TB010299

doi 10.1098/rstb.2000.
0769
Geometric visual hallucinations, Euclidean

symmetry and the functional architecture
of striate cortex
Paul C. Bresslo¡1, Jack D. Cowan2*, Martin Golubitsky3,

Peter J. Thomas4 and Matthew C. Wiener5
1
Department of Mathematics, University of Utah, Salt Lake City, UT 84112, USA
2
Department of Mathematics, University of Chicago, Chicago, IL 60637, USA
3
Department of Mathematics, University of Houston, Houston,TX 77204-3476, USA
4
Computational Neurobiology Laboratory, Salk Institute for Biological Studies, PO Box 85800, San Diego, CA 92186-5800, USA
5
Laboratory of Neuropsychology, National Institutes of Health, Bethesda, MD 20892, USA
This paper is concerned with a striking visual experience: that of seeing geometric visual hallucinations.
Hallucinatory images were classi¢ed by Klu«ver into four groups called form constants comprising
(i) gratings, lattices, fretworks, ¢ligrees, honeycombs and chequer-boards, (ii) cobwebs, (iii) tunnels,
funnels, alleys, cones and vessels, and (iv) spirals. This paper describes a mathematical investigation of
their origin based on the assumption that the patterns of connection between retina and striate cortex
(henceforth referred to as V1) ö the retinocortical map ö and of neuronal circuits in V1, both local and
lateral, determine their geometry.
In the ¢rst part of the paper we show that form constants, when viewed in V1 coordinates, essentially
correspond to combinations of plane waves, the wavelengths of which are integral multiples of the width
of a human Hubel ^Wiesel hypercolumn, ca. 1.33^2 mm. We next introduce a mathematical description of
the large-scale dynamics of V1 in terms of the continuum limit of a lattice of interconnected hyper-
columns, each of which itself comprises a number of interconnected iso-orientation columns. We then
show that the patterns of interconnection in V1 exhibit a very interesting symmetry, i.e. they are invariant
under the action of the planar Euclidean group E(2) ö the group of rigid motions in the plane ö
rotations, re£ections and translations. What is novel is that the lateral connectivity of V1 is such that a
new group action is needed to represent its properties: by virtue of its anisotropy it is invariant with
respect to certain shifts and twists of the plane. It is this shift ^ twist invariance that generates new
representations of E(2). Assuming that the strength of lateral connections is weak compared with that of
local connections, we next calculate the eigenvalues and eigenfunctions of the cortical dynamics, using
Rayleigh ^Schro«dinger perturbation theory. The result is that in the absence of lateral connections, the
eigenfunctions are degenerate, comprising both even and odd combinations of sinusoids in , the cortical
label for orientation preference, and plane waves in r, the cortical position coordinate.`Switching-on' the
lateral interactions breaks the degeneracy and either even or else odd eigenfunctions are selected. These
results can be shown to follow directly from the Euclidean symmetry we have imposed.
In the second part of the paper we study the nature of various even and odd combinations of eigen-
functions or planforms, the symmetries of which are such that they remain invariant under the particular
action of E(2) we have imposed. These symmetries correspond to certain subgroups of E(2), the so-called
axial subgroups. Axial subgroups are important in that the equivariant branching lemma indicates that
when a symmetrical dynamical system becomes unstable, new solutions emerge which have symmetries
corresponding to the axial subgroups of the underlying symmetry group. This is precisely the case studied
in this paper. Thus we study the various planforms that emerge when our model V1 dynamics become
unstable under the presumed action of hallucinogens or £ickering lights. We show that the planforms
correspond to the axial subgroups of E(2), under the shift ^ twist action. We then compute what such
planforms would look like in the visual ¢eld, given an extension of the retinocortical map to include its
action on local edges and contours. What is most interesting is that, given our interpretation of the
correspondence between V1 planforms and perceived patterns, the set of planforms generates represent-
atives of all the form constants. It is also noteworthy that the planforms derived from our continuum
model naturally divide V1 into what are called linear regions, in which the pattern has a near constant
orientation, reminiscent of the iso-orientation patches constructed via optical imaging. The boundaries of
such regions form fractures whose points of intersection correspond to the well-known `pinwheels'.
*
Author for correspondence (cowan@math.uchicago.edu).
Phil. Trans. R. Soc. Lond. B (2001) 356, 299^330 299 & 2001 The Royal Society
Received 18 April 2000 Accepted 11 August 2000
300 P. C. Bresslo¡ and others Geometric visual hallucinations
To complete the study we then investigate the stability of the planforms, using methods of nonlinear
stability analysis, including Liapunov ^Schmidt reduction and Poincarë ^ Lindstedt perturbation theory.
We ¢nd a close correspondence between stable planforms and form constants. The results are sensitive to
the detailed speci¢cation of the lateral connectivity and suggest an interesting possibility, that the cortical
mechanisms by which geometric visual hallucinations are generated, if sited mainly in V1, are closely
related to those involved in the processing of edges and contours.
Keywords: hallucinations; visual imagery; £icker phosphenes; neural modelling;
horizontal connections; contours
`. . . the hallucination is . . . not a static process but a boards; (II) cobwebs; (III) tunnels and funnels, alleys,
dynamic process, the instability of which re£ects an cones, vessels; and (IV) spirals. Some examples of class I
instability in its conditions of origin' (Klu«ver (1966), form constants are shown in ¢gure 1, while examples of
p. 95, in a comment on Mourgue (1932)). the other classes are shown in ¢gures 2^4.
Such images are seen both by blind subjects and in
1. INTRODUCTION sealed dark rooms (Krill et al. 1963). Various reports
(a) Form constants and visual imagery (Klu«ver 1966) indicate that although they are di¤cult to
Geometric visual hallucinations are seen in many situa- localize in space, and actually move with the eyes, their
tions, for example, after being exposed to £ickering positions relative to each other remain stable with respect
lights (Purkinje 1918; Helmholtz 1924; Smythies 1960) to such movements. This suggests that they are generated
after the administration of certain anaesthetics (Winters not in the eyes, but somewhere in the brain. One clue on
1975), on waking up or falling asleep (Dybowski 1939), their location in the brain is provided by recent studies of
following deep binocular pressure on one's eyeballs visual imagery (Miyashita 1995). Although controversial,
(Tyler 1978), and shortly after the ingesting of drugs such the evidence seems to suggest that areas V1 and V2, the
as LSD and marijuana (Oster 1970; Siegel 1977). Patterns striate and extra-striate visual cortices, are involved in
that may be hallucinatory are found preserved in petro- visual imagery, particularly if the image requires detailed
glyphs (Patterson 1992) and in cave paintings (Clottes & inspection (Kosslyn 1994). More precisely, it has been
Lewis-Williams 1998). There are many reports of such suggested that (Ishai & Sagi 1995, p. 1773)
experiences (Knauer & Maloney 1913, pp. 429^430): `[the] topological representation [provided by V1] might
subserve visual imagery when the subject is scrutinizing
Ìmmediately before my open eyes are a vast number of
attentively local features of objects that are stored in
rings, apparently made of extremely ¢ne steel wire, all
memory'.
constantly rotating in the direction of the hands of a
clock; these circles are concentrically arranged, the inner- Thus visual imagery is seen as the result of an interaction
most being in¢nitely small, almost pointlike, the between mechanisms subserving the retrieval of visual
outermost being about a meter and a half in diameter. memories and those involving focal attention. In this
The spaces between the wires seem brighter than the respect it is interesting that there seems to be competition
wires themselves. Now the wires shine like dim silver in
between the seeing of visual imagery and hallucinations
parts. Now a beautiful light violet tint has developed in
them. As I watch, the center seems to recede into the
(Knauer & Maloney 1913, p. 433):
depth of the room, leaving the periphery stationary, till `. . . after a picture had been placed on a background and
the whole assumes the form of a deep tunnel of wire rings. then removed `Ì tried to see the picture with open eyes.
The light, which was irregularly distributed among the In no case was I successful; only [hallucinatory] visionary
circles, has receded with the center into the apex of the phenomena covered the ground'''.
funnel. The center is gradually returning, and passing the
position when all the rings are in the same vertical plane, Competition between hallucinatory images and after-
continues to advance, till a cone forms with its apex images was also reported Klu«ver 1966, p. 35):
toward me . . . . The wires are now £attening into bands Ìn some instances, the [hallucinatory] visions prevented
or ribbons, with a suggestion of transverse striation, and the appearance of after-images entirely; [however] in
colored a gorgeous ultramarine blue, which passes in most cases a sharply outlined normal after-image
places into an intense sea green. These bands move rhyth- appeared for a while . . . while the visionary phenomena
mically, in a wavy upward direction, suggesting a slow were stationary, the after-images moved with the eyes'.
endless procession of small mosaics, ascending the wall in
single ¢les. The whole picture has suddenly receded, the As pointed out to us by one of the referees, the fused
center much more than the sides, and now in a moment, image of a pair of random dot stereograms also seems to
high above me, is a dome of the most beautiful mosiacs, be stationary with respect to eye movements. It has also
. . . . The dome has absolutely no discernible pattern. But been argued that because hallucinatory images are seen
circles are now developing upon it; the circles are as continuous across the midline, they must be located at
becoming sharp and elongated . . . now they are rhombics higher levels in the visual pathway than V1 or V2
now oblongs; and now all sorts of curious angles are (R. Shapley, personal communication.) In this respect
forming; and mathematical ¢gures are chasing each other there is evidence that callosal connections along the V1/
wildly across the roof . . . .'
V2 border can act to maintain continuity of the images
Klu«ver (1966) organized the many reported images into across the vertical meridian (Hubel & Wiesel 1967).
four classes, which he called form constants: (I) gratings, All these observations suggest that both areas V1 and
lattices, fretworks, ¢ligrees, honeycombs and chequer- V2 are involved in the generation of hallucinatory
Phil. Trans. R. Soc. Lond. B (2001)

Geometric visual hallucinations P. C. Bresslo¡ and others 301
(a) (a)
(b) (b)
Figure 2. (a) Funnel and (b) spiral hallucinations generated

by LSD. Redrawn from Oster (1970).
Figure 1. (a) `Phosphene' produced by deep binocular
pressure on the eyeballs. Redrawn from Tyler (1978).
(a)
(b) Honeycomb hallucination generated by marijuana.
Redrawn from Clottes & Lewis-Williams (1998).
images. In our view such images are generated in V1 and

stabilized with respect to eye movements by mechanisms
present in V2 and elsewhere. It is likely that the action of
such mechanisms is rapidly fed back to V1 (Lee et al.
1998). It now follows, because all observers report seeing
Klu«ver's form constants or variations, that those
properties common to all such hallucinations should yield
information about the architecture of V1. We therefore
investigate that architecture, i.e. the patterns of connec-
tion between neurons in the retina and those in V1,
together with intracortical V1 connections, on the (b)
hypothesis that such patterns determine, in large part,
the geometry of hallucinatory form constants, and we
defer until a later study, the investigation of mechanisms
that contribute to their continuity across the midline and
to their stability in the visual ¢eld.
(b) The human retinocortical map

The ¢rst step is to calculate what visual hallucinations
look like, not in the standard polar coordinates of the
visual ¢eld, but in the coordinates of V1. It is well estab-
lished that there is a topographic map of the visual ¢eld
in V1, the retinotopic representation, and that the central
region of the visual ¢eld has a much bigger representation
in V1 than it does in the visual ¢eld (Sereno et al. 1995). Figure 3. (a) Funnel and (b) spiral tunnel hallucinations
The reason for this is partly that there is a non-uniform generated by LSD. Redrawn from Siegel (1977).

1
R ,
(w0 erR )2
where w0 and e are constants. Estimates of w0 0:087
and e 0:051 in appropriate units can be obtained from
published data (Drasdo 1977). From the inverse square
law one can calculate the Jacobian of the map and hence
V1 coordinates fx, yg as functions of visual ¢eld or retinal
coordinates frR , R g. The resulting coordinate transforma-
tion takes the form

e
x ln 1 rR ,
e w0
rR R
y ,
w0 erR
Figure 4. Cobweb petroglyph. Redrawn from Patterson
(1992). where and are constants in appropriate units.
Figure 5 shows the map.
The transformation has two important limiting cases:
(i) near the fovea, erR 5w0, it reduces to
distribution of retinal ganglion cells, each of which r
x R,
connects to V1 via the lateral geniculate nucleus (LGN). w0
This allows calculation of the details of the map (Cowan r
y R R,
1977). Let R be the packing density of retinal ganglion w0
cells per unit area of the visual ¢eld, the corresponding
density per unit surface area of cells in V1, and rR ; R and (ii), su¤ciently far away from the fovea, erR w0, it
retinal or equivalently, visual ¢eld coordinates. Then becomes
R rR drR dR is the number of ganglion cell axons in a er
retinal element of area rR drR dR . By hypothesis these x ln R ,
e w0
axons connect topographically to cells in an element of
V1 surface area dx dy, i.e. to dx dy cortical cells. (V1 is y R.
e
assumed to be locally £at with Cartesian coordinates.)
Empirical evidence indicates that is approximately Case (i) is just a scaled version of the identity map, and case
constant (Hubel & Wiesel 1974a,b), whereas R declines (ii) is a scaled version of the complex logarithm as was ¢rst
from the origin of the visual ¢eld, i.e. the fovea, with an recognized by Schwartz (1977). To see this, let
inverse square law (Drasdo 1977) zR xR iyR rR expiR , be the complex representation
(a) π /2 (b)
y
π /2 π /2
π 0 π 0 x
3π /2 3π /2
3π /2
(c) y
π /2 π /2
π 0
x
3π /2 3π /2
Figure 5. The retinocortical map: (a) visual ¢eld; (b) the actual cortical map, comprising right and left hemisphere transforms;
(c) a transformed version of the cortical map in which the two transforms are realigned so that both foveal regions correspond to
x 0.

(a) (a)
(b)
(b)
Figure 6. Action of the retinocortical map on the funnel form

constant. (a) Image in the visual ¢eld; (b) V1 map of the image.
Figure 7. Action of the retinocortical map on the spiral form
constant. (a) Image in the visual ¢eld; (b) V1 map of the image.
of a retinal point (xR , yR ) (rR , R ), then z x iy
ln( rR expiR ) ln rR iR . Thus x ln rR , y R. generate more complex hallucinations in the visual ¢eld
such as chequer-boards. Similar results are found in a
(c) Form constants as spontaneous cortical patterns reduced single-population model provided that the inter-
Given that the retinocortical map is generated by the actions are characterized by a mixture of short-range
complex logarithm (except near the fovea), it is easy to excitation and long-range inhibition (the so-called
calculate the action of the transformation on circles, rays, `Mexican hat distribution').
and logarithmic spirals in the visual ¢eld. Circles of
constant rR in the visual ¢eld become vertical lines in V1, (d) Orientation tuning in V1
whereas rays of constant R become horizontal lines. The Ermentrout ^ Cowan theory of visual hallucinations
Interestingly, logarithmic spirals become oblique lines in is over-simpli¢ed in the sense that V1 is represented as if it
V1: the equation of such a spiral is just R a ln rR were just a cortical retina. However, V1 cells do much
whence y ax under the action of zR ! z. Thus form more than merely signalling position in the visual ¢eld:
constants comprising circles, rays and logarithmic spirals most cortical cells signal the local orientation of a contrast
in the visual ¢eld correspond to stripes of neural activity edge or bar ö they are tuned to a particular local orienta-
at various angles in V1. Figures 6 and 7 show the map tion (Hubel & Wiesel 1974a). The absence of orientation
action on the funnel and spiral form constants shown in representation in the Ermentrout ^ Cowan model means
¢gure 2. that a number of the form constants cannot be generated
A possible mechanism for the spontaneous formation of by the model, including lattice tunnels (¢gure 42), honey-
stripes of neural activity under the action of hallucinogens combs and certain chequer-boards (¢gure 1), and cobwebs
was originally proposed by Ermentrout & Cowan (1979). (¢gure 4). These hallucinations, except the chequer-
They studied interacting populations of excitatory and boards, are more accurately characterized as lattices of
inhibitory neurons distributed within a two-dimensional locally orientated contours or edges rather than in terms of
(2D) cortical sheet. Modelling the evolution of the system contrasting regions of light and dark.
in terms of a set of Wilson ^ Cowan equations (Wilson & In recent years, much information has accumulated
Cowan 1972, 1973) they showed how spatially periodic about the distribution of orientation selective cells in V1,
activity patterns such as stripes can bifurcate from a and about their pattern of interconnection (Gilbert 1992).
homogeneous low-activity state via a Turing-like Figure 8 shows a typical arrangement of such cells,
instability (Turing 1952). The model also supports the obtained via microelectrodes implanted in cat V1. The ¢rst
formation of other periodic patterns such as hexagons panel shows how orientation preferences rotate smoothly
and squares ö under the retinocortical map these over V1, so that approximately every 300 mm the same

(a) 1 2 (a)
3
(b) y
1
3
(b)
(c) y
2
Figure 8. (a) Orientation tuned cells in V1. Note the

constancy of orientation preference at each cortical location
(electrode tracks 1 and 3), and the rotation of orientation
preference as cortical location changes (electrode track 2).
(b) Receptive ¢elds for tracks 1 and 3. (c) Expansion of the
receptive ¢elds of track 2 to show the rotation of orientation
preference. Redrawn from Gilbert (1992).
Figure 9 (a) Distribution of orientation preferences in
macaque V1 obtained via optical imaging. Redrawn from
preference reappears, i.e. the distribution is -periodic in
Blasdel (1992). (b) Connections made by an inhibitory
the orientation preference angle. The second panel shows interneuron in cat V1. Redrawn from Eysel (1999).
the receptive ¢elds of the cells, and how they change with
V1 location. The third panel shows more clearly the
rotation of such ¢elds with translation across V1. III of V1, and to some extent in layer V (Rockland &
How are orientation tuned cells distributed and inter- Lund 1983), tend to be orientated along the direction of
connected ? Recent work on optical imaging has made it their cell's preference (Gilbert 1992; Bosking et al. 1997),
possible to see how the cells are actually distributed in V1 i.e. they run parallel to the visuotopic axis of their cell's
(Blasdel 1992), and a variety of stains and labels has orientation preference. These horizontal connections arise
made it possible to see how they are interconnected almost exclusively from excitatory neurons (Levitt &
(G. G. Blasdel and L. Sincich, personal communication), Lund 1997; Gilbert & Wiesel 1983), although 20%
(Eysel 1999; Bosking et al. 1997). Figures 9 and 10 show terminate on inhibitory cells and can thus have signi¢cant
such data. Thus, ¢gure 9a shows that the distribution of inhibitory e¡ects (McGuire et al. 1991).
orientation preferences is indeed roughly -periodic, in There is some anatomical and psychophysical evidence
that approximately every 0.5 mm (in the macaque) there (Horton 1996; Tyler 1982) that human V1 has several
is an iso-orientation patch of a given preference, and times the surface area of macaque V1 with a hypercolumn
¢gure 10 shows that there seem to be at least two length- spacing of ca. 1.33^2 mm. In the rest of this paper we
scales: work with this length-scale to extend the Ermentrout ^
Cowan theory of visual hallucinations to include orienta-
(i) local ö cells less than 0.5 mm apart tend to make
tion selective cells. A preliminary account of this was
connections with most of their neighbours in a
described in Wiener (1994) and Cowan (1997).
roughly isotropic fashion, as seen in ¢gure 9b, and
(ii) lateral ö cells make contacts only every 0.5 mm or
so along their axons with cells in similar iso- 2. A MODEL OF V1 WITH ANISOTROPIC LATERAL
orientation patches. CONNECTIONS
In addition, ¢gure 10 shows that the long axons which (a) The model
support such connections, known as intrinsic lateral or The state of a population of cells comprising an iso-
horizontal connections, and found mainly in layers II and orientation patch at cortical position r 2 R2 at time t is

(a) would be that it is a continuum version of a lattice of

hypercolumns. However, a potential di¤culty with this
interpretation is that the e¡ective wavelength of many of
the patterns underlying visual hallucinations is of the
order of twice the hypercolumn spacing (see, for example,
¢gure 2), suggesting that lattice e¡ects might be impor-
tant. A counter-argument for the validity of the
continuum model (besides mathematical convenience) is
to note that the separation of two points in the visual
¢eldövisual acuityö(at a given retinal eccentricity of
r 0R ), corresponds to hypercolumn spacing (Hubel &
Wiesel 1974b), and so to each location in the visual ¢eld
there corresponds to a representation in V1 of that loca-
tion with ¢nite resolution and all possible orientations.
The activity variable a(r, , t) evolves according to a
generalization of the Wilson ^ Cowan equations (Wilson &
Cowan 1972, 1973) that takes into account the additional
internal degree of freedom arising from orientation
preference:
(b) Z Z
@a(r, , t)
a(r, , t) w(r, jr0 , 0 )
@t 0 R2
dr0 d0
a(r0 , 0 , t) h(r, , t), (1)

where and are decay and coupling coe¤cients,
h(r, , t) is an external input, w(r, jr0 , 0 ) is the weight
of connections between neurons at r tuned to and
neurons at r0 tuned to 0, and z is the smooth nonlinear
function
1
z , (2)
1e (z )
for constants and . Without loss of generality we may

subtract from z a constant equal to 1 e 1 to obtain
the (mathematically) important property that 0 0,
which implies that for zero external inputs the homoge-
neous state a(r, , t) 0 for all r, , t is a solution to
equation (1). From the discussion in ½ 1(d), we take the
Figure 10. Lateral connections made by cells in (a) owl pattern of connections w(r, jr0 , 0 ) to satisfy the
monkey and (b) tree shrew V1. A radioactive tracer is used to following properties (see ¢gure 1).
show the locations of all terminating axons from cells in a
central injection site, superimposed on an orientation map
obtained by optical imaging. Redrawn from G. G. Blasdel and (i) There exists a mixture of local connections within a
L. Sincich (personal communication) and Bosking et al. (1997). hypercolumn and (anisotropic) lateral connections
between hypercolumns; the latter only connect
characterized by the real-valued activity variable elements with the same orientation preference. Thus
a(r, , t), where 2 0, ) is the orientation preference of in the continuum model, w is decomposed as
the patch. V1 is treated as an (unbounded) continuous 2D w(r, jr0 , 0 ) wloc ( 0 )(r r0 )
sheet of nervous tissue. For the sake of analytical tract- 0
ability, we make the additional simplifying assumption wlat (r r , )( 0 ), (3)
that and r are independent variables ö all possible with wloc ( ) wloc ().
orientations are represented at every position. A more (ii) Lateral connections between hypercolumns only
accurate model would need to incorporate details join neurons that lie along the direction of their
concerning the distribution of orientation patches in the (common) orientation preference . Thus in the
cortical plane (as illustrated in ¢gure 9a). It is known, for continuum model
example, that a region of human V1 ca. 2:67 mm2 on its
surface and extending throughout its depth contains at wlat (r, ) w(R
^ r), (4)
least two sets of all iso-orientation patches in the range
045, one for each eye. Such a slab was called a with
hypercolumn by Hubel & Wiesel (1974b). If human V1 as Z 1
a whole (in one hemisphere) has a surface area of ca. w(r)
^ g(s)(r sr0 ) (r sr0 )ds, (5)
0
3500 mm2 (Horton 1996), this gives approximately 1300
such hypercolumns. So one interpretation of our model where r0 (1, 0) and R is the rotation matrix

φ0
θ0
φ0
local connections
φ0
Figure 12. Example of an angular spread in the anisotropic

lateral connections lateral connections between hypercolumns with respect to
both space (0 ) and orientation preference (0 ).
space, respectively, whereas 0 determines the (spatial)

range of the isotropic local connections.
hypercolumn (b) Euclidean symmetry

Suppose that the weight distribution w satis¢es
equations (7) and (8). We show that w is invariant under
the action of the Euclidean group E(2) of rigid motions in
the plane, and discuss some of the important conse-
Figure 11. Illustration of the local connections within a quences of such a symmetry.
hypercolumn and the anisotropic lateral connections between
hypercolumns. (i) Euclidean group action
The Euclidean group is composed of the (semi-direct)
x cos sin x product of O(2), the group of planar rotations and re£ec-
R .
y sin cos y tions, with R2 , the group of planar translations. The
action of the Euclidean group on R2 S1 is generated by
The weighting function g(s) determines how the
strength of lateral connections varies with the distance s (r, ) (r s, ) s 2 R2
of separation.We take g(s) to be of the particular form (r, ) (R r, ) 2 S1 (9)
(r, ) (r, ),

2
g(s) 2lat 1=2
exp s2 /2lat 2
Alat 2^lat
2
1=2
where is the re£ection (x1 , x2 )7 !(x1 , x2 ) and R is a

exp s2 /2^lat
2
, (6) rotation by .
The corresponding group action on a function
with lat 5^lat and Alat 41, which represent a combi-
a:R2 S1 ! R where P (r, ) is given by
nation of short-range excitation and long-range
inhibition. This is an example of the Mexican hat a(P) a( 1
P), (10)
distribution. (Note that one can view the short-
range excitatory connections as arising from patchy for all _ R2 and the action on w(PjP 0 ) is
2 O(2)
local connections within a hypercolumn.)
w(PjP 0 ) w( 1
Pj 1
P 0 ).
It is possible to consider more general choices of weight
distribution w that (i) allow for some spread in the The particular form of the action of rotations in
distribution of lateral connections (see ¢gure 12), and equations (9) re£ects a crucial feature of the lateral
(ii) incorporate spatially extended isotropic local inter- connections, namely that they tend to be orientated along
actions. An example of such a distribution is given by the the direction of their cell's preference (see ¢gure 11). Thus,
following generalization of equations (3) and (4): if we just rotate V1, then the cells that are now connected
w(r, jr0 , 0 ) wloc ( 0 )loc (jr r0 j) at long range will not be connected in the direction of
0 their preference. This di¤culty can be overcome by
w(R
^ r r)lat ( 0 ), (7)
permuting the local cells in each hypercolumn so that
with lat ( ) lat (), lat () 0 for jj40, and cells that are connected at long range are again
loc (jrj) 0 for r40. Moreover, equation (5) is modi- connected in the direction of their preference. Thus, in
¢ed according to the continuum model, the action of rotation of V1 by
Z 0 Z 1 corresponds to rotation of r by while simultaneously
sending to . This is illustrated in ¢gure 13. The
w(r)
^ p() g(s)(r sr ) (r sr )dsd,
0 0 action of re£ections is justi¢ed in a similar fashion.
(8)
(ii) Invariant weight distribution w
with r (cos (), sin ()) and p( ) p(). The para- We now prove that w as given by equations (7) and (8) is
meters 0 and 0 determine the angular spread of lateral invariant under the action of the Euclidean group de¢ned
connections with respect to orientation preference and by equations (9). (It then follows that the distribution

2
two points P, Q ∈ R × [0, π ) rotation by θ = π /6
1 1
0.5 0.5
(rQ , φQ)
(r'Q, φ'Q)
θ
0 0
− 0.5 − 0.5
(rp , φp) (r'P, φ'P)
−1 −1 Figure 13. Action of a rotation by
−1 − 0.5 0 0.5 1 −1 − 0.5 0 0.5 1 : (r, ) ! (r0 , 0 ) (R r, ).
satisfying equations (3)^(5) is also Euclidean invariant.) symmetry group E(2) (Golubitsky et al. 1988). The equiv-
Translation invariance of w is obvious, i.e. ariance of the operator F with respect to the action of
E(2) has major implications for the nature of solutions
w(r s, jr0 s, 0 ) w(r, jr0 ,0 ).
bifurcating from the homogeneous resting state. Let be
Invariance with respect to a rotation by follows from a bifurcation parameter. We show in ½ 4 that near a point
for which the steady state a(r, , ) 0 becomes unstable,
w(R r, jR r0 , 0 ) there must exist smooth solutions to the equilibrium equa-
wloc ( 0 )loc (jR r r0 j) tion Ga(r, , ) 0 that are identi¢ed by their
symmetry (Golubitsky et al. 1988). We ¢nd solutions that
w(R
^ R (r r0 ))lat ( 0 )
are doubly periodic with respect to a rhombic, square or
wloc ( 0 )loc (jr r0 j) w(R
^ r)lat ( 0 ) hexagonal lattice by using the remnants of Euclidean
w(r, jr , ). 0 0 symmetry on these lattices. These remnants are the (semi-
direct) products G of the torus T 2 of translations modulo
Finally, invariance under a re£ection about the x-axis the lattice with the dihedral groups D2 , D4 and D6, the
holds because holohedries of the lattice. Thus, when a(r, , ) 0
becomes unstable, new solutions emerge from the
w(r, jr0 , 0 ) wloc ( 0 )loc (jr r0 j)
instability with symmetries that are broken compared
^ (r r0 ))lat ( 0 )
w(R with G. Su¤ciently close to the bifurcation point these
wloc ( 0 )loc (jr r0 j) patterns are characterized by (¢nite) linear combinations
w(R
^ r0 ))lat ( 0 ) of eigenfunctions of the linear operator L D0 G
(r
obtained by linearizing equation (1) about the homoge-
wloc ( 0 )loc (jr r0 j) neous state a 0. These eigenfunctions are derived in ½ 3.
w(R
^ (r r0 ))lat ( 0 )
w(r, jr0 , 0 ). (c) Two limiting cases
For the sake of mathematical convenience, we restrict
We have used the identity R R and the conditions our analysis in this paper to the simpler weight distribu-
wloc ( ) wloc (), lat ( ) lat (), w(r)
^ w(r).
^ tion given by equations (3) and (4) with w^ satisfying
either equation (5) or (8). The most important property of
(iii) Implications of Euclidean symmetry w is its invariance under the extended Euclidean group
Consider the action of on equation (1) for h(r, t) 0: action (9), which is itself a natural consequence of the
1
anisotropic pattern of lateral connections. Substitution of
@a( P, t) equation (3) into equation (1) gives (for zero external
@t Z inputs)
1
a( P, t) w( 1 PjP 0 )a(P 0 , t)dP 0 @a(r, , t)
R2 S1
Z @t
a( 1
P, t) w(Pj P 0 )a(P 0 , t)dP 0 Z
d0
Z R2 S1 a(r, , t) wloc ( 0 )a(r, 0 , t)
0
a( 1
P, t) w(PjP 00 )a( 1 P 00 , t)dP 00 , Z
R2 S1 wlat (r r0 , )a(r0 , , t)dr0 , (11)
R2
since d 1 P dP and w is Euclidean invariant. If we
rewrite equation (1) as an operator equation, namely, where we have introduced an additional coupling para-
da meter that characterizes the relative strength of lateral
Fa Ga 0, interactions. Equation (11) is of convolution type, in that
dt
the weighting functions are homogeneous in their respec-
then it follows that F a F a. Thus F commutes with tive domains. However, the weighting function wlat (r, )
2 E(2) and F is said to be equivariant with respect to the is anisotropic, as it depends on . Before proceeding to

analyse the full model described by equation (11), it is

useful to consider two limiting cases, namely the ring
model of orientation tuning and the Ermentrout ^ Cowan
model (Ermentrout & Cowan 1979).
4
(i) The ring model of orientation tuning
The ¢rst limiting case is to neglect lateral connections
completely by setting 0 in equation (11). Each point r
3
in the cortex is then independently described by the so-
called ring model of orientation tuning (Hansel &
σ (a)
Sompolinsky 1997; Mundel et al. 1997; Ermentrout 1998; 2
Bresslo¡ et al. 2000a):
Z
@a(r, , t)
a(r, , t) wloc ( 0 ) 1
@t 0
0 d0
a(r, , t) . (12)

0 π /2 π
Linearizing this equation about the homogeneous state orientation φ
a(r, , t) 0 and considering perturbations of the form
a(r, , t) elt a(r, ) yields the eigenvalue equation Figure 14. Sharp orientation tuning curves for a Mexican hat
Z weight kernel with loc 208, ^ loc 608 and Aloc 1. The
d0 tuning curve is marginally stable so that the peak activity a at
la(r, ) a(r, ) wloc ( 0 )a(r, 0 ) . each point in the cortical plane is arbitrary. The activity is
0
truncated at 0 in line with the choice of 0 0.
Introducing
P the Fourier series expansion a(r, )
m zm (r) e2im c:c: generates the following discrete ¢gure 14, which is obtained by taking wloc () to be a
dispersion relation for the eigenvalue l: di¡erence of Gaussians over the domain /2, /2:
l 1 Wm lm , (13) 2
wloc () 2loc 1=2
exp( 2 /2loc
2
) Aloc 2^loc
2
1=2
exp( 2 ^2
/2loc ), (15)
where 1 dz/dz evaluated at z 0 and
X with loc 5^loc and Aloc 41.
wloc () Wn e2ni . (14)
The location of the centre (r) of each tuning curve is
n2Z
arbitrary, which re£ects the rotational equivariance of
Note that because wloc () is a real and even function of , equation (12) under the modi¢ed group action
W m Wm W m . : (r, ) ! (r, ). Moreover, in the absence of
Let Wp maxfWn , n 2 Z g and suppose that p is lateral interactions the tuned response is uncorrelated
unique with Wp 40 and p51. It then follows from across di¡erent points in V1. In this paper we show how
equation (13) that the homogeneous state a(r, ) 0 is the presence of anisotropic lateral connections leads to
stable for su¤ciently small , but becomes unstable when periodic patterns of activity across V1 in which the peaks
increases beyond the critical value c /1 Wp due to of the tuning curve at di¡erent locations are correlated.
excitation of linear eigenmodes of the form
a(r, ) z(r)e2ip z(r)e 2ip , where z(r) is an arbitrary (ii) The Ermentrout ^ Cowan model
complex function of r. It can be shown that the saturating The other limiting case is to neglect the orientation
nonlinearities of the system stabilize the growing pattern label completely. Equation (11) then reduces to a one-
of activity (Ermentrout 1998; Bresslo¡ et al. 2000a). In population version of the model studied by Ermentrout &
terms of polar coordinates z(r) Z(r)e2i(r) we have Cowan (1979):
a(r, ) Z(r) cos (2p (r)). Thus at each point r in Z
@
the plane the maximum (linear) response occurs at the a(r, t) a(r, t) wlat (r r0 ) a(r0, t) dr0 . (16)
@t O
orientations (r) k/p, k 0, 1, : : :, p 1 when p 6 0.
Of particular relevance from a biological perspective In this model there is no reason to distinguish any direc-
are the cases p 0 and p 1. In the ¢rst case there is a tion in V1, so we assume that wlat (r r0 ) ! wlat (jr r0 j),
bulk instability in which the new steady state shows no i.e. wlat depends only on the magnitude of r r0 . It can
orientation preference. Any tuning is generated in the be shown that the resulting system is equivariant with
genicocortical map. We call this the `Hubel ^Wiesel' respect to the standard action of the Euclidean group in
mode (Hubel & Wiesel 1974a). In the second case the the plane.
response is unimodal with respect to . The occurrence of Linearizing equation (16) about the homogeneous state
a sharply tuned response peaked at some angle (r) in a and taking a(r, t) elt a(r) gives rise to the eigenvalue
local region of V1 corresponds to the presence of a local problem
contour there, the orientation of which is determined by Z
the inverse of the double retinocortical map described in la(r) a(r) 1 wlat (jr r0 j)a(r0 )dr0 ,
½ 5(a). An example of typical tuning curves is shown in O

Z
which upon Fourier transforming generates a dispersion
d0
relation for the eigenvalue l as a function of q jkj, i.e. lu() u() 1 wloc ( 0 )u(0 )
0

l e
1 W(q) l(q), w
e lat (k, 'u() . (19)
where W(q) e we lat (k) is the Fourier transform of Here we lat (k, ) is the Fourier transform of wlat (r, ).
wlat (jrj). Note that l is real. If we choose wlat (jrj) to be in Assume that wlat satis¢es equations (4) and (5) so that
the form of a Mexican hat function, then it is simple to the total weight distribution w is Euclidean invariant. The
establish that l passes through zero at a critical para- resulting symmetry of the system then restricts the struc-
meter value c signalling the growth of spatially periodic ture of the solutions of the eigenvalue equation (19):
patterns with wavenumber qc , where W(q e c)
e
maxq fW(q)g. Close to the bifurcation point these (i) l and u() only depend on the magnitude q jkj of
patterns can be represented as linear combinations of the wave vector k. That is, there is an in¢nite degen-
plane waves eracy due to rotational invariance.
X (ii) For each k the associated subspace of eigenfunctions
a(r) ci exp(iki r),
i
Vk fu( ')eikr c:c: : u( )u() and u 2 Cg

with jki j qc . As shown by Ermentrout & Cowan (1979) (20)
and Cowan (1982), the underlying Euclidean symmetry
of the weighting function, together with the restriction to decomposes into two invariant subspaces
doubly periodic functions, then determines the allowable
combinations of plane waves comprising steady-state solu- Vk Vk Vk , (21)
tions. In particular, stripe, chequer-board and hexagonal corresponding to even and odd functions, respec-
patterns of activity can form in the V1 map of the visual tively
¢eld. In this paper we generalize the treatment by
Ermentrout & Cowan (1979) to incorporate the e¡ects of Vk fv 2 Vk : u( ) u()g
orientation preference ö and show how plane waves of
and
cortical activity modulate the distribution of tuning
curves across the network and lead to contoured patterns. Vk fv 2 Vk : u( ) u()g: (22)
As noted in greater generality by Bosch Vivancos
3. LINEAR STABILITY ANALYSIS et al. (1995), this is a consequence of re£ection invar-
The ¢rst step in the analysis of pattern-forming iance, as we now indicate. That is, let k denote
instabilities in the full cortical model is to linearize equa- re£ections about the wavevector k so that k k k.
tion (11) about the homogeneous solution a(r, ) 0 and Then k a(r, )a(k r, 2' )u(' )eikr c:c.
to solve the resulting eigenvalue problem. In particular, Since k is a re£ection, any space that it acts on
we wish to ¢nd conditions under which the homogeneous decomposes into two subspaces, one on which it acts
solution becomes marginally stable due to the vanishing as the identity I and one on which it acts as I.
of one of the (degenerate) eigenvalues, and to identify the
marginally stable modes. This will require performing a Results (i) and (ii) can also be derived directly from
perturbation expansion with respect to the small para- equation (19). For expanding the -periodic function u()
meter characterizing the relative strength of the aniso- as a Fourier series with respect to
tropic lateral connections. X
u() An e2ni , (23)
n2Z
(a) Linearization
We linearize equation (11) about the homogeneous state and setting wlat (r, ) w(R
^ r) leads to the matrix
and introduce solutions of the form a(r, , t) elt a(r, ). eigenvalue equation
This generates the eigenvalue equation
X
Z lAm Am 1 Wm Am W ^ m n (q)An , (24)
d0 n2Z
la(r, ) a(r, ) 1 wloc ( 0 )a(r, 0 )
0 with Wn given by equation (14) and
Z
Z Z
wlat (r r0 , )a(r0 , )dr0 . (17) d
R2 W^ n (q) e 2in e iqx cos ()y sin () w(r)dr
^ .
0 R 2
Because of translation symmetry, the eigenvalue equation (25)
(17) can be written in the form It is clear from equation (24) that item (i) holds. The
decomposition of the eigenfunctions into odd and even
a(r, ) u( ')eikr c:c: (18) invariant subspaces (see equation (21) of item (ii)) is a
consequence of the fact that w(r)
^ is an even function of x
with k q(cos ', sin ') and and y (see equation (5)), and hence W ^ n (q).
^ n (q) W

An z1 n, 1 A(1)

n
2
A(2)
n ..., (28)
λ1 δλ
where n, m is the Kronecker delta function. We substitute
µσ 1 ∆W
these expansions into the matrix eigenvalue equation (26)
0 and systematically solve the resulting hierarchy of
λm equations to successive orders in using (degenerate)
perturbation theory. This analysis, which is carried out in
Appendix A(a), leads to the following results: (i) l l
λ m'
for even () and odd ( ) solutions where to O ( 2 )
l h i
W1 W ^ 0 (q) W^ 2 (q)
β =0 β << 1 1
Figure 15. Splitting of degenerate eigenvalues due to
X W ^ m 1 (q) W ^ m1 (q)2
2
anisotropic lateral connections between hypercolumns. W1 Wm
m50, m61
G (q), (29)
(b) Eigenfunctions and eigenvalues
and (ii) u() u () where to O ( )
The calculation of the eigenvalues and eigenfunctions
of the linearized equation (17), and hence the derivation X
u () cos(2) u
m (q) cos(2m), (30)
of conditions for the marginal stability of the homoge- m50, m61
neous state, has been reduced to the problem of solving X
the matrix equation (24), which we rewrite in the more u () sin(2) um (q) sin(2m), (31)
convenient form m41
X
l ^ m n (q)An . with
Wm Am W (26)
1 n2Z ^ 1 (q)
W
u
0 (q) ,
We exploit the experimental observation that the intrinsic W1 W 0
lateral connections appear to be weak relative to the local (32)
connections, i.e. W ^ W. Equation (26) can then be ^ m 1 (q) W
W ^ m1 (q)
u
m (q) , m41.
solved by expanding as a power series in and using W1 Wm
Rayleigh ^Schro«dinger perturbation theory.
(i) Case 0 (c) Marginal stability

In the limiting case of zero lateral interactions equation Suppose that G (q) has a unique maximum at
(26) reduces to equation (13). Following the discussion of q q 6 0 and let qc q if G (q )4G (q ) and
the ring model in ½ 2(c), let Wp maxfWn , n 2 Z g40 qc q if G (q )4G (q ). Under such circumstances,
and suppose that p 1 (unimodal orientation tuning the homogeneous state a(r, ) 0 will become margin-
curves). The homogeneous state a(r, ) 0 is then stable ally stable at the critical point c /1 G (qc ) and the
for su¤ciently small , but becomes marginally stable at marginally stable modes will be of the form
the critical point c /1 W1 due to the vanishing of X
N
the eigenvalue l1. In this case there are both even and a(r, ) ci eiki r u( 'i ) c:c:, (33)
odd marginally stable modes cos(2) and sin(2). i1
(ii) Case 40 where ki qc (cos 'i , sin 'i ) and u() u () for qc q.
If we now switch on the lateral connections, then there The in¢nite degeneracy arising from rotation invariance
is a q-dependent splitting of the degenerate eigenvalue l1 means that all modes lying on the circle jkj qc become
that also separates out odd and even solutions. Denoting marginally stable at the critical point. However, this can
the characteristic size of such a splitting by l O ( ), be reduced to a ¢nite set of modes by restricting solutions
we impose the condition that l 1 W, where to be doubly periodic functions. The types of doubly peri-
odic solutions that can bifurcate from the homogeneous
W minfW1 Wm , m 6 1g: state will be determined in ½ 4.
As a speci¢c example illustrating marginal stability let
This ensures that the perturbation does not excite states
w(r)
^ be given by equation (5). Substitution into equation
associated with other eigenvalues of the unperturbed
(25) gives
problem (see ¢gure 15). We can then restrict ourselves to
Z Z 1
calculating perturbative corrections to the degenerate d
W ^ n (q) e 2in g(s) cos(sq cos )ds .
eigenvalue l1 and its associated eigenfunctions. Therefore,
0 0
we introduce the power series expansions
Using the Jacobi Ânger expansion
l
W1 l(1) 2 (2)
l ..., (27)
1 X
1
cos(sq cos ) J0 (sq) 2 ( 1)m J2m (sq) cos(2m),
and m1

(a)
1.0
0.98 0.98
0.96 0.96
µ even
0.94
µ 0.94
odd
0.92
odd 0.92
0.9
even
qc
0.90 qc
1 2 3 4 5
q 1 2 3 4 5
q
(b)
1
q− ^
Figure 17. Same as ¢gure 16 except that W(q) satis¢es
0.8 equation (36) with 0 /3 rather than equation (34). It can
q+
be seen that the marginally stable eigenmodes are now even
0.6 functions of .
q
0.4
An interesting question concerns under what circum-
0.2 stances can even patterns be excited by a primary
instability rather than odd, in the regime of weak lateral
interactions. One example occurs when there is a su¤-
0.2 0.4 0.6 0.8 1 cient spread in the distribution of lateral connections
Alat
along the lines shown in ¢gure 12. In particular, suppose
that w(r)
^ is given by equation (8) with p() 1 for 40
Figure 16. (a) Plot of marginal stability curves (q) for g(s) and zero otherwise. Equation (34) then becomes
given by the di¡erence of Gaussians (equation (6)) with
lat 1, ^ lat 3, Alat 1 and 0:4W1 . Also set
Z 1
/1 W1 1. The critical wavenumber for spontaneous W^ n (q) ( 1)n sin(2n0 ) g(s)J2n (sq)ds. (36)
2n0 0
pattern formation is qc . The marginally stable eigenmodes are
odd functions of . (b) Plot of critical wavenumber q for To ¢rst order in the size of the gap between the odd
marginal stability of even ( + ) and odd (7) patterns as a
and even eigenmodes at the critical point qc is deter-
function of the strength of inhibitory coupling Alat . If the
inhibition is too weak then there is a bulk instability with mined by 2W^ 2 (qc ) (see equation 29). It follows that if
0 4/4 then W ^ 2 (q) reverses sign, suggesting that even
respect to the spatial domain.
rather than odd eigenmodes become marginally stable
with Jn (x) the Bessel function of integer order n, we ¢rst. This is con¢rmed by the marginal stability curves
derive the result shown in ¢gure 17.
Z 1
W^ n (q) ( 1)n g(s)J2n (sq)ds. (34) (i) Choosing the bifurcation parameter
0 It is worth commenting at this stage on the choice of
bifurcation parameter . One way to induce a primary
Next we substitute equation (6) into (34) and use standard instability of the homogeneous state is to increase the
properties of Bessel functions to obtain global coupling parameter in equation (29) until the
2 2 2 2 critical point c is reached. However, it is clear from
^ ( 1)n lat q q equation (29) that an equivalent way to induce such an
Wn (q) exp In lat
2 4 4 instability is to keep ¢xed and increase the slope 1 of
^ 2 2 ^ 2 2 the neural output function . The latter could be achieved
lat q q by keeping a non-zero uniform input h(r, , t) h0 in
Alat exp In lat , (35)
4 4 equation (1) so that the homogeneous state is non-zero,
a(r, , t) a0 6 0 with 1 0 (a0 ). Then variation of the
where In is a modi¢ed Bessel function of integer order n. input h0 and consequently 1, corresponds to changing
The resulting marginal stability curves (q) the e¡ective neural threshold and hence the level of
/1 G (q) are plotted to ¢rst order in in ¢gure 16a. The network excitability. Indeed, this is thought to be one of
existence of a non-zero critical wavenumber qc q at the possible e¡ects of hallucinogens. In summary, the
c (qc ) is evident, indicating that the marginally mathematically convenient choice of as the bifurcation
stable eigenmodes are odd functions of . The inclusion of parameter can be reinterpreted in terms of biologically
higher-order terms in does not alter this basic result, at meaningful parameter variations. It is also possible that
least for small . If we take the fundamental unit of hallucinogens act indirectly on the relative levels of inhi-
length to be ca. 400 mm; then the wavelength of a pattern bition and this could also play a role in determining the
is 2(0:400)/qc mm, ca. 2:66 mm at the critical wave- type of patterns that emerge ö a particular example is
number qc 1 (see ¢gure 16b). discussed below.

Substituting these expansions into the matrix eigenvalue

1.0 equation (26) and solving the resulting equations to
successive orders in leads to the following results:
0.98
l X W ^ m (q)2
^ 0 (q)
W0 W 2
O( 3 ) (39)
0.96 1 m40
W0 Wm
µ 0.94 G0 (q)
0.92 and
X
0.9 u() 1 u0m (q) cos(2m O( 2 ), (40)
qc m40
with
1 2 3 4 5
q ^ m (q)
W
u0m (q) . (41)
Figure 18. Plot of marginal stability curve 0 (q) for a bulk W0 Wm
instability with respect to orientation and g(s) given by the
di¡erence of Gaussians (equation (6)) with lat 1, ^lat 3,
Substituting equation (40) into equation (33) shows that
Alat 1:0, 0:4W0 and /1 W0 1. The critical wave- the marginally stable states are now only weakly depen-
number for spontaneous pattern formation is qc . dent on the orientation , and to lowest order in simply
correspond to the spatially periodic patterns of the
Ermentrout ^ Cowan model. The length-scale of these
(d) The Ermentrout ^ Cowan model revisited patterns is determined by the marginal stability curve
The marginally stable eigenmodes (equation (33)) 0 (q) /1 G0 (q), an example of which is shown in
identi¢ed in the analysis consist of spatially periodic ¢gure 18.
patterns of activity that modulate the distribution of The occurrence of a bulk instability in orientation
orientation tuning curves across V1. Examples of these means that for su¤ciently small the resulting cortical
contoured cortical planforms will be presented in ½ 4 and patterns will be more like contrasting regions of light and
the corresponding hallucination patterns in the visual dark rather than a lattice of oriented contours (see ½ 4).
¢eld (obtained by applying an inverse retinocortical map) However, if the strength of lateral connections were
will be constructed in ½ 5. It turns out that the resulting increased, then the eigenfunctions (40) would develop a
patterns account for some of the more complicated form signi¢cant dependence on the orientation . This could
constants where contours are prominent, including then provide an alternative mechanism for the generation
cobwebs, honeycombs and lattices (¢gure 4). However, of even contoured patterns ö recall from ½ 3(c) that only
other form constants such as chequer-boards, funnels and odd contoured patterns emerge in the case of a tuned
spirals (¢gures 6 and 7) comprise contrasting regions of instability with respect to orientation, unless there is
light and dark. One possibility is that these hallucinations signi¢cant angular spread in the lateral connections.
are a result of higher-level processes ¢lling in the
contoured patterns generated in V1. An alternative expla-
nation is that such regions are actually formed in V1 itself 4. DOUBLY PERIODIC PLANFORMS
by a mechanism similar to that suggested in the original As we found in ½ 3 (c) and ½ 3 (d), rotation symmetry
Ermentrout ^ Cowan model. This raises the interesting implies that the space of marginally stable eigenfunctions
issue as to whether or not there is some parameter regime of the linearized Wilson ^ Cowan equation is in¢nite-
in which the new model can behave in a similar fashion dimensional, that is, if u()eikr is a solution then so is
to the `cortical retina' of Ermentrout & Cowan (1979), u( ')eiR' kr . However, translation symmetry suggests
that is, can cortical orientation tuning somehow be that we can restrict the space of solutions of the nonlinear
switched o¡ ? One possible mechanism is the following: Wilson ^ Cowan equation (11) to that of doubly periodic
suppose that the relative level of local inhibition, which is functions. This restriction is standard in many treatments
speci¢ed by the parameter Aloc in equation (15), is of spontaneous pattern formation, but as yet it has no
reduced (e.g. by the possible (indirect) action of halluci- formal justi¢cation. There is, however, a wealth of
nogens). Then W0 maxfWn , n 2 Z g rather than W1, evidence from experiments on convecting £uids and
and the marginally stable eigenmodes will consist of chemical reaction ^ di¡usion systems (Walgraef 1997), and
spatially periodic patterns that modulate bulk instabiliti- simulations of neural nets (Von der Malsburg & Cowan
ties with respect to orientation. 1982), which indicates that such systems tend to generate
To make these ideas more explicit, we introduce the doubly periodic patterns in the plane when the homoge-
perturbation expansions neous state is destabilized. Given such a restriction, the
associated space of marginally stable eigenfunctions is
l then ¢nite-dimensional. A ¢nite set of speci¢c eigen-
W0 l(1) 2 (2)
l ..., (37)
1 functions can then be identi¢ed as candidate planforms,
in the sense that they approximate time-independent
and solutions of equation (11) su¤ciently close to the critical
point where the homogeneous state loses stability. In this
An zn,0 A(1)
n
2
A(2)
n .... (38) section we construct such planforms.

Table 1. Generators for the planar lattices and their dual Table 2. Eigenfunctions corresponding to shortest dual wave
lattices vectors
lattice ``1 ``2 k1 k2 lattice a(r, )
square (1, 0)p (0, 1)p (1, 0) (0, 1) p square c1 u()eik1 r c2 u( /2)eik2 r c:c:
1
hexagonal (1, 1/ 3) (0, 2/ 3) (1, 0) 2
( 1, 3) hexagonal c1 u()eik1 r c2 u( 2/3)eik3 r
rhombic (1, cot ) (0, cosec ) (1, 0) (cos , sin ) c3 u( 2/3)eik3 r c:c:
rhombic c1 u()eik1 r c2 u( )eik2 r c:c:
(a) Restriction to doubly periodic solutions

Let L be a planar lattice; that is, choose two linearly (b) Planforms
independent vectors ` 1 and ` 2 and let We now use an important result from bifurcation
L f2m1 ` 1 2m2 ` 2 : m1 , m2 2 Zg: theory in the presence of symmetries, namely, the equi-
variant branching lemma (Golubitsky et al. 1988). For our
Note that L is a subgroup of the group of planar trans- particular problem, the equivariant branching lemma
lations. A function f :R2 S1 ! R is doubly periodic implies that generically there exists a (unique) doubly
with respect to L if periodic solution bifurcating from the homogeneous state
for each of the axial subgroups of GL under the action
f (x ` , ) f (x, ), (9) ö a subgroup S GL is axial if the dimension of the
for every ` 2 L . Let be the angle between the two basis space of vectors that are ¢xed by S is equal to unity. The
vectors ` 1 and ` 2. We can then distinguish three types of axial subgroups are calculated from the actions presented
lattice according to the value of : square lattice ( /2), in Appendix A(b) (see Bresslo¡ et al. (2000b) for details)
rhombic lattice (055/2, 6 /3) and hexagonal and lead to the even planforms listed in table 3 and the
( =3). After rotation, the generators of the planar odd planforms listed in table 4. The generic planforms
lattices are given in table 1 (for unit lattice spacing). can then be generated by combining basic properties of
Restriction to double periodicity means that the the Euclidean group action (equation (9)) on doubly peri-
original Euclidean symmetry group is now restricted to odic functions with solutions of the underlying linear
the symmetry group of the lattice, GL HL T _ 2, where eigenvalue problem. The latter determines both the
HL is the holohedry of the lattice, the subgroup of O 2) critical wavenumber qc and the -periodic function u().
that preserves the lattice, and T 2 is the two torus of In particular, the perturbation analysis of ½ 3(c) and
planar translations modulo the lattice. Thus, the holo- ½ 3(d) shows that (in the case of weak lateral interactions)
hedry of the rhombic lattice is D2, the holohedry of the u() can take one of three possible forms:
square lattice is D4 and the holohedry of the hexagonal (i) even contoured planforms (equation (30)) with
lattice is D6. Observe that the corresponding space of u() cos(2),
marginally stable modes is now ¢nite-dimensionalöwe (ii) odd contoured planforms (equation (31)) with
can only rotate eigenfunctions through a ¢nite set of u() sin(2),
angles (for example, multiples of /2 for the square lattice (iii) even non-contoured planforms (equation (40)) with
and multiples of /3 for the hexagonal lattice). u() 1.
It remains to determine the space KL of marginally
stable eigenfunctions and the action of GL on this space. Each planform is an approximate steady-state solution
In ½ 3 we showed that eigenfunctions either reside in Vk a(r, ) of the continuum model (equation (11)) de¢ned on
(the even case) or Vk (the odd case) where the length of the unbounded domain R2 S1 . To determine how these
k is equal to the critical wavenumber qc . In particular, solutions generate hallucinations in the visual ¢eld, we
the eigenfunctions have the form u( ')eikr where u is ¢rst need to interpret the planforms in terms of activity
either an odd or even eigenfunction. We now choose the patterns in a bounded domain of V1, which we denote by
size of the lattice so that eikr is doubly periodic with M R. Once this has been achieved, the resulting
respect to that lattice, i.e. k is a dual wave vector for the patterns in the visual ¢eld can be obtained by applying
lattice. In fact, there are in¢nitely many choices for the the inverse retinocortical map as described in ½ 5(a).
lattice size that satisfy this constraint ö we select the one The interpretation of non-contoured planforms is
for which qc is the shortest length of a dual wave vector. relatively straightforward, since to lowest order in the
The generators for the dual lattices are also given in solutions are -independent and can thus be directly
table 1 with qc 1. The eigenfunctions corresponding to treated as activity patterns a(r) in V1 with r 2 M. At the
dual wave vectors of unit length are given in table 2. It simplest level, such patterns can be represented as
follows that KL can be identi¢ed with the m-dimensional contrasting regions of light and dark depending on
complex vector space spanned by the vectors whether a(r)40 or a(r)50. These regions form square,
(c1 , : : :, cm ) 2 C m with m 2 for square or rhombic triangular, or rhombic cells that tile M as illustrated in
lattices and m 3 for hexagonal lattices. It can be shown ¢gures 19 and 20.
that these form GL -irreducible representations. The The case of contoured planforms is more subtle. At a
actions of the group GL on KL can then be explicitly given location r in V1 we have a sum of two or three
written down for both the square or rhombic and hexa- sinusoids with di¡erent phases and amplitudes (see tables 3
gonal lattices in both the odd and even cases. These and 4), which can be written as a(r, ) A(r) cos2
actions are given in Appendix A(b). 20 (r). The phase 0 (r) determines the peak of the

Table 3. Even planforms with u( ) u()

(The hexagon solutions (0) and () have the same isotropy subgroup, but they are not conjugate solutions.)
lattice name planform eigenfunction
square even square u() cos x u( /2) cos y

even roll u() cos x
rhombic even rhombic u() cos(k1 r) u( ) cos(k2 r)
even roll u() cos(k1 r)
hexagonal even hexagon (0) u() cos(k1 r) u( /3) cos(k2 r) u( /3) cos(k3 r)
even hexagon () u() cos(k1 r) u( /3) cos(k2 r) u( /3) cos(k3 r)
even roll u() cos(k1 r)
Table 4. Odd planforms with u( ) u()
lattice name planform eigenfunction
square odd square u() cos x u( /2) cos y

odd roll u() cos x
rhombic odd rhombic u() cos(k1 r) u( ) cos(k2 r)
odd roll u() cos(k1 r)
hexagonal odd hexagon u() cos(k1 r) u( /3) cos(k2 r) u( /3) cos(k3 r)
triangle u() sin(k1 r) u( /3) sin(k2 r) u( /3) sin(k3 r)
patchwork quilt u( /3) cos(k2 r) u( /3) cos(k3 r)
odd roll u() cos(k1 r)
orientation tuning curve at r (see ¢gure 14). Hence the and the visual ¢eld is that the periodicity of the angular
contoured solutions generally consist of iso-orientation retinal coordinate R implies that the y-coordinate in V1
regions or patches over which 0 (r) is constant, but the satis¢es cylindrical periodic boundary conditions (see
amplitude A(r) varies. As in the non-contoured case, these ¢gure 5). This boundary condition should be commensu-
patches are either square, triangular or rhombic in shape. rate with the square, rhombic or hexagonal lattice
However, we now show each patch to be represented by a associated with the doubly periodic planforms.
locally orientated contour centred at the point of maximal
amplitude A(rmax ) within the patch. The resulting odd and (a) The double retinocortical map
even contoured patterns are shown in ¢gures 21 and 22 for An important consequence of the introduction of orien-
the square latttice, in ¢gures 23 and 24 for the rhombic tation as a cortical label is that the retinocortical map
latttice and in ¢gures 25 and 26 for the hexagonal lattice. described earlier needs to be extended to cover the
Note that our particular interpretation of contoured plan- mapping of local contours in the visual ¢eld ö in e¡ect to
forms breaks down in the case of an odd triangle on a treat them as a vector ¢eld. Let R be the orientation of
hexagonal lattice: the latter comprises hexagonal patches such a local contour, and its image in V1. What is the
in which all orientations are present with equal magni- appropriate map from R to that must be added to the
tudes. In this case we draw a `star' shape indicating the map zR ! z described earlier? We note that a line in V1
presence of multiple orientations at a given point, see of constant slope tan is a level curve of the equation
¢gure 26b. Note that this planform contains the well-
known `pinwheels' described by Blasdel (1992). f (x, y) y cos x sin ,
where (x, y) are Cartesian coordinates in V1. Such a line

5. FROM CORTICAL PATTERNS TO VISUAL has a constant tangent vector
HALLUCINATIONS
@ @
In ½ 4 we derived the generic planforms that bifurcate v cos sin :
@x @y
from the homogeneous state and interpreted them in
terms of cortical activity patterns. In order to compute The pre-image of such a line in the visual ¢eld, assuming
what the various planforms look like in visual ¢eld the retinocortical map generated by the complex
coordinates, we need to apply an inverse retinocortical logarithm is obtained by changing from cortical to retinal
map. In the case of non-contoured patterns this can be coordinates via the complex exponential, is
carried out directly using the (single) retinocortical map
introduced in ½ 1(b). However, for contoured planforms it f (x, y) ! f~ (rR , R ) R cos log rR sin ,
is necessary to specify how to map local contours in the
visual ¢eld as well as position ö this is achieved by the level curves of which are the logarithmic spirals
considering a so-called double retinocortical map.
Another important feature of the mapping between V1 rR (R ) A exp(cot ()R ).

(a) (a)
(b)
(b)
Figure 21. Contours of even axial eigenfunctions on the

Figure 19. Non-contoured axial eigenfunctions on the square square lattice: (a) square, (b) roll.
lattice: (a) square, (b) roll.
(a)
(a)
(b)
(b)
Figure 20. Non-contoured axial eigenfunctions on rhombic Figure 22. Contours of odd axial eigenfunctions on the square
and hexagonal lattices: (a) rhombic, (b) hexagonal. lattice: (a) square, (b) roll.

(a) (a)
(b) (b)
Figure 23. Contours of even axial eigenfunctions on the Figure 25. Contours of even axial eigenfunctions on the
rhombic lattice: (a) rhombic, (b) roll. hexagonal lattice: (a) -hexagonal, (b) 0-hexagonal.
(a)
(a)
(b) (b)
Figure 24. Contours of odd axial eigenfunctions on the Figure 26. Contours of odd axial eigenfunctions on the
rhombic lattice: (a) rhombic, (b) roll. hexagonal lattice: (a) triangular, (b) 0-hexagonal.

the far edge of the periphery, so we have no constraint on

yR
φ the patterns in this direction.
Recall that each V1 planform is doubly periodic with
φR respect to a spatial lattice generated by two lattice vectors
` 1 , ` 2 . The cylindrical periodicity is thus equivalent to
requiring that there be an integral combination of lattice
rR vectors that spans Y 96 mm in the y direction with no
change in the x direction:

θR 0 2m ` 2m ` . (43)
1 1 2 2
xR 96
If the acute angle of the lattice 0 is speci¢ed, then the

wavevectors ki are determined by the requirement
Figure 27. The geometry of orientation tuning.

1, i j
ki ` j (44)
0, i 6 j:
It is easy to show that the tangent vector corresponding
to such a curve takes the form The integral combination requirement limits which wave-
@ @ lengths are permitted for planforms in the cortex. The
v~ rR cos( R ) rR sin( R ) : length-scale for a planform is given by the length of the
@xR @yR
lattice vectors j`` 1 j j`` 2 j : j`` j:
Thus the retinal vector ¢eld induced by a constant vector 96
¢eld in V1 twists with the retinal angle R and stretches j`` j p . (45)
with the retinal radius rR. It follows that if R is the m21 2m1 m2 cos(0 ) m22
orientation of a line in the visual ¢eld, then
The commonly reported hallucination patterns usually
R R , (42) have 30^40 repetitions of the pattern around a circumfer-
ence of the visual ¢eld, corresponding to length-scales
i.e. local orientation in V1 is relative to the angular coor- ranging from 2.4^3.2 mm. Therefore, we would expect
dinate of visual ¢eld position. The geometry of the above the critical wavelength 2/qc for bifurcations to be in this
setup is shown in ¢gure 27. range (see ½ 3(c)). Note that when we rotate the planform
The resulting double map fzR , R g ! fz, g has very to match the cylindrical boundary conditions we rotate
interesting properties. As previously noted, the map k1 and hence the maximal amplitude orientations 0 (r)
zR ! z takes circles, rays and logarithmic spirals into by
vertical, horizontal and oblique lines, respectively. What
about the extended map? Because the tangent to a circle at 1 m2 j``j
cos sin( ) 0
0
:
a given point is perpendicular to the radius at that point, Y 2
for circles, R R /2, so that /2. Similarly, for
The resulting non-contoured planforms in the visual ¢eld
rays, R R, so 0. For logarithmic spirals we can
obtained by applying the inverse single retinocortical
write either R a ln rR or rR expbR . In retinal coor-
map to the corresponding V1 planforms are shown in
dinates we ¢nd the somewhat cumbersome formula
¢gures 29 and 30.
brR sin R ebR cos R Similarly, the odd and even contoured planforms
tan R : obtained by applying the double retinocortical map are
brR cos R ebR sin R
shown in ¢gures 31 and 32 for the square lattice, in
However, this can be rewritten as tan (R R ) a, so ¢gures 33 and 34 for the rhombic lattice, and in
that in V1 coordinates, tan a. Thus we see that the ¢gures 35 and 36 for the hexagonal lattice.
local orientations of circles, rays and logarithmic spirals, One of the striking features of the resulting (contoured)
measured in relative terms, all lie along the cortical visual planforms is that only the even planforms appear
images of such forms. Figure 28 shows the details. to be contour completing and it is these that recover the
remaining form constants missing from the original
(b) Planforms in the visual ¢eld Ermentrout ^ Cowan model. The reader should compare,
In order to generate a visual ¢eld pattern we split our for example, the pressure phosphenes shown in ¢gure 1
model V1 domain M into two pieces, each running with the planform shown in ¢gure 35a, and the cobweb
72 mm along the x direction and 48 mm along the y of ¢gure 4 with that of ¢gure 31a.
direction, representing the right and left hemi¢elds in the
visual ¢eld (see ¢gure 5). Because the y coordinate corre-
6. THE SELECTION AND STABILITY OF PATTERNS
sponds to a change from /2 to /2 in 72 mm, which
meets up again smoothly with the representation in the It remains to determine which of the various planforms
opposite hemi¢eld, we must only choose scalings and we have presented above in our model are actually stable
rotations of our planforms that satisfy cylindrical periodic for biologically relevant parameter sets. So far we have
boundary conditions in the y direction. In the x direction, used a mixture of perturbation theory and symmetry to
corresponding to the logarithm of radial eccentricity, we construct the linear eigenmodes (equation (33)) that are
neglect the region immediately around the fovea and also candidate planforms for pattern forming instabilities. To

(a) (b) y
(i)
π /2 π /2
π /2
x
Figure 28. Action of the single

3π /2 −π /2
and double maps on logarithmic
spirals. Dashed lines show the
local tangents to a logarithmic
spiral contour in the visual ¢eld,
(ii) y and the resulting image in V1.
Since circle and ray contours in
π /2 π /2
the visual ¢eld are just special
cases of logarithmic spirals, the
3π /2 x same result holds also for such
contours. (a) Visual ¢eld; (b)
striate cortex, (i) single map,
3π /2 −π /2 (ii) double map.
(a) (a)
(b) (b)
Figure 29. Action of the single inverse retinocortical map on

Figure 30. Action of the single inverse retinocortical map
non-contoured square planforms: (a) square, (b) roll.
on non-contoured rhombic and hexagonal planforms:
(a) rhombic, (b) hexagonal.
determine which of these modes are stabilized by the
nonlinearities of the system we use techniques such as
Liapunov ^Schmidt reduction and Poincarë ^ Lindstedt the bifurcation point). The symmetries of the system
perturbation theory to reduce the dynamics to a set of severely restrict the allowed forms (Golubitsky et al.
nonlinear equations for the amplitudes ci appearing in 1988); however, the coe¤cients in this form are inher-
equation (33) (Walgraef 1997). These amplitude equa- ently model dependent and have to be calculated expli-
tions, which e¡ectively describe the dynamics on a ¢nite- citly.
dimensional centre manifold, then determine the selec- In this section we determine the amplitude equation for
tion and stability of patterns (at least su¤ciently close to our cortical model up to cubic order and use this to

(a) (a)
(b)
(b)
Figure 32. Action of the double inverse retinocortical map on

odd square planforms: (a) square, (b) roll.
Figure 31. Action of the double inverse retinocortical map on
even square planforms: (a) square, (b) roll.
equation is equivalent to the eigenvalue equation (10)
investigate the selection and stability of both odd patterns with l 0, c and jkj qc so that
satisfying u( ) u() and even patterns satisfying
X
N
u( ) u(). A more complete discussion of stability a1 (r, , t) cj (t)eikj r u( 'j ) c:c:, (46)
and selection based on symmetrical bifurcation theory, j1
which takes into account the possible e¡ects of higher-
order contributions to the amplitude equation, will be with kj qc ( cos 'j , sin 'j ). Requiring that the O (e2 )
presented elsewhere (Bresslo¡ et al. 2000b). and O (e3 ) equations in the hierarchy be self-consistent
then leads to a solvability condition, which in turn
(a) The cubic amplitude equation generates evolution equations for the amplitudes cj (t)
Assume that su¤ciently close to the bifurcation point at (see Appendix A(c)).
which the homogeneous state a(r, ) 0 becomes
marginally stable, the excited modes grow slowly at a rate (i) Square or rhombic lattice
O(e2 ) where e2 c. One can then use the method of First, consider planforms (equation (46)) corresponding
multiple-scales to perform a Poincarë ^ Lindstedt pertur- to a bimodal structure of the square or rhombic type
bation expansion in e. First, we Taylor expand the (N 2). That is, take k1 qc (1, 0) and k2 qc ( cos(),
nonlinear function a appearing in equation (11), sin()), with /2 for the square lattice and 055/2,
6 /3 for a rhombic lattice. The amplitudes evolve
a 1 a 2 a2 3 a3 : : : , according to a pair of equations of the form
dc1 2
where 1 0 0, 2 00 0/2, 3 000 0/3!, etc. Then c1 L 0 jc1 j 2 jc2 j2 ,
we perform a perturbation expansion of equation (11) dt (47)
dc2
with respect to e by writing c2 L 0 jc2 j
2
2 jc1 j2 ,
dt
a ea1 e2 a2 : : : , where L c measures the deviation from the
critical point, and
and introducing a slow time-scale t e2 t. Collecting
terms with equal powers of e then generates a hierarchy 3 j3 j (3)
' G ('), (48)
of equations as shown in Appendix A(c). The O (e) 1

(a)
(a)
(b) (b)
Figure 33. Action of the double inverse retinocortical map on Figure 35. Action of the double inverse retinocortical map on
even rhombic planforms: (a) rhombic, (b) roll. even hexagonal planforms: (a) -hexagonal, (b) 0-hexagonal.
(a) (a)
(b) (b)
Figure 34. Action of the double inverse retinocortical map on Figure 36. Action of the double inverse retinocortical map on
odd rhombic planforms: (a) rhombic, (b) roll. odd hexagonal planforms: (a) triangular, (b) 0-hexagonal.

for all 04'5 with (v) Even non-contoured planforms

Z Substituting the perturbation expansion of the eigen-
d
G(3) (') u( ')2 u()2 . (49) function (equation (40)) for even non-contoured plan-
0 forms into equations (52) and (49) gives, respectively,
X
G(2) 1 32 2 u0m (qc )2 cos(2m/3) O ( 3 ), (57)
(ii) Hexagonal lattice m40
Next consider planforms on a hexagonal lattice with
N 3, '1 0, '2 2/3, '3 2/3. The cubic ampli- and
tude equations take the form X
G(3) () 1 2
u0m (qc )2 1 2 cos(2m) O ( 3 ),
dcj 2 2
m40
cj L 0 jcj j 2 2=3 (jcj1 j jcj 1 j2 ) cj 1 cj1 , (58)
dt
(50) with the coe¤cients u0n de¢ned by equation (41).
where j 1, 2, 3 mod 3, 2=3 is given by equation (49)
for ' 2/3, and (b) Even and odd patterns on square or rhombic
lattices
2
G(2) ,
p (51) We now use equation (47) to investigate the selection
1 1 W1 and stability of odd or even patterns on square or
with rhombic lattices. Assuming that 40 and L40, the
Z following three types of steady state are possible for arbi-

d trary phases 1 , 2.
G(2) u()u( 2/3)u( 2/3) . (52)

0 (i) the homogeneous state: c1 c2 0. p
p
In deriving equation (50) we have assumed that the (ii) rolls: c1 L/ 0 ei 1 , c2 0 or c1 0, c2 L= 0 ei 2.
p
neurons are operating close to threshold such that (iii) squares or rhombics: c1 L/ 0 2 ei 1 ,
p
i
2 O (e). c2 L/ 0 2 e 2 .
The basic structure of equations (47) and (50) is
universal in the sense that it only depends on the under- The non-trivial solutions correspond to the axial plan-
lying symmetries of the system and on the type of bifurca- forms listed in tables 3 and 4. A standard linear stability
tion that it is undergoing. In contrast, the actual values of analysis shows that if 2 4 0 , then rolls are stable
the coe¤cients ' and are model-dependent and have whereas the square or rhombic patterns are unstable. The
to be calculated explicitly. Moreover, these coe¤cients are opposite holds if 2 5 0 . These stability properties
di¡erent for odd and even patterns because they have persist when higher-order terms in the amplitude
distinct eigenfunctions u(). Note also that because of equation are included (Bresslo¡ et al. 2000b).
symmetry the quadratic term in equation (50) must Using equations (48), (54), (56) and (58) with 3 j3 j/1
vanish identically in the case of odd patterns. 1, we deduce that for non-contoured patterns
2 0 1 O ( ),
(iii) Even contoured planforms
Substituting the perturbation expansion of the eigen- and for (odd or even) contoured patterns
function (equation (30)) for even contoured planforms 1 2 cos(4)
into equations (52) and (49) gives 2 0 O ( ).
8
G(2) 34 u
2 (qc ) u 2
0 (qc ) O ( ) (53) Hence, in the case of a square or rhombic lattice we have
the following results concerning patterns bifurcating from
G(3) () 18 2 cos(4) 4 u 2
3 (qc ) cos(4) O ( ) the homogeneous state close to the point of marginal
(54) stability (in the limit of weak lateral interactions):
(i) For non-contoured patterns on a square or rhombic
with the coe¤cients u
n de¢ned by equation (32).
lattice there exist stable rolls and unstable squares.
(ii) For (even or odd) contoured patterns on a square
(iv) Odd contoured planforms
lattice there exist stable rolls and unstable squares. In
Substituting the perturbation expansion of the eigen-
the case of a rhombic lattice of angle 6 /2, rolls
function (equation (31)) for odd contoured planforms into
are stable if cos(4)4 1/2 whereas -rhombics are
equations (52) and (49) gives, respectively,
stable if cos(4)5 1/2, that is, if /655/3.
G(2) 0, (55) It should be noted that this result di¡ers from that obtained
and by Ermentrout & Cowan (1979) in which stable squares
were shown to occur for certain parameter ranges (see also
G(3) () 18 2 cos(4) 4 u3 (qc ) cos(4) O ( 2 ), Ermentrout (1991)). We attribute this di¡erence to the
(56) anisotropy of the lateral connections incorporated into the
current model and the consequent shift ^ twist symmetry of
with the coe¤cients un de¢ned by equation (32). Note the Euclidean group action. The e¡ects of this anisotropy
that the quadratic term in equation (50) vanishes identi- persist even in the limit of weak lateral connections, and
cally in the case of odd patterns. preclude the existence of stable square patterns.

qc
rolls
0.2
RA
w+(q)
f (q)
0.1
u3+(q) 0-hexagons
C
q
1 2 3 4 5
−0.05
π -hexagons
Figure 37. Plot of the even eigenfunction coe¤cient u 3 (q) of

equation (32) as a function of wavenumber q. Also plotted is
the O( ) contribution to the even eigenvalue expansion,
equation (29), w (q) W^ 0 (q) W
^ 2 (q). The peak of w (q) µc
µ
determines the critical wavenumber qc (to ¢rst order in ).
Same parameter values as ¢gure 17.
Figure 38. Bifurcation diagram showing the variation of the
(c) Even patterns on a hexagonal lattice amplitude C with the parameter for even hexagonal and roll
Next we use equations (50) and (51) to analyse the patterns with 40. Solid and dashed curves indicate stable
stability of even planforms on a hexagonal lattice. On and unstable solutions, respectively. Also shown is a secondary
branch of rectangular patterns, RA. Higher-order terms in
decomposing ci Ci ei i , it is a simple matter to show that
the amplitude equation are needed to determine its stability.
two P of the phases i are arbitrary while the sum
3
i1 i and the real amplitudes Ci evolve according
to the equations (0-hexagons) whereas the maxima are located on an
equilateral hexagonal lattice for 50 (-hexagons).
dCi
LCi Ci1 Ci 1 cos 3
0 Ci 2 2 2
2=3 (Ci1 Ci 1 )Ci
Both classes of hexagonal planform have the same D6
dt axial subgroup (up to conjugacy), see table A2 in
(59) Appendix A(b). One can also establish that rolls are
and unstable versus hexagonal structures in the range
d X
3
Ci1 Ci
1
sin , (60) 2
dt Ci 05L5 . (62)
i1 0 2 2=3 2
with i, j 1, 2, 3 mod 3. It immediately follows from
equation (60) that the stable steady-state solution will Hence, in the case of a hexagonal lattice, we have the
have a phase 0 if 40 and a phase if 50. following result concerning the even patterns bifurcating
From equations (48), (54) and (58) with 3 j3 j/1 1 from the homogeneous state close to the point of marginal
we see that stability (in the limit of weak lateral interactions). For
even (contoured or non-contoured) patterns on a
2 2=3 0 1 O ( 2 ), hexagonal lattice, stable hexagonal patterns are the ¢rst
to appear (subcritically) beyond the bifurcation point.
for even non-contoured patterns, and Subsequently, the stable hexagonal branch exchanges
2 u3 (qc ) O ( 2 ), stability with an unstable branch of roll patterns, as
2=3 0
shown in ¢gure 38.
for even contoured patterns. In the parameter regime Techniques from symmetrical bifurcation theory can be
where the marginally stable modes are even contoured used to investigate the e¡ects of higher order terms in the
planforms (such as in ¢gure 17) we ¢nd that u 3 (qc )40. amplitude equation (Buzano & Golubitsky 1983): in the
This is illustrated in ¢gure 37. case of even planforms the results are identical to those
Therefore, 2 2=3 4 0 for both the contoured and non- obtained in the analysis of Bënard convection in the
contoured cases. Standard analysis then shows that (to absence of midplane symmetry. For example, one ¢nds
cubic order) there exists a stable hexagonal pattern that the exchange of stability between the hexagons and
Ci C for i 1, 2, 3 with amplitude (Busse 1962) rolls is due to a secondary bifurcation that generates
h q i rectangular patterns.
1
C jj 2 4 0 4 2=3 L , (61)
2 0 4 2=3 (d) Odd patterns on a hexagonal lattice
over the parameter range Recall that in the case of odd patterns, the quadratic
term in equation (50) vanishes identically. The homoge-
2 22 0 2=3 neous state now destabilizes via a (supercritical) pitchfork
5L5 2 . bifurcation to the four axial planforms listed in tables 3
4 0 4 2=3 0 2 2=3
and 4. In this particular case it is necessary to include
The maxima of the resulting hexagonal pattern are higher-order (quartic and quintic terms) in the amplitude
located on an equilateral triangular lattice for 40 equation to completely specify the stability of these

u3−(q)
0.4
qc
0.2
w−(q)
H,T
f (q)
q
1 2 3 4 5 C
−0.2 PQ
−0.4 R
Figure 39. Plot of the odd eigenfunction coe¤cient u3 (q) of

equation (32) as a function of wavenumber q. Also plotted is
the O( ) contribution to the odd eigenvalue expansion,
µc
µ
equation (29), w (q) W ^ 0 (q) W^ 2 (q). The peak of w (q)
determines the critical wavenumber qc (to ¢rst order in ). Figure 40. Bifurcation diagram showing the variation of the
Same parameter values as ¢gure 16. amplitude C with the parameter for odd patterns on a
hexagonal lattice. Solid and dashed curves indicate stable and
various solutions, and to identify possible secondary unstable solutions, respectively. Either hexagons (H) or
bifurcations. Unfortunately, one cannot carry over triangles (T) are stable (depending on higher-order terms in
the amplitude equation) whereas patchwork quilts (PQ ) and
previous results obtained from the study of the Bënard
rolls (R) are unstable. Secondary bifurcations (not shown)
convection problem with midplane symmetry, even may arise from higher-order terms (Bresslo¡ et al. 2000b).
though the corresponding amplitude equation is identical
in structure at cubic order (Golubitsky et al. 1984): higher-
order contributions to the amplitude equation will di¡er Cowan started with a general neural network and sought
in the two problems due to the radically di¡erent actions the minimal restrictions necessary to produce hallucina-
of the Euclidean group and the resulting di¡erences in tion patterns, the current model incorporates data gath-
the associated axial subgroups.1 The e¡ects of such contri- ered over the past two decades to show that common
butions on the bifurcation structure of odd (and even) hallucinatory images can be generated by a biologically
cortical patterns will be studied in detail elsewhere plausible architecture in which the connections between
(Bresslo¡ et al. 2000b). Here, we simply describe the more iso-orientation patches in V1 are locally isotropic, but
limited stability results that can be deduced at cubic non-locally anisotropic. As we, and Ermentrout and
order. A basic question concerns which of the four odd Cowan before us show, the Euclidean symmetry of such
planforms on a hexagonal lattice (hexagons, triangles, an architecture, i.e. the symmetry with respect to rigid
patchwork quilts and rolls) are stable. It turns out that if motions in the plane, plays a key role in determining
2 2=3 4 0 , then rolls are stable, whereas if 2 2=3 5 0 which patterns of activation of the iso-orientation patches
then either hexagons or triangles are stable (depending appear when the homogeneous state becomes unstable,
upon higher-order terms). Equations (49) and (56) with presumed to occur, for example, shortly after the action
3 j3 j/1 1 imply that of hallucinogens on those brain stem nuclei that control
2 u3 (qc ) O ( 2 ). (63) cortical excitability.
2=3 0
There are, however, two important di¡erences between
In the parameter regime where the marginally stable the current work and that of Ermentrout and Cowan in
modes are odd contoured planforms (such as in ¢gure 16) the way in which the Euclidean group is implemented:
we ¢nd that u3 (qc )50, and thus 2 2=3 5 0. This is
illustrated in ¢gure 39. Hence, in the case of a hexagonal (i) The group action is di¡erent and novel, and so the
lattice we have the following result concerning the odd way in which the various subgroups of the Euclidean
patterns bifurcating from the homogeneous state close to group are generated is signi¢cantly di¡erent. In
the point of marginal stability (in the limit of weak particular, the various planforms corresponding to
lateral interactions). For odd (contoured) patterns on a the subgroups are labelled by orientation preference,
hexagonal lattice there exist four primary bifurcation as well as by their location in the cortical plane. It
branches corresponding to hexagons, triangles, patchwork follows that the eigenfunctions that generate such
quilts and rolls. Either the hexagons or the triangles are planforms are also labelled in such a fashion. This
stable (depending on higher-order terms through a adds an additional complication to the problem of
secondary bifurcation) and all other branches are calculating such eigenfunctions and the eigenvalues
unstable. This is illustrated in ¢gure 40. to which they belong, from the linearized cortical
dynamics. Assuming that the non-local lateral or
horizontal connections are modulatory and weak
7. DISCUSSION
relative to the local connections, we show how the
This paper describes a new model of the spontaneous methods of Rayleigh ^Schro«dinger degenerate pertur-
generation of patterns in V1 (seen as geometric hallucina- bation theory can be used to compute, to some
tions). Whereas the earlier work of Ermentrout and appropriate level of approximation, the requisite

eigenvalues and eigenfunctions, and therefore the weakly biased inputs to a hypercolumn can trigger a
planforms. Given such eigenfunctions we then make sharply tuned response such that, under the
use of Poincarë ^ Lindstedt perturbation theory to combined action of many interacting hypercolumns,
compute the stability of the various planforms that plane waves labelled for orientation preference can
appear when the homogeneous state becomes emerge. The resulting planforms correspond to
unstable. contoured patterns and to the remaining form
(ii) Because we include orientation preference in the constants described by Klu«ver ö honeycombs and
formulation, we have to consider the action of the chequer-boards, and cobwebs. Interestingly, all but
retinocortical map on orientated contours or edges. the square planforms are stable, but there do exist
In e¡ect we do this by treating the local tangents to hallucinatory images that correspond to square
such contours as a vector ¢eld. As we discussed, this planforms and it is possible that these are just transi-
is carried out by the tangent map associated with tional forms.
the complex logarithm, one consequence of which is (ii) Another conclusion to be drawn from this analysis is
that , the V1 label for orientation preference, is not that the circuits in V1, which are normally involved
exactly equal to orientation preference in the visual in the detection of oriented edges and the formation
¢eld, R, but di¡ers from it by the angle R , the and processing of contours, are also responsible for
polar angle of receptive ¢eld position. We called the the generation of the hallucinatory form constants.
map from visual ¢eld coordinates frR , R , R g to V1 Thus, we introduced in ½ 2(a) a V1 model circuit in
coordinates fx, y, g a double map. Its possible which the lateral connectivity is anisotropic and
presence in V1 is subject to experimental veri¢cation. inhibitory. (We noted in ½ 1(d) that 20% of the
If the double map is present, then elements tuned to (excitatory) lateral connections in layers II and III
the same relative angle should be connected with of V1 end on inhibitory inter-neurons, so the overall
greater strength than others; if only the single map action of the lateral connections could become
frR , R g ! fx, yg obtains, then elements tuned to the inhibitory, especially at high levels of activity.) As we
same absolute angle R should be so connected. If, demonstrated in ½ 3(c) the mathematical conse-
in fact, the double map is present, then elements quences of this is the selection of odd planforms, but
tuned to the same angle should be connected these do not form continuous contours (see ½ 5(b)).
along lines at that angle in V1. This would support This is consistent with the possibility that such
Mitchison and Crick's hypothesis on connectivity in connections are involved in the segmentation of
V1 (Mitchison & Crick 1982) and would be consis- visual images (Li 1999). In order to select even plan-
tent with the observations of G. G. Blasdel (personal forms, which are contour forming and correspond to
communication) and Bosking et al. (1997). In this seen form constants, it proved su¤cient to allow for
connection, it is of interest that from equation (42) it deviation away from the visuotopic axis by at least
follows that near the vertical meridian (where most 458 in the pattern of lateral connections between
of the observations have been made), changes in iso-orientation patches. These results are consistent
closely approximate changes in R. However, a with observations that suggest that there are two
prediction of the double map is that such changes circuits in V1, one dealing with contrast edges, in
should be relatively large and detectable with optical which the relevant lateral connections have the
imaging, near the horizontal meridian. anisotropy found by G. G. Blasdel and L. Sincich
(personal communication) and Bosking et al. (1997),
The main advance over the Ermentrout ^ Cowan work
and another that might be involved with the proces-
is that all the Klu«ver form constants can now be obtained
sing of textures, surfaces and colour contrast, and
as planforms associated with axial subgroups of the
which has a much more isotropic lateral connec-
Euclidean group in the plane, generated by the new
tivity Livingstone & Hubel 1984). One can interpret
representations we have discovered. There are several
the less anisotropic pattern needed to generate even
aspects of this work that require comment.
planforms as a composite of the two circuits.
(i) The analysis indicates that under certain conditions There are also two other intriguing possible
the planforms are either contoured or else non- scenarios that are consistent with our analysis. The
contoured, depending on the strength of inhibition ¢rst was referred to in ½ 3(d). In the case where V1 is
between neighbouring iso-orientation patches. If operating in the Hubel ^Wiesel mode, with no
such inhibition is weak, individual hypercolumns do intrinsic tuning for orientation, and if the lateral
not exhibit any tendency to amplify any particular interactions are not as weak as we have assumed in
orientation. In normal circumstances, such a prefer- our analysis, then even contoured planforms can
ence would have to be supplied by inputs from the form. The second possibility stems from the observa-
LGN. In this case, V1 can be said to operate in the tion that at low levels of V1 activity, lateral
Hubel ^Wiesel mode (see ½ 2(c)). If the horizontal interactions are all excitatory (Hirsch & Gilbert
interactions are still e¡ective, then plane waves of 1991), so that a bulk instability occurs if the homo-
cortical activity can emerge, with no label for orien- geneous state becomes unstable, followed by
tation preference. The resulting planforms are called secondary bifurcations to patterned planforms at the
non-contoured, and correspond to a subset of the critical wavelength of 2:4^3:2 mm, when the level of
Klu«ver form constants: tunnels and funnels, and activity rises and the inhibition is activated. In
spirals. Conversely, if there is strong inhibition many cases secondary bifurcations tend to be asso-
between neighbouring iso-orientation patches, even ciated with complex eigenvalues, and are therefore

(a)
Figure 41. Tunnel hallucination generated by LSD. Redrawn (b)

from Oster (1970).
Hopf bifurcations (Ermentrout & Cowan 1980) that

give rise to oscillations or propagating waves. In
such cases it is possible for even planforms to be
selected by the anistropic connectivity and odd
planforms by the isotropic connectivity. In addition,
such a scenario is actually observed: many subjects
who have taken LSD and similar hallucinogens
report seeing bright white light at the centre of the
visual ¢eld, which then explodes into a hallucinatory
image (Siegel & Jarvik 1975) in ca. 3 s, corre-
sponding to a propagation velocity in V1 of ca.
2:4 cm s 1, suggestive of slowly moving epileptiform Figure 42. (a) Lattice^tunnel hallucination generated by
activity (Milton et al. 1995; Senseman 1999). marijuana. Reproduced from Siegel (1977), with permission
(iii) One of the major aspects described in this paper is from Alan D. Eiselin. (b) A simulation of the lattice^tunnel.
the presumed Euclidean symmetry of V1. Many
systems exhibit Euclidean symmetry, but what is
novel here is the way in which such a symmetry is
generated. Thus, equation (9) shows that the
symmetry group is generated, in large part, by a
translation or shift fr, g ! fr s, g followed by a
rotation or twist fr, g ! fR r, g. It is the ¢nal
twist ! that is novel, and which is required
to match the observations of G. G. Blasdel and
L. Sincich (personal communication) and Bosking
et al. (1997). In this respect it is of considerable
interest that Zweck & Williams (2001) have
introduced a set of basis functions with the same Figure 43. Complex hallucination generated by LSD.
shift ^ twist symmetry as part of an algorithm to Redrawn from Oster (1970).
implement contour completion. Their reason for
doing so is to bind sparsely distributed receptive
¢elds together functionally, to perform Euclidean segment of medial superior temporal cortex (MSTd)
invariant computations. It remains to explicate the that process radial motion, are involved in the
precise relationship between the Euclidean invariant generation of such images, via a feedback to V1
circuits we have introduced here, and the Euclidean (Morrone et al. 1995).
invariant receptive ¢eld models introduced by Similarly, the lattice ^ tunnel shown in ¢gure 42a
Zweck and Williams. is more complicated than any of the simple form
(iv) Finally, it should also be emphasized that many constants shown earlier. One intriguing possibility is
variants of the Klu«ver form constants have been that such images are generated as a result of a
described, some of which cannot be understood in mismatch between the planform corresponding to
terms of the simple model we have introduced. For one of the Klu«ver form constants, and the underlying
example the tunnel image shown in ¢gure 41 structure of V1. We have (implicitly) assumed that V1
exhibits a reversed retinocortical magni¢cation, and has patchy connections that endow it with lattice
corresponds to images described in Knauer & properties. It should be clear from ¢gures 9 and 10
Maloney (1913). It is possible that some of the that such a cortical lattice is somewhat disordered.
circuits beyond V1, for example, those in the dorsal Thus one might expect some distortions to occur

when planforms are spontaneously generated in such ^ ^ 2 (q) z1

W0 (q) W (1) z1
a lattice. Figure 42b shows a computation of the l . (A4)
W^ 2 (q) ^ 0 (q)
W z 1 z 1
appearance in the visual ¢eld of a hexagonal roll on a
square lattice, when there is a slight incommensur- Equation (A4) has solutions of the form
ability between the two.
As a last example we show in ¢gure 43 another ^ 0 (q) W
l(1) W ^ 2 (q), (A5)
hallucinatory image triggered by LSD. Such an
image does not ¢t very well as a form constant. z 1 z1 , (A6)
However, there is some secondary structure along the where plus and minus denote the even and odd solutions.
main (horizontal) axis of the its major components. We have used the result W ^ 2W ^ 2. The O ( ) terms in
(This is also true of the funnel and spiral images equation (A1) for which m 6 1 generate the corre-
shown in ¢gure 2, also triggered by LSD.) This sponding ¢rst-order amplitudes
suggests the possibility that at least two di¡ering
length-scales are involved in their generation, but ^ m 1 (q)z1 W
W ^ m1 (q)z 1
this is beyond the scope of the model described in the A(1)
m . (A7)
W1 Wm
current paper. It is of interest that similar images
have been reported following stimulation with £ick-
ering light (Smythies 1960). (ii) O ( 2 ) terms
The O ( 2 ) contribution to equation (A1) for m 1 is
The authors wish to thank Dr Alex Dimitrov, Dr Trevor Mundel X
and Dr Gary Blasdel for many helpful discussions. The authors W^ 1 n (q)A(1) ^
n W0 (q) l(1) A(1) ^ (1) (2)
1 W2 (q)A 1 l z1 .
also wish to thank the referees for a number of helpful comments n61
and Alan D. Eiselin for permission to reproduce his artwork in Combining with the analogous equation for m 1 yields
¢gure 41. This work was supported in part by grant 96-24 from
the James S. McDonnell Foundation to J.D.C. The research of
the matrix equation
M.G. was supported in part by NSF grant DMS-9704980. M.G. ^
wishes to thank the Center for Biodynamics, Boston University, W0 (q) l(1) W ^ 2 (q) A(1)
1

B1 (q)

for its hospitality and support. The research of P.C.B. was sup- ,
^ 2 (q)
W ^ 0 (q) l(1)
W A(1)1 B 1 (q)
ported by a grant from the Leverhulme Trust. P.C.B. wishes to
thank the Mathematics Department, University of Chicago, for (A8)
its hospitality and support. P.C.B. and J.D.C. also wish to thank
Professor Geo¡rey Hinton FRS and the Gatsby Computational where
Neurosciences Unit, University College, London, for hospitality X
and support. P.J.T. was supported, in part, by NIH grant T-32- B1 (q) l(2) z1 ^ 1 n (q)A(1)
W n . (A9)
MH20029. n61
Multiplying both sides of equation (A8) on the left by

APPENDIX A z 1 , z1 ) and using equation (A4) implies that B1 (q) 0.
(a) Perturbation expansion of the eigenfunctions This, together with equation (A7), determines the second-
We summarize here the derivation of equations (29)^ order contribution to the eigenvalue:
(31). This involves solving the matrix equation X W ^ 1 m (q) W
^ 1m (q)2
l(2) . (A10)
l X m61;m50
W1 W m
Wm A m W^ m n (q)An , (A1)
1 n2Z Having obtained l(2) , we can then use equations (A8) and
(A5) to obtain the result
using a standard application of degenerate perturbation
theory. That is, we introduce the perturbation expansions A(1)1 A(1)
1 . (A11)
l The unknown amplitudes z1 and A(1) 1 are determined by
W1 l(1) 2 (2)
l ..., (A2)
1 the overall normalization of the solution.
Finally, combining equations (A2), (A5), (A6), and
(A10) generates equation (29). Similarly, combining equa-
An z1 n, 1 A(1)
n
2
A(2)
n ..., (A3) tions (A3), (A6), (A7), (A11) and (23) yields the pair of
equations (30) and (31).
and substitute these into the eigenvalue equation (26). We
then systematically solve the resulting hierarchy of (b) Construction of axial subgroups
equations to successive orders in . We sketch how to construct the axial subgroups from
the irreducible representations of the holohedry HL corre-
(i) O ( ) terms sponding to the shortest dual wave vectors as given in
Setting m 1 in equation (A1) yields the O ) equation table 2. By rescalings we can assume that the critical
wavenumber qc 1 and that the doubly periodic func-
W ^ 2 (q)z
^ 0 (q)z1 W 1 l(1) z1 . tions are on a lattice L whose dual lattice L is generated
by wave vectors of length unity. There are two types of
Combining this with the conjugate equation m 1 we irreducible representations for each lattice corresponding
obtain the matrix equation to the cases u() odd and u() even. We derive the explicit

Table A1. Torus action on GL -irreducible representation Table A3. Axial subgroups when u( ) u()
2
(O(2) is generated by [0, 2 ]2 T and rotation by ( on the
lattice torus action rhombic lattice, 2 on the square lattice, and 3 on the
hexagonal lattice). The points (1, 1, 1) and ( 1, 1, 1)
square (e2i1 c1 , e2i2 c2 ) have the same isotropy subgroup (D6 ), but are not conjugated
hexagonal (e2i1 c1 , e2i1 c2 , e 2i(1 2 )
c3 ) by a group element; therefore, the associated eigenfunctions
rhombic (e2i1 c1 , e2i2 c2 ) generate di¡erent planforms.)
lattice subgroup S ¢x(S) name

2
Table A2. Left, D2 T action on rhombic lattice (centre,
_ 2 action on square lattice; right, D6 T 2 action on
D4 T square D4 (, ) (1, 1) even square
O(2) Z 2 () (1, 0) even roll
hexagonal lattice)
rhombic D2 ( , ) (1, 1) even roll
(For u() even, e 1; for u() odd, e 1.) O(2) (1, 0) even rhombic
hexagonal D6 (, ) (1, 1, 1) even hexagon (0)
D2 action D4 action D6 action D6 (, ) ( 1, 1, 1) even hexagon ()
O(2) Z 2 () (1, 0, 0) even roll
1 (c1 , c2 ) 1 (c1 , c2 ) 1 (c1 , c2 , c3 )
(c1 , c2 ) (c2 , c1 ) (c2 , c3 , c1 ) Table A4. Axial subgroups when u( ) u()
e(c2 , c1 ) 2 (c1 , c2 ) 2 (c3 , c1 , c2 ) 2
e(c2 , c1 ) 3 (c2 , c1 ) 3 (c1 , c2 , c3 ) (O(2) is generated by [0, 2 ]2 T and rotation by ( on the
e(c1 , c2 ) 4 (c2 , c3 , c1 ) rhombic lattice, 2 on the square lattice, and 3 on the
hexagonal lattice).)
e(c2 , c1 ) 5 (c3 , c1 , c2 )
2 e(c1 , c2 ) e(c1 , c3 , c2 )
3 e(c2 , c1 ) e(c2 , c1 , c3 ) lattice subgroup S ¢x(S) name
2 e(c3 , c2 , c1 )
3 e(c1 , c3 , c2 ) square D4 (12 , 12, ) (1, 1) odd square
4 e(c2 , c1 , c3 ) O(2) Z 2 (2 12 , 0) (1, 0) odd roll
5 e(c3 , c2 , c1 ) rhombic D2 ( , 12 , 12, ) (1, 1) odd rhombic
2ii1 2i2 2i1 O(2) Z 2 (2 12 , 0) (1, 0) odd roll
[1 , 2 ] (e c1 , e c2 ) (e c1 , e 2i2 c2 ,
2i(1 2 ) hexagonal Z 6 () (1, 1, 1) odd hexagon
e c3 )
D3 (, 2 ) (i, i, i) triangle
D2 (, 3 ) (0, 1, 1) patchwork quilt
O(2) Z 4 (3 12 , 0) (1, 0, 0) odd roll
action of GL on these subspaces and determine the axial
subgroups.
The action of the torus T 2 on the subspace KL is ally stable, the excited modes grow slowly at a rate O (e2 )
derived as follows: write 2 T 2 as where e2 c. We use the method of multiple scales to
21 ` 1 22 ` 2 : derive the cubic amplitude equations (47) and (50).
Using the fact that ki `j ij , the result of the trans- (i) Multiple-scale analysis
lation action is given in table A1. We begin by rewriting equation (11) in the more
The holohedries HL are D4, D6 and D2 for the square, compact form
hexagonal and rhombic lattices, respectively. In each
case, the generators for these groups are a re£ection and da
a w a, (A12)
a rotation. For the square and hexagonal lattices, the dt
re£ection is , the re£ection across the x-axis where
r x, y). For the rhombic lattice, the re£ection is . with
The counterclockwise rotation , through angles /2, /3 Z
and , is the rotation generator for the three lattices. The w a wloc ( 0 )a(r, 0 ; t)d0
action of HL for the various lattices is given in table A2. 0
Z
Finally, for each of the six types of irreducible represen-
tations, we compute the axial subgroups ö those isotropy wlat (r r0 , )a(r0 , ; t)dr0 , (A13)
R2
subgroups S that have one-dimensional ¢xed-point
subspaces ¢x(S), in each irreducible representation. The c e2 . Taylor expanding the nonlinear function a
computations for the square and rhombic lattices are appearing in equation (A12) gives
straightforward because we can use the T 2 action in
table A1 to assume, after conjugacy, that c1 and c2 are real
a 1 a 2 a2 3 a3 : : : ,
and non-negative. The computations on the hexagonal
lattice are more complicated (Bresslo¡ et al. 2000b). The
results, up to conjugacy, are listed in tables A3 and A4. where 1 0 0, 2 00 0/2, 3 000 0/3!, etc. Then
we perform a perturbation expansion of equation (A12)
(c) Derivation of the amplitude equation with respect to e by writing
Assume that su¤ciently close to the bifurcation point at
which the homogeneous state a(r, ) 0 becomes margin- a ea1 e2 a2 : : : ,

and introducing a slow time-scale t e2 t. Collecting would then give a quadratic (rather than a cubic)
terms with equal powers of e then generates a hierarchy amplitude equation describing the growth of unstable
of equations of the form hexagonal patterns. In the case of odd patterns
hvl jw a21 i 0 and no restriction on 2 is required.
La1 0, However, for ease of exposition we treat the odd and even
(A14)
Lan bn , n41, cases in the same way.
where (iii) Amplitude equations

La a c 1 w a, In order to explicitly derive the amplitude equations
(47) and (50) from the solvability condition (A18), we
and need to evaluate inner products of the form hvl jw an1 i.
Since vl is a solution to the linear equations (A14), it
b2 c 2 w a21 , (A15) follows that

da1 hvl jw an1 i hw vl jan1 i hvl jan1 i. (A19)
b3 c 3 w a31 2c 2 w a1 a2 1 w a1 . c 1
dt
(A16) Thus, substituting equation (A17) into the left-hand side
of equation (A18) and using (A19) shows that

(ii) Solvability conditions da1 dc
vl 1 w a1 1 G(1) l 1 W01 G^ (1) cl
The ¢rst equation in the hierarchy is equivalent to the dt dt
eigenvalue equation (26) with l 0, c and jkj qc. (A20)
Therefore, the relevant classes of solution are of the form
(equation (46)): with G(1) , G^ (1) O ( ). The -dependent factors appearing
on the right-hand side of equation (A20) are eliminated
X
N
a1 (r, , t) cj (t)eikj r u( 'j ) c:c: (A17) from the ¢nal amplitude equations by an appropriate
j1 rescaling of the time t and a global rescaling of the ampli-
tudes cj . Similarly,
Following ½ 4 we restrict solutions to the space of doubly
X
3
periodic functions on a square or rhombic lattice (N 2) hvl ja21 i G(2) ci cj (ki kj kl ) (A21)
or a hexagonal lattice (N 3). Next we de¢ne the inner i, j1
product of two arbitrary functions a(r, ), b(r, )
and
according to h i
Z Z X
d hvl ja31 i3cl G(3) (0)jcl j2 2 G(3) ('j 'l )jcj j2 , (A22)
hajbi a(r, )b(r, ) dr j6l
O 0
with G(2) and G(3) given by equations (57) and (58). Note
where O is a fundamental domain of the periodically
from equation (A21) that the inner product hvl ja21 i is only
tiled plane (whose area is normalized to unity). The
non-vanishing when N 3 P (corresponding to hexagonal
linear operator L is self-adjoint with respect to this inner
planforms) since we require N j1 kj 0. One possible set
product, i.e. hajLbi hLajbi. Therefore, de¢ning
of wave vectors is kj qc (cos('j ), sin('j )) with
vl (r, ) eikl r u( 'l ), '1 0, '2 2/3, '3 2/3. Also note that if u() is
an odd eigenfunction then it immediately follows that
we have hvl jLan i hLvl jan i 0 for n 2, 3, . . . . Since G(2) 0.
Lan bn , we obtain a hierarchy of solvability conditions Finally, we substitute equations (A19), (A20), (A21) and
(A22)p into (A18) and perform the rescaling
hvl jbn i 0.
ecl ! (1 W1 G^ (1) )cl . After an additional rescaling of
time we obtain the amplitude equations (47) for N 2
From equation (A15) the lowest order solvability condition and (50) for N 3.
is c 2 hvl jw a21 i 0. It turns out that in the presence of
lateral interactions the inner product hvl jw a21 i can be
non-vanishing (in the case of even patterns), which leads ENDNOTE
to a contradiction when 2 6 0. This can be remedied by 1
Interestingly, there does exist an example from £uid
assuming that 2 e02 O (e2 ) and considering the dynamics where the modi¢ed Euclidean group action
modi¢ed solvability condition hvl je 1 b2 b3 i 0. This (equation (9)) arises (Bosch Vivancos et al. 1995).
generates the equation

da1 REFERENCES
vl 1 w a1 c 3 hvl jw a31 i c 02 hvl jw a21 i.
dt Blasdel, G. G. 1992 Orientation selectivity, preference, and
(A18) continuity in monkey striate cortex. J. Neurosci. 12,
3139^3161.
An alternative approach to handling the non-vanishing of Bosch Vivancos, I., Chossat, P. & Melbourne, I. 1995 New
the inner product hvl jw a21 i would be to expand the planforms in systems of partial di¡erential equations with
bifurcation parameter as c e1 e2 2 . . . . This Euclidean symmetry. Arch. Ration. Mech. Analysis 131, 199^224.

Bosking, W. H., Zhang, Y., Scho¢eld, B. & Fitzpatrick, D. Hubel, D. H. & Wiesel, T. N. 1974b Uniformity of monkey striate
1997 Orientation selectivity and the arrangement of hori- cortex: a parallel relationship between ¢eld size, scatter, and
zontal connections in tree shrew striate cortex. J. Neurosci. magni¢cation factor. J. Comp. Neurol. 158, 295^306.
17, 2112^2127. Ishai, A. & Sagi, D. 1995 Common mechanisms of visual
Bresslo¡, P. C., Bresslo¡, N. W. & Cowan, J. D. 2000a imagery and perception. Science 268, 1772^1774.
Dynamical mechanism for sharp orientation tuning in an Klu«ver, H. 1966 Mescal and mechanisms and hallucinations.
integrate-and-¢re model of a cortical hypercolumn. Neural University of Chicago Press.
Comput. 12, 2473^2512. Knauer, A. & Maloney,W. J.1913 A preliminary note on the psychic
Bresslo¡, P. C., Cowan, J. D., Golubitsky, M. & Thomas, P. J. action of mescalin, with special reference to the mechanism of
2000b Scalar and pseudoscalar bifurcations: pattern forma- visual hallucinations. J. Nerv. Ment. Dis. 40, 425^436.
tion in the visual cortex. (Submitted.) Kosslyn, S. M. 1994 Image and brain. Cambridge, MA: MIT
Busse, F. H. 1962 Das stabilita«tsverhalten der zellular konvektion bei Press.
endlicher amplitude. PhD dissertation, University of Munich, Krill, A. E., Alpert, H. J. & Ost¢eld, A. M. 1963 E¡ects of a
Germany. Rand Report LT-66-19. [English translation by hallucinogenic agent in totally blind subjects. Arch. Ophthalmol.
S. H. Davis.] 69, 180^185.
Buzano, E. & Golubitsky, M. 1983 Bifurcation involving the Lee, T. S., Mumford, D., Romero, R. & Lamme, V. A. F. 1998
hexagonal lattice and the planar Bënard problem. Phil. Trans. The role of primary visual cortex in higher level vision. Vision
R. Soc. Lond. A 308, 617^667. Res. 38, 2429^2454.
Clottes, J. & Lewis-Williams, D. 1998 The shamans of prehistory: Levitt, J. B. & Lund, J. 1997 Contrast dependence of contextual
trance and magic in the painted caves. New York: Abrams. e¡ects in primate visual cortex. Nature 387, 73^76.
Cowan, J. D. 1977 Some remarks on channel bandwidths for Li, Z. 1999 Pre-attentive segmentation in the primary visual
visual contrast detection. Neurosci. Res. Program Bull. 15, 492^517. cortex. Spatial Vision 13, 25^50.
Cowan, J. D. 1982 Spontaneous symmetry breaking in large Livingstone, M. S. & Hubel, D. H. 1984 Speci¢city of intrinsic
scale nervous activity. Int. J. Quantum Chem. 22, 1059^1082. connections in primate primary visual cortex. J. Neurosci. 4,
Cowan, J. D. 1997 Neurodynamics and brain mechanisms. In 2830^2835.
Cognition, computation, and consciousness (ed. M. Ito, Y. Miyashita McGuire, B. A., Gilbert, C. D., Rivlin, P. K. & Wiesel, T. N.
& E. T. Rolls), pp. 205^233. Oxford University Press. 1991 Targets of horizontal connections in macaque primary
Drasdo, N. 1977 The neural representation of visual space. visual cortex. J. Comp. Neurol. 305, 370^392.
Nature 266, 554^556. Milton, J., Mundel, T., an der Heiden, U., Spire, J.-P. &
Dybowski, M. 1939 Conditions for the appearance of Cowan, J. D. 1995 Traveling activity waves. In The handbook of
hypnagogic visions. Kwart. Psychol. 11, 68^94. brain theory and neural networks (ed. M. A. Arbib), pp. 994^997.
Ermentrout, G. B. 1991 Spots or stripes. Proc. R. Soc. Lond. A 434, Cambridge, MA: MIT Press.
413^428. Mitchison, G. & Crick, F. 1982 Long axons within the striate
Ermentrout, G. B. 1998 Neural networks as spatial pattern cortex: their distribution, orientation, and patterns of connec-
forming systems. Rep. Prog. Phys. 61, 353^430. tion. Proc. Natl Acad. Sci. USA 79, 3661^3665.
Ermentrout, G. B. & Cowan, J. D. 1979 A mathematical theory Miyashita, Y. 1995 How the brain creates imagery: projection to
of visual hallucination patterns. Biol. Cyber. 34, 137^150. primary visual cortex. Science 268, 1719^1720.
Ermentrout, G. B. & Cowan, J. D. 1980 Secondary bifurcation Morrone, M. C., Burr, D. C. & Vaina, L. M. 1995 Two stages of
in neuronal nets. SIAM J. Appl. Math. 39, 323^340. visual processing for radial and circular motion. Nature 376,
Eysel, U. 1999 Turning a corner in vision research. Nature 399, 507^509.
641^644. Mourgue, R. 1932 Neurobiologie de l'hallucination. Brussels:
Gilbert, C. D. 1992 Horizontal integration and cortical Lamertin.
dynamics. Neuron 9, 1^13. Mundel, T., Dimitrov, A. & Cowan, J. D. 1997 Visual cortex
Gilbert, C. D. & Wiesel, T. N. 1983 Clustered intrinsic connec- circuitry and orientation tuning. In Advances in neural informa-
tions in cat visual cortex. J. Neurosci. 3, 1116^1133. tion processing systems, vol. 9 (ed. M. C. Mozer, M. I. Jordan &
Golubitsky, M., Swift, J. W. & Knobloch, E. 1984 Symmetries T. Petsche), pp. 887^893. Cambridge, MA: MIT Press.
and pattern selection in Rayleigh ^ Bënard convection. Physica Oster, G. 1970 Phosphenes. Sci. Am. 222, 83^87.
D 10, 249^276. Patterson, A. 1992 Rock art symbols of the greater southwest. Boulder,
Golubitsky, M., Steward, I. & Schae¡er, D. G. 1988 Singularities CO: Johnson Books.
and groups in bifurcation theory II. Berlin: Springer. Purkinje, J. E. 1918 Opera omnia, vol. 1, pp. 1^62. Prague: Society
Hansel, D. & Sompolinsky, H. 1997 Modeling feature selectivity for Czech Physicians.
in local cortical circuits. In Methods of neuronal modeling, 2nd Rockland, K. S. & Lund, J. 1983 Intrinsic laminar lattice
edn (ed. C. Koch & I. Segev), pp. 499^567. Cambridge, MA: connections in primate visual cortex. J. Comp. Neurol. 216,
MIT Press. 303^318.
Helmholtz, H. 1924 Physiological optics, vol. 2. Rochester, NY: Schwartz, E. 1977 Spatial mapping in the primate sensory
Optical Society of America. projection: analytic structure and relevance to projection.
Hirsch, J. D. & Gilbert, C. D. 1991 Synaptic physiology of Biol. Cyber. 25, 181^194.
horizontal connections in the cat's visual cortex. J. Neurosci. Senseman, D. M. 1999 Spatiotemporal structure of depolariza-
11, 1800^1809. tion spread in cortical pyramidal cell populations evoked by
Horton, J. C. 1996 Pattern of ocular dominance columns in di¡use retinal light £ashes. Vis. Neurosci. 16, 65^79.
human striate cortex in strabismic amblyopia. Vis. Neurosci. 13, Sereno, M. I., Dale, A. M., Reppas, J. B., Kwong, K. K.,
787^795. Belliveau, J. W., Brady, T. J., Rosen, B. R. & Tootell,
Hubel, D. H. & Wiesel, T. N. 1967 Cortical and callosal connec- R. B. H. 1995 Borders of multiple visual areas in humans
tions concerned with the vertical meridian in the cat. J. revealed by functional magnetic resonance imaging. Science
Neurophysiol. 30, 1561^1573. 268, 889^893.
Hubel, D. H. & Wiesel, T. N. 1974a Sequence regularity and Siegel, R. K. 1977 Hallucinations. Sci. Am. 237, 132^140.
geometry of orientation columns in the monkey striate cortex. Siegel, R. K. & Jarvik, M. E. 1975 Drug-induced hallucinations
J. Comp. Neurol. 158, 267^294. in animals and man. In Hallucinations: behavior, experience and

theory (ed. R. K. Siegel & L. J. West), pp. 81^161. New York: Wiener, M. C. 1994 Hallucinations, symmetry, and the structure
Wiley. of primary visual cortex: a bifurcation theory approach. PhD
Smythies, J. R. 1960 The stroboscopic patterns. III. Further thesis, University of Chicago, IL, USA.
experiments and discussion. Br. J. Psychol. 51, 247^255. Wilson, H. R. & Cowan, J. D. 1972 Excitatory and inhibitory
Turing, A. M. 1952 The chemical basis of morphogenesis. Phil. interactions in localized populations of model neurons.
Trans. R. Soc. Lond. B 237, 32^72. Biophys. J. 12, 1^24.
Tyler, C. W. 1978 Some new entoptic phenomena. Vision Res. 181, Wilson, H. R. & Cowan, J. D. 1973 A mathematical theory of
1633^1639. the functional dynamics of cortical and thalamic nervous
Tyler, C. W. 1982 Do grating stimuli interact with the hyper- tissue. Kybernetik 13, 55^80.
column spacing in the cortex? Suppl. Invest. Ophthalmol. 222, Winters, W. D. 1975 The continuum of CNS excitatory states
254. and hallucinosis. In Hallucinations: behavior, experience and
Von der Malsburg, C. & Cowan, J. D. 1982 Outline of a theory theory (ed. R. K. Siegel & L. J. West), pp. 53^70. New York:
for the ontogenesis of iso-orientation domains in visual Wiley.
cortex. Biol. Cyber. 45, 49^56. Zweck, J. W. & Williams, L. R. 2001 Euclidean group invariant
Walgraef, D. 1997 Spatio-temporal pattern formation. New York: computation of stochastic completion ¢elds using shiftable ^
Springer-Verlag. twistable functions. Neural Comput. (In the press.)

TB010299

Uploaded by

Copyright:

Available Formats

TB010299

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

TB010299

Uploaded by

Copyright:

Available Formats

doi 10.1098/rstb.2000.

Geometric visual hallucinations, Euclidean

Paul C. Bresslo¡1, Jack D. Cowan2*, Martin Golubitsky3,

Phil. Trans. R. Soc. Lond. B (2001)

Figure 2. (a) Funnel and (b) spiral hallucinations generated

images. In our view such images are generated in V1 and

(b) The human retinocortical map

Phil. Trans. R. Soc. Lond. B (2001)

Phil. Trans. R. Soc. Lond. B (2001)

Figure 6. Action of the retinocortical map on the funnel form

Phil. Trans. R. Soc. Lond. B (2001)

Figure 8. (a) Orientation tuned cells in V1. Note the

Phil. Trans. R. Soc. Lond. B (2001)

(a) would be that it is a continuum version of a lattice of

for constants and . Without loss of generality we may

Phil. Trans. R. Soc. Lond. B (2001)

Figure 12. Example of an angular spread in the anisotropic

space, respectively, whereas 0 determines the (spatial)

hypercolumn (b) Euclidean symmetry

where  is the re£ection (x1 , x2 )7 !(x1 , x2 ) and R is a

Phil. Trans. R. Soc. Lond. B (2001)

Phil. Trans. R. Soc. Lond. B (2001)

analyse the full model described by equation (11), it is

Phil. Trans. R. Soc. Lond. B (2001)

Vk fu( ')eikr  c:c: : u(  )u() and u 2 Cg

Phil. Trans. R. Soc. Lond. B (2001)

An  z1 n, 1  A(1)

(i) Case 0 (c) Marginal stability

Phil. Trans. R. Soc. Lond. B (2001)

Phil. Trans. R. Soc. Lond. B (2001)

Substituting these expansions into the matrix eigenvalue

Phil. Trans. R. Soc. Lond. B (2001)

lattice ``1 ``2 k1 k2 lattice a(r, )

(a) Restriction to doubly periodic solutions

Phil. Trans. R. Soc. Lond. B (2001)

Table 3. Even planforms with u( )  u()

lattice name planform eigenfunction

square even square u() cos x  u( /2) cos y

Table 4. Odd planforms with u( )  u()

lattice name planform eigenfunction

square odd square u() cos x u( /2) cos y

where (x, y) are Cartesian coordinates in V1. Such a line

Phil. Trans. R. Soc. Lond. B (2001)

Figure 21. Contours of even axial eigenfunctions on the

Phil. Trans. R. Soc. Lond. B (2001)

Phil. Trans. R. Soc. Lond. B (2001)

the far edge of the periphery, so we have no constraint on

If the acute angle of the lattice 0 is speci¢ed, then the

Phil. Trans. R. Soc. Lond. B (2001)

Figure 28. Action of the single

Figure 29. Action of the single inverse retinocortical map on

Phil. Trans. R. Soc. Lond. B (2001)

Figure 32. Action of the double inverse retinocortical map on

Phil. Trans. R. Soc. Lond. B (2001)

Phil. Trans. R. Soc. Lond. B (2001)

for all 04'5 with (v) Even non-contoured planforms

Phil. Trans. R. Soc. Lond. B (2001)

Figure 37. Plot of the even eigenfunction coe¤cient u 3 (q) of

Phil. Trans. R. Soc. Lond. B (2001)

Figure 39. Plot of the odd eigenfunction coe¤cient u3 (q) of

for constants and . Without loss of generality we may

space, respectively, whereas 0 determines the (spatial)

where is the re£ection (x1 , x2 )7 !(x1 , x2 ) and R is a

Vk fu( ')eikr c:c: : u( )u() and u 2 Cg

An z1 n, 1 A(1)

(i) Case 0 (c) Marginal stability

lattice ``1 ``2 k1 k2 lattice a(r, )

Table 3. Even planforms with u( ) u()

square even square u() cos x u( /2) cos y

Table 4. Odd planforms with u( ) u()

square odd square u() cos x u( /2) cos y

If the acute angle of the lattice 0 is speci¢ed, then the

for all 04'5 with (v) Even non-contoured planforms

Figure 37. Plot of the even eigenfunction coe¤cient u 3 (q) of

when planforms are spontaneously generated in such ^ ^ 2 (q) z1