Ancestral ecological regime shapes reaction to food limitation in the Least Killifish, Heterandria formosa

Abstract Populations with different densities often show genetically based differences in life histories. The divergent life histories could be driven by several agents of selection, one of which is variation in per‐capita food levels. Its relationship with population density is complex, as it depends on overall food availability, individual metabolic demand, and food‐independent factors potentially affecting density, such as predation intensity. Here, we present a case study of two populations of a small live‐bearing freshwater fish, one characterized by high density, low predation risk, low overall food availability, and presumably low per‐capita food levels, and the other by low density, high predation risk, high overall food availability, and presumably high per‐capita food levels. Using a laboratory experiment, we examined whether fish from these populations respond differently to food limitation, and whether size at birth, a key trait with respect to density variation in this species, is associated with any such differential responses. While at the lower food level growth was slower, body size smaller, maturation delayed, and survival reduced in both populations, these fitness costs were smaller in fish from the high‐density population. At low food, only 15% of high‐density fish died, compared to 75% of low‐density fish. This difference was much smaller at high food (0% vs. 15% mortality). The increased survival of high‐density fish may, at least partly, be due to their larger size at birth. Moreover, being larger at birth enabled fish to mature relatively early even at the lower food level. We demonstrate that sensitivities to food limitation differ between study populations, consistent with selection for a greater ability to tolerate low per‐capita food availability in the high‐density population. While we cannot preclude other agents of selection from operating in these populations simultaneously, our results suggest that variation in per‐capita food levels is one of those agents.

Although the lower food level resulted in slower growth, smaller body size, delayed 23 maturation and reduced survival in both populations, especially survival to maturity showed a 24 highly significant population x food-level interaction. At low food, 75% of fish from the low-25 density population died, compared to only 15% of fish from the high-density population. This 26 difference was much smaller at high food (15% vs. 0% mortality), and was mediated, at least 27 partly, through a larger size at birth of fish from the high-density regime. While we cannot 28 preclude other agents of selection from operating differently in the study populations, we 29 demonstrate that selection at higher density confers a greater ability to cope with low per-30 capita food availability. 31 Introduction 35 Ecologists have long been interested in whether animal populations experiencing different 36 regimes of population density will evolve adaptive differences in response to these conditions 37 (Berec et  One of the challenges in understanding the incidence and importance of density-dependent 44 evolution is identifying the agents of selection behind adaptation to different population 45 densities. The contrast between lower and higher density can include a host of ecological 46 differences, depending upon the specific case. Higher density can mean, among other factors, 47 lower per-capita food levels, a higher rate of stressful social interactions, higher rates of 48 pathogen transmission, greater competition for mates, and, in some systems, higher rates of 49 accumulation of waste products (Berec et al. 2018;Than et al. 2020). Laboratory studies with 50 Drosophila have shown that nearly all of these agents can be acting (Joshi et al. 1996 Understanding how selection acts differently in different density regimes is not a mere 54 exercise in detail. The responses to density-dependent selection and their consequences for 55 subsequent population dynamics depend on the demographic details of selection, for example 56 regarding the age classes primarily affected (Charlesworth 1994; Engen and Saether 2016;57 Trexler 1986; Warner et al. 1991), with the result that the mere observation of consistent 83 differences in density need not imply consistent differences in per-capita resource 84 availability. Only experiments can assess whether adaptation to lower per-capita food levels 85 is an outcome of natural selection operating at high population densities. 86 The Least Killifish, Heterandria formosa, offers an excellent opportunity to investigate this 87 issue. This fish is found throughout the lower coastal plain of the southeastern United States. 88 Populations in north Florida occur in a wide variety of habitats, from freshwater springs to 89 lakes, ponds, and swamps in river bottoms (MacRae and Travis 2014). In all habitats, H. 90 formosa occupies the shallow littoral zone and is a primary consumer (Aresco et al. 2015). 91 Adults are small, with male standard length (SL, the distance from the tip of the snout to the 92 hypural plate in the tail) varying, typically, from 9 mm to 15 mm and female SL varying from 93 10 to 25 mm. Reproduction is placental; females provide almost all nourishment for their 94 embryos after fertilisation. 95 Prior work has shown that population density is dramatically higher and per-capita predation 96 risk lower at Wacissa River (WR) than Trout Pond (TP; Leips and Travis 1999; MacRae and 97 Travis 2014; Richardson et al. 2006). Size at maturity is smaller, fecundity is higher, and 98 average offspring size lower in TP, distinctions based in genetic differences (Leips et al. 99 2000). The different life histories appear to be adaptations to the different regimes of density 100 and predation risk that each population experiences, with higher predation and lower density 101 favouring earlier maturation, greater fecundity and smaller individual offspring. While there 102 are many abiotic differences between the two locations, there is no evidence that the two 103 populations differ in their responses to them (Hale and Travis  Mescocosm experiments implicate one or more factors associated with the two populations' 107 characteristic differences in population density as major agents of selection. Experiments 108 manipulating population density (Leips et al. 2009;Leips et al. 2013) and genetic 109 composition (Leips et al. 2000) showed that the two populations differed in their response to 110 density, with the reproductive traits of individuals from the Wacissa River, the low-111 predation/high-density population, being less sensitive to the depressant effects of density and 112 the degree of sensitivity being proportional to the initial dosage of Wacissa River alleles 113 (Leips et al. 2000). 114 It is unclear if the reduced sensitivity to higher density in the WR fish reflects an adaptation 115 to lower per-capita food levels. Trout Pond is the more productive habitat (Aresco and Travis,116 unpublished data), suggesting that the higher density of WR fish, which is a function of lower 117 predation risk (MacRae and Travis 2014; Richardson et al. 2006), does create lower per-118 capita food levels. However, laboratory experiments indicate that social stress is at least as 119 strong an effect on female reproduction rate as reduced food levels (Leatherbury and Travis 120 2019). In this paper we report the results of testing whether adaptation to different per-capita 121 food levels plays a role in how these two populations have adapted to different density 122 regimes. 123

125
Experimental design 126 We collected adults from Trout Pond, Wakulla County, Florida, and from Wacissa River, 127 Jefferson County, Florida, in September 1993 and used them to establish breeding colonies in 128 the laboratory. We housed the adults from each population in two 76-litre aquaria per 129 population and collected F1 offspring as they were born. Those offspring were raised in 38-130 litre aquaria until maturity, when we removed females and mated them to males from 131 different aquaria to minimise inbreeding. We isolated pairs in small aquaria and collected 132 their offspring at birth for the experiment. We began the experiment in February 1994. 133 The experiment was a factorial design in which we raised offspring to maturity from either 134 population (TP or WR) at one of two food levels, designated "high" or "low." We set the 135 initial daily "high" food ration as 4 mg and "low" as 1 mg. We increased the food ration as 136 individuals grew to a maximum by day 63 of 20 mg/day at "high" and 5 mg/day at "low." 137 "Low" food was always set as one-quarter of "high." We considered a male to be mature 138 when his gonopodium (modified anal fin used as an intromittent organ) was fully formed; we 139 considered a female to be mature when the characteristic black spot appeared on her anal fin. 140 Sample sizes are provided in Table 1. 141

142
We measured each individual fish at birth, at 14 days, 28 days, 42 days, and at sexual 143 maturity. We used standard length (SL) as a measure of body size. To measure each 144 individual's SL, we removed the fish from its aquarium, placed it in a small petri dish against 145 a standard grid, photographed it, and returned it to its aquarium. No individual died as a result 146 of this procedure. Standard length and dry mass were strongly correlated (Pearson's product-moment correlation: r = 0.76, t 18 = 5.01, p < 0.0001, n = 20 fry measured at birth). 148 Photographic length measurements were highly repeatable (repeatability r = 0.99, effect of 149 individual identity in a simple ANOVA: F 36,37 = 381.3, p < 0.0001, n = 37 individuals that 150 were measured twice). 151

152
Size at birth was approximately normally distributed, as judged from a quantile-quantile plot 153 (QQ-plot) and a two-sided Kolmogorov-Smirnov test (D = 0.10, p = 0.42). We therefore 154 analysed it using a generalised linear mixed model with Gaussian errors. As predictors we 155 included the population of origin (Trout Pond vs. Wacissa River), the experimental food level 156 (high vs. low), maternal size (continuous) and the interaction between population and 157 maternal size. The sole random effect was maternal identity. It is not possible to determine 158 the sex of a juvenile from external morphology and so the sex of fish that died before 159 attaining sexual maturity remains unknown. Given the differential survival of fish from TP 160 and WR, and of fish assigned to the high-and low-food treatment, sex as a categorical 161 predictor with three levels (female, male, unknown) was confounded with both population 162 identity and food level. We therefore tested for a size difference between males and females 163 in a separate model using only the individuals of known sex. All remaining predictors in this 164 model were as before. As the QQ-plot identified four data points as potential outliers (size at 165 birth > 8.0 mm), we fitted additional models after excluding these data points. 166 Survival to sexual maturity was analysed using a generalised linear mixed model with 167 binomial errors. As fixed effects we included the population of origin, the experimental food 168 level, size at birth (continuous), and maternal size (continuous), and as random effects, 169 maternal identity. We were unable to include the interaction between population identity and 170 food level, because the main effects of population (92.5% surviving fish from WR vs. 55.5% from TP) and of the food treatment (92.5% surviving fish at high food vs. 55.5% at low food) 172 were exactly the same, causing the model to suffer from computational problems. We 173 therefore used a Chi-square test, which proved robust to the mirror-image effects of 174 population and food level, to compare the number of dead and surviving fish among the four 175 experimental groups. To identify the cells in the contingency table that accounted for most of 176 the difference between expected and observed values, we computed the relative contribution 177 of each cell to the total Chi-square score as 100 * 2 3 2 , where r was the Pearson residual and 178 3 2 the Chi-square score. 179 To analyse size during adolescence and at sexual maturity, we performed two analyses. The 180 first one was a generalised linear mixed model with size at 14 days as the dependent variable, 181 which included all individuals that survived to and were measured at that age (n = 66). Size at 182 14 days was approximately normally distributed (QQ-plot; D = 0.14, p = 0.17), so models 183 with Gaussian errors were fitted. For predictors, we used the population identity, food level, 184 their interaction, maternal size, and size at birth, plus maternal identity as a random effect. 185 Originally, we included interactions between food level and size at birth, between population 186 and size at birth, and between population and maternal size; we dropped these effects from 187 the final model because they had no impact on the response (elimination criterion: ratio of 2 188 over its degrees of freedom < 1 in log-likelihood ratio tests comparing a model with and 189 without the interaction in question). We could not include sex as a predictor due to a small 190 but non-negligible number of fish with unknown sex (n = 8). Again, the five fish that were 191 largest at 14 days (> 9 mm) appeared to be outliers, based on a QQ-plot. To explore their 192 influence on results, we also fitted a model without them. 193 We additionally analysed the three juvenile sizes and size at maturity with a repeated 194 measures analysis of variance, using function "aov" in R (R Core Team 2020). We included only individuals that survived to maturity and were measured on all four occasions (n = 55). 196 All these individuals were of known sex. The predictors were population, food level, age 197 category, sex, size at birth, maternal size, all the two-way interactions between age and these 198 predictors (except the age x maternal size interaction), and maternal identity. Our formula 199 contained individual identity as the single error term, to account for repeated measures of 200 fish. The error term specified two error strata, with appropriate models fitted within each 201 stratum. We tested the effects of age and all its interactions within individuals, while we 202 tested all the other effects between individuals. The low number of survivors that were from 203 Trout Pond, kept at the lower food level, and had no missing data (n = 3) meant that we were 204 unable to fit the interaction between food level and population. We also initially fitted, and 205 then dropped from the final model due to non-significance, the interactions between age and 206 maternal size, between population and size at birth, between food level and sex, and between 207 population and sex (elimination criterion: F-value < 1). 208 The age at sexual maturity was approximately normally distributed (QQ-plot; D = 0.11, p = 209 0.52), and so we analysed this variable with generalised linear mixed models with Gaussian 210 errors. Our final model included the population, food level, size at birth, sex, and the 211 interaction between food level and size at birth as fixed effects, and maternal identity as a 212 random effect. We did not fit more predictors to ensure sufficient statistical power given the 213 reduced sample size in this analysis (n = 57 fish with known age at maturity). We therefore 214 did not include non-experimental (i.e. purely correlative) predictors in our model that, upon 215 visual inspection, appeared unrelated to the age at sexual maturity (e.g. maternal size, size at 216 maturity, and the interaction between size at birth and population). In addition, we could not 217 fit the interactions between population and food level, population and sex, and food level and 218 sex, because in each of these cases one or several treatment levels had insufficient sample 219 sizes (i.e. < 10 data points). Three data points (age at maturity < 36 days and > 85 days) showed slight deviations from normality on the QQ-plot, and we fitted an additional model 221 after excluding them. 222 We assessed the fit of all models that converged with diagnostic plots. The reference levels 223 for categorical predictors were Trout Pond, high food, female. In models that included 224 maternal identity as a random effect, we obtained a p-value for maternal identity with log-225 likelihood ratio tests comparing the full model to one without random effects. 226 We performed all statistical analyses in R 4.0.0 (R Core Team 2020). Unless stated 227 differently, we fitted generalised linear mixed models using function "glmmTMB" in R-228 package "glmmTMB" (Brooks et al. 2017). Values are given as mean ± SD. Binomial 229 standard errors for survival rates were computed according to Zar (1996) where p was the proportion of fish that survived and n the total number of fish. Scatter plots 231 were prepared using R-package "beeswarm" (Eklund 2016) and depict individual data points 232 superimposed on boxplots. White triangles on boxplots show group means. 233

235
Size at birth 236 Fish from Wacissa River were about 15% larger at birth than fish from Trout Pond (b = 4.04, 237 z = 2.25, p = 0.0245, Fig. 1a). They were also slightly larger when assigned to the high-238 rather than the low-food treatment, although this effect is coincidental, as fish were measured 239 before they were fed for the first time (b = -0.36, z = -3.08, p = 0.0021, Fig. 1b). There was 240 no overall effect of maternal size on offspring size at birth (b = 0.09, z = 1.44, p = 0.15), but 241 in fish from WR larger offspring tended to be born to smaller mothers (interaction between 242 maternal size and population: b = -0.14, z = -1.78, p = 0.07, Fig. 1c). Maternal identity, 243 included as a random effect, explained a significant amount of variation in offspring size at 244 birth ( 1 2 = 9.77, p = 0.0018). We tested for a size difference between the sexes using only the 245 subset of fish with known sex, and found that males and females did not differ in size at birth 246 (b = -0.25, z = -1.62, p = 0.11, Fig. 1d). The inclusion of sex as a predictor left the remaining 247 model results largely unchanged (data not shown). 248 Some of these results were contingent on four exceptionally large fry (> 8.0 mm), which were 249 identified as potential outliers in a QQ-plot. When we excluded these fry, the population of 250 origin still had a strong effect (b = 2.94, z = 2.23, p = 0.0258, Fig. 1a), but the coincidental 251 effect of food level disappeared (b = -0.17, z = -1.53, p = 0.13, Fig. 1b). Moreover, the main 252 effect of maternal size became significant, indicating that larger mothers were likely to 253 produce larger offspring (b = 0.10, z = 2.12, p = 0.0340, Fig. 1c). This relationship was 254 evident in fish from TP; in fish from WR, there was still a slight tendency for larger mothers 255 to have smaller offspring (interaction between maternal size and population: b = -0.10, z = -256 1.73, p = 0.08). When using the outlier-free dataset, maternal identity did not predict size at birth ( 1 2 = 2.36, p = 0.12). The size difference between males and females remained non-258 significant (b = -0.02, z = -0.16, p = 0.88, Fig. 1d). 259

260
Twenty-six percent of fish died before reaching sexual maturity. The mortality rate was 261 highest in the two weeks after birth (16.3%), intermediate between 14 and 28 days of age 262 (7.5%), and lowest between 28 days and the age when fish attained maturity (2.5%). Survival 263 to maturity was significantly higher among fish from WR (92.5%) than among fish from TP 264 (55.0%, b = 2.95, z = 2.91, p = 0.0037, Fig. 2a). An effect of the same exact size was caused 265 by the food treatment, with 92.5% surviving fish under high-and only 55.5% under low-food 266 conditions (b = -2.94, z = -3.57, p = 0.0004, Fig. 2a). The association between size at birth 267 and survival was non-significant (b = 0.17, z = 0.27, p = 0.79). However, once population 268 identity was removed from the list of predictors, a larger size at birth was associated with 269 increased survival (b =1.22, z =2.55, p = 0.0107, Fig. 2b), showing that the higher survival 270 rate of fish from WR was, at least in part, mediated through a larger size at birth. In addition, 271 survival tended to be slightly higher in fish born to larger mothers (b = 0.28, z = 1.67, p = 272 0.10, Fig. 2c). Maternal identity did not explain any variation in offspring survival ( 1 2 = 273 0.00, p = 1.00). 274 Fish from TP were more sensitive to food limitation than fish from WR: at the lower food 275 level, 75% of fish from TP died, compared to only 15% of fish from WR (Fig. 2a). Mortality 276 rates were much lower at the higher food level and not so different between populations: 15% 277 in TP and 0% in WR. This population identity x food level interaction was evident in a Chi-278 square test ( 3 2 = 34.29, p < 0.0001), of which the residuals showed an excess of dead 279 (Pearson residual r = 4.26) and concomitant lack of surviving fish (r = -2.54) that were from 280 TP and kept at the lower food level, and a lack of dead fish that were from WR and kept at the higher food level (r = -2.29). Accordingly, the cells in the contingency table that  282 contributed the most to the total Chi-square score were dead fish / TP / low-food (52.8%), 283 surviving fish / TP / high-food (18.8%), and dead fish / WR / high-food (15.3%). Together, 284 these cells accounted for 86.9% of the observed difference between expected and observed 285 frequencies. 286

287
Fish from WR were about 15% larger than fish from TP at 14 days of age (b = 0.41, z = 2.13, 288 p = 0.0330), and fish on the high food ration were about 14% larger at this age than fish on 289 the low food ration (b = -0.43, z = -2.22, p = 0.0264, Fig. 3a). There was no interaction 290 between population and food level (b = -0.30, z = -1.26, p = 0.21, Fig. 3a). At that young age, 291 an individual's size was still strongly correlated with its size at birth (b = 1.00, z = 10.75, p < 292 0.0001, Fig. 3b), and tended to be positively associated with maternal size (b = 0.05, z = 1.78, 293 p = 0.08). The effect of maternal identity was non-significant ( 1 2 = 1.08, p = 0.30). Sex could 294 not be included in the model, but a visual inspection of the data tentatively suggests that 295 males and females did not differ in size at age 14 days, while fish that died as juveniles may 296 potentially have been relatively small (Fig. 3c). 297 When excluding the five largest fish (> 9 mm), identified as potential outliers, the influence A joint analysis of all juvenile sizes and size at sexual maturity (model results in Table 2) 304 showed that fish from TP were always smaller (significant main effect of population), even at maturity, yet did not grow more slowly than fish from WR (non-significant age x population 306 interaction, Fig. 4a), suggesting that population differences in juvenile size stemmed from the 307 differential size at birth. Fish did grow more slowly under low-food conditions (significant 308 main effect of food, and age x food interaction, Fig. 4b). The overall effect of sex was non-309 significant, owing to the fact that males and females did not visibly differ in size at 14, 28, 310 and 42 days of age, but once they reached maturity males were considerably larger than 311 females (significant age x sex interaction, Fig. 4c). Juvenile size also showed a positive 312 statistical association with size at birth, which was most pronounced at 14 days and gradually 313 disappeared at older ages, until no effect was left at sexual maturity (significant main effect 314 of size at birth, and age x size-at-birth interaction, Fig. 4d). Neither maternal size nor 315 maternal identity had a bearing on somatic growth (Table 2). 316

317
Fish from the two populations did not differ from each other in age at maturity (b = -2.77, z = 318 -0.96, p = 0.34, Fig. 5a), but attained sexual maturity about 29% later when fed a low-319 quantity diet (b = 73.78, z = 3.10, p = 0.0019, Fig. 5b). The interaction between population 320 and food level could not be analysed statistically due to the poor survival of fish from TP 321 under low-food conditions (n = 4). On average, males matured substantially later than 322 females (64 ± 13 days vs. 48 ± 13 days: b = 19.83, z = 8.21, p < 0.0001, Fig. 5c). Although 323 there was no main effect of size at birth (b = -2.33, z = -1.34, p = 0.18), its interaction with 324 the feeding regime was significant (b = -8.56, z = -2.43, p = 0.0153, Fig. 5d), indicating that 325 fish kept at the lower food level reached sexual maturity earlier when they were born 326 relatively large. Maternal identity did not explain any variation ( 1 2 = 0.00, p = 1.00). The 327 exclusion of two females that matured very early (29 and 32 days, respectively) and of a male 328 that matured very late (90 days), which were identified as potentially influential data points in a QQ-plot, did not change these results (data not shown). The limited sample size (57 fish 330 survived to sexual maturity) prevented us from fitting additional predictors, but a visual 331 inspection of the data suggested that age at maturity was independent of maternal size (r = -332 0.04, t 55 = -0.29, p = 0.77; Fig. 5e) and size at maturity (r = 0.18, t 55 = 1.39, p = 0.17; Fig. 5f). 333

334
Our results suggest that the divergent life histories of Least Killifish living in habitats with 335 contrasting predation regimes and population densities result, in part, from local adaptation to 336 different per-capita food levels. We reached this conclusion from a full-factorial laboratory 337 experiment in which fish from two populations were raised to sexual maturity at one of two 338 food levels. We found that fish from Wacissa River, characterised by low-predation/high-339 density conditions, were larger throughout the experiment and had higher survival to 340 maturity, especially, but not only, at the lower food level, than fish from Trout Pond, a high-341 predation/low-density population. The increased survival of WR fish was, at least partly, 342 mediated through a larger size at birth. Of the traits measured in our study, only the age at 343 maturity showed no difference between populations. Experimentally reduced per-capita food 344 availability led to smaller body sizes, slower growth rates, delayed maturation, and decreased 345 survival. Size at birth proved a key success factor also in relation to food: a larger size at birth 346 enabled fish to mature relatively early even when they were kept at the lower food level. 347 Lastly, we found that size at birth and juvenile growth were identical in males and females, 348 but that females matured earlier and hence smaller. 2008). In this experiment, age at maturity did not differ, on average, between fish from TP 354 and those from WR, a result that has been found before (Hale and Travis 2015). Also the 355 smaller size at maturity of fish from TP is consistent with earlier studies (Hale and Travis  356 2015; Leips et al. 2000), as is the older age at maturation and thus larger size of males compared to females (Hale and Travis 2015). Field data using otolith ring counts as 358 indicators of age also suggest that females mature before males (J. Travis, unpublished data). 359 Maternal effects, manifested through variation among individual females in the growth and 360 development of their offspring, diminished gradually during adolescence and were no longer 361 detectable by the time that fish matured. This is similar to patterns in other studies of fish 362 species in which maternal effects were important in early life but became less so with 363 offspring age (Heath and Blouw 1998;Venney et al. 2020). This was in contrast to the 364 longer-lasting effects of the differential sizes at birth between females from the two 365 populations. In the comparison between the populations, the smaller initial size of TP 366 offspring carried through to a smaller size at maturity, even though the actual somatic growth 367 rates of immature fish did not differ between the populations. This result illustrates the 368 importance of distinguishing variation in maternal effects within populations from those that 369 might be observed between populations. 370 The maternal effect within populations, manifested through size at birth, was important in a 371 very particular manner. At low food levels, individuals that were larger at birth matured 372 earlier. This was a substantial effect: an individual that was smaller than 6.5 mm at birth 373 matured ~16 days later than one that was larger than 7.0 mm at birth (see Fig. 5d). This 374 relationship was weaker at high food levels (difference of ~10 days). This result reinforces 375 the conclusion drawn by Leips et al. (2013), who found that larger size at birth was inversely 376 correlated with age at maturity, especially when fish were raised in competitive conditions. 377 The pattern suggests that the larger size at birth in WR fish is an adaptation to low per-capita 378 food levels produced through some combination of lower primary productivity and higher 379 population density. 380 The most direct evidence that density-dependent selection has been driven at least in part by 381 different levels of per-capita food availability was the stark difference in survival between TP 382 and WR fish at the lower food level. There was a 75% mortality rate for TP fish at low food, 383 compared with a 15% mortality rate for WR fish. At the higher food level, mortality rates 384 were more similar between populations, 15% for TP fish and 0% for WR fish. This 385 interpretation, that WR fish are adaptively more suited to lower food levels, is bolstered by 386 the contributions of the individual cells to the Chi-square score, with the number dead in TP 387 at low food being the major component. Survival was positively associated with size at birth, 388 suggesting that the reduced survival of TP fish resulted at least partially from their smaller 389 average size at birth. From previous work we know that fish from these populations show a 390 trade-off in offspring size and offspring number, with TP females having more, but smaller, 391 offspring than WR females (Leips et al. 2000). Taken together, this suggests that, in TP fish, 392 the production of many small offspring, rather than fewer large ones, pays off because the 393 higher primary productivity of their habitat allows also small fish to survive, while the larger 394 number of offspring ensures that at least some offspring survive to maturity despite the high 395 predation risk. 396 These results do not preclude other agents of selection from operating differently in the 397 contrasting density conditions in WR and TP. Social stress is important in TP fish 398 (Leatherbury and Travis 2019) but we do not know if WR fish are less sensitive to this source 399 of stress. We also have no information on pathogen incidence and virulence in each 400 population. However, demonstrating that selection at higher density confers a greater ability 401 to cope with low food levels is a step forward in moving the study of density-dependent 402 selection toward a more biological and less phenomenological focus (cf. Engen et al. 2020).   Survival was reduced at the lower food level, particularly in fish from Trout Pond, and 560 positively associated with size at birth and potentially maternal size. In (a), sample size is 561 20 fish for each vertical bar, and error bars are binomial standard errors (± 1 SE). Note that 562 survival was 100% for fish from Wacissa River kept at high food, and the binomial standard 563 error consequently zero. 564

Figure 3
565 Size at 14 days was reduced at the lower food level and in fish from Trout Pond, and 566 was highly correlated to size at birth. The relationship between size at 14 days and sex (c) 567 is shown for illustrative purposes only, as the low number of fish that did not survive to 568 sexual maturity and hence could not be sexed (n = 8) precluded its inclusion in the statistical 569 model. 570  During adolescence and at maturity, body size was lower for fish from Trout Pond, for 572 fish fed a low-quantity died, and, at maturity, for females. The effect of size at birth (d) gradually disappeared as fish grew older. This analysis only included fish that reached sexual 574 maturity (n = 55). 575

Figure 5
576 The age at sexual maturity was delayed at the lower food level, especially in fish that 577 were small at birth, and was earlier in females than in males. There was no effect of 578 population of origin. The relationships of age at maturity with maternal size (e) and with size 579 at maturity (f) are shown for illustrative purposes only, and were not included in the statistical 580 model. 581  The analysis includes standard lengths measured at 14 days, 28 days, 42 days, and at sexual 606 maturity. Only individuals that survived to maturity and were measured on all four occasions 607 were included (n = 55). Standard length was measured in mm. Italics: p < 0.05. Df: degrees 608 of freedom, Sum Sq: sum of squares, Mean Sq: mean squares, SD: standard deviation.