#### Study species

The Serbian spruce, *Picea omorika* (Pančić) Purkyne, is an endemic conifer tree and Tertiary relict of the European flora. Range of distribution of this species is rather narrow (approx. 10.000 km^{2}), occupying exclusively the middle and upper courses around the river Drina. The extant *P. omorika* populations are scarce and not abundant. In total, the species consists of about 10 natural populations, varying in size from large to small in Bosnia, and about 20, mostly smaller populations, in Serbia (Fukarek, 1951).

For this study we selected six natural *P. omorika* populations (see Table 1 for localities), each consisting of more than 30 individuals. These populations encompassed almost the full range of the species distribution. To avoid including individuals with common lineage, seed cones were sampled separately from at least 20 randomly chosen, non-adjacent trees within each population. Because *P. omorika* is a wind-pollinated outcrossing plant, it is likely that the ovules produced by an individual tree are fertilized by pollen from many different paternal plants. If so, a sample of the seed collected from a particular naturally pollinated tree will represent a maternal half-sib family.

Table 1. Location and sample size (no. of maternal families) of six natural *Picea omorika* populations included in the study. #### Experimental set-up

Seeds from all six populations were pooled into a synthetic population comprising 117 maternal families in order to enlarge the range of phenotypic variation available for selection. Prior to planting, 50 seeds from each family were spread over the wet filter paper (soaked with 2% fungicide Venturin–S 50; Župa, Kruševac, YU) in Petri dishes and kept in the dark (2 °C, 30 days) in order to synchronize germination. The incubated seeds were then exposed to room temperature where the germination began after 2–3 days. Sixteen offspring of each maternal family were planted into 300-mL pots in two densities: a single individual plant per pot (10 replicates) and three individuals per pot (two replicates). Pots were filled with a mixture of humus, peat and sand (1:1:1). The potted seedlings were transferred to a growth room where each pot was placed at a randomly chosen position on a shelf. The distance between the centres of neighbouring pots was approximately 9 cm. The ambient temperature in the growth-room was kept at 21/16 °C (light/dark) with a 16-h photoperiod. The light was provided by a set of four Philips TLD 36-W/33 fluorescent tubes. At the onset of the experiment, average photosynthetically active radiation (PAR) above the growing plants amounted 110 μmol m^{–2} s^{–1}, while the R:FR ratio (quantum flux ratio between 665 and 735 nm) was 8.20, i.e. far above the normal, sunlight, ratio of 1.0. The plants were regularly bottom-watered and fertilized every 10 days with a 0.5% water-soluble fertilizer (Floravit, N : P : K 12 : 6 : 6). To minimize the position effects (genotype–environment correlation) the pots were rotated every 2 days.

On the 240th day after the germination, plants were harvested individually. Eleven seedling traits were recorded on each seedling. Table 2 gives the acronyms and a short description for each trait measured. Based upon the functional roles that distinct traits play within a plant organism (Berg, 1960; Armbruster *et al*., 1999), the recorded seedling traits of *P. omorika* were placed into three functional groups: epicotyl traits (involved in resources aquisition), bud traits (associated with resource allocation) and performance traits (fitness components). Except for total plant dry weight, all the readings were taken non-destructively on each individual seedling. To estimate total plant biomass, harvested seedlings were weighed after being oven-dried (72 h, 70 °C). Total plant dry weight was used as a measure of performance.

Table 2. Morphological traits measured. #### Statistical analyses

The trait means based on all data points in the sample were calculated separately for each density treatment using a GLM procedure in SAS (SAS, 1989). Significance of the differences in trait means between densities was tested by a standard Student’s *t*-test. Following Schlichting (1986), plasticity was measured as the absolute value of the difference in family mean phenotype between density environments.

Within each density, phenotypic correlations among trait pairs were calculated as the Pearson correlations of the individual data points. Genetic correlations within treatments were estimated as the Pearson product-moment correlations of the family means, *r*_{m} (Via, 1984):

where COV_{m(}_{xy}_{)} is the covariance among the family means of the traits *x* and *y*, and *V*_{m(}_{x}_{)(}_{y}_{)} are the variances among the family means of traits *x* and *y*. The *r*_{m} correlations are only an approximation of the additive genetic correlations. A fraction of the within-family (special environmental) (co)variance component is included in the (co)variance among family means:

where *n* is the number of siblings per family. When maternal half-sib families (MHS) are used, the family mean correlations can be inflated by common maternal effects included into the phenotypic covariance between maternal sibships:

where *V*_{A} represents the additive genetic variance, *V*_{Gm} is the variance due to genetic maternal effects and *V*_{Ec} is the variance due to environmental maternal effects (Lynch & Walsh, 1998). The maternal effects are often difficult to isolate and, moreover, these effects are not decreased by increasing the number of individuals (*n*) per family. One advantage of using the family mean approach is that *r*_{m} is a true product-moment correlation, which, unlike the variance component correlation, cannot exceed ±1.0; and its significance is tested using standard tables of critical values for the sample correlation coefficients. Analogously to the genetic correlations within densities, plasticity correlations were calculated as the Pearson correlations between the trait plasticity values (Schlichting, 1986). Confidence intervals and standard errors of all correlations based on bootstrap resampling were calculated to test the hypothesis that correlations were significantly different from 1 or – 1.

Variance component correlations across density environments were estimated using the assumptions of the SAS mixed-model factorial ANOVA (Fry, 1992; Fry *et al*., 1996). In this approach, the genetic correlation between a trait expressed in two density environments can be written as

where *V*_{F,hl} is the variance component due to the random family main effect computed as the covariance of family means across levels of the fixed factor – density environment, and *V*_{F,h} and *V*_{F,l,} are the variance components due to family effects obtained from separate ANOVAs within high- and low-density treatments (REML option of VARCOMP procedure), respectively. To test whether the correlation *r*_{g} differs significantly from unity, the data from each density treatment were first standardized such that the variance components explained by the family effect were equal to one, and then were subjected to a mixed-model factorial ANOVA performed on the entire dataset. A significant family–density interaction in the ANOVA on the transformed data provides evidence that the corresponding across-environment genetic correlation is less than one.

To estimate the intensity of phenotypic selection acting on seedling traits, standardized selection gradients (Lande & Arnold, 1983) were computed by regressing relative fitness on all traits simultaneously, after standardizing each one to a mean of 0 and a variance of 1 (*z*-transformation; Sokal & Rohlf, 1981). Absolute fitness was transformed to a relative one by dividing each absolute value by the average absolute fitness of all plants in the synthetic population. Linear and quadratic terms in the regression were used to test for directional and stabilizing/disruptive selection, respectively.

The standardized directional selection gradients were calculated by a multiple regression analysis of relative fitness on the standardized trait values, while the stabilizing/disruptive selection gradients were obtained from a multiple regression analysis of the relative fitness on the standardized traits and their squares (PROC GLM in SAS). Selection gradients estimated in a multiple regression of relative fitness on the standardized trait values are, in fact, the coefficients of a partial regression which quantify (in units of phenotypic standard deviation) the effect of each trait on relative fitness, holding other traits fixed (Lande & Arnold, 1983).

To estimate the cost of plasticity, we implemented a multiple regression analysis as suggested by Scheiner & Berrigan (1998). This method is based on the following statistical model:

where *W* is the absolute fitness for an individual in one of the environments, *X* represents the trait value in that environment and *plX* is the plasticity of the trait. Significant regression coefficients β_{1} and β_{2} measure direct selection on the trait value and account for the linear and non-linear component of selection, respectively. The regression coefficient β_{3} describes how the ability to be plastic affects fitness, once the direct effects are taken into account. A cost of plasticity appears as a significant negative regression coefficient for the *plX* term. This term encompasses both maintenance and production costs. Regression coefficients for the interaction terms, β_{4} and β_{5}, measure additional production costs. A significant positive regression coefficient for the interaction term (*X* * *plX* or *X*^{2} * *plX*) would indicate that production costs are greater for more plastic genotypes. This regression is calculated for each density environment separately.

A sequential Bonferroni procedure was also applied to all *P*-values in order to correct for multiple comparisons (Rice, 1989).