Chapter One
Introduction

Outline

In this thesis I explore some of the most interesting aspects of biology - natural selection and mimicry; perceptual biases, learning and sexual selection leading to speciation. I investigate all these factors by studying and carrying out experiments on polymorphic butterfly species. One is the well-known species Papilio dardanus (the Mocker Swallowtail), which has many geographic races in Africa, each with a different balance of mimetic and non-mimetic morphs. The other is Argynnis paphia (the Silver-Washed Fritillary), which has two morphs, the proportions of which vary across its range. The genetic control of the morphs in both species is quite well understood, and preliminary behavioural experiments have been carried out indicating that males show morph preferences in their choice of mates.

In this chapter, I introduce the background to each aspect of the thesis. Chapter 2 is an investigation of the mate colour preferences of Argynnis paphia, attempting to clarify the basis of the preferences previously found for bright orange flowers and mates. In Chapter 3, which introduces and provides background material on Papilio dardanus for the subsequent chapters, I analyse the colours used by the species and its models, and also the visual sensitivities of the butterflies, in order to gain an idea of how the butterflies may appear to each other and to their predators. In Chapter 4 I determine the feeding colour preferences of Papilio dardanus, and also investigate the role of learning and social influences on their flower choice. Chapter 5 concerns the mate preferences of Papilio dardanus, following on from the fieldwork of Cook et al. (1994) in which wild males were shown to have a preference for black and white females over other morphs. I investigate the basis for this preference, including the role of learning. In Chapter 6 I describe work on the palatability of Papilio dardanus in comparison with its models. Finally, in Chapter 7, I construct a mathematical model of the balance of morphs in a population, taking into account the results of the previous chapters, to demonstrate how the balanced polymorphism may be maintained.

Each chapter is written to stand on its own, and includes a short introduction to the literature surrounding each topic. I have not attempted to review the literature exhaustively where good review articles already exist, but have recommended the reader to articles which do this where possible.

Mimicry - the eye of the beholder

The subject of mimicry has long been a matter of great debate among biologists. H. W. Bates formally introduced the concept of mimicry in 1861 (only two years after the publication of Darwin's 'Origin of Species') in a reading at the Linnean Society of London after 11 years studying butterflies in the Amazon. It has still to be satisfactorily defined.

The problem with the definition of mimicry is the range of phenomena which it can include. One of the most controversial issues is the position of crypsis. Vane-Wright probably came closest to a satisfactory definition in 1980 when he said:

"Mimicry involves an organism (the mimic) which simulates signal properties of a second living organism (the model) which are perceived as signals of interest by a third living organism (the operator), such that the mimic gains fitness as a result of the operator identifying it as an example of the model."

In this definition, Vane-Wright demonstrates an important principle - that the message of the signal is determined by the psychology of the signal receiver (or 'operator' to use Vane-Wright's terminology). This view point makes a distinction between crypsis and mimicry since in crypsis an organism is trying to avoid the transmission of signals to a receiver who may be a predator whilst a mimetic organism is actively attempting to signal to any potential predator receivers (Silberglied, 1977). Different receivers, however, have different states-of-mind which affect the perceived information in a signal. This means that it is impossible to classify a species with regard to its mimicry or crypsis without taking into account the potential signal receivers. For example, Rothschild (1981) describes how burnet moths (Zygaena sp.), with their classic black and red warning colours, can easily be overlooked as seed pods of the nectar plant Vicia cracca. An entomologist can suddenly see individual moths which they had previously seen as seed pods from further away, and she also describes how, as the entomologist gets older, the moths are more frequently overlooked as pods. This demonstrates how small differences in the signal receiver's perception can alter the classification of a signal. This is an important theme in this thesis as I explore how butterflies receive and perceive various signals (specifically with regard to the role of colour in food and mate choice).

Mimicry is classically divided into two types: Batesian mimicry, in which the mimic is palatable; and Mullerian mimicry, in which the mimic is unpalatable. However, there is naturally a 'palatability spectrum' (Brower et al., 1968) which means that there is a considerable grey area in which species cannot be classified as either Batesian or Mullerian mimics, the extremes of the spectrum. Palatability varies between individuals of a species, often depending upon their foodplant (Brower et al., 1968), and the palatability also depends on the predator concerned. A good example of variable palatability is described again by Rothschild (1981): cuckoos (Cuculus canorus) easily recognise black-and-yellow hypsid caterpillars and similar mimics, but whilst as young fledglings they assiduously avoid them (as they prove highly indigestible and harmful to the young birds), as adults they actively seek them out and show no ill-effects of eating even very large numbers of them.

Batesian and Mullerian mimics have different population dynamics. If the mimic is palatable, then the model's own protection is endangered. This means that the mimics must always be rare with respect to the models as the risk to them is positively density-dependent (as their density with respect to their model increases, so does the attack rate on them as predators find more of the similar-looking individuals palatable). If the mimic is as unpalatable as the model then they both gain equally from the similarity. The risk to unpalatable, aposematic morphs or species is negatively density-dependent (as their density decreases, the frequency of attacks on them increases as predators are less likely to recognise the signal of unpalatability). This difference theoretically allows a distinction between the two systems to be made in the field, and this has been tested by Benson (1972) who captured individuals from a natural population of the butterfly Heliconius erato and altered their patterns in one of two ways. Recapture data showed that the rarer of the two altered patterns had a lower survival rate, suggesting that the species was a Mullerian mimic. Gordon (1987) showed by standard capture/re-capture that the commonest forms of Hypolimnas misippus had a lower survival rate than the rarer forms, and other evidence showed that this was likely to be due to increased predation, suggesting that its mimicry is Batesian.

However, since a species may differ in palatability to different predators or across its range (due to changes in foodplants, which often provide the toxic chemicals which make a butterfly unpalatable), just as the signals an individual gives out may have different meanings to different individual predators, and so it may not be possible to classify a species' mimicry in such general terms. One of the species studied in this thesis, Papilio dardanus is often cited as a classic example of a polymorphic Batesian mimic. However, there is a little evidence to suggest that this may not be entirely true, and that the species may possess some chemical protection. This is the subject of Chapter 6.

Perception and bias

Mimicry is only one example of signalling, and the importance of signal receiver psychology holds for all signals. Throughout the natural world, aposematic coloration has converged on bold red and yellow (sometimes white) markings, often with black. These colours are very obvious against a wide range of backgrounds, and appear to exploit the senses of predators to achieve universal conspicuousness. The exact reason why defended prey are almost always conspicuous is debated, but they appear to be exploiting the psychology of the predators in some way to decrease the chance of attack (which would not be beneficial to either predator or prey). It may be that the colours are more readily associated with unpalatability (Rothschild, 1984), that the conspicuous colours are more easily remembered (Gittleman & Harvey, 1980; Turner, 1984), that the novel pattern will be easier to form an association of unpalatability with (Turner, 1975), or that the conspicuousness reduces accidental attacks by increasing the 'detection distance' (Guilford, 1986). Other hypotheses are reviewed in Guilford, 1990. In fact, as in most natural systems, it is likely that many factors are acting at once, different ones being more important with different predator species and individuals. In each case, however, it is the way in which the predator perceives and interprets the environment which determines the signals used by the prey.

It is not just the psychology of animals which affects the meaning of a signal. The perceptual systems themselves can have 'inherent bias' (Arak & Enquist, 1993) for certain colours and patterns. If a bias exists, it is easy to see how this is likely to be 'exploited' by a signal. Such biases could be very important in initiating a signal - which may subsequently become very important in sexual selection, being subject to Fisherian run-away selection, or, as recently suggested, acting as a means of driving sympatric speciation (Turner and Burrows, 1995). Arak and Enquist refer to these biases for forms which may not actually exist in nature as 'hidden preferences', and describe how these allow for the evolution of novel forms of stimulus which may prove even more striking to the receiver than the existing forms to which it reacts. They used simple artificial neural networks to demonstrate this phenomenon, although the method that they used was unrealistic and could not readily be applied to living systems. Such simple neural network designs have been criticised, for example by Dawkins and Guilford (1995), for being unrealistic in their interpretation.

There are many ways in which the whole area of using artificial neural networks can be misleading. Most of the papers claiming to simulate, in some way, a biological property of perceptual systems (e.g. Arak & Enquist, 1993; Enquist & Arak, 1993; Johnstone, 1994) use an input layer ('artificial retina'), a single hidden layer, and a single cell output layer neural network. They present the input layer with categorical data (each cell has the value of either 1 or 0), and train the network using only a few stimuli (either a positive or negative). They then present the network with numerous other stimuli and assess the response of the network to each. Artificial neural networks are not designed for this sort of task. They do not simulate neurones in any meaningful way, and are merely a means of non-linear regression on continuous data (fitting a curved surface to a multi-dimensional dataset). In this case, the dataset may be seen as in multi-dimensional space (one dimension for each input), but the points being used are at the corners of a hypercube (since the input can be either 0 or 1, but never anywhere along the continuum between these extremes). The neural network is only trained to interpolate from two data points (the two training stimuli), and is then asked to predict the values at corners of the hypercube (by presenting it with other datapoints at corners and recording the response). This explanation of the task presented demonstrates how misleading the results can be. The 'hidden preferences' which may appear to have arisen in the network are caused by the fact that the task is to regress a complex surface on the basis of only a few data points. However, the theory of 'hidden preferences' could be demonstrated by using artificial neural networks - the rest of the hypersurface which is untrained could represent the area of possible stimuli which is not under selection.

Apart from these fundamental problems with the experimental design, particular flaws in some cases have lead to false conclusions being drawn from the results. Enquist and Arak (1993) attempted to use an artificial neural network model to illustrate the evolution of exaggerated male traits (in this case long tails). They used a 6 x 6 node input layer, each node being connected to each of 10 hidden cells, which in turn were all connected to an output cell. It is therefore apparent that the way in which the 36 input cells are arranged (and thus the apparent 'pattern' projected onto them) is irrelevant. For the network, the 'pattern' is simply an array of 36 binary numbers (Enquist and Arak rather misleadingly describe this as 'a network representing the recognition system of a female bird'). The network is then trained in 'a procedure that mimics the process of natural selection of recognition systems over evolutionary time'. This is done by starting with a random series of weightings on all the connections, recording the output value (which is normalised between 0 and 1), and then 'mutating' a tenth of the weightings, and comparing the output value. The better network was retained and the process repeated. This was done until the probability of the network producing an output value of less than 0.5 for a 'correct' stimulus was less than 10^-5. The result of such a 'threshold' in the training procedure means that there is no pressure on the network producing an output of 0.6 to a 'correct' stimulus to improve on this (and vice versa). Hence the training is weak, and the results at the end of the training period indicate this (response to 'correct' stimulus 0.60 and 0.61, to 'incorrect' stimulus 0.42 and 0.41 for networks I and II respectively). In addition the 'incorrect' stimulus has 5 '1' inputs (out of the total of 36), whilst the 'correct' stimulus has 6. Thus it only takes a majority of the weightings in the network to become positive for the network to become, effectively, a kind of counting machine - the greater the number of '1' inputs the greater the output, and given the weak training this could easily become selected for (as the output is only judged on whether or not it is greater than 0.5). Enquist and Arak only published five of the test images together with the responses of their two networks. Four gave 'positive' responses, and these had 7,8,8, and 9 '1' inputs (two of them are shaped to represent increasingly long tails, and one is a shape which does not at all resemble the training pattern). One gave a 'negative' response, and had only 4 '1' inputs. Enquist and Arak interpret this as a demonstration of 'peak shift' behaviour (where a slight super-stimulus gives an exaggerated positive result and vice versa), claiming that "this proves that the recognition mechanism itself exerts selection pressure on the signal; in this case there is a 'bias' in the mechanism favouring males with longer tails". In fact it is entirely possible that the results they obtained were due to the network giving a greater output with an increased input (due to an increased number of positive weightings translating the increased number of positive inputs into a greater output value).

The same training procedure was employed by Arak & Enquist (1993) in which they attempted to illustrate the presence of two hidden preferences in two different neural networks representing "the recognition system of a hypothetical insect species faced with the problem of discriminating between flowers which differ in profitability". The networks in this case were trained with only two stimuli:

It is obvious that the only differences between these two arrays are nodes 8 and 18, and that the networks will therefore be trained to discriminate between the two arrays on the basis of these two nodes. Arak and Enquist discovered that two of their networks (in saying that "many other networks exist that can achieve the task just as well" they hint that these two were chosen from a greater number in order to illustrate their point) appeared to discriminate between the arrays in different ways - one on the basis of the status of node 8 and one on that of node 18. However, this result itself demonstrates the ineffectiveness of the training routine. Again, the threshold method was used (at the end of training the two networks gave output values of 0.61 and 0.62 for the 'correct' stimulus and 0.41 and 0.38 for the 'incorrect' stimulus), and this has resulted in the two networks becoming stuck at local maxima in their response, rather than finding the global maximum response (which would be to use the status of both nodes 8 and 18 as a discrimination mechanism). Other training mechanisms, which use genetic algorithms (ranking a population of networks relative to each other, selecting at each generation from those which had the highest outputs to a 'correct' stimulus combined with the lowest outputs to an 'incorrect' stimulus, and using 'mutations' to move from local maxima) are much more stringent, and will eventually converge on a global maximum. The fact that two of Arak and Enquists networks show very different responses only serves to highlight the fact that they have not converged on the best discrimination mechanism, and therefore that the training has not been successful. The networks are only performing a mathematical curve-fitting procedure through the data and hence there should be a 'best-fit' which all networks would converge upon when correctly trained. Arak and Enquist conclude "Only if the network is initially trained using all possible input patterns as exemplars can a perfect recognition system evolve". Although this conclusion (illustrating how perceptual biases may arise in natural recognition systems) may be correct, the method they have employed to illustrate it seems weak. The choice of such a threshold training mechanism does not even simulate natural selection processes in any way. Natural selection is a continuous process, resulting from competition between individuals and not related to the ability to reach a threshold, or stopping at that point.

By contrast, Johnstone (1994) used a similar artificial neural network structure (4 x 5 inputs, 5 hidden cells), but a different training method to illustrate female preference for symmetrical traits. He trained five networks (again "representing the recognition system of a female bird") to a set of either 4 or 5 patterns, mutating a twentieth of the connections each time. However, after the mutation each time he chose only one network from the population for 'replication', and he ranked them by "determining the mean value of its response to the patterns in the training set, and the mean value of its response to a set of randomly generated patterns, and subtracting the latter value from the former". The network with the highest score was then chosen to be mutated to form a new population of five competing networks. This was repeated 50 times. This training procedure appears to have been very successful since his results showed a clear pattern in preference emerging after training in 1000 networks which must have converged on the same 'result'.

His five training patterns were:

Each had 6 '1' inputs, and it can be seen that the informative nodes are 6 & 7, 10 & 11, 14 & 15, and 18 & 19. Johnstone found that networks trained to respond to all 5 patterns, and those which were only trained on patterns 1,2,4 and 5, all showed a preference for pattern 3, followed by patterns 2 & 4, and then by patterns 1 & 5. This is probably the result of the networks converging on the same system of recognition. At the crucial nodes, the networks are using the best guess they can at the value. At nodes 6 & 7 three of the four training sets (ignoring 3) have a '1', so the networks are looking for a '1' here, at 18 & 19 three of the four have a '0' so a '0' is the best indicator of a positive stimulus. At 10 & 11 and 14 & 15 the four stimuli are half '1' and half '0', so without pattern 3 the network might be expected to ignore these nodes as uninformative. When tested on other patterns, the networks trained on all 5 patterns showed an even greater response to another pattern:

Johnstone explained this preference as being caused by the untrained pattern overlapping more of the training set average. In real terms, more networks are using the information of a '1' in nodes 14 and 15 as indicators of a positive stimulus than might be expected given that 60% of the time the positive stimuli have a '0' in these positions. Perhaps this is due to the negative stimuli presented during training. Whatever the reason, Johnstone's statement that "the above results thus indicate that when males of a species possess paired ornaments, selection for female recognition of appropriate mates can lead to biases favouring symmetrical males" appears to be rather an unsubstantiated claim. There is no evidence to suggest that behaviour illustrated by artificial neural networks closely models that shown in real systems.

Despite these criticisms of the methods used to demonstrate the possible existence of 'perceptual biases' and 'hidden preferences' in natural recognition systems, the theory remains sound. Recently, 'hidden preferences' have been invoked as a possible cause of the dramatic speciation of the cichlid fish in African lakes (Galis and Metz, 1998), and some cases of biases appear to have been shown experimentally. Basolo (1990, 1991, 1995) demonstrated a preference for swords in swordless platyfish (Xiophophorus maculatus and Xiophophorus variatus), whose genus also includes swordfish who possess such a sword. Phylogenies of the genus disagree as to whether the presence of a sword (and the preference for it) are a primitive characteristic, with the sword subsequently being lost by two platyfish clades; or whether the preference for swords is a pre-existing primitive condition, and that the two swordfish clades have subsequently evolved swords. Ryan & Rand (1993) showed the existence of perceptual biases in Tungara frogs (Physalaemus spp.), where, when the 'chuck' portion of the male call of Physalaemus pustulosus (which stimulated one of the female inner ear receptors) was added to the call of other species, such as Physalaemus coloradorum, the latter was rendered more attractive to conspecific females. Thus it appears that Physalaemus females have a perceptual bias for the 'chuck' which has been exploited by male Physalaemus pustulosus, but which has not yet evolved in other species. Another example of this seemed to be possible when Cook et al. (1994) described a preference in male Papilio dardanus (the African Mocker Swallowtail) for the black and white female morph over the black and yellow morph and the black, orange and yellow/white morph. Given that males show this preference, and yet only one part of the range of one race (the Western area of the range of race dardanus) can be considered to have gone to fixation, with all females being of the black and white (hippocoonides) morph, it must be assumed that the other morphs have some advantage so as to keep the population in a balanced polymorphism. Since the hippocoonides morph is considered to be a Batesian mimic, it could be positive frequency-dependent predator pressure which keeps this balance. Whatever the advantage to the other morphs, given that the polymorphism is balanced the males should theoretically have no morph preference. It appears, therefore, that the hippocoonides females are intrinsically more attractive to the wild males for some reason. I investigate the basis of this in Chapters 3 & 5.

Learning

The psychology of the receiver can be altered in many ways, all of which can affect how signals are perceived and interpreted. For example, the physiological state of the receiver, such as how hungry it is, may make some signals appear to be more obvious or demanding than others - a starving animal may not recognise a predator when it can also see food. One of the most important ways in which a receiver's perception can change is through learning. Predators learn to avoid aposematic animals and form a 'search image' for their prey. Butterflies have been shown to learn to recognise rewarding flowers (see Chapter 4, Introduction). The ability to learn allows a much greater plasticity in response from receivers. Although flower preferences can be innate, the ever-changing range of flowers available to nectar-feeding animals means that a degree of flexibility in preference is highly advantageous. Such an ability may extend to the recognition of individual characteristics, and allow individual recognition in a species. This can be highly advantageous in species where the same individuals may be encountered again (such as in territorial and social animals) as it gives the receiver further, background information about the content of the signal and allows a more sophisticated interpretation of it. I investigate the role of learning in flower choice in Papilio dardanus in Chapter 4.

Returning to the case of Papilio dardanus and mate choice, the preferred hippocoonides morph is also the most common in the race studied. In a polymorphic species it may be advantageous for individuals to learn to recognise the patterns of potential mates so that they can avoid wasting resources chasing butterflies of the wrong species or sex. If this were the case, then a male may learn to recognise the pattern of his first mate and subsequently choose females of the same morph more frequently. Since hippocoonides is the most common morph, experienced males are most likely to have come across this morph as their first mate, and thus have acquired a preference for hippocoonides in their subsequent matings. I investigate the possibility of learning being involved in mate choice in Papilio dardanus and also in the British species Argynnis paphia (the Silver-washed Fritillary) which has two female morphs and in which there is a similar preference for the more common morph (Magnus, 1958).

The study of Papilio dardanus

Papilio dardanus (Brown), the Mocker Swallowtail, is a classic example of a polymorphic and mimetic butterfly. It is found in about 13 geographically distinct races in sub-Saharan Africa, including Madagascar and the Comoro Islands, and throughout its range the males are yellow and black with characteristic 'swallowtails' on the hind wings, with only slight variations in the degree of black patterning. The females, however, are polymorphic, and display many morphs in differing proportions according to the geographical race. Many of these morphs are mimetic, and their distribution corresponds to the distribution of their model. See Appendix 1 for a summary of the races and morphs.

In the early part of the century there was great debate as to how mimicry could evolve according to Darwin's theory of natural selection. It was evident that forms intermediate between the original non-mimetic form and the perfect mimic would have no selective advantage. It also became evident that crosses between mimetic forms never appeared to produce intermediate forms, making it seem unlikely that there were many genes involved, each with small effects. Biologists such as Punnett (1915) and Goldschmidt (1945) argued that evolution proceeded in large jumps, due to mutation, and that natural selection only acted to perfect the mimicry after a likeness had been achieved by random change. Others (such as Fisher, 1927; Carpenter, 1946; and Ford, 1953) believed that random mutation could not have produced so many mimetic species and forms, and their ideas formed a 'two-phase' theory. According to this, a mutation causing any resemblance (particularly in colour) would spread in a population since predators often generalise, and even a slight increase in the chance of avoiding predators would be advantageous. Then there would be selection for 'modifier genes' which perfect the mimicry and which are closely linked, thus intermediate forms due to interbreeding of morphs is avoided. An article by Clarke et al. (1995) details the discussion in much greater detail.

With its many races and morphs, Papilio dardanus provided an excellent opportunity for further understanding of the genetics involved in mimicry. In the 1950s Sir Cyril Clarke and Philip Sheppard began to breed Papilio dardanus in Britain, after discovering that the butterflies can be mated easily by hand. Through many years of breeding and meticulous note-taking, they pieced together the genetics of many races, and published their results in a series of four papers (1959, 1960a, 1960b, 1962). By making crosses between races (especially the non-mimetic races of Madagascar and the Comoro Islands) they not only discovered the main genes involved, but also investigated the presence of local modifier genes keeping the mimicry perfect. The system turned out to be relatively simple - the morphs were controlled by 'supergenes', made up of closely linked genes and their modifiers, and these were ranked in a dominance heirarchy (with some codominance and a few heterozygote forms).

This genetic work is now being followed up by a molecular genetic approach by Vane-Wright et al. at the Natural History Museum, London, where they are trying to understand the evolution of the various races and morphs of Papilio dardanus (Vane Wright et al., in press). There is still a great deal of debate about how the various races and morphs that are seen in the species today came about (see Chapter 7 for a discussion of this), and it is hoped that the molecular genetic analysis will shed light on this. The work in this thesis should complement this phylogenetic work, studying the way in which the evolution may have been driven. In a wider context, the results should illustrate some of the ways in which the psychology of the signal receiver can have a profound effect on the evolution of species and on the evolution of mimicry.

References

Arak, A. & Enquist, M. 1993. Hidden preferences and the evolution of signals. Phil. Trans. R. Soc. Lond. B., 340, 207-213.

Basolo, A.L. 1990. Female preference pre-dates the evolution of the sword in swordtail fish. Science, 250, 808-810.

Basolo, A.L. 1991. Male swords and female preferences. Science 253, 1426-1427.

Basolo, A.L. 1995. A further examination of a pre-existing bias favouring a sword in the genus Xiphophorus. Anim. Behav. 50, 365-375.

Bates, H.W. 1861. Contributions to an insect fauna of the Amazon Valley. Trans. of the Linn. Soc. [Offprint]

Bates, H.W. 1864. The Naturalist on the River Amazonas. J. Murray, London.

Benson, W.W. 1972. Natural selection for Mullerian mimicry in Heliconius erato in Costa Rica. Science, 176, 936-939.

Brower, L.P., Ryerson, W.N., Coppinger, L.L., & Glazier, S.C. 1968. Ecological chemistry and the palatability spectrum. Science, 161, 1349-51.

Carpenter, G.D.H. 1946. Mimetic Polymorphism. Nature 158, 277-9.

Clarke, C.A., Clarke, F.M.M., Gordon, I.J. Mimicry and other controversial topics in East African Lepidoptera. 1995. J. East Afr. Nat. Hist. 84, 3-19.

Clarke, C.A. & Sheppard, P.M. 1959. The genetics of Papilio dardanus, Brown. I race cenea from South Africa. Genetics 44, 1347-1358.

Clarke, C.A. & Sheppard, P.M. 1960a. The genetics of Papilio dardanus, Brown. II races dardanus, polytrophus, meseres and tibullus. Genetics 45, 439-457.

Clarke, C.A. & Sheppard, P.M. 1960b. The genetics of Papilio dardanus, Brown. III race antinorii from Abyssinia and race meriones from Madagascar. Genetics 45, 683-698.

Clarke, C.A. & Sheppard, P.M. 1962. The genetics of Papilio dardanus, Brown. IV data on race ochracea, race flavicornis, and further information on races polytrophus and dardanus. Genetics 47, 909-920.

Cook, S.E., Vernon, J.G., Bateson, M., Guilford, T. 1994. Mate choice in the polymorphic African swallowtail butterfly, Papilio dardanus: male-like females may avoid sexual harassment. Anim. Behav., 47, 389-397.

Dawkins, M. and Guilford, T. 1995. An exaggerated preference for simple neural network models of signal evolution? Proc. R. Soc. Lond. B., 261, 357-360.

Enquist, M. & Arak, A. 1993. Selection of exaggerated male traits by female aesthetic senses. Nature, 361, 446-448.

Fisher, R.A. 1927. On some objections to mimicry theory; statistical and genetic. Trans. R. Ent. Soc. Lond. 75, 269-78.

Ford, E.B. 1953. The genetics of polymorphism in the Lepidoptera. Advanc. Genet. 5, 43-87.

Galis, F. & Metz, J.A.J. 1998. Why are there so many cichlid species?. TREE, 13(1), 1-2.

Gittleman, J.L. & Harvey, P.H. 1980. Why are distasteful prey not cryptic? Nature, 286, 149-150.

Goldschmidt, R. 1945. Mimetic Polymorphism, a controversial chapter of Darwinism. Quarterly Review of Biology, 20, 147-164.

Gordon, I.J. 1987. Natural selection for rare and mimetic colour pattern combinations in wild populations of the diadem butterfly, Hypolimnas misippus L.. Biol. J. Linn. Soc., 31, 1-23.

Guilford, T. 1986. How do warning colours work: Conspicuousness may reduce recognition errors in experienced predators. Anim. Behav., 34, 286-288.

Guilford, T. 1990. The evolution of aposematism; in D.L. Evans and J.O. Schmidt (eds), Insect Defenses - Adaptive mechanisms and strategies of prey and predators, 23-63. State University of New York Press.

Johnstone, R.A. 1994. Female preference for symmetrical males as a by-product of selection for mate recognition. Nature, 372, 172-5.

Magnus, D. 1958. Experimentelle Untersuchungen zur Bionomie und Ethologie des Kaisermantels. Z. Tierpsychol. 15, 397-426.

Poulton, E.B. 1890. The colours of animals: their meaning and use especially considered in the case of insects. London, England.

Punnett, R.C. 1915. Mimicry in Butterflies. Cambridge, England.

Rothschild, M. 1981. The mimicrats must move with the times. Biol. J. Linn. Soc., 16 (1), 21-23.

Rothschild, M. 1984. Aide memoire mimicry; Ecol. Entomol., 9, 311-319.

Ryan, M.J. & Rand, A.S. 1993. Sexual selection and signal evolution: the ghost of biases past, Phil. Trans. R. Soc. Lond. B, 340, 187-195.

Silberglied, R.E. 1977. Communication in the Lepidoptera. In How animals communicate, ed. Sebeok, T.A. Bloomington, Indiana Univ. 362-402.

Turner, J.R.G. 1975. A tale of two butterflies. Nat. Hist., 84, 28-37.

Turner, J.R.G. 1984. Mimicry: the palatability spectrum and its consequences. in R.I. Vane-Wright and P.R. Ackery (eds), The Biology of Butterflies. Symp. Roy. Ent. Soc. Lond., 11, 141-161. Academic Press, London.

Turner, G. F. & Burrows, M. T. 1995. A model of sympatric speciation. Proc. R. Soc. Lond. B, 260, 287-292.

Vane-Wright, R.I. 1980. On the definition of mimicry. Biol. J. Linn. Soc., 13(1), 1-6.

Vane-Wright, R.I., Raheem, D.C., Cieslak, A., Vogler, A.P. In press. Evolution of the mimetic African swallowtail butterfly Papilio dardanus: molecular data confirm relationships with P. phorcas and P. constantinus. Biol. J. Linn. Soc.