Broca's arrow: Evolution, prediction, and language in the brain


  • David L. Cooper

    Corresponding author
    • Krasnow Institute for Advanced Study, Mail Stop 2A1, George Mason University, Fairfax, VA 22030
    Search for more papers by this author
    • Fax: 703-993-4325

    • Mr. Cooper is a former University Scholar at Princeton University, a retired U.S. Army officer, a senior federal bureaucrat, and a PhD candidate in neuroscience at George Mason University in Fairfax, VA. He is the author of Linguistic Attractors: The Cognitive Dynamics of Language Acquisition and Change.


Brodmann's areas 44 and 45 in the human brain, also known as Broca's area, have long been associated with language functions, especially in the left hemisphere. However, the precise role Broca's area plays in human language has not been established with certainty. Broca's area has homologs in the great apes and in area F5 in monkeys, which suggests that its original function was not linguistic at all. In fact, great ape and hominid brains show very similar left-over-right asymmetries in Broca's area homologs as well as in other areas, such as homologs to Wernicke's area, that are normally associated with language in modern humans. Moreover, the so-called mirror neurons are located in Broca's area in great apes and area F5 in monkeys, which seem to provide a representation of cause and effect in a primate's environment, particularly its social environment. Humans appear to have these mirror neurons in Broca's area as well. Similarly, genetic evidence related to the FOXP2 gene implicates Broca's area in linguistic function and dysfunction, but the gene itself is a highly conserved developmental gene in vertebrates and is shared with only two or three differences between humans and great apes, five between humans and mice, and eight between humans and songbirds. Taking neurons and portions of the brain as discrete computational segments in the sense of constituting specific Turing machines, this evidence points to a predictive motor and conceptual function for Broca's area in primates, especially for social concepts. In human language, this is consistent with evidence from typological and cognitive linguistics. Anat Rec (Part B: New Anat) 289B:9–24, 2006. © 2006 Wiley-Liss, Inc.


Paul Broca's 1861 demonstration of linguistic specialization in the left hemisphere of the human brain (Broca,1861a,1861b,1861c,1861d) created a sensation in the medical and scientific world, no doubt enhanced by his presentation of key findings onstage to an audience of fellow scientists. His association of stroke patient Tan's difficulties with the posterior half of the second and third left frontal gyri (circonvolutions) was subsequently refined to Brodmann areas 44 and 45, a refinement very likely assisted by the patient's brain preservation: it can be examined to this day. His clear and compelling correlation of specific symptoms to identifiable structures has likewise been refined and expanded considerably, but remains justifiably famous as a critical tipping point in the history of neuroscience.

Tan's difficulties, which Broca called an aphemia (lack of speech), are now known as Broca's aphasia (also lack of speech, but taken from the verb stem and not the noun). Patients with Broca's aphasia have deficits both in language production and in language comprehension. Tan himself generally only spoke the phrase “tan tan,” which was the source of his nickname, while simultaneously demonstrating by gestures that he could comprehend the language of others. Higher-performing patients generally produce speech telegraphically, in very short bursts, and delivered with a great deal of effort. In terms of comprehension, they have difficulty with repeating or understanding phrases with unusual or complex word order, such as the English passive, and they also have difficulty repeating long and complex words accurately (Gazzaniga et al.,2002). Broca ascribed these problems to “the motor image of the word” (Kolb and Whishaw,1990: p. 580).

While evidence from patients with Broca's aphasia as well as evidence from brain activations indicate that Broca's area is important for processing syntactic information (Caplan et al.,2000), other areas in the brain, such as Wernicke's area, including portions of Brodmann areas 22, 41, and 42 (Just et al.,1996), or the anterior portion of the superior temporal gyrus, Brodmann area 22 (Dronkers,1996), are also implicated. These areas lie along the perisylvian area of the brain, which is a highly conserved area in terms of human genetic expression (Thompson et al., 2000). Their linkage overall has been taken as support for the classic Lichtheim-Geschwind triangle model with a motor processing area (Broca's), an auditory processing area (Wernicke's), and an unlocated conceptual area (Lichtheim,1885; Geschwind,1967). However, their common association with the task of processing syntactic information may also indicate that language processing is a very complex and multifaceted task, and that the triangle model is too simple.

In fact, despite being the first such area identified and probably the one most widely studied, the precise correlation of Broca's area to language and language processing is still a matter of some controversy. This is exacerbated considerably by the fact that Broca's area in humans corresponds to similar brain regions in primates. Moreover, the famous FOXP2 mutation in humans that reproduces symptoms very similar to Broca's aphasia also affects much wider areas in the human brain, while FOXP2 has nearly identical homologs not only in primates, but in vertebrates generally (Vargha-Khadem et al.,2005).

This article will trace these various structural, functional, and genetic connections in an effort to see if there are indeed any underlying unifying threads and to bound the contribution of Broca's area as precisely as possible. To do that, it will organize the available anatomical, genetic, behavioral, and functional evidence within a computational framework very similar to the one David Marr used to great effect in his study of vision nearly 3 decades ago. This is summarized at Table 1. This will perhaps help to bring some of the earlier issues into focus and define precisely what Broca's area is “for.”

Table 1. An implicit computational framework
 Computational TheoryRepresentation and AlgorithmHardware Implementation
  1. This is inspired by the computational framework in Marr (1982: 25) Marr's categories appear in the title row. Corresponding questions from the perspective of implicit computational steps appear in the second row. This perspective looks at neurons, their subcomponents, and their ensembles within neural networks as states within a Turing computational process. The transitions between states, equivalent to Turing operational steps, are performed by cellular and network operations. This implicit computational perspective seeks to construct and perform Turing machine-like functions without an explicit program written beforehand.

Implicit Turing ComputationWhat is the goal of the computation in terms of Turing machine configurations and operations?What are the states and possible transformations available to the neuron or neural network cell assembly Turing machine? Discuss developmental stages as well as functions of the mature organism.What are the structural and procedural elements used by the cell or neural network cell assembly?

Marr employed a general framework to help focus his work on human visual processing, but its computational focus was not on the detailed implementation within a neural network but on the transformation of image intensity maps into three-dimensional object representations using the full panoply of mathematics and digital computational power available at the time. While human brains developed all these things, in their natural state none of this would be available, nor is any of it available to the other species that share homologous structures. Like all other brain areas, Broca's area must acquire and refine its functions solely from the organism interacting with its environment.

Table 1 attempts to reformulate Marr's framework in a way that is closer to Turing's initial computational framework (Turing,1936). The goal of the computation will remain very much the same—a strategy to handle inputs and transform them to the appropriate outputs—but representations, algorithms, and physical implementation in Turing's terms require that we specify initial and final configuration states, a system of signals (Turing used the word “symbol” but that will have a different implication below), and a set of operations that act on the configurations in the context of the signal at hand. If these conditions are met, it is possible to specify an initial configuration (the initial state and an active signal) and predict the subsequent behavior (an operation bringing about a final state). This depiction in terms of states, signals, and operations can be used within a neuron, as well as across collections of them—looking at channel distributions, skeletal proteins, and reactions to electrical or chemical signals at the cellular level, and at connectivity properties, priming, resonance, firing, and connectivity adjustments for neural ensembles. This permits valid inferences about computation within the brain without imputing activity to elaborate processes or even explicit knowledge that may not exist except as an artifact of human civilization.

In this framework, the first question about Broca's area then becomes the final one of identifying its role within an implicit computational strategy. In the discussion below, the nature of the evidence is best suited to working through these issues from right to left, from hardware to strategy.

Figure 1 provides a context to provide more focus to this series of questions. It exploits the ideas of symbolic representations from Peirce (1868) and cognitive linguistics from Langacker (1998) and others and tries to lay out the steps necessary for the functions of learning, inference, and symbolic representations that we associate with language. The following discussion will attempt to identify where the computations made by Broca's area fall, and how much of the diagram they cover.

Figure 1.

Conceptual Sketch of the Computations Involved in Learning and Inference. Learning, and its computational underpinnings, begins with external entities, which can be distinguished by the senses. External entities may be parts of a system, or a bounded set of entities and the rules that relate them to each other. Representations of external entities are also entities, which are distinguishable internal states. The internal entities that represent external entities are indices. When these indices are combined to reflect—or represent—the same rules that bind the external entities, they become symbols. This means that symbols form internal representational systems. These sets of relations are encoded as linguistic construction, which are further explained in Figure 2. For computational purposes, each index is characterized by a sensory measurement, and organized into data sets. Algorithms performed on data provide inferences, which, when validated, become knowledge. Learning occurs when this knowledge changes the set of indices. There is a parallel linguistic chain (depicted by dashed lines), where the data is interpreted (now called information), and this information is encoded as description or concepts, which can form hypotheses, which can be validated and become knowledge as well. Information is also the set of interpreted data that provides the validation in the linguistic chain.

Figure 2.

Causal Chains and Linguistic Phenomena. Perception of causation requires an initiator and an endpoint, with no presumption of telekinesis. For mental events, which require a “theory of mind,” it is possible for a mental stimulus to affect another mental state. A mental initiator can also affect a physical endpoint, which may initiate other physical effects. Similarly, physical causes can have an influence on mental states, or, sometimes by an intermediate physical mechanisms, they also affect a physical endpoint. These casual relationships are reflected directly in how human language describes the equivalent events. Thematic roles are explicitly encoded in various human languages. The causal role immediately precedes the event described by the verbal segment. Passive agents are also expressed external to the verb segment when there is a passive construction. “Subjects” are logical initiators of verbal actions. Their expression may vary by language. For example, in accusative languages such as English, or other European languages, this is uniquely coded for the verb “X.” The comitative role is performed by the entity that participates in this causal chain as the initiator. Means, manner, and instrumental roles are informative about how the verbal action is implemented. The “object” is the logical recipient of the verbal action. In ergative languages, this role receives the same marker as the “subject” in intransitive sentences, coded as an absolutive “X.” Thus, across languages, either end of the verbal segment can receive the focus of the coding system. Results, benefactive (good result), malefactive (bad result), and recipient roles follow the span of the verbal segment. The observation of physical causation, which also corresponds to basic verbal descriptions, is common to the primate goal-oriented perception in mirror neurons.

Figure 1 depicts entities that can be external or internal to the organism. Internal entities are representations. Indexes are internal representations that distinguish between external entities. When external entities form a system of some sort, the corresponding indexes provide the basis for a reference framework for them. When the internal representations form a corresponding system of rules and relationships to the external entities, we have symbolic representations, following Peirce. Cognitive linguistics posits that language is built up from symbolic ensembles, which tend to be expressed as sets of constructions rooted in external experience and reinforced by intensive and focused interactions between adults and children as language is acquired (Tomasello,2003). The diagram shows both nonlinguistic inference (in solid lines) and linguistic processes (in dashed lines). Both pathways rely on characterizations of external entities by means of measurements, organized into data structures. Both pathways feed an inferential cycle and ultimately result in the updated representations that amount to learning.


In Marr's framework, the first step in understanding a given computational role is to establish what the computation is “for.” However, for Broca's area, that was the puzzle Broca first brought up, and the issue that has yet to be resolved. Instead, the process must be inductive and begin with the “hardware,” or the neural anatomy underlying what Broca's area does. The best line of evidence to determine this comes from primates, beginning with humans, as they are the primates who speak.

In humans, Broca's area is correlated with Brodmann areas 44 and 45, including portions of the pars triangularis and pars opercularis of the inferior frontal gyrus of the brain. This lies across the anterior ascending ramus of the lateral, or Sylvian fissureopposite the temporal lobe, which is an auditory association area. It is adjacent to the premotor cortex, appropriately near the areas that control oral and facial features, particularly the lips and tongue. Many descriptions refer to the perisylvian area of the brain, which is a shorthand term for the areas on either side of the lateral fissure, including Broca's area toward the front and above the fissure, and Wernicke's area toward the back and below it. When studying genetic expression in fraternal and identical twins, Thompson et al. (2001) showed very clearly that the perisylvian area is highly conserved in both groups, whereas many other regions coexpressed in identical twins show divergent genetic expression in the fraternal twins. Clearly, Broca's area is appropriately placed for a linguistic function, is adjacent to other areas that are also appropriately placed for other aspects of linguistic processing, and seems to be part of the brain where the genetic code is highly conserved, indicating an important role in individual survival.

Wernicke himself, while describing another kind of aphasia that is the polar opposite of Broca's, also noted the direct connection of the area now known as Wernicke's area to Broca's area by way of the angular gyrus and the arcuate fasciculus. Because Wernicke's area is adjacent to the auditory cortex in Heschl's gyri, he ascribed an auditory memory role to it. Wernicke's aphasia describes a condition where the patient is fluent, but produces nonsensical insertions of word forms (Wernicke,1874). The two kinds of aphasia are contrasted here in examples drawn from Gazzaniga et al. (2002: p. 385–387):

Broca's aphasia:

Spontaneous speech: “Son … university … smart … boy … good … good.”

Listening—Prompt: “The boy was hit by the girl. Who hit whom?”

Response: “Boy hit girl.”

Repeating—Prompt: “Chrysanthemum.”

Response: “Chrysa … mum … mum …”

Wernicke's aphasia:

“I called my mother on the television and did not understand the door.”

Thus, we have physically connected areas in the brain, where damage to the endpoints corresponds to polar opposite dysfunction. At one end, Broca's, the patient makes sense but has much difficulty with even short sequences, while at the other end, Wernicke's, the patient is completely fluent but makes no sense. This highly connected region corresponds to the perisylvian region as well.

Using a new statistical method called voxel-based lesion-symptom mapping, Bates et al. (2003) provides strong support for a functional connection along this physical pathway. Summing across 101 patients with lesions on the left side of their brains, they found that those who had difficulty with fluency in language showed a very high correlation between the symptoms and their associated lesions along the gray matter connections in the insula and the arcuate fasciculus lying between Wernicke's and Broca's areas. Similarly, they found the highest correlation of symptoms to lesions for language comprehension in the middle temporal gyrus, in the vicinity of Wernicke's area. These observations essentially confirm the association of Broca's aphasia to fluency and Wernicke's aphasia to word comprehension, but it is very interesting to note that the highest symptom-location correlations for fluency lie on the pathway itself and posterior to Broca's area in the cortex, while the correlation of cortical lesions to Wernicke's aphasia is much stronger. This pattern pointing to the pathways and connections to Broca's area as well as the cortical area itself will be repeated when we turn to the FOXP2 gene later.

As for Broca's area itself, using a novel observer-independent method for measuring cell densities, Amunts et al. (1999) focused on the detailed structure of the cortical layers in 10 human brains, taking thousands of profiles across five male and five female subjects. They reported a number of key observations, confirming, for example, the left-biased lateralization that is normally imputed to language functions. Interestingly, while all five male subjects showed a left-over-right asymmetry for area 44, which is adjacent to area 6 in the motor cortex, only three of the female subjects did. Neither group showed a similar asymmetry for area 45, which is further forward than area 44 and borders on frontal areas of the cortex such as area 10. Their other findings also show clear structural differences between areas 44 and 45, as well as significant differences between subjects. Thus, the size of area 44 on the left could vary by a factor of 10 between subjects, while the differences between subjects for both areas were greater than the differences between hemispheres in a single subject. Of particular concern to anyone conducting brain imaging studies, they reported the absence of any large-scale landmarks that definitively identify Broca's area or these subdivisions.

Amunts et al. (1999) also described some cellular details in their descriptions of the layers in areas 44 and 45. Both areas have conspicuous large pyramidal cells deep in layer III and in layer V. Layers II and III do not have a distinct boundary, and layer VI has a very low cell density. Layer IV is almost nonexistent in area 44 and described as “dysgranular.” In area 45, layer IV is much more clearly visible, but less distinct than adjacent cortical areas, such as area 10. Area 44, which shows considerable left-over-right asymmetry with respect to its volume, also shows a left-over-right asymmetry in cellular density: the cells in the left hemisphere had a smaller separation than those on the right. Here again, area 45 did not show a similar asymmetry in cellular separations. Amunts and colleagues also confirm earlier observations that areas 44 and 45 are further subdivided (Economo and Koskinas,1925), but were not able to provide definitive maps of the subareas.

The tremendous variation in the brain volume in area 44 is very interesting because its absence in area 45 suggests some kind of functional difference between the two areas. It is also interesting, however, because both portions of Broca's area are conserved in genetic expression, both for humans and in primates generally. Highly localized variation of this magnitude would probably not be the result of noise in genetic transcription, which one would expect to have a more regular pattern across affected areas in individuals. If, on the other hand, it is related to noise in genetic translation, it might indicate that this variation depends on persistent activity subsequently reinforced by a learning process and hence rule out a predetermined structure for area 44.

The lack of major unambiguous landmarks delineating Broca's area and its subdivisions should provide a major warning when trying to understand imaging data. Without resolution sufficient to measure cellular densities in individual cortical layers, we simply cannot distinguish between areas 44 and 45 from image data, much less one of their anterior or posterior subareas, and precise boundaries between areas 44 and 45 and regions outside Broca's area are equally indistinct. The only other method to calibrate brain images to this level would be to create thousands of microscopic cortical profiles dissected from subjects who had previously been used in the imaging experiments—the ultimate longitudinal study. Basically, we can only make general conclusions broadly associated with Broca's area from images when using current imaging technology.

Nevertheless, it is possible to learn a great deal about Broca's area even within these constraints. For example, there are homologous regions in other primates, which clearly provided some selective advantage other than human language production. An area equivalent to area 45 is present in monkeys (Preuss and Goldman-Rakic,1991). Chimpanzees have equivalents to areas 44 and 45 (Carroll,2003), and a recent cytoarchitectural and electrostimulation study revealed an equivalent to area 44 in macaques as well (Petrides et al.,2005). Moreover, the Broca homolog in great apes shows the same kind of left-over-right asymmetry found in area 44 in humans (Cantalupo and Hopkins,2001).

Consequently, this left-biased asymmetry clearly cannot be related just to human language. In fact, there is evidence for left-hemispheric dominance in all the early hominins, as well as modern chimapanzees, bonobos, and gorillas (Hopkins and Leavens,1998). This dominance corresponds to right-handedness, pointing gestures, and vocalizations. Similarly, there is a parallel left-asymmetrical extension of the planum temporale, adjacent to Wernicke's area in humans, found in Homo habilis, Homo erectus, and Homo neanderthalensis (Holloway,1980). This left-asymmetric extension is also found in chimpanzees (Gannon et al.,1998; Hopkins et al.,1998). Thus, the asymmetry existed throughout the perisylvian region, including the areas homologous to Broca's, prior to the development of language capable of symbolic representation, and correlated with nonlinguistic factors.

However, these left-biased asymmetries are not identical across primates. For example, left lateralization in Broca's area in great apes is evident down to the level of fine structure (Amunts et al.,1999), yet the lateralization in gross structure of the planum temporale does not correspond to a similar asymmetry in minicolumn size and connectivity in chimpanzees, while it is found in humans (Buxhoeveden et al.,2001). This produces an overall pattern that implies that the perisylvian area has been important to primates for millions of years, based on gross structure. On the other hand, fine-structure asymmetry, which probably has an activity-related component, appears in great apes for Broca's area at one end of the perisylvian region, while it does not at the other end, at the planum temporale. This would make the fine-structure changes near Wernicke's area, the planum temporale, and the auditory cortex in general more likely locations for language-specific changes than Broca's area, where fine-structure changes already took place before the appearance of language.

What does Broca's area do in nonspeaking primates that makes it important for survival? Kohler et al. (2002) shed some important light on this question by examining the firing patterns of individual visuomotor “mirror neurons” in area F5 of macaques, the homolog to Broca's area in humans. These mirror neurons were already associated with action-related perception that required viewing both an agent, such as a hand or a mouth, and an object manipulated by the agent (Gallese et al.,1996). These neurons are thought to be important in planning and execution of movement. Kohler et al. (2002) showed that these neurons are multifunctional and also react to the sounds produced by objects on which the monkey performed an action.

It is easy to understand how an area that links all of these functions might be important to individual survival. Gestures are important to social animals, and this area is well situated to contemplate and plan such gestures. Understanding agent-object relationships is also important in resolving objects of interest in the environment, as well as predicting what might become of them if they are manipulated. The auditory connection would make this more powerful, as such a capacity provides a multidimensional representation of sensory data, as well as feedback in both visual and audio modes.

Mirror neurons may have an even closer tie to linguistic performance than multimodal correlation. These neurons in area F5 appear to code goal-oriented movement of the hand and mouth (Rizzolatti and Camarda et al.,1988; Murata et al.,1997; Rizzolatti et al.,2000). Some of these mirror neurons are highly specific, coding particular types of grasping movements, for example, but most of them are active under much broader sets of stimuli and appear to generalize across classes of particular instances (Rizzolatti et al.,2001). Moreover, understanding of these gestures may very well be accomplished by mapping the visual representations onto their motor representations, creating a type of “motor knowledge.” Importantly, there is very good evidence, both direct and indirect, that humans have mirror neurons, and that Broca's area responds similarly to area F5 when humans undergo experiments on arm and hand actions (Grafton et al.,1996; Rizzolatti et al.,1996; Decety,1997; Grèzes,1998). These experiments show the left-over-right asymmetry associated with primate anatomy already noted and associate Broca's area very clearly with “meaningful” rather than “meaningless” gestures.

Returning briefly to Figure 1 and jumping ahead to the question of the computational purpose for Broca's area, this evidence helps to localize the particular functions that might be performed there in monkeys, great apes, and humans. First, Broca's area is clearly a place where multimodal information converges: visual, auditory, and motor. In Figure 1, that makes it a candidate for applying algorithms of some sort to juxtaposed data sets, both along the verbal and nonverbal paths. That leads to inferential or predictive steps that precede the acquisition of new knowledge and the completion of learning steps. In assessing the evidence for and against competing hypotheses related to the primate mirror system, Rizzolatti et al. (2001) point out that the main weakness in the “visual hypothesis”—whereby actions are understood solely on the basis of their visual inputs, without reference to motor representations—is that there is no mechanism for validation of the meaning of the observed action. By contrast, “motor knowledge” provides the mechanism for validating and understanding gestures under the “direct matching hypothesis.” Thus, we have a prediction step and a validation step, which are adjacent in Figure 1, as well as adjacent in the flow of data in the brain: In the case of Broca's area, the inputs to the adjacent Brodmann's area 6 in the premotor cortex as well as the subcortical areas that coordinate fluent, multicomponent motor activity in subcortical areas.

These same areas will be prominent when we examine the expression of the FOXP2 gene next: another element in the “hardware” box in Table 1. In this case, because the brain employs physical representations and because these are laid down during development, the patterns of gene expression will help to outline the kinds of representations and algorithms associated with Broca's area.


In reviewing evidence from genes, it is often necessary to distinguish between the gene itself, which consists of a string of nucleic acids, and the products of that gene, expressed as amino acids and peptide chains. Much genetic evidence comes from animal studies, so it is also useful to distinguish between the human and nonhuman forms of a gene and its products. There is a shorthand to convey these distinctions, summarized in Table 2. Italics appear when we are talking about the gene itself. Human forms of genes and their products are always in upper case. Table 2 uses the FOXP2 gene as an example, since it is the gene that will concern us the most.

Table 2. Shorthand expressions for genes and gene products
 GeneGene Expression (Protein)
  1. Genes are in italics and human forms for genes and gene products are in upper case. Here FOXP2, a gene product, would have an extra site for protein kinase C phosphorylation, while FoxP2 would not. FOXP2 would code for that site, while FoxP2 would not.


The FOXP2 gene was isolated thanks to a point mutation in the KE family in which afflicted members have problems with fluency and grammar. However, just as the parallel anatomical patterns and behaviors related to Broca's area in great apes and area F5 in monkeys rule out an exclusive tie between Broca's area and human language, the evidence also excludes an exclusive tie between FOXP2 and language. It is highly conserved, among the 5% most highly conserved genes in human-rodent pairings, for example, which is consistent with brain areas with ancient associations. Moreover, the human form of the gene differs from the chimpanzee version at only two amino acids. The two changes occurred in the last 4 to 6 million years, after the branching of the hominid line from its common ancestor with the chimpanzee. This is twice the expected mutation rate, providing further evidence for intense evolutionary pressure on human ancestors at that time. In fact, these two changes, in association with colocated alleles on the seventh chromosome, show evidence of an evolutionary “sweep” no more than 200,000 years ago (Enard et al.,2002). That is, the evidence points to a small, important change related to Broca's area, language, and to the brain.

Just as the differing evidence of left-biased asymmetry across humans and great apes for Broca's area and the planum temporale indicates that more than one set of changes underlies the emergence of language, evidence of development in hominins and great apes also shows that humans emerged from more than just one change, no matter how important. Compared to chimpanzees, for example, humans have an immature skull shape and size. This accounts for its relatively large size in humans, but it is unlikely that these differences stem from a single source. For example, human and chimpanzee growth rates differ substantially through adolescence (Gould,1977). Humans likewise developed more slowly and had more immature features than the earlier hominids (Dean et al.,2001; Rice,2001), and differences between humans and Neanderthals also arose early in child development (Ponce de León and Zollikofer,2001). Generally, comparative evidence indicates a mosaic pattern of developmental traits, and not a simple change of rates or the acquisition of a single new trait (Moggi-Cecchi,2001). Similarly, language is likely the outcome of a mosaic of changes as well, but those associated with FOXP2 were clearly important.

FOXP2 belongs to a family of “forkhead box” proteins, which regulate the expression of their respective DNA sequences by means of a three-winged helical structure (Carlsson and Mahlapouu,2002). FOX genes invariably have arginine at site 553, while mutant FOXP2, as found in the KE family, substitutes histidine, which is adjacent to another histidine in the third helix (Lai et al.,2001). The analogous mutation at that site in FOXC1 causes a critical loss of function (Saleem et al.,2003).

As noted earlier, the structure of FOXP2 is highly conserved. There are only two differences between humans on the one hand and chimpanzees and gorillas on the other (threonine to asparagine at site 303 and asparagine to serine at site 325), three between humans and orangutans, and five between humans and mice (Zhang et al.,2002). Of particular interest when we take up comparisons of functions at Broca's area to similar functions in songbirds later, there are only eight differences between humans and zebra finches, making the protein 98% identical (Haesler et al.,2004). The human-specific change at site 325 probably created a substrate for phosphorylation by protein kinase C (Enard et al.,2002). This particular prediction is based on an artificial neural network method that can estimate a protein's structure with an accuracy that exceeds 70% and generally lies between 75% and 82% (Rost,1996; Sun,1997). This new substrate may well be the small but important change that enabled Broca's area and related areas, particularly in the limbic system and cerebellum, to function as they do in language. On the other hand, a new phosphorylation substrate may simply be an example of the general upregulation of gene expression in the brain that characterizes the differences between humans and chimpanzees (Preuss et al.,2004).

FOXP2 is one of a subfamily of FOXP proteins that has at least three other members. Proteins in the subfamily contain four signature domains: a DNA-binding winged helix, a leucine zipper, a zinc finger, and a polyglutamine tract (Wang et al.,2003). All of them appear to be highly conserved, with the winged-helix the most divergent in structure. Generally, all seem to function by repressing genetic transcription, especially during development. The leucine zipper appears to foster dimerization of Foxp proteins, and both DNA binding and dimerization may be required for these proteins to function. Possibly as a result, Foxp1 and Foxp2 are expressed in mice in different ways in the epithelial tissue of airway branches in lungs, in motor pathways in the brain, in the outer mesoderm of the intestines, and in the outflow tract of the atria of the heart (Shu et al.,2001). Foxp1 and Foxp3 are coexpressed in lymphoid cells. Foxp4 overlaps the Foxp1/Foxp2 pattern in lung, intestine, and neural tissues as well, which further supports a complex pattern for transcription control (Lu et al.,2002).

The zinc finger domain that appears in Fox proteins in all species with Fox genes can modify repression in either direction. For example, while a zinc finger domain normally provides one of the mechanisms for repression, when tested in yeast cells, this domain seems to do the opposite: Foxp2 transcription activity triples when fused to the GAL4 binding domain (Li et al.,2004). The polyglutamine region, larger in Foxp2 than Foxp1, seems to modulate repression activity as well. The phosphorylation site in human FOXP2 provides yet more capacity for varying the activity of the forkhead domain. In general, this set of complex interactions, whether by various combinations of dimerization, or by other means of cooperative or antagonistic control, is representative of the forkhead box family of genes. In fact, the number of types of Fox genes in an organism is directly correlated with that organism's complexity: humans have more than 40 kinds of FOX genes overall. They are related to a wide variety of developmental disorders, including the linguistic difficulties of the KE family stemming from a mutation in FOXP2 (Carlsson and Mahlapuu,2002).

Keeping this evidence in mind, we can turn to the representational level in Table 1. During development, genes cause the basic structures in the brain to arise, where specific connections between inputs and outputs are laid down, and where adjacent regions provide the probability for lateral connectivity as the organism matures. Since sense organs measure specific types of input data, and connected or adjacent areas are primarily affected by these specific information types, this essentially organizes inputs into data sets, thereby providing the physical framework for information processing—the basis of implicit computation of sensory data.

Gene transcription is subject to two principal sources of noise that can affect the gene regulation function: noise internal to the cell and external noise. Quantitative assessment of these two sources indicates that the internal noise is subject to very rapid decay, so that the principal influences on single gene regulation are the biochemical factors that trigger gene transcription, together with external noise and only those internal factors that change slowly relative to an entire cell cycle (Rosenfeld et al.,2005). In gene networks, including the gene cascades that proceed as a cell develops, even the small perturbations at the local noise level, however, can have major effects downstream (Elowitz et al.,2002; Pedraza and van Oudenaarden,2005). Translational noise is also a source of variation in phenotypes, so the combination of these factors probably led to selection pressure in favor of inefficient translation of genes even when efficient translation would consume less energy, since inefficient translation would lead to lower fluctuations (noise) in protein concentrations in the cell. One well-known example of this inefficient translation is cyclic AMP (Ozbudak et al.,2002).

At the same time, cell-to-cell variations can depend on noise at the transcription level, so that cell populations demonstrate extended bistable states in gene expression. Thus, noise in genetic cascades may very well play a significant role in cell phenotype variation, or cell differentiation as well (Blake et al.,2003). Taken further, this means that incorporation of noise in regulation is an evolvable trait that can help maintain a balance between the fidelity of gene expression and in creating cellular diversity (Raser and O'Shea,2004). Autocatalysis, or positive feedback, demonstrably contributes to control of cellular function in the context of these bistable states (Becksei et al.,2001), while negative feedback tends to function in homeostatic adjustments (Becksei and Serrano,2000) and can produce effects five times faster than when negative autoregulation is absent (Rosenfeld et al.,2002). Bistability of this sort essentially creates an all-or-nothing switching mechanism, illustrated in another well-known example by the activity of MAP kinase in the cell cycle (Ferrell and Machleder,1998).

What are some potential implications for FOXP2, which is active during the development of brain structures, and which normally functions by repressing genetic transcription? In the cortex, the first layers to form from the neural preplate are the lower layers, then the upper ones (Sidman and Rakic,1973; Marin-Padilla,1999). FOXP2 appears only in layers IV to VI (Maviel et al.,2004), so it is expressed only during the first half of that process and thus affects the layers that communicate outside the cortex. In Broca's area, layer V contains numerous and noticeably large pyramidal cells (Amunts et al.,1999). If mutant FOXP2 fails to trigger the proper conditions for layer V, this would have implications for the establishment of the requisite connections to other brain areas, as well as the composition of the layer itself. This may explain the relatively low levels of gray matter to the limbic system in members of the KE family affected by the gene mutation. The areas of the limbic system and the cerebellum affected by the mutation may also be subject to similar failures in switching from one stable state to another during development.

Genes also control the migration of cell types during development, thereby contributing to the fine structural detail of the brain's implicit computational framework. In the formation of the six-layered cortex, for example, the DCX protein appears to play a key role. It is located preferentially in the growth cones of developing neurons and activated by c-Jun N-terminal kinase (JNK) phosphorylation (Gdalyahu et al.,2004). Different mechanisms are involved for the migration of neurons when guided by glial cells. In this case, the control mechanism appears to involve the Par6α protein, as well as protein kinase C (PKCζ, in this case) and γ-tubulin as a cytoskeletal component (Solecki et al.,2004). This is the kind of mechanism that is likely to reveal what is happening with FOXP2, because the normal human form of the gene adds a phosphorylation site for protein kinase C. The involvement of protein kinase C is potentially quite significant, as its presence would shift the noise parameters associated with the gene and possibly create new possibilities for migration as well as cell differentiation. Protein kinase C (PKCγ) is critical in a developmental cascade in cerebellum when climbing fiber cells are culled to leave one climbing fiber synapse to one Purkinje cell in circuits that control fine motor movement (Kano et al.,1995).

While genes, by virtue of their control of cell differentiation and migration, set down a framework for data processing that specifies what data types interact with each other, the neural constituents of the brain must take that input, use it, and learn from it. Development sets the conditions, while the components of the active brain must then follow through to create an individual's actual competence at any given task. This is where Broca's area and possibly its subregions and even individual neurons would play a computational role.


The timescale for this activity also shifts to both faster and longer-term processes. Migrating cerebellar neurons have average velocities of 3–4 μm/min (Maviel et al.,2004), while migrating cells in rat neocortex could achieve their final position in several days (Bai et al.,2003). By contrast, the actual processing based on the results of all this movement and positioning takes place in tens and hundreds of milliseconds, while learning processes set in over different time courses from minutes to days, months, and years.

The processes at this level are very different from genetic regulation, of course, but they also are essentially stochastic and lead to very similar patterns of stability. For example, sensory changes that affect potassium conductances can shift Purkinje cells from a bistable spiking mode to either of two single stable spiking states (Loewenstein et al.,2005). Bistable behavior is also easy to find in assemblies of neurons. Networks with N-methyl-D-aspartate (NMDA)-mediated recurrent synapses, consisting of pyramidal cells and interneurons that provide the feedback, show bistable behavior that shifts to a single stable state either when γ-aminobutyric acid (GABA) conductance prompted by the interneurons shifts above a critical threshold, or when α-amino-3-hydroxy-5-methyl-4-isoxazole propionic acid (AMPA)-mediated conductance shifts the state of the network in the other direction (Lisman et al.,1998).

Models of working memory consisting of recurrent excitatory networks with simpler leaky integration neurons also produce this behavior (Durstewitz et al.,2000). Other studies of bistability in networks show that the hysteresis loops that create these bistable-reactive regions also allow sensitivity by the network to its recent history on input patterns (Pouget and Latham,2002). In terms that would match a Turing-like framework for network computation, maximization of Gaussian mutual information in the presence of noise turns out to provide the stable computational behavior and sparse coding found in biological neural networks as well (Linsker,1993). In this case, the network's problem is simply to capture the characteristics of an input array using local processes, as we would require within the implicit Turing framework.

Activity-dependent processes in excitatory neurons show just this kind of bistable behavior in experimental studies of CA1 hippocampal neurons (Lisman et al.,2002). While the AMPA and NMDA receptors in these neurons function at timescales of far less than a second, NMDA receptors, anchored by PSD95 connected to AMPA anchored by assemblies containing actinin, actin, and SAP97, all phosphorylated by calcium/calmodulin-related kinase II (CaMKII), provide an energy-efficient, bistable “switch” that persists on the order of days. At normal local concentrations of Ca2+, this configuration, which relies on an autophosphorylated state of CaMKII, shows only 10% dephosphorylation of the kinase after 45 hr (Lisman and Zhabotinsky,2001).

Processes that persist longer than this would presumably require the rearrangement or reconfiguration of synapses, rather than adjustments at the channel level. However, this NMDA-AMPA/CaMKII mechanism would be a reasonable one to impute to the numerous and prominently large pyramidal cells in layer III in Broca's area. Layer III is also predominantly involved in remote, rather than recent, memories (Frankland and Bontempi,2005), so activation of these neurons in language tasks would probably account for the fMRI patterns that can distinguish native from learned second languages.

Whether the other structural changes to synapse formation or transformation are needed for long-term memories is still an open question. If they are needed, some form of genetic translation would probably also be required, if not transcription as well. Incidentally, FOXP2 is not implicated in any of these functions, as FOXP2 is not expressed in layer III. Perhaps this is the reason that the KE family members with the FOXP2 mutation show fewer difficulties with comprehension than with speech production (Belton et al.,2003b).


Much like humans, many songbirds learn their vocal patterns by copying the vocal cues they hear. Unlike humans, birds do not have a six-layered cortex, so there is no question of an exact homolog to Broca's area, but they do express Foxp1 and Foxp2 in a manner strikingly similar to human fetuses (Teramitsu et al.,2004). Instead of a cortex, birds have a pallium, from the Latin for a cloak, which describes how it forms a cover over the subpallial areas that bear somewhat closer correspondences to mammalian subcortical areas. The parallel structures for bird song are located in avian pallial and subpallial areas, as well as the homologous sets of nuclei within the dorsal thalamus. These avian brain areas help provide sensorimotor integration, as well as skilled, coordinated movement, strikingly similar to the cognitive and motor capacities related to Broca's area. Since the eight differences between Foxp2 in zebra finches and FOXP2 in humans include sites 303 and 325 (zebra finches have the same amino acids as chimpanzees), there is no question that the human phosphorylation substrate for protein kinase C provided the mechanism for parallel developmental and behavioral traits, but the other changes in these birds may have resulted in similar results by alternate means.

In humans, there are both cognitive and motor pathways associating the cortex with the basal ganglia and the cerebellum (Middleton and Strick,2000). Both are implicated in KE family sufferers from the FOXP2 mutation (Vargha-Khadem et al.,2005). The principal subcortical areas implicated in human cognitive and motor circuits by FOXP2 expression are the caudate nucleus and putamen, the substantia nigra pars reticulata and globus pallidus internal segment, as well as the medial dorsal, ventral anterior, and other nuclei of the thalamus on the cognitive loop, while the cerebellum (lobules VIIB, VIIIB, as well as the inferior olivary complex and red nucleus), the dentate nucleus, and the medial dorsal, ventral lateral, and other nuclei of the thalamus are on the motor loop. It is exceedingly interesting to note here that knockout mice lacking a form of protein kinase C (PKCγ) that occurs in climbing fiber cells are capable of learning simple motor skills, but are impaired in the smooth coordination of those same skills, such as in walking or balancing on a narrow object (Chen et al.,1995). With these mice, the usual paring back of synapses between climbing fiber cells and Purkinje cells does not take place during development, so that more than one climbing fiber cell will have synpases with the same Purkinje cell in adults. In normal mature mice, there is a one-to-one correlation. This is an intriguing correlation of the natural repression of multiple neural pathways during development, protein kinase C (implicated by site 325 in FOXP2, possibly to added effect), and the smooth coordination of compound motor activity, such as that required for fluent speech.

As for songbirds, there are evident correspondences between these subcortical areas implicated by FOXP2 and the subpallial areas involved in the avian song cycle (Jarvis et al.,2005). For example, the anterior (cognitive) loop in both involves the striatum and thalamus. In birds, it contains the lateral area X (LAreaX) in the striatum, which passes signals to the dorsal lateral nucleus of the medial thalamus (DLM). In zebra finches, Foxp2 is expressed in area X during the critical period for song learning. In adult canaries, it is expressed in area X seasonally, when song production is unstable. Its expression in other birds varies similarly, indicating its association with vocal plasticity (Haesler et al.,2004). Similarly, the motor loop also contains the thalamus, in this case the nucleus uvaeformis.

Other correspondences are more notional, but the subcortical analogies to avian subpallial areas seem very strong. To extend the analogies into the equivalent of the cortex for songbirds, there are also areas in the avian pallium that participate in both the auditory and motor pathways—the higher vocal center (HVC), and the robust nucleus of the arcopallium (RA). Another key area in the cerebrum on the cognitive loop is the lateral magnocellular nucleus of the anterior nidopallium (LMAN). These may play the computational role of Broca's area. They are involved in the moment-to-moment modulation of syllables in the songs of zebra finches (Kao et al.,2005).

Birds are also capable of acquiring “syntax.” White-crowned sparrows, when exposed to their native song in two syllable phrases, were able to learn the entire song sequence despite never hearing the entire song from end to end. Exposed to the song when the syllable pairs were in reverse order, the sparrows learned the song backward (Rose et al.,2004).

To summarize briefly, Broca's area was first identified with language deficits related to sequences (syntax) and motor control. It has structural homologs in primates, where the area is associated with execution of motor functions, especially gestures, and cross-modal sensorimotor perception. It has a functional homolog with songbirds, which learn and execute complex acoustic sequences. A point mutation in the FOXP2 gene also ties the area to a gene that is expressed in a highly conserved form in all these animals. That gene is normally a developmental repressor intimately involved in the development of complex pathways in the brain, intestines, lungs, and heart. In the brain, it shows a consistent pattern of expression in the cognitive and motor-related area in the basal ganglia or their equivalent.

Thus, Broca's area is not just for language, any more than the FOXP2 gene is, although the human form of the gene may well contain a small but important shift that was necessary for language. Mirror neuron functions in primates and Foxp2 expression in vertebrates help to shed some light on the next representational question in Table 1, as well as the ultimate question on what Broca's area is “for.” In other words, this evidence, despite the series of questions that still remain unresolved, helps us understand what roles Broca's area plays in brain functions, how these are accomplished, and how the change unique to humans affected them.

The abnormalities associated with the KE family help to clarify some of these issues even more. FOXP2 is expressed in Broca's area and Brodmann area 6 in the motor cortex, as well as the subcortical areas listed above. In addition, the KE point mutation results in reduced gray matter connecting these areas to other areas in the brain, especially in the caudate nucleus, the cerebellum, and the left and right inferior frontal gyrus. Interestingly, it is also associated with increased gray matter in the planum temporale (Belton et al.,2003a). These differences would indicate no dysfunction related to the communication of auditory information to Broca's area, but decreased information flowing from it to motor control areas, consistent with the KE family problems with fluency and fine motor movement in the face and mouth.

This pattern is further consistent with expression of Foxp2 in mammal fetuses, where the mRNA signal appears on the inner cortical plate and is limited to the tissue below the granule cells in layer IV and especially in layer VI (Ferland et al.,2003; Lai et al.,2003; Takahashi et al.,2003). Thus, the gene appears in the layers that connect to subcortical areas, so that damage from mutations would also occur there. In fact, KE family members with the point mutation show significant lack of activity in Broca's area and the putamen between Broca's area and the limbic system in verbal generation tasks. Possibly in partial compensation, they show heightened activity in verbal tasks in areas further to the rear of the cortex and in both hemispheres (Liégeois et al.,2003). This provides yet more evidence for a role for Broca's area in motor control related to language in humans.


Besides language, meaningful patterns and sequences related to music also trigger responses to areas in human brains that are generally associated with language. In an fMRI study, which provides good spatial resolution, but not very good temporal resolution, subjects exposed to musical sequences ending in discordant notes reacted significantly more throughout the perisylvian region than when exposed to note sequences that behaved according to rules of tonality with which they were familiar (Koelsch et al.,2002). Magnetic encephalography (MEG), which has excellent temporal resolution, provides good evidence that Broca's area is involved in processing this “musical syntax” at different timescales. For example, “in key” and discordant tones produce distinctly different levels of activity, and Broca's area is particularly active in the case of discordant tones, with the reaction occurring approximately 200 msec after the tones are heard (Maess et al.,2001). These are consistent with data on linguistic perception, where gaps in a “tone group” are perceived from 80 to about 240 msec and ignored otherwise (Butcher,1981). In sequential processing, when confronted with a key phrase that was grammatical in a simple sentence, grammatical in an embedded sentence, and inserted ungrammatically into a third sentence, subjects showed a similar electrical activity (a P600 event-related potential/ERP) beginning at 200–300 msec, and reaching a maximum amplitude at 800–900 msec for the ungrammatical structure. Incongruous tones in musical sequences showed a virtually identical pattern (Patel et al.,1998). Reaction times and amplitudes tend to be proportional to the “distance” from anticipated musical values, similar to the hierarchy of reaction times in language as phrases become difficult or impossible to interpret (Koelsch et al.,2000; Patel,2003).

Broca's area also appears to have an important role in memory. For language, bilingual subjects show distinctly different fMRI activation patterns for a given language depending on whether their second language was acquired simultaneously with their first one, or later when they were adults. In the case of subjects who learned two languages at the same time, activation patterns in Broca's area overlap considerably. When the second language was acquired in adulthood, the two areas are distinct (Kim et al.,1997).

Broca's area thus makes associations over times less than a second and stores patterns acquired over a lifetime. This implies a function requiring both short- and (very) long-term memory. Recent and remote memories have different processes associated with them. In the case of episodic memories, these are related to cortical-hippocampal networks (Frankland and Bontempi,2005). The expression of c-fos, which is activity-dependent, in cortical layers demonstrates these different processes and shows distinct differences in storing spatial memories in the parietal cortex of mice (Maviel at el.,2004). That pattern is essentially identical to the expression of Foxp2 with respect to cortical layers: Foxp2 is expressed predominantly in layers V and VI, which are also the layers showing the greatest c-fos activation and implying a tie to recent memories; Foxp2 is not expressed in layers II and III and hardly expressed in layer IV, where the lack of c-fos activity implied a significant relationship to remote memories.

Broca's area is also involved with a number of memory tasks apart from language. This returns to the multimodal associations with the area noted earlier in monkeys. In humans, Broca's area is involved in both spatial and object memory, in storage tasks and in executive tasks (Smith and Jonides,1999).

Considering the conceptual sketch at Figure 1, the simplest computational association we can make with Broca's area that is consistent with its connections to sequences, multimodal inputs, and motor executive functions would seem to be the transformation of multimodal input data sets into motor or conceptual inferences. These transformed data sets would be validated (“understood”) in adjacent premotor areas or used to perform motor commands. For humans, the conceptual inferences would essentially be parsed constructions, which may also be validated by subsequent inputs.


Transforming sequential data into predictive inferences requires choices, which require the presence of the appropriate bi- or multistable switching mechanisms to make them. Both neural models and experiments on similar cells to those that are prominent in Broca's area show that such bimodal behavior is to be expected there. This implies that the area may very well function as the place where such predictive constructs are stored, and where already stored constructs are evoked by current inputs or for planning future activity. For language, laying down the long-term storage infrastructure is part of language acquisition; the moment-to-moment use of that structure to understand or make utterances is part of real-time language processing.

Figure 2 helps to show precisely what this might mean. It is adapted from Croft's causal-order hypothesis linked to his analysis of causation types. This may help to understand the different patterns of case expression across numerous languages (Croft,1991). He is one of the “other” cognitive linguists mentioned earlier in connection with Langacker, and the creator of radical construction grammar, which dispenses with the need for a formalist set of syntactic rules for explaining linguistic phenomena (Croft,2001). Tomasello's psychological theories for language acquisition rely on these types of “construction” theories (Marin-Padilla,1999), so together this body of work provides the necessary theoretical framework to explain language acquisition and processing in humans.

Figure 2 attempts to provide a predictive framework for how different languages might go about signaling real-world causal relationships and their related descriptions in terms of case marking, such as the use of accusative case for direct objects in many European languages. For example, in the English sentence “George hit him,” “him” is descended from an accusative form. To explain some of the less obvious annotations as well, the “X's” in the diagram refer to two different candidates as points of focus for linguistic signaling. In English, the pattern follows the accusative format. In ergative languages such as Dyirbal in Australia, the pattern is absolutive. Logical subjects and objects are the same, but the way they are signaled is different. However, the diagram shows that the choice made by the respective speech community falls at one end or the other of the segment that is encoded by verbs.

For computational purposes, Figure 2 shows that this complex array of interrelated signals can be mapped back to a simple linear pattern of cause and effect that can be observed by any intelligent organism in its environment: monkeys can see these patterns even if they cannot describe them. This pattern lies beneath the multimodal perceptions monkeys make when looking at and listening to their environment and comprehending the gestures of their fellow monkeys (Gallese et al.,1996; Kohler et al.,2002), and these aspects are associated with F5, the very early equivalent of Broca's area in primate evolution.

Tomasello (1999) posits a simple perceptual framework like this for language acquisition. He further requires the recognition in growing children that they and the adults they are listening to have similar mental states and goals, and that the adults make focused, identifiable speech sequences (constructions) within that framework. Gestures figure prominently in this framework as well, along with joint attention that is certainly more difficult for other primates. It is also interesting that the difficulty of Broca's aphasia patients and KE family members with the English passive construction corresponds to a reversal of the underlying causal arrow in Figure 2 in the arguments of the construction. These patients will often say that “George hit Bill” when they hear “George was hit by Bill,” which is consistent with the “primate core” interpretation of the order of entities but a wrong interpretation of the construction semantically. In any event, associating Broca's area within this multimodal framework, as well as calling on it to organize speech inputs along parallel lines, is not a very large leap.

Do these associations account for language in general? Reference to the conceptual sketch in Figure 1 would rule out such a broad role for Broca's area, in addition to the extensive evidence tying the entire perisylvian area to language processing. The symbolic representations corresponding to external entities would be far more likely in the parietal and temporal lobe structures linked to Wernicke's area, for example.

Downstream from the data organization step in Figure 1 is the inferential computation that enters the feedback loop for learning, which would have a motor variant related to predictive estimates for motor actions. Are all the functions on the feedback loop performed in Broca's area, or in other areas to which it is connected? Connections to the premotor cortex would indicate that the validation function is performed downstream. For motor sequences, the more likely candidate for validation would be the cerebellum, where error estimates in motor feedback loops are a very likely output (Marr,1969; Albus,1971; Ito,1989). For the cortex, the anterior cingulate cortex seems to function as an error detector (Paus et al.,1998; Koski and Paus,2000; Brown and Braver,2005). It has connections in cognitive, motor, and arousal pathways (Paus,2001), as well as to the prefrontal cortex (Barbas,1997) that overlap many of the language connections we have already reviewed. These connections include the motor cortex (Dum and Strick,1991; Morecraft and Van Hoesen,1992), vocal regions (Barbas et al.,1999), and the limbic region as well (Montaron and Buser,1988; Barbas and De Olmos,1990; Barbas et al.,1991; Kunishio and Haber,1994).

Figure 1 also has an interpretation step associated with language downstream from data ordering, but prior to full description and conceptual inference. Given the apparent role of other regions for symbolic representation, this function also is likely performed in other brain regions, particularly in the temporal cortex. In fact, the interpretation function is likely to be so complex that to trace all its facets would be a search for the neural correlates of semantic interpretation, which would be on a par with the attempt to establish the neural correlates of consciousness (Koch,2004).

Returning to Marr's “hardware” category (Marr,1982), there is also a network level where functions can be analyzed in terms of noise (once again), priming, resonance, and the structural changes that adjust the connections among various components. Pulvermüller (2002) has provided a very complete description of linguistic structures at this level. His analysis, not surprisingly, comes back repeatedly to the perisylvian region in the brain. He also presents extensive support for the idea of “word webs,” which incorporate widely distributed activations across relevant portions of the cortex that would provide a very good first approximation of symbolic representation or the neural correlates of semantic interpretation. All of the word web activations would occur well into the time frame necessary for fMRI imaging, so there is extensive imaging data available for this analysis, which he reviews very thoroughly.

That predictive computational niche, between the formulation of organized data and the validation of possible inferences, is the equivalent to the center of the cyber-spider's web in Figure 1 and what remains for Broca's area. In cognitive linguistic terms, this is probably sufficient to account for syntactic phenomena, illustrated in Figure 3. The original sentence is from an Old English homily by Ælfric in the 11th century, quoted because he predeceased any controversy about linguistic theory. It means, “He thought, if he killed them all, that the one he sought could not escape.” The first version of this sentence in Figure 3 is from Modern English. The second and third variants are the original Old English version, and a translation into modern German, which project the equivalent basic clauses into the final finished sentence.

Figure 3.

Projections of Basic Clausal Constructions onto Grammatical Modern English, Old English, and Modern German Variants of the Same Sentence. a) Modern English, b) Old English, c) Modern German. The upper sets of blocks chart basic constructions that are merged into the final sentence. The semantic content is identical for all three languages.

This sentence, which contains some of the case data explained in Figure 2, can be accounted for in terms of superposition and a minimalization computation step for parallel processes in connection with “bracket” signals used in Old English (Cooper,1999) and modern German. Modern English clusters its verbs but still shows the crossover “movement” of elements that are not easily accounted for by derivational trees. Each shows a decomposition into discrete constructions, as in Croft (2001) and Tomasello (2003). The Old English pieces are outlined below:

Đohte þæt Y: (He) thought that …

(This requires recognition of other people as thinking beings in addition to being a highly frequent construction, as Tomasello would probably point out.)

hē ofslōge hī ealle: he killed (slew) them all

hē ne ætburste: he (could) not escape

hē sōhte þone ānne: he sought the one

one ānne: “the one” is accusative, a direct object; se ān is nominative, a subject, like hē.)

gif … þanne: if … then

The arrows show that there are several superpositions, and that additional signals are built into the sentence, presumably to make it easily comprehensible (Ælfric was a famous preacher and educator in his day): “bracket” constructions, which put forms of the verb toward the rear of some Old English clauses to signal subordination, would reflect requirements of the minimalization step for large aggregations like this one, as would the change in case for “the one.” The position of “the one” also indicates a minimalization step in modern English. Modern German marks the end of subordinate clauses by placing the finite portion of the verb there, very similar to Old English. Generally, the minimalization step uses a higher-order construction to incorporate inputs from basic constructions.

It is easy to see that any disruption of sequencing would make these sentences impossible to produce. As with PKCγ-knockout mouse models with cerebellar dysfunctions that prevent the combination of individual movements into fluent sequences, here we would have similar disruptions that would first cause the failure to integrate basic constructions into larger ones, and if taken further, possibly even the inability to put together the basic constructions themselves.

That would describe many symptoms of Broca's aphasia, as well as the inability of KE family members with the FOXP2 mutation to produce this kind of utterance. Perhaps this illustrates the effect of that single point mutation and points to the specific function for Broca's area.

So what is Broca's area “for”? It is clearly linked to language, although its ancient connections in primates prove that it had other functions well before human language was even possible. It has genetic associations that implicate it in fluent combinations of motor movements, and these genetic associations are even more ancient than an equivalent to Broca's area. In computational terms, it appears to occupy a critical niche between organized sensory observations and predictive knowledge about the environment, and this becomes especially important in primates with respect to gestures and in humans with respect to language.


Special thanks to Ann Butler and Jim Olds of the Krasnow Institute for Advanced Study for their helpful suggestions and corrections, and particularly to the latter for allowing the author to follow his nose. The author also acknowledges in memoriam William G. Moulton, from Princeton, who helped him put together his original inquiries into computation, information theory, and language, which eventually led from dialectology and neostructural linguistics to cognitive neuroscience.


The importance of Broca's Area within the primate mirror system has recently been confirmed and expanded in Nelissen, et al.,2005, which provides fMRI evidence for the macaque homologs for BA 44 and BA 45. These areas favor gestural information over object identification, but also distinguish their focus between had gestures in the homolog to BA 45 and actions taken by an acting person in the homolog to BA 44, representing the action and its context, respectively. This may make BA 45 a location where action and object identification information coverage.