The combined effect of color and odor on flower choice behavior of bumble bees in flower mimicry systems

Abstract

Food-deceptive flowers are pollinated by animals that expect a reward but are cheated. Such plants profit from their similarity to rewarding plants and should develop signals that hinder discrimination. We use artificial rewarding model flowers and nonrewarding mimicking flowers that present similar visual cues. We test how additional scent cues change flower choice of the mimic by bumble bees (Bombus terrestris) in two situations: (1) both flower types are simultaneously present and can be compared by the pollinator, and (2) both flower types are encountered successively in the absence of each other. We find that in situation 1, discrimination learning is greater if scents are used as cues for identifying the mimic, whether the mimic has a different scent or if it is scentless while the model is scented. In situation 2, a generalization task, a scented mimic is avoided faster than a scentless one. Discrimination of the mimic is poorest if it has the same scent as the model, thus demonstrating a potential for scent mimicry, which has not yet been proved to exist among differently rewarding flowers. Thus, the best strategy for a mimic would be to have the same scent as the model, but this strategy may not be used due to evolutionary constraints. Alternatively, if there are several potential models, then having no scent would be a better strategy than mimicking just one of the models. In situation 1 flower discrimination by color cues is enhanced in the mere presence of scent, compared to unscented controls, even if the scent does not provide a distinguishable cue itself. The results indicate that the presence of scent may enhance color discrimination by improving attention towards visual cues and/or that combined color/odor cues may lead to better memory formation and retrieval.

Floral signals consist in most cases of visual and olfactory cues, forming a multimodal combined stimulus. Learning these stimuli is fundamental for the observed choice behavior of flower visitors. However, although little is known about the rules that govern memory formation and retrieval of combined stimuli, one may speculate that flowers are adapted to exploit those rules with their signals. In this study we explicitly ask questions that arose from the observation of flower pollination syndromes concerning signal learning and memory in hymenopteran flower visitors. We do this by experimentally testing the effect of odor cues on color discrimination and generalization in bumble bees. We derive these questions directly from the study of flower mimicry syndromes under natural conditions.

Many flowers emit bouquets of volatiles that humans perceive as characteristically “flower-like.” However, these constitute hundreds of different compounds which give most floral scents a unique composition (Dobson, 1994). Scents allow bees to discriminate flowers from nonfood objects and they provide cues for discrimination among different species. Scent production is therefore adaptive for a flower that is designed to achieve a directed pollen transfer to conspecifics, because it should be discriminated from all other flower species and should allow flower-constant behavior by the pollinator (Free, 1970; Waser, 1986). However, food-deceptive flowers differ in this respect. They are pollinated by animals that expect a reward at the flowers but are deceived. Food-deceptive plants utilize the pollinators of rewarding plants flowering nearby, and are assumed to have increased reproductive success if they cannot be discriminated from them (Roy and Widmer, 1999).

Food deception is particularly common among orchids (Ackerman, 1986; Dafni, 1984; Roy and Widmer, 1999). Specific cases of mimicry have been reported mostly with respect to visual cues, in which pollinators were observed to forage on single rewarding species and appear to mistake the orchid for their food plant (Orchis israelitica; Dafni and Ivri, 1981a; Orchis pallens; Vöth, 1982). Such a mimetic relationship was also inferred from the similarity in floral color signals between the food-deceptive orchid Cephalantera rubra and the rewarding species of Campanula (Campanulaceae) with which it commonly co-flowers (Nilsson, 1983a). There are also cases of less specific mimicry in which an orchid, Orchis boryi, mimics a variety of different rewarding plants (Gumbert and Kunze, 2001). Although the similarity to any of the rewarding models was very high, the orchid's visitation rate correlated with perceptual similarity of color, indicating that bees broadly generalized from the floral signals of their food plants to the mimic (Gumbert and Kunze, 2001).

So far, no cases of scent mimicry have been reported for food-deceptive species although it occurs in cases of sexual mimicry (Borg-Karlsson, 1990; Dafni, 1984; Dettner and Liepert, 1994). Chemical analysis of floral scents revealed that several orchids produce volatile compounds (Kaiser, 1993; Nilsson, 1980, 1983a, b, 1984), but there are many food-deceptive orchids that give off scents not perceptible to humans and only negligible amounts of floral volatiles (Kunze J, unpublished data). Rewarding orchids, on the other hand, often produce strong scents (e.g., Orchis coriophora, Platanthera spp. Hiemantoglossum hircinum; Kaiser, 1993).

The imitation of a model plant is never perfect in floral mimicry, which means that the pollinators do indeed after visits to the nonrewarding flowers notice that they have been cheated and subsequently learn to discriminate and avoid the mimic (Heinrich, 1975; Little, 1983; Nilsson, 1980, 1983b). This behavior was also observed in controlled experiments with artificial rewarding and nonrewarding flowers, where negative frequency-dependent selection towards the nonrewarding flower types was shown (Smithson and Macnair, 1996, 1997). If the pollination success of food-deceptive flowers depends on their pollinators' learning behavior, selection should favor floral characters that disrupt the learning process. Accordingly, the hypothesis was raised that the absence of a floral scent may be an adaptive characteristic for these plants because it may cause greater difficulties for the pollinator to discriminate them from rewarding flowers (Dafni, 1984; Heinrich, 1979). However, so far there is only indirect evidence for scents influencing discrimination or generalization learning of food-deceptive flowers by bees. Bees generally learn to associate floral signals, such as colors and scents, with a reward obtained at a flower, and there are differences in learning performance in response to visual and olfactory stimuli (Menzel, 1985, 1990, 1999). Scents are learned faster than colors, and scents appear to be more salient stimuli than colors. If a scent is presented with visual cues as a compound stimulus, scent is the prominent stimulus and superimposes itself onto the memory of the other cues (Bogdany, 1978).

In this study we experimentally test the hypothesis that odor cues enhance the similarity of mimic to model and thus affect the flower choice behavior of the pollinators. This would demonstrate the potential for odor mimicry in such a system. We also try to answer the question whether a food-deceptive flower that does not offer any olfactory cues receives more visits from flower-visiting bumble bees. Thus the often-observed scentlessness in many food-deceptive orchids may reflect the selective advantage of being scentless.

We use artificial flowers of different colors and scents assuming that a rewarding plant (“model”) is associated with a food-deceptive plant (“mimic”). Model and mimic are chosen such that they are rather similar to each other in respect to color, so that a bumble bee would be able to discriminate both to a low degree solely on the basis of visual cues as is probably the case in natural mimicry systems. We compare learning rates of initially naive bumble bees in discriminating the mimic in favor of the model based on color and scent cues. In particular we test the hypothesis that bumble bees discriminate the model from the mimic better if they are differently scented than if only the model is scented while the mimic is scentless. We also test whether generalization from model to mimic is influenced by the presence or absence of scent cues and their particular quality with respect to the model's olfactory cues. As we shall see, odor signals have indeed a substantial impact on color choice behavior, indicating that attention to small differences in color signals can be modulated by odor cues. We discuss our findings with respect to theoretical learning rules. But even more interestingly, applying our laboratory findings to the conditions found in the flower meadow, we not only demonstrate the high potential for odor mimicry but also show that producing no olfactory cues may depict an adaptive character for food-deceptive plants.

METHODS

Two experiments were performed. Experiment 1 consisted of a discrimination task where the rewarding model and the nonrewarding mimic were simultaneously present. Experiment 2 represents a generalization test, divided into a training session, in which the bees learned to forage on the model first, and a successive test phase in which the mimic was presented in the absence of the model but along with new alternative test flowers.

For both experiments bumble bees (Bombus terrestris L.) were purchased from STB Control (Aarbergen, Germany) and kept in a flight cage before and during the entire experiment. The only food sources, apart from the reward offered in the experiments were unscented sucrose solution in a transparent uv-transmitting perspex feeder, and pollen either spread on a sheet of cardboard or poured directly into the hive. The animals were thus flower-naive such that they had had no contact with real flowers or other colored or scented nectar sources. Animals were marked individually with colors or numbered tags. The flight cage was either connected to the experimental arena with a tunnel such that individually trained animals could fly from the flight cage to the experimental arena (Experiment 1) or bumble bees were transported in a transparent plastic cup from the flight cage to the experimental arena (Experiment 2). In the arena 124 (Experiment 1) or 68 (Experiment 2) artificial flowers were arranged in a hexagonal grid at a distance of 10 cm between each flower and its six neighbors. Each flower position was numbered. Flowers consisted of a round transparent Perspex stalk, 4.5 cm in height. A 2 × 2 cm cardboard square was attached at the top and covered with colored papers from the HKS—N series (Hostmann-Steinberg K+E Druckfarben, H. Schmincke and Co., Germany). The Perspex stalk poked through a hole in the center of the cardboard and had a small depression on the top, which contained either sugar water (model) or plain water (mimic).

The two main test colors, blue 1 (HKS—N 59) and blue 2 (HKS—N 51), were chosen on the basis of being perceptually similar for a bumble bee, in order to achieve a low discrimination and a slow learning rate. In Experiment 2, additional colors were used: human green (HKS—N 64), human violet (HKS—N 37) and human orange (HKS—N 22). The reflectance spectra of all colors and the corresponding loci in the bumble bee color space are shown in Figure 1. Flowers were presented on a human gray background (HKS—N 93). The color opponent coding model of color vision (Backhaus, 1991) was used to measure how different the tested colors were for bumble bees. Photoreceptor sensitivity functions of Bombus terrestris were taken from physiological measurements (evened out from Peitsch et al., 1992). The spectral reflectance of all HKS—N papers were measured with a spectral flash photometer at 300-700 nm (SR-01 Groebel, Ettlingen, Germany, resolution 1 nm) against a white standard (Menzel and Shmida, 1993). Color loci were calculated according to Backhaus (1991) and plotted in the color opponent coding diagram (Figure 1). The photoreceptors were assumed to be adapted to the human gray background which was chromatic to a bee due to its uv absorption. In this diagram, perceptual differences between colors relate to distances between loci in the color space, measured in a city-block metric.

Figure 1

Color loci of the experimental colors in the color opponent coding diagram (Backhaus, 1991). Spectral reflectance properties were measured using flash photometry between 300 and 700 nm (see details in Menzel and Shmida, 1993). The line shows the spectral boundary which is obtained by connecting the loci of spectral lights, marked by the dots in steps of 10 nm.

Two different scents were used in the experiments: rose oil and clove oil. Both scents were diluted with pentane 1:200 and a 3 μl (rose) or a 2 μl (clove) drop was pipetted onto the colored cardboard of the artificial flowers immediately before the experimental run. In pre-experiments this concentration was found to be perceptible by bumble bees. A more intense scent on the flowers resulted in aversive behavior such that the bumble bees refused to take the sugar water from the scented flowers. This effect is slightly stronger with clove than with rose, which is why different amounts of scents were chosen.

The position number of each flower visited was announced by the observer on a tape recorder while the bee was foraging. Visits to the same flower were only recorded as different visits if another flower had been visited in the meantime. A visit was counted if the bumble bee touched the flower with several legs and significantly arrested its flight. Cases in which the bees touched the flowers with their front legs only without alighting on the flower were not recorded as visits.

Experiment 1

Discrimination learning experiment—model and mimic are presented together in mixed arrays

In the pretraining phase individually marked bumble bees were trained to fly into the training arena by presenting drops of sugar water on a gray cardboard that was continuously moved through the tunnel into the experimental arena. They were then trained to discriminate between the model and the mimic with rather similar bluish colors (balanced between individuals to be either blue 1 or blue 2) that were presented simultaneously in the experimental arena. 3 μl 50% sucrose solution was pipetted into half of the 124 flowers (model flowers) and water into the other half (mimic flowers). The positions of rewarding and unrewarding flowers were randomized and varied for each individual but kept constant for the test phase which lasted two bouts for each animal. Rewarding and unrewarding flowers had, depending on the experiment, no scents, the same or different scents, but were always different in color.

Once a bee had learned to fly into the arena it was caught and gently placed onto one of the rewarding flowers, and fed with sucrose solution to satiation. This was repeated once after the animal flew back into the hive. After this training procedure the test phase started in which the animals started to forage in the flower array. After a bumble bee had completed its first foraging flight it flew back to the hive. For each animal two consecutive test flights were recorded. The rewarding flowers were replaced by fresh ones between the flights and filled as described above.

The different groups are named in the following way: The model was named as 0+ if it had no scent and A+ if was scented. The mimic was named 0- if it had no scent and A- if it had the same scent as the model, but B- if it had a different scent. A and B do not represent specific scents, as the two scents used in the experiments were always balanced. We ran five different experimental groups in Experiment 1 consisting of 12 bumble bees each (Figure 2):

Figure 2

Relative number of choices to the rewarding flowers (models) of bumble bees in the discrimination task in Experiment 1 for different experimental groups. A+ / 0-, model scented and mimic unscented; A+ / A-, model and mimic with same scent; A+ / B-, model and mimic with different scent; 0+ / 0-, model and mimic not scented; 0+ / A-, model unscented and mimic scented. Model and mimic always differed in color (blue 1 or blue 2). Each group consisted of 12 bees. Statistical tests of significance are presented in Table 1. An overview of the group differences is given in the additional table below the figure.

• Group A+/0- (Models were scented and mimics were unscented).

• Group A+/A- (Models and mimics were scented with the same scent).

• Group A+/B- (Models and mimics were scented with different scents).

• Control group 0+/0- (Models and mimics were unscented).

• Control group 0+/A- (Models were unscented and mimics were scented).

All experiments were balanced with respect to colors and scents; this means that half of the individuals were tested with blue 1 as the rewarding color and half of the individuals were tested with color blue 2 as the rewarding color. Similarly, in experiments with the groups A+/0-, A+/A-, and A+/B- half of the bees were tested with rose as the scent of the rewarding flowers and half of the bees with clove. The two sensory modalities were also balanced against each other in such a way that blue 1 was paired with rose for some animals and with clove for others. In control group 0+/A- half of the animals were also tested with each scent on the unrewarding flowers. No scent-specific differences were observed and therefore scents were not balanced in Experiment 2.

Experiment 2

Generalization experiment—model and mimic are present sequentially

In Experiment 2 the bees were first trained to distinguish the model against an alternative flower type offering less reward. In the test phase the mimic was presented against a rewarding novel alternative test type. The experiment was divided into a pretraining, training, and test phase.

In the pretraining phase bumble bees were carried from the flight cage to the experimental arena where 68 flowers were presented simultaneously, half of them being highly rewarding model flowers and the other half an alternative flower type of lower reward. Since in Experiment 1 no color specific differences were found we did not balance colors in Experiment 2. Thus the model plant's color was always blue 1 (4 μl of 50% sucrose, scent present or absent) and the alternative training color was green 4 μl of 20% sucrose, no scent). The bee was gently placed on one model flower, where it was fed ad libitum with sugar water (50%) and set back into the flight cage. This was repeated once. In the training phase the bee foraged freely in the flower array, where it learned to discriminate the highly rewarding models from the less rewarding alternative flower type. After the training run, the bee was returned to the flight cage. The procedure was repeated once with a different random distribution of freshly filled and, depending on the experimental group, freshly scented training flowers. In the test phase the bee was placed into the array where half of the flowers were mimics (the test color was either blue 2, blue 1, or orange, scent was present or absent) and the other half were a rewarding alternative test flower type (the test color was always violet, 4 μl of 50% sucrose, no scent).

We performed tests with seven groups consisting of six bumble bees each that differed with respect to flowers scents and training/test colors (Figure 3):

Figure 3

Avoidance learning of bumble bees in Experiment 2. The relative amount of incorrect choices (with standard errors) of the nonrewarding mimic is shown. Groups differed with respect to scent cues added in training or testing. Group A+ / 0-, model (blue 1) scented and mimic (blue 2) unscented. A+ / A-, model (blue 1) and mimic (blue 2) with the same scent. A+ / B-, model (blue 1) and mimic (blue 2) with different scents. 0+ / 0-, model (blue 1) and mimic (blue 2) not scented. 0+ / 0- (same color), model (blue 1) and mimic (blue 1) not scented. A+ / A- (same color), model (blue 1) and mimic (blue 1) with the same scent. 0+ / 0- (distinct color), model (blue 1) and mimic (orange) not scented. Each group consisted of 6 bees, and 50 choices were counted. Statistical tests of significance are presented in Table 2. Group differences are depicted in the additional table.

• Group A+/0-: The model (blue 1) was scented and the mimic (blue 2) was unscented.

• Group A+ / A-: The model (blue 1) and the mimic (blue 2) were scented with the same scent.

• Group A+ / B-: The model (blue 1) and the mimic (blue 2) were scented with different scents.

• Control group 0+ / 0-: The model (blue 1) and the mimic (blue 2) were unscented.

• Control group 0+ / 0- (same color): The model (blue 1) and the mimic (blue 1) were unscented.

• Control group A+ / A- (same color): The model (blue 1) and the mimic (blue 1) were scented with the same scent.

• Control group 0+ / 0- (distinct color): The model (blue 1) and the mimic (orange) were unscented.

Data analysis and statistics

Learning curves in Experiments 1 and 2 depict averages of relative number of visits to the model (Experiment 1 and training in Experiment 2) or the rewarding test flower type (test in Experiment 2) for each group, with standard errors. Maximal performance level (MPL) describes the percentage of correct choices once the learning reached saturation, that is, the animal arrived at a learning plateau. Because a learning plateau was reached depending on the task and set up after a different number of total visits we measured the MLP after 150 total visits in Experiment 1, after 100 total visits for the training in Experiment 2 and after 30 total visits in the test phase of Experiment 2. In cases where for one of the experimental groups saturation was not reached, the final performance level was taken as MPL.

Differences between groups in Experiment 1 were compared over the time course with a repeated-measures ANOVA using Statistica 5.5dt (Statsoft, 1993). Taking group treatment as an independent between-subjects factor and time as an independent within-subjects factor (measured after 30, 60, 90, 120, and 150 total visits) and the relative number of visits to the rewarding flower type for each individual as a dependent variable. The effect of group treatment was thereafter determined by ANOVA after each interval of visits separately, and differences between groups were determined by least significant difference (LSD) post hoc comparisons.

For the training in Experiment 2 intervals of 10 total visits each were chosen from 1 to 100. For the test situation in Experiment 2 statistical analysis was carried out for the group differences after 30 visits, because the choice behavior was earlier saturated than in Experiment 1 or motivation of the bumble bees in some groups dropped rapidly during the test. Differences between groups were tested by ANOVA taking group treatment as an independent factor and the relative number of visits to the alternative test flower as a dependent variable. Group differences were determined by LSD post hoc comparisons.

RESULTS

Experiment 1

Bumble bees learned to discriminate mimic from model on the basis of color cues alone when no scent was present (control group 0+ / 0-); learning curves for all successive groups are depicted in Figure 2. For statistical differences in learning performance see Table 1. As there was a significant interaction between time and group treatment (repeated-measurements ANOVA: Rao's R (24, 88) = 2.69 p = 0.0004), group differences had to be tested for each interval separately. In all intervals there was a significant difference between groups (ANOVA results, 1-30: F = 9.84 p <.0001; 31-60: F = 5.69 p =.0007; 61-90: F = 9.55 p <.0001; 91-120: F = 5.19 p =.0014; 121-150: F = 5.77 p =.0009; 151-180: F = 6.04 p =.0011). The acquisition curve rises slowly and MPL is at 78% and thus relatively low. This indicates that both flower types are rather similar to bumble bees. When model and mimic are distinctly scented (group A+/B-) the acquisition curve is steeper and MPL reaches 95%. As expected, adding a different scent cue to the color alone leads to faster discrimination learning and discrimination at a high level (compare control group 0+/0- versus group A+/B-).

Table 1

P-values of the LSD post hoc comparison of the ANOVA test for Experiment 1: group differences based on the relative number of visits to the alternative test flower for each individual in the respective group

	Intervals of total flower visits
Group comparisons	1-30	31-60	61-90	91-120	121-150
A+/0- versus A+/B-	.3540	.6439	.4481	.8153	.8936
A+/0- versus A+/A-	.0070	.0203	.3108	.7708	.9900
A+/A- versus A+/B-	.0004	.0593	.0833	.5873	.9080
0+/0- versus A+/A-	.1008	.0683	.0003	.0011	.0014
0+/A- versus A+/0-	.6421	.0593	.4981	.3708	.0855
A+/0- versus 0+/0-	<.0001	.0001	<.0001	.0007	.0009
A+/B- versus 0+/0-	<.0001	.0004	<.0001	.0003	.0003
A+/B- versus 0+/A-	.1665	.1498	.1581	.2466	.519
0+/0- versus 0+/A-	.0002	.0238	.0001	.0069	.0882

If the model was scented but the mimic was not (group A+/0-), the results are quite similar and learning performance does not significantly differ between the groups (A+/0- versus A+/B- Table 2). Thus the task of discriminating between a scented model and a mimic is as difficult for bumble bees when the mimic presented a scent that differed from the model's as it is when the mimic presented no scent. Having no scent instead of a different one is not advantageous in this situation. In the reverse case, where only the mimic was scented and the model was not (control group II), MPL is slightly lower than in group A+/0- (87% compared to 95%, A+/0- versus control group 0+/A- Table 1) but differences were not significant.

Table 2

P-values of the LSD post hoc comparison of the ANOVA test in Experiment 2 after 30 visits: group differences based on the relative number of visits to the alternative test flower for each individual in the respective group

	A+/0-	A+/A-	A+/B-	0+/0-	0+/0- (same color)	A+/A- (same color)	0+/0- (distinct color)
A+/0-		.0176	.0017	.2949	.2092	.0696	.0001
A+/A-	.0176		.0000	.0021	.2294	.5342	.0000
A+/B-	.0017	.0000		.0468	.0001	.0000	.4364
0+/0-	.2949	.0021	.0468		.0331	.0092	.0072
0+/0- (same color)	.2092	.2294	.0001	.0331		.5548	.0000
A+/A- (same color)	.0696	.5342	.0000	.0092	.5548		.0000
0+/0- (distinct color)	.0001	.0000	.4364	.0072	.0000	.0000

In group A+/A-, where model and mimic had the same scent, and scent cues could therefore not be used to tell the two flowers apart, we expected either a performance similar to that of 0+/0- (no scent on both flower types) or even reduced discrimination due to the higher overall signal similarity. The opposite behavior was observed: acquisition was steep, with MPL higher than in control group 0+/0- (A+/A- versus control group 0+/0-). We tested for differences between groups during the progression and observed that the differences between A+/A- and control group 0+/0- became significant only after 61-90 interval of total visits, but that they then remained significant. This indicates that the mere presence of a scent at both flower types enhances the bumble bees' behavioral discrimination performance. For the rewarding plant it turns out to be advantageous to bear scents when confronted with mimics. We compared group A+/0- with control group 0+/0- and saw a significant change in the respective learning curves (see Table 1: A+/0- versus control group 0+/0-); bumble bees make fewer visits to the mimic if the model is scented. However, for the case where a potential model is scentless it turned out to be better for the mimic to be scentless as well: visitation rate in control group 0+/0- with the scentless mimic is higher than in control group 0+/A-, with a scented mimic (compare control groups 0+/0- versus 0+/A-).

Experiment 2

Training

Bumble bees learned to discriminate the highly rewarding model from the unscented alternative flower type of lower reward faster if the model was scented (group A+/0- and control group A+/A- [same color]) than if it was unscented (control groups 0+/0-, 0+/0- [same color] and 0+/0- [distinct color]; Figure 4). Bumble bees reached MPL within the first 10 flower visits if the model was scented, and it was only after 40 visits that no significant difference between scented and nonscented groups was detectable. Therefore we determined that the scents used were perceived and do indeed influence the learning behavior. We hereby replicate the observation from Experiment 1, where the addition of a scent also improved the discrimination (Experiment 1: group A+/0- versus control group 0+/0-, Figure 2).

Figure 4

Learning function of bumble bees in the training phase of Experiment 2. Bumble bees had to distinguish the model (color: blue 1) from a less rewarding alternative training flower type (color: green). Groups differed with respect to scent cues added to the model; filled circles: scent present, squares: no scent cues present. The alternative training flower type was not scented in both groups. The relative amount of rewarded choices to the model plant is shown with standard error. Unpaired t test was used for calculating the significance between groups.

Test

In the test phase, bumble bees always started to visit the mimic (except in group A+/B- and control group 0+/0- [distinct color], where the choice level was already at 50% during the first 10 visits). After several visits the insects learned to restrict their visits to the rewarding alternative test flower type. The learning function observed differs between different experimental groups (ANOVA results tested after 30 total visits, F = 11.35 p <.0001) and reflects the influence of scent and color cues in this generalization task (see learning functions in Figure 3 and Table 2 for ANOVA results of post hoc group comparisons). In group A+/B-, where the mimic provided a scent different from that of the model, the generalization from the model to the mimic is reduced. On the contrary, in group A+/A-, in which model and mimic had the same scent, avoidance learning was significantly slower (compare groups A+/A- versus A+/B-, Table 2). Bumble bees of this group needed more visits in order to avoid the mimic.

In group A+/0-, in which the model was scented but the mimic was not, the change observed from mimic to the alternative rewarding flower type was intermediate as compared to those observed in groups A+/A- and A+/B-. The test flowers in this group differed from the training flower in the following aspects: the color, though rather similar, was different, no reward was present, and the flower was not scented. Bumble bees of this group learned to avoid the nonrewarding flower, but it took them significantly longer, compared to bumble bees of group A+/B- (A+/0- versus A+/B-, Table 2). This indicates that the absence of scent on the nonrewarding flower in this experimental task impairs the generalization from model to mimic in this learning process.

We used control groups 0+/0- and 0+/0- (same color) to test how similar the colors are to the bumble bees and how quickly they can learn to switch to the rewarding flower type solely by means of color discrimination. In control group 0+/0- (same color) exactly the same flower type was presented as was used in training (blue 1, unscented). In control group 0+/0- we used blue 2 as a test color instead (as we did in the experimental groups A+/0-, A+/A-, and A+/B-). Although the two colors are very similar to the color vision system of the bumble bee, the animals in control group I were quicker at learning to avoid the nonrewarding flower if this flower type was signaled with the different color (Figure 3).

However, due to high color similarity between the trained and the tested colors in control group 0+/0- (same color; blue 1 versus blue 2), bees still performed poorly in this task. We therefore asked whether a color with a greater difference in the color space of the bumble bee (i.e., a better discriminability and/or lower generalization from the training color) could increase performance. To do so we used a new and conspicuously different color as a test color in control group 0+/0- (distinct color). In this task bumble bees trained to blue 1 should now avoid the nonrewarding flower type in the test, which was given an orange color. As expected, this was much easier for the bumble bees and their learning curve was the steepest of all groups (Figure 3).

We next compared different experimental groups in order to answer the following questions.

How does a learned scent influence choice (group A+/0- versus control group 0+/0-)? We compare two groups where the rewarding training color is either scented (group A+/0-) or unscented (control group 0+/0-). The nonrewarding flower type in the test situation is unscented in both groups. We do not find a significant difference between these experimental setups (Table 2).

What kind of behavior is observed if the training and test color are the same and if both are unscented in the first group (control group 0+/0- [same color]), where in another group both flower types were scented with the same scent (control group A+/A- [same color])? The results of these groups are not significantly different (Table 2), thus showing that an additional but cueless scent does not affect performance in this task.

Furthermore we searched for differences in the generalization performance between flowers with the same scent in this situation of Experiment 2 (control group A+/A- [same color] versus A+/A-). In one case training and test flower types showed identical color and scent signals (blue 1 vs. blue 1, control group A+/A- [same color]); in the other group (group A+/A-) training and test flowers were scented with the same scent but their colors differed slightly from each other (blue 1 versus blue 2). We observed no significant difference (Table 2) between these groups, indicating that the scent cue dominates generalization and suppresses color information.

If the mere presence of scent cues were to have any impact on generalization (as we showed for a discrimination learning task in Experiment 1) we would expect lower generalization in Experiment 2, hence faster learning performance in those situations, where training and test flower types, although different in color, bear the same scent (group A+/A-), as compared to a group where the two colors had to be learned and generalization was tested without any scent cues. We observed the opposite: if the same scent was present during training and testing, generalization between the two colors was greater compared to the group where no scent was present in either of the flower types (control group 0+/0- versus group A+/A-, Table 2). This could be explained by a higher overall similarity between training and test flower in group A+/A-, leading to more generalization.

DISCUSSION

We tested the influence of olfactory cues presented in combination with color signals on bumble bees' flower choice behavior towards nonrewarding flowers. We assumed that a mimic would have a single rewarding model, from which it is distinguishable but similar with respect to visual signals. We observed that bumble bees discriminated differently between the model and the mimic depending on the presence and quality of additional olfactory cues, and we also observed that the finding differed according to whether mimic and model are presented together (Experiment 1) or successively (Experiment 2). These situations simulate not only two natural growth conditions, but also reflect different learning paradigms, namely, discrimination learning and generalization. We set up the experiments in such a way that model and mimic could be only poorly discriminated on the basis of color cues alone and the mimic was therefore only barely avoided. In the search for an adequate strategy of nonrewarding plants in a food mimicry system we interpret our main experimental findings for a natural situation. We observed that:

• The generalization from a scented model to a mimic was strongest if the latter had the same scent as the model. This holds true for both experiments.

• If the scented model was presented simultaneously with the mimic (Experiment 1), the mimic was avoided at the same rate if it had either no scent or a scent different from that of the model.

• If the animal was first trained to discriminate the model against a less rewarding alternative flower and was subsequently tested for discrimination between the mimic and a second alternative rewarding flower in the absence of the model (Experiment 2) it initially visited the mimic and switched more readily from the mimic to the rewarding alternative flower from differently scented mimics than from scentless mimics. Thus generalization from the model to the mimic and the resulting flower choice behavior was impaired by the absence of scent.

Based on our observations, which strategy would be the best for a food-deceptive flower to employ? If the model were scentless, the best strategy would be to be scentless, too, because a scentless mimic would be discriminated less well than a scented one. If the model were scented, which is certainly more common under natural conditions, the best strategy for the mimic would be to have the same scent, followed by that of being scentless, and the worst strategy, at least in a generalization task of a transfer situation (as in Experiment 2), would be to have a scent that differs from that of the model. Thus, based on the behavioral experiments, we expect that there exists a potential for odor mimicry based on the imitation of a model's scent cues.

Scent-based mimicry has never been observed in food-deceptive flowers although it occurs in other mimicry systems (Borg-Karlson, 1990; Dafni, 1984; Dettner and Liepert, 1994). In cases of specialized floral mimicry, models are imitated by visual cues (Dafni and Ivri, 1981a; Nilsson, 1983b; Vöth, 1982) or tactile cues (Dafni and Ivri, 1981b). Even in the case of fungi imitating flowers signal imitation appears to be limited to visual cues (Raguso and Roy, 1998; Roy, 1993; Roy and Raguso, 1997). Roy and Raguso (1997) interpreted the absence of olfactory mimicry as an adaptive strategy of the mimic to enhance flower constancy; however, in their example, the mimicking “ pseudoflower” provides a reward and is therefore not necessarily avoided. More data are needed to answer the question whether scent-based mimicry was simply overlooked because it is more difficult to detect than visual mimicry or whether it is limited by physiological or phylogenetic constraints.

Whether a scent different from that of the model reduced visitation rates at the mimic was dependent on the learning paradigm: in the differential conditioning paradigm of Experiment 1 both flower types could be well discriminated as long as there were scent cues available to recognize each type (presence or absence of scent, different scents). Here, the animal learned to distinguish the scented model from the mimic at the same rate if it was scentless or differently scented. This learning paradigm matches a natural situation in which a mimic would grow close to the model and the animal could repeatedly switch between the two flower types. If model and mimic grow in spatially separated patches, the learning paradigm corresponds more closely to the generalization paradigm of Experiment 2. In this case, the mimic had a greater advantage from being scentless (i.e., it received more visits) than from having a scent different from that of the model.

For a rewarding plant the best signaling strategy in order to achieve a high degree of flower constancy would always be to produce a distinctive scent, rather than producing none at all or a scent identical to others. This is particularly important if the plant provides a high reward as compared to others, because the plant is expected to lose attractiveness if it is confused with others of lower profitability to the pollinator (Ferdy et al., 1998). A rewarding plant invests resources in order to be pollinated; the pollinator associates these resources with floral signals (Menzel, 1985). Flower constant pollinators search for these signals more consistently and plants differing in their floral signals are discriminated better (Chittka et al., 1997; Heinrich, 1979). If flower signals are designed to enhance flower constancy, as is generally assumed (Feinsinger, 1987; Gumbert et al., 1999; Waser, 1983), then floral scents should be used for making a flower more distinctive. Bees use floral scents to discriminate among different species (Ackermann et al., 1997; Free, 1970; Galen and Kevan, 1983) and, accordingly, most flowers that provide rewards have a perceptible scent.

We thus may conclude for natural mimicry systems that, although scent mimicry would be the most effective signaling strategy enabling a mimic to attract pollinators away from a specific model it may not be the most reliable strategy in unpredictable and constantly changing plant communities. Floral mimics cannot influence the presence of surrounding model plants and, with the exception of few cases of one-to-one floral mimicry (Dafni and Ivry, 1981a; Nilsson, 1983b), most mimics rely on a general similarity to several models (Ackerman, 1981; Bernhardt and Burns-Balogh, 1986; Brown and Kodric-Brown, 1979; Dafni, 1983, 1984; Dafni and Calder, 1987; Gumbert and Kunze, 2001; Little, 1983; Thien and Marcks, 1972). For them it may be much easier to imitate other species solely in respect to color, because the sensory space of color is much less complex than that of scents and similarity among co-flowering species is common (Gumbert et al., 1999). While deceptive flowers could certainly imitate one particular scent very well, as was successfully demonstrated for sexual mimics (Borg-Karlson, 1990; Borg-Karlson et al., 1993; Schiestel et al., 1999), a food-deceptive mimic could probably only imitate a single model species, due to the different combinations of many volatile components in most floral scents. In this case, the mimic would be completely dependent on the presence of its specialized model because the similarity to all other plants would be strongly decreased. Most food-deceptive species appear to have chosen a more general mimicry strategy and, in this case, being scentless is certainly the best option.

Implications for learning theory of mulitmodal signals and color discrimination in bees

Apart from the results that may be directly related to pollination strategies of nonrewarding plants our experimental findings raise some more general questions concerning learning and memory processes. Interestingly, bees were able to discriminate the mimic from the model faster when both flower types were scented with the same scent, as compared to a situation in which both were unscented. This was unexpected because the presence of scent in both flower types could not be used to tell them apart and color information was the only cue that could be utilized by the bee. Thus it appears that the mere presence of scent enhances the discriminability of the same pair of colors. How can such an effect be interpreted?

Two explanations come to mind: First, color discrimination may be limited by attention, and attention may be increased in the presence of scent. If the scent could be perceived from a larger distance than visual stimuli then flowers would be detected earlier and the time period would be increased in which the bee could pay attention to the stimuli to be evaluated for learning and flower choice. Colors were long thought to be long-distance signals, whereas scents were believed mainly to govern the final landing stage of a flower visit (von Aufsess, 1960; Dobson, 1987, 1994; von Frisch, 1914; Lunau, 1992). Recently it was shown that colors are actually only perceived from quite a short distance, when the colored object subtends a visual angle of 15° or larger (Giurfa et al., 1996). Thus, it may well be that scent is perceived before color information is available to the animal. However, in a recent study Odell et al. (1999) showed that bumble bees foraging on differently colored morphs of Antirrhinium (Scrophulariaceae) did not change their foraging behavior according to scent differences. Furthermore, Ackermann et al. (1997) found that floral fragrance phenotypes did not affect visitation patterns.

Second, the discrimination task in our case may have been limited to a greater extend by memory processes. Either discrimination learning or memory retrieval or both may be enhanced for multicomponent stimuli, as has been described for many animals. Multimodal learning was studied in many different animals (Burne and Rogers, 1997; Pointer and Bond, 1998; Roper and Marples, 1997; Rosenzweig, 1983) and it has been shown that signals involving different modalities were learned and memorized better than signals of a single modality (Rowe, 1999). Also, in honeybees Gerber and Smith (1998) found that pretraining to a visual stimulus increased subsequent scent cue learning. They proposed a different neural reinforcement system for visual and odor cues. From all these findings one can conclude that multimodal stimuli do not necessarily follow the learning rules of unimodal compound stimuli (Gerber and Smith, 1998; Rowe, 1999). A phenomenon predicted by learning theories and observed in behavioral experiments is that learning in discrimination tasks is altered when multicomponent stimuli are used (Rescorla and Wagner, 1972; see review in Pearce, 1994). Both negative effects, such as overshadowing or blocking of one of the stimuli components by the other and facilitatory effects, frequently occur in learning tasks (Pearce, 1994), and it is debatable whether this is related to attention phenomena or is an intrinsic process of memory formation and retrieval.

It is not clear which mechanism is responsible for the enhancement of color discrimination in the presence of scent. Calculating and evaluating the similarity of two compound stimuli is still difficult. Whether two or more components of one stimulus act as a new configural stimulus or are still evaluated by their elements remains unclear for most animal studies. The Rescorla-Wagner model (Rescorla and Wagner, 1972) predicts for learning multicomponent stimuli that the addition of a common stimulus in certain discrimination tasks enhances discrimination (Pearce, 1994: 591). This theoretical assumption has never been observed in real behavioral experiments (Pearce, 1994) but underlies our experiments. However, our observations indicate that color choice is controlled by more than color vision alone. If attention or memory formation and/or retrieval do indeed affect the weighting of color information as an appetitive signal, then the color similarity functions as derived from a pure color vision model may not be exclusively limited by sensory input. For discriminating food-deceptive plants under natural conditions, we find that floral signals must be evaluated with regard to their multicomponent nature. Whether attention is the influencing factor in our discrimination/generalization task remains unclear, but it is conceivable that scents may enhance attention for other discrimination tasks, such as for color. Alternatively, scents may increase color evaluation by facilitating memory and retention processes.

We thank K. Faust, K. Göllner, K. Huckauf, B. Komischke, O. Neuber, and A. Perlewitz for their help with the experiments. We are most grateful to R. Menzel, B. Gerber, and N. Hempel de Ibarra and two anonymous referees for helpful comments on earlier drafts of the manuscript. This study was supported by the Deutsche Forschungs-gemeinschaft (grant to Randolf Menzel Me 365/22-3). M. Wurm helped us improve the English.