Avoiding costly fights can help conserve energy needed to survive rapid environmental change. Competitor recognition processes help resolve contests without escalating to attack, yet we have limited understanding of how they are affected by resource depletion and potential effects on species coexistence. Using a mass coral mortality event as a natural experiment and 3,770 field observations of butterflyfish encounters, we test how rapid resource depletion could disrupt recognition processes in butterflyfishes. Following resource loss, heterospecifics approached each other more closely before initiating aggression, fewer contests were resolved by signalling, and the energy invested in attacks was greater. In contrast, behaviour towards conspecifics did not change. As predicted by theory, conspecifics approached one another more closely and were more consistent in attack intensity yet, contrary to expectations, resolution of contests via signalling was more common among heterospecifics. Phylogenetic relatedness or body size did not predict these outcomes. Our results suggest that competitor recognition processes for heterospecifics became less accurate after mass coral mortality, which we hypothesise is due to altered resource overlaps following dietary shifts. Our work implies that competitor recognition is common among heterospecifics, and disruption of this system could lead to suboptimal decision-making, exacerbating sublethal impacts of food scarcity.

(a) Study system

Taxon. Coral reef butterflyfishes (genus Chaetodon) offer a model field system to examine how changes in resource availability can affect competitor recognition processes for several reasons. First, butterflyfishes show clear signalling (e.g., tail-up while head angled down, erection of the dorsal fin) and attacks during both conspecific and heterospecific encounters [23, 24], and changes in resource availability can lead to changes in aggression both within and among species [7, 25]. Secondly, butterflyfishes use predominantly visual cues for species recognition [26-28]. Thirdly, individuals of many butterflyfish species form stable mated pairs [24], meaning that conspecific aggression is overwhelmingly driven by access to food rather than reproduction. Fourth, the phylogeny of butterflyfishes is well resolved [29], which allows us to test the hypothesis that behavioural responses involved in the competitor recognition process can be predicted by phylogenetic relatedness of species [4, 8].

Field Sites. We observed encounters among butterflyfishes at five regions across the central Indo-Pacific: Iriomote (Japan; 123.7° E, 24.4° N), Christmas Island (Indian Ocean; 105.6° E, 10.4° S), Luzon (the Philippines; 120.8° E, 13.7° N), Aceh (Indonesia; 95.1–95.3° E, 5.4–5.9° S) and Bali (Indonesia; 115.6° E, 8.4° S). We visited reefs up to 12 months either side of a global mass coral bleaching event [7]. We recorded data on 17 reefs in total, comprising of 3 to 4 sampled reefs per region. Reefs were separated by > 1 km non-reef patches (corallivorous butterflyfish territories are generally < 0.2 km2) [30] to sample different populations. For further information on field sites and data collection, see Keith et al [7].

(b) Data collection

We used a well-established protocol to observe butterflyfishes on snorkel or SCUBA depending on depth and visibility [31, 32]. Focal individuals were followed at a distance of 2–4 m for five minutes following an acclimation period (~1 minute) to check that the individual was responding naturally (that is, feeding). Butterflyfish are often found in a mated pair, so to ensure independence of sampled individuals, only one individual of each pair was recorded. To reduce the risk of selecting the same fish as a focal individual twice, we used a U-shaped search pattern and attempted to observe one individual from every pair present on the reef. When a congeneric individual from outside the mated pair came within 1 m of the focal individual, we assumed they were aware of each other’s presence with the potential to interact, and therefore recorded an encounter.

Proximity was recorded as the smallest distance observed between two individuals during an encounter. Encounter outcomes were recorded as passive, where neither individual showed a discernible change in behaviour, or aggressive. Aggressive encounters were further subdivided into those that involved signalling only and those that escalated to a chase [32]. Both forms of aggression are linked strongly to competition over food resources, and as the majority of our observed individuals were in a pair, the possibility that they were engaging in courtship displays with individuals outside of their pair was minimal [32]. Each new observer underwent training by an experienced observer (either J-P.H. or S.A.K.) until recorded data were identical to ensure standardization. Behaviour was unlikely to have been affected by diver presence [33].

(c) Statistical analysis

Data used for analysis were restricted to species pairs with at least five encounters across five different focal individuals. For some analyses, data were restricted further to include only species that were present with these minimum sample sizes in both conspecific and heterospecific encounters. This reduced dataset ensured that any differences were not driven solely by a larger and more variable pool of species in the heterospecific encounters. Note that due to this requirement to use a reduced dataset, statistical models with more parameters (e.g., generalised linear mixed effects models) were not appropriate. The three Philippines reefs did not experience significant coral mortality as a result of bleaching and are therefore not included in the data for after the coral mortality event. All analyses were done in R v.3.6.1. [34].

Prediction 1: Individuals approach heterospecifics less closely than conspecifics. Proximity distances were estimated to the nearest 25 cm in the field (0-24 cm, 25-49 cm, 50-74 cm, 75-100 cm) and converted to dummy variables (1-4) for analysis. To account for non-independence of repeated samples, we calculated the mean of the distance categories from the dummy variables for each individual across its conspecific and heterospecific encounters separately. Use of the mean is appropriate for ordinal data in this case because the numeric difference between each category is equal. We then used a permutation based two-tailed Mann Whitney U test from the coin package [35] with a Monte Carlo-derived approximate distribution to determine whether the mean proximity during conspecific encounters was significantly different from the mean proximity during heterospecific encounters. Data were logged for plotting purposes only.

Prediction 2: Signalling is more common between conspecifics than heterospecifics. We calculated the proportion of aggressive encounters that involved visual signalling for conspecific and heterospecific pairs both before and after bleaching. To deal with non-independence of samples due to repeated measures within individuals (i.e., multiple encounters per individual), we bootstrapped the data 1000 times, each time sampling one encounter only per individual. For each bootstrapped dataset, we tested whether the frequency of signalling (rather than escalation to attack) for conspecific and heterospecific encounters, and before and after coral mortality within those groups, were significantly different from expected using chi-squared permutation tests from the coin package [35], which is robust to small or skewed sample sizes. We calculated the mean chi-squared statistics and p values, and their 95% confidence intervals, across bootstrapped datasets.

Prediction 3: Attack intensity is more variable between heterospecifics than conspecifics. We used the coefficient of variation (CV) to quantify variation in chase distances across heterospecific and conspecific encounters relative to the mean. We use this approach because the mean chase distance is higher for conspecifics, as we would expect from the literature and theory, so a measure relative to the mean is essential. To minimise the influence of rare long chases, we grouped all that were ³10 m. We tested whether there was a significant difference in the CV between conspecific and heterospecific chase distances using the modified signed-likelihood ratio test for equality of CVs (MLSR), which is robust to differences in sample size [36], from the R package cvequality [37].

Prediction 4: Proximity, signalling and variation in attack intensity can be predicted by phylogenetic relatedness and difference in body size. To determine whether approach proximity, signalling proportion and variation in chase distance could be predicted by phylogenetic relatedness, we used branch length between each species pair in the phylogeny [29], calculated with the ape package [38]. We also tested whether body size could predict these behaviours because it can be a cue for individuals to identify competitors. Body size differences were calculated from the species level trait maximum body length downloaded from Fishbase [39], which is appropriate because Chaetodontids achieve 68-92% of their full adult body size in the first year and do not differ perceptibly as adults [40]. Conspecifics were excluded to ensure the result could be interpreted as a nuanced representation and was not overwhelmed by zeros (i.e., for phylogenetic distance). We generated separate regression models for mean proximity, signalling proportion and variation in chase distance as dependent variables (species pairs: n = 107 proximity; n = 24 signalling; n = 16 for chase variation) and checked QQ plots to ensure assumptions were met for proximity and attack intensity. Signalling was modelled using a binomial GLM to account for proportional dependent data, and McFadden’s R2 index was used to assess predictive ability.