in the hybrid zone (Pearson & Rohwer, 2000), song and its role in interspecific interactions have not been assessed in detail.
Our study aimed to help determine (a) whether different song traits evolve independently or in concert in this system, (b) which selective forces act most strongly on which song trait, and (c) how cultural evolution and genetic evolution interact through song learning and production to shape song trait evolution. Our overarching hypothesis was that song modules of the two species will show separate and distinct geographic patterns due to the influence of selective pressures for effective territorial aggression and for effective mate attraction. Here, we report (a) how different acoustic features, based on two production modules, of the songs of Hermit and Townsend's warblers vary across a hybrid zone and (b) the results from an experiment investigating the social function of song modules that may partially explain the observed patterns of spatial change.

| ME THODS
Over the course of four consecutive breeding seasons, we recorded songs of Hermit warbler, Townsend's warbler, and hybrid males across the Olympic hybrid zone. We conducted a production-based quantitative acoustic analysis and performed a geographic cline analysis to investigate how song traits vary across the focal hybrid zone. In parallel, we used playback of song to study the response of male and female Hermit, Townsend's, and hybrid warblers. The playback study allowed us to investigate the social function (i.e., which song features are used in mating versus territoriality) and the social value (i.e., what trait values are associated with stronger or weaker response for a given social function) of different song features across the hybrid zone.
To form baseline expectations, we first sought to establish whether song differed between species. We assessed song recordings from Hermit and Townsend's warblers from two databases: xeno-canto.org and the Cornell University Lab of Ornithology Macaulay Library (see Appendix S2 for accession numbers). We visually inspected spectrograms of 141 Townsend's warbler and 46 Hermit warbler songs that were recorded during the breeding season away from any hybrid zone. Because the occurrence of multinote versus single-note introductory syllables differs between these species (Morrison & Hardy, 1983), we categorized each song into one of these two groups. We then compared relative proportions of song categories across the two species and used these results to formulate our research design (see Appendix S1).
From May to July 2015-2018, the Olympic hybrid zone (described in Rohwer & Wood, 1998) and surrounding allopatric regions were surveyed for singing Hermit, Townsend's, and hybrid warbler males (Figure 1). A roving approach was used to provide a random sample of individuals. Roads, trails, and off-trail routes were followed in vehicles or on foot until a singing individual was heard; that individual was then located, and high-quality record- identification and study using a playback speaker with the observer located over 10 m away. Geographic coordinates were recorded with a handheld GPS. Resampling was avoided primarily by tracking these highly territorial birds in the canopy and ensuring that each subsequent sample was well removed from the previous. Previously sampled regions were avoided within and across seasons.

| Plumage scoring
Hermit and Townsend's warblers were discriminated visually on the basis of their different black and yellow plumage patterning, which was readily visible at a distance. Hybrids had intermediate plumage, and we used an established protocol for attaining a hybrid index score for individuals in the Hermit-Townsend's warbler hybrid zone (Rohwer & Wood, 1998). This identification scheme can be reliably employed in the field and represents genetic background (Pearson, 1997(Pearson, , 2000. Briefly, for each bird, clearly defined plumage patches (midflank, low flank, bib, face, extent of yellow on breast, intensity of yellow on breast, crown, and back color) were scored from 0 (Hermit-like) to 1 (Townsend's-like) and then averaged. Using binoculars and digital photography, we followed previous protocol with one exception: Face coloration was included in the overall score.
This trait was omitted in Rohwer and Wood (1998) because it was considered a single-locus trait, but subsequent research in congeners indicates that it may be a multilocus trait (Brelsford et al., 2017).
Furthermore, recent evidence suggests that in these warblers, face color is tightly genetically linked to crown and breast coloration . Since our dataset includes many individuals with intermediate cheek patch coloration, we feel confident that including all measured plumage patches in our scoring system contributes positively to our ability to infer genetic background.

| Acoustic analyses
From each recording, the single song with the greatest acoustic quality was selected for analysis (high signal-to-noise ratio) and extracted as a separate short.wav file using Praat software. Each short file was band-pass-filtered between 2.5 and 9.5 kHz to reduce background noise while maintaining full acoustic integrity across the song frequency range. These files were then segmented by syllable using RavenPro v1.5 by highlighting each syllable in a joint spectrogram and waveform view and compiling a "selection table" that includes start and end times for each selection. Syllables were defined as a single continuous sound separated from other sounds by a silent period longer than 0.015 s or as a group of sounds separated from each other by less than 0.015 s and separated from other sounds by more than 0.015 s. This definition was used to identify vocalizations most likely to be made during a single expiratory event; 0.015 s is a value that approximates the shortest recorded inspiratory event ("minibreath") of any bird species (e.g., Goller & Daley, 2001;Hartley & Suthers, 1989). Two separate acoustic analyses were conducted: syllable morphology and rhythm.
The acoustic morphology of syllables was analyzed by taking au- ). These measurements were then included in a linear discriminant analysis ("lda" function in R). The lda was trained on the data from allopatric populations of Hermit and Townsend's warblers, and then applied to the full dataset, including samples from both allopatric (i.e., removed from the hybrid zone) and sympatric (i.e., within the hybrid zone) populations. As described in Results, syllable morphology LD1 scores reflect a complex set of acoustic features (see Figure S5 for song spectrogram showing LD1 scores for each syllable).
Rhythm was analyzed using custom code in R that assessed components of songs relating to the pattern of sound and silence (fol- Then, the proportions were normalized by dividing each proportion by the maximum proportion. These normalized proportions were summed, and the sum was divided by the total number of syllables in the song. The resulting value yielded the syllable regularity. The same procedure was utilized to produce the silent period regularity, substituting silent period lengths for syllable lengths. A linear discriminant function was then applied, following the same procedure that was used for syllable morphology. As described in Results, songs that receive more positive rhythm LD1 scores have fewer syllables, longer silent periods between syllables, and a lower syllable repetition rate with less consistent syllable duration. We quantified the degree of convergence in two ways: one as the proportion of sympatric birds predicted to be Hermit (for syllable morphology) or Townsend's (for rhythm) by linear discriminant classification, and one as the proportion of sympatric birds with LD1 scores that were negative (for syllable morphology) or positive (for rhythm).
To assess the statistical validity of our linear discriminant-based findings, we conducted a randomization test. Briefly, we replicated our syllable morphology and rhythm data but randomly assigned each sample to a species identity. For each song module, we then trained 10,000 linear discriminant functions on simulated data, repeating above methods, and reported the two measures of convergence. We then determined the proportion of the distribution of simulated convergence statistics (here, the null distribution), which were equal to or more extreme than the values that we found with our real data. This proportion functions as a p-value (Whitlock & Schluter, 2009).
It has been previously described that Hermit and Townsend's warblers have two main song types, which are "performance-encoded" (Byers 1995) or distinguished by their apparent use (Spector 1992, Janes & Ryker, 2006, Janes and Ryker 2016. We conducted a thorough song-type analysis and did not find strong evidence to support this distinction of songs (Love and Goller unpublished). We therefore include all song forms in the present study.

| Geographic cline analyses
To visualize and quantify trait change across the hybrid zone, we conducted three one-dimensional geographic cline analyses. The plumage-based hybrid index scores and the first linear discriminant scores produced by the syllable morphology and rhythm linear discriminant functions served as trait data for the three analyses.
Obtaining clines based on each of these traits allowed us to compare how each song trait evolves across the hybrid zone relative to the plumage trait, a reliable indicator of genetic background. If a song trait cline differs from the plumage cline in width, then we can infer that there is a difference between the strength of selection acting on plumage versus that acting on the song trait. If the clines differ in center location, then we can infer introgression between the two traits. Lastly, if the data cannot be effectively modeled by a cline, then it suggests that either parent species do not differ in this trait or that atypical patterns of inheritance are present.
We used the hzar package in R, following a protocol outlined in Derryberry et al. (2014) for quantitative traits and using default settings, no parameter constraints, and a single cycle of fit requests (two sequential model fittings per analysis) to reach a maximumlikelihood cline, which gives a representation of the variation in a trait over a geographic area. Since these analyses relied on grouped sampling events, we binned our samples by every kilometer following a north-south axis. We used a direct north-south axis as our single geographic dimension, which fits the species and plumage score distribution across our study area and any other possible straightline transect (see Results) and is in line with past studies (Rohwer et al., 2001;Rohwer & Wood, 1998).

| Playback experiment
We conducted a playback experiment in order to evaluate the social function of song traits in this system. By varying the rhythm and syllable morphology trait values of playback songs across trials and recording response magnitude of each individual, we intended to identify which features of song are socially salient in terms of module (rhythm versus syllable morphology) and trait value (Hermitlike versus Townsend's-like). By testing both males and females, we intended to determine the relative importance of song traits in functioning as a mating signal in male-female interactions or as an aggressive signal in male-male interactions.
The playback experiment was conducted in May-July 2016, 2017, and 2018. Locations were determined through a roving approach of the entire study area. Individuals were located by their song and then tracked to their perch. After an individual was located and its song recorded as above, playback was initiated. The playback song used was randomly selected from a suite of our previously recorded songs, which were selected for features described in the following text. Prior to use in the experiment, these songs were edited in the Praat software to remove any background noise with a band-pass filter and spectral subtraction, using a window length of 0.025 s and a smoothing bandwidth of 1.0. After filtering, mean intensity of the song was scaled to 75 db. Peak intensity was not normalized in order to preserve any potentially informative differences between songs.
Preliminary analysis of song databases (xeno-canto.org and Macaulay Library) showed a clear pattern in allopatric speciesspecific song: Allopatric Townsend's warblers use single-note introductory syllables, while allopatric Hermit warblers use multinote introductory syllables (see Appendix S1). With this in mind, we established a suite of 8 songs, each sung by a different bird, to use for playback that varied in introductory syllable morphology and species identity of the singer. 8 songs were used for playback (see To refine our design for the playback experiment, we conducted a pilot study (see Appendix S1). All trials included in our dataset used the following design. These species live in the forest canopy, so playback was initiated under the perched bird, immediately adjacent to the perch tree, in an exact spot selected for the availability of perches at continuously decreasing distances from the playback speaker, which was placed on the ground. Next to the speaker, we placed a clay mount, one side being painted with the plumage pattern of a Townsend's warbler, and the other side painted with the plumage pattern of a Hermit warbler, which was oriented so as to not favor one side over the other from the most readily available perches. 10-m, 3-m, and 1-m distances from the speaker were measured with a premeasured lightweight pole, which was removed before playback began and marked by naturally available objects.
With the experimenter standing over 10 m away in a location that allowed for observation of the full study area, playback was commenced. Songs were randomly selected for playback with a balanced design. A song was played at a rate of once every 3.5 s for 10 min, and magnitude of response to playback was continuously recorded.
After the playback period was complete, photographs and additional song recordings were taken. Many individuals were tested a second time with a different playback song a minimum of 30 min after the end of the first playback period, an amount of time that was sufficient for the focal bird to resume normal behavior. Order of playback response was recorded (i.e., 1st or 2nd playback response) in the field and was subsequently included as a random effect in analyses. Since they are more difficult to detect and locate in the field, we could not reliably target females in the same way as males.
Still, whenever we had direct confirmation that a female was in the immediate vicinity of the male-directed playbacks described above, we recorded their behavior and plumage score (using the methods for males but calibrated for female plumage traits) as well. We included in our analyses only female responses, which were uninterrupted by a responding male. Females were occasionally observed showing interest in playback, but their potential approach was interrupted by the male who was the primary target of our playback. In this case, we observed that the responding male appeared to chase the female, sometimes aggressively, a behavior that suggested mate guarding by the male (Birkhead, 1979). This observation also suggests that response from the female to playback may indicate an intent to explore extra-pair mating opportunities rather than to defend the territory. Sample sizes were much lower than for males (n = 32 total female responses recorded; 10 Hermit, 6 hybrid, and 16 Townsend's). Female responses were generally of a less aggressive nature than male responses. In contrast to our observations of males, we did not observe any female make contact with the playback speaker or mount, and their flights and behaviors appeared less determined. These observations again suggest that female response to playback did not reflect territory defense. We categorized female response more broadly as close (less than 3 m from the playback speaker), far (over 3 m away, but clearly altering their behavior by approaching the playback speaker), and ignore (did not clearly alter their behavior by approaching the playback speaker). For an additional direct comparison of male-to-female response, we scaled and centered the female close response data and combined them with the separately scaled male response data, then conducted an ANOVA with a sex-by-song rhythm interaction term.
In all analyses that used sympatry/allopatry status, we defined individuals as residing in sympatry when they were within 40 km of a heterospecific (or hybrid). Those that were more than 40 km from a heterospecific (or hybrid) were defined as residing in allopatry.

| RE SULTS
Hermit and Townsend's warbler males use different songs in allopatry but sing similar songs near the center of the hybrid zone.
General patterns of trait change across the hybrid zone suggest overall convergence, with notable convergence on Townsend's-like song rhythm, though the specifics of the patterns of change in the two song modules are different. The playback experiment revealed that females respond differently to song rhythm, while males show little difference in response to different song forms.

| Acoustic analysis
Preliminary analysis of online databases of Hermit and Townsend's warbler songs yielded 141 Townsend's and 46 Hermit warbler songs.
Visual inspection of these spectrograms showed a clear trend toward predominant use of multinote syllables by Hermit warblers, while Townsend's warblers use single-note syllables in the first song phrase (see Appendix S1). This observation is consistent with the findings of Morrison and Hardy (1983). We used these results to frame the rest of our study.
One part of our analysis of songs recorded in and immediately around the hybrid zone assessed song rhythm. Hermit warblers and Townsend's warblers use different song rhythms in allopatry, but song rhythm converges to Townsend's-like values in areas of sympatry (Figure 3), suggesting a pattern of introgression. Coefficients of the first linear discriminant (LD1) in the linear discriminant analysis are shown in Table 1. Songs that receive more positive rhythm LD1 scores have fewer syllables, longer silent periods between syllables, and a lower syllable repetition rate with less consistent syllable dura-

tion. In allopatry, rhythm LD1 scores readily distinguished between
Hermit and Townsend's warbler song. Townsend's warblers primarily sang songs with rhythm LD1 scores greater than zero (peak estimated at LD1 = 0.97), and Hermit warbler songs scored on rhythm We compared these convergence statistics with the distribution of simulated statistics in randomization tests. Using the rhythm LDA classification-based statistic, we found that 9,652/10,000 trials had a lesser degree of convergence than was found in the real data (p = .029). Using the rhythm LD1 score-based convergence metric, an extreme majority of trials (9,926/10,000; p-value = 0.0074) had a lower proportion of positive sympatric rhythm LD1 scores than was found in our dataset. In particular, since LD1 scores reflect acoustic traits more directly than LDA classification, the results of these tests

| Geographic cline analysis
Song rhythm showed a clinal transition across the hybrid zone that is shifted to the south and less steep when compared to the plumage ID cline (Figure 4). In contrast, syllable morphology displayed negligible clinal change across the hybrid zone (Figure 4, bottom). Onedimensional geographic cline analysis of plumage identity scores showed a cline with center at 5270800m N (UTM Zone 10N) and

| Playback experiments
If individuals of these species use either introductory syllable morphology or song rhythm as a cue in territorial disputes or mate attraction, then we expected these tendencies to be expressed during song playback experiments. We found that categorical parameters (single/multiple introductory syllable morphology and species identity of the singer), around which we designed our playback, did not significantly affect response, but song rhythm LD1 scores did.
As described in Methods, composite response scores were generated by applying a PCA to raw response data. The composite response scores (PC1) show similar patterns to those of the raw data (see Appendix S1). A linear mixed-effects model was constructed to assess the relationship between male response to song and the following explanatory variables: species identity of the responding bird, allopatry/sympatry status, introductory syllable morphology category of the playback song, and species identity of the playback song singer ("lmer," lme4 package, p-values from Satterthwaite's method).
Only parent species male responses were considered in these analyses. Order of playback, Julian date, and stimulus song were included as random effects. Species identity of the responding bird showed a significant relationship (Townsend's warblers show greater response than Hermit warblers; df = 122.51, t = 2.46, p = .015), as did allopatry status (overall response is lower in sympatry than in allopatry; df = 123.81, t = 3.08, p = .0026). Interestingly, neither syllable morphology category nor species identity of the singer of the playback song had a significant effect on response (df = 5.19, t = −0.607, p = .57 and df = 5.40, t = 1.30, p = .25, respectively).
To further characterize these relationships, additional statistical analyses were conducted. Response to playback song was lower in the hybrid zone compared with that found in allopatric populations, and this trend appears to be heavily driven by a reduced response In addition to analyzing responses to song categories (single-vs.
multinote syllables and species ID of singer), we also assessed the effect of average syllable morphology LD1 and rhythm LD1 scores on responses by males and females. We found that one of these features of song was more informative to response magnitude.  females did (F = 6.39, df = 30, adjusted r-squared = 0.15, p = .017) ( Figure 5a). To confirm this sex difference in response, we conducted an ANOVA on response with a sex-by-rhythm LD1 interaction term (sex: F = 0, p = 1; rhythm: F = 2.08, p = .15; sex*rhythm:

| D ISCUSS I ON
Our focus in this study of song in an avian hybrid zone was on the interactions among two primary sources of variation in birdsong (rhythm and syllable morphology) and two primary selective regimes F I G U R E 4 Geographic cline analysis. Samples were averaged every 1 km. Black line represents the maximum-likelihood cline. Shaded region is 95% credible cline region. Red and blue lines indicate trait cline-based center and edges of hybrid zone. Plots produced with HZAR package (Derryberry et al., 2014) 0 predicted by theory to be active in such a system (selection for effective mate attraction and for efficient territory defense). We found that song traits change across the hybrid zone in a bipartite manner; song rhythm converges on Townsend's-like values, while syllable morphology converges on intermediate values (Figures 3 and   4). The directions of this two-part change may be explained by the result of a playback experiment: Females responded more strongly toward song with Townsend's-like song rhythm (high rhythm LD1 scores), while the responses of neither sex appeared to differ based on syllable morphology ( Figure 5). While uneven sample sizes limit our confidence to draw strong conclusions, we interpret this result to suggest that song rhythm is a main target of selection for effective mate attraction. If further study indicates that our interpretation is correct, rhythm and syllable morphology respond to different selective regimes, and song evolves in accordingly complex patterns.
Song rhythm and syllable morphology arise through two largely distinct song production modules. It is this independence, which allows the two aspects of one signal to evolve separately from one another. Song rhythm is produced through song-associated respiratory patterning, which is thought to be largely innate in many species (Araki et al., 2016;Love et al., 2019;Marler & Sherman, 1983Thorpe, 1958), while syllable morphology arises through syringeal motor gesturing, which is more likely to be learned to a greater extent (Ali et al., 2013;Love et al., 2019;Marler & Sherman, 1983. Irrespective of their different directions of change, the remarkably different shapes of change of rhythm and syllable morphology in our study (Figure 4) may reflect the respective inheritance modes of these two song production modules. Rhythm follows a clinal transition characteristic of vertical genetic inheritance (Mundinger, 1982, Barton and Hewitt, 1985, Isler et al. 2005, Nyári, 2007. In contrast, syllable morphology follows a nonclinal "mosaic" pattern, which is consistent with locally mediated cultural transmission in learned vocalizations (Podos & Warren, 2007), but which could potentially be the result of processes unrelated to vocal learning (Odom & Mennill, 2012). Further study is therefore warranted to confirm these findings. Learning of local song features is an important factor regulating song development (e.g., Akçay et al., 2013;Nordby et al., 2001), and so we suggest that cultural transmission of syllable morphology is a likely interpretation of the mosaic pattern of syllable morphology in this system.
Our suggestion that unlearned song rhythm plays a greater role in mate attraction than does learned syllable morphology is of great interest in the context of the evolution of vocal learning. Mate attraction has been suggested as a key contributor to the evolution of vocal learning, which in some species allows individuals to develop large repertoires that could aid in mate attraction (Catchpole and Slater, 2003). In particular, if confirmed by direct study of song development in these species, our results subtly suggest that unlearned song features could be more important than learned song features for mate attraction in this system. While studies in some other species find that learned song components are important in mate attraction (e.g., Hernandez et al., 2008;Spencer et al., 2003), further research into the role of unlearned song components in mate attraction for vocal learning species is warranted. Our observation that rhythm appears less likely to be learned than syllable morphology fits with data from other species (Love et al. 2019), providing a potential starting point for such future research.
We hypothesized that song modules would show separate and distinct patterns of change across the hybrid zone because of the interaction of selective pressures for effective territorial aggression and for effective mate attraction. While we found general support for our hypothesis, the specific patterns that we found were somewhat unexpected. Reinforcement theory predicts that in the case of reduced hybrid fitness, which exists in this system (Pearson, 2000;Pearson & Rohwer, 2000), mating signals should diverge. We did not find evidence of divergence between species for any aspect of song, and thus no evidence for reinforcement acting on song in this system. Pearson (2000) hypothesized that female preference for Townsend's warbler traits drives hybridization. Our results provide some support for this hypothesis. The apparent female tendency to respond more strongly to Townsend's-like song rhythm may explain convergence of this feature to Townsend's-like values, which could potentially act in opposition to reinforcing selection. In the absence of such a preference, we would expect reinforcement to drive divergence of song rhythm between species. However, both the overall pattern of convergence of song and the convergence in syllable morphology are consistent with convergent agonistic character displacement (Grether et al., 2017).
Past research has suggested that Townsend's males are more aggressive than Hermit males and that species differences in aggression could contribute to movement of this hybrid zone (Pearson, 2000;Pearson & Rohwer, 2000). Our results suggest that male Townsend's warblers showed higher response magnitude than male Hermit warblers, but that difference was only evident in allopatric populations.
In sympatric regions, we did not find that response magnitude differed between the two species (Appendix S1, Figure S4).
To what degree variation in the reliance on learning of song modules differentially drives phenotypic change needs to be investigated further, and our suggested mechanisms present testable hypotheses. Indeed, studies of other avian hybrid zones find similar complex patterns of trait evolution characterized by a clinal convergent transition in song rhythm-related features paired with a nonclinal pattern of syllable morphology-related features (e.g., Lipshutz et al., 2017;Secondi et al., 2003). It may be the case that differences in reliance on learning between song production modules are observable in other birdsong hybrid zones. However, since many studies use a multidimensional scaling approach to song analysis, the distinction between learned and genetically determined song traits may go unnoticed or the patterns uncovered in the process of analysis may be difficult to interpret.
Given the theoretical predictions of reinforcement and convergent agonistic character displacement, our conclusion about selective forces in this system is well supported by our data. However, we cannot directly exclude other possible explanations for the observed patterns. Our data do not allow us to assess whether genetic covariance of song traits with other targets of selection or potential differences between species in their respective vocal learning program may also have played a role in the observed geographic patterns of song. In addition, our interpretation that female response to playback reflects mating attraction informs our view that selective forces for effective mate attraction act on song rhythm. If female response is motivated differently, such as by territory defense, then it suggests that females and males weigh their levels of territorial aggression based on different male song features and that selection for effective territory defense may act separately through males and females. The Townsend's-Hermit warbler hybrid system would be an appropriate venue for further exploration of all of these possibilities, and the results of our study provide testable predictions.
Song is viewed as an important mating signal under strong sexual selection (Collins, 2004;Searcy & Nowicki, 2005). Song learning is expected to accelerate divergence of song between species.
Rapid divergence of song between diverging species should, then, establish reproductive barriers quickly during the speciation process. Indeed, it has been hypothesized that the rapid diversification of vocal learning songbirds may be due to rapid evolution of song (Lachlan & Servedio, 2004). However, this hypothesis has been subject to continued discussion (Baptista & Trail, 1992;Olofsson et al., 2011;Seddon & Tobias, 2007;Slabbekoorn & Smith, 2002;Verzijden et al., 2012;Yeh, 2018Yeh, , 2019. Overall, this study questions the notion that vocal learning accelerates speciation through rapid cultural changes in song. In the pair of recently diverged species in the current study, we find evidence that aspects of song that appear more innate may play a stronger role in mate attraction, and therefore, divergence in these aspects would be more likely to constitute a barrier to reproduction between species. It remains to be determined whether this observation applies to other taxa.

This research was supported by George R. Riser Travel and Research
Awards to JL. We thank Natasha Verzhbitskiy for support and for the clay mount. Mandy Holmgren with the Institute for Bird Populations and Scott Gremel with the National Park Service provided logistical support and thoughtful discussion that greatly benefitted the project. We thank Matthew Dubeau, Roger Hoffman, and Patti Happe at Olympic National Park and Joel Nowak at Olympic National Forest for help with permitting. We thank Sievert Rohwer for early encouragement. Several reviewers provided helpful critiques that improved the quality of this work.

CO N FLI C T O F I NTE R E S T
None declared.

DATA AVA I L A B I L I T Y S TAT E M E N T
Data supporting our results are included in the supplement and on Dryad at https://doi.org/10.5061/dryad.n5tb2 rbvk. Additional supporting data are available upon request.