0% found this document useful (0 votes)
241 views11 pagesEnergy Levels and Wave Functions
Research journal about Energy levels and Wave Functions of Bloch electrons
Uploaded by
Tanu HonggonegoroCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
0% found this document useful (0 votes)
241 views11 pagesEnergy Levels and Wave Functions
Research journal about Energy levels and Wave Functions of Bloch electrons
Uploaded by
Tanu HonggonegoroCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
PHYSICAL REVIEW B
VOLUME 14, NUMBER 6 IS SEPTEMBER 1976
Energy levels and wave functions of Bloch electrons in rational and irrational magnetic fields*
Douglas R. Hofstadter?
Physics Depariment Univers of Oregon. Eugene, Oregon 97403
‘Received 9 February 1976)
An effective singletand Hamiltonian representing a crystal electron in uniform magnetic fed is constructed
fiom the tight-binding form of a Bloch tand by replacing 4K by the operator B— eA/e. The resultant
Schrédinger equation becomes 4 ritedifference equation whose eigenvalues can be computed by & matrix
‘method. The magnetic Mux which passes through 2 lattice cel, divided by a Mux quantum, yields a
‘dimensionless parameter whose rationality or irrationality highly influences the nature of the computed
spectrum. The graph of the spectrum over a wide range of “ration” feds i plated. A recursive structre is
scovered in the graph, which enables « number of theorems to be proven, bearing particulary on the
‘question of contin. The recurve structure is not unlike that predicted by Azbe using a continued
fraction for the dimensionless parameter. An iterative algnthm for deriving the clustering pattern of the
‘magnetic subbands is given, which folloms from the recursive structure. From ths algorithm, the nature ofthe
Spectrum at an “rational” field can be deduced; itis seen to be an uncountable but measurezero set of
points (4 Cantor st). Despite thee-feature, it shown thatthe graph is continuous asthe magnetic Fld
‘aries. It is aso shown how a spectrum with simplified properties can be derived from the rigorously derived
spectrum, by introducing a spread inthe field values, This spectrum satisfies all the intutively desirable
properties of a spectrum. The spectrum here presented is shown to agree with that predicted by A. Rauh ina
‘completely diferent model for crystal electrons in a magnetic fel. A new type of magnetic “superattice” i
introduced, constructed so that its unit cell intercepts precisely one quantum of ux. Ii shown that this ell
presents the periodicity of solution ofthe difrence equation. Iti also shown how this supelaice allows
the determination of the wave function at nonlatie sites Evidence is offered that the wave fonctions
belonging to irational fields are everywhere defined and ate continuous in this model, wheres those
Ielonging to ational elds are only defined on a dicrete set of points. A method for investigating these
predictions experimentally is sketched.
1 INTRODUCTION
‘The problem of Bloch electrons in magnetic
fields is a very peculiar problem, because it is
one of the very few places in physics where the
difference between rational numbers and irrational
‘mumbers makes itself felt.? Common sense tells
us that there can be no physical effect stemming
from the irrationality of some parameter, because
an arbitrarily small change in that parameter
‘would make it rational — and this would create
some physical effect with the property of being
everywhere discontinuous, which is unreasonable.
‘The only alternative, then, is to show that a theory
which apparently distinguishes between rational
and irrational values of some parameter does so
only in a mathematical sense, and yields physical
observables which are nevertheless continuous. It
is the purpose of this paper to present a method
Which effects such a reconciliation of “rational
and “irrational” magnetic fields. ‘The method is
illustrated in a maximally simple model of the
physical situation, but the ideas which arise are,
it is to be hoped, applicable to more realistic
‘models of the physical situation.
I, DERIVATION OF THE DIFFERENCE EQUATION
Briefly, then, the model involves a two-dimen
‘sional square lattice of spacing a, Immersed in a
‘uniform magnetic field Hf perpendicular to it. We
restrict our considerations to what happens to a
single Bloch band when the field is applied. This
1s one strong simplifying feature of the model; the
next is that we postulate the following tight-bind~
ing form for the Bloch energy function:
W(R)=2,(cosh,a cosh)
Perhaps the most difficult step to justify on phys-
teal grounds isthe following one, which {shall
fefer to as the "Peters subetitulon™: we replace
‘ik in the above function by the operator § -eA/e
(being the vector potential), to create an opera
for cut or WH), which we tun teat ap an effective
Snglo-band laeltontan, Work to jstly this eu
Stitation has been dove.
When this substitution is made, the etfective
omiltonian ls oeen to cntan translation opera
tors explap,/) and exolap,/., Depending onthe
tugs chonen, there are, th alton, certain
Blase factors dependent on te magnetic fed
229 2200 DOUGLAS R. HOFSTADTER
strength, which multiply the translation opera-
tors. If the Landau gauge — A= 11(0,x,0) — is,
chosen, then only the translations along y are
multiplied by phases. From now on, we assume
this gauge. Now when the effective Hamiltonian is
introduced into a time-independent Schrodinger
‘equation with a two-dimensional wave function, the
following eigenvalue equation results:
EgL¥le+ays)+ 8c =a, y)+ eH Y(x, 9 +0)
peters, y a)
Note how the wave function at (x, ) 4s Linked to
its four nearest neighbors in the lattice. Tt is con-
venient to make the substitutions
ey)
xema, yena, B/E=€.
It is furthermore reasonable to assume plane
wave behavior inthe y direction, since the coeffi-
cients inthe above equation only involve x. There
fore, we write
oma, na)= ego)
Finally we introduce the parameter about which
all the fuss is made.
a=0'H/2xthe/e)
Notice that « is dimensionless, being the ratio
of flux through a lattice cell to one flux quantum.
(he author is indebted to Professor F. Bloch for
pointing out that this parameter can be interpreted
4s the ratio of two characteristi periods of this
problem: one isthe period of the motion of an
slectron ina state with crystal momentum 2zh/a,
‘which is a'n/2af the other is the reciprocal of
the eyclotron frequency etf/me.) A valNe of «
implies an enormous magnetic field (on the order
of a billion gauss), Ifthe lattice spacing is typical
of real crystals (onthe order of 2A). Despite
this, wo are going to be interested inthe results
for such values of a (fora treatment of smaller
values of a in this same equation, sce Ref. 8.)
‘With all these substitutions, our Schrdinger
‘equation turns into a one-dimensional difference
equa
(m+ 1)+ g(m~1) + 2cos(2mma— v)g(m)= ex(m).
a
‘This equation is sometimes called “Harper's”
tion sand has been studied by a number of author
a
IIL, CALCULATION OF THE SPECTRUM AND THE
RATIONALITY CONDITION
Another way of writing Bq. (1) is
(cor) . (‘ aie ae)
‘The 2x2 matrix is called “A(m).” When a product
of m successive A matrices is multiplied with the
two vector ((1),(0)), the result is the two vector
(g(+1),a0m))..‘The physical condition which
must be imposed on the wave function (Le., the
function g) is boundedness, for all m. ‘This trans~
lates into a condition on the products of successive
‘A matrices. Now if the A matrices are periodic
in m (which they may very well be, since mn enters
only under 2 cosine), then long products of A ma-
trices consist essentially in repetitions of one
block of A matrices, whose length is the period
in m. Let us assume that the A matrices are in-
deed periodic in m, with period q. This is a re-
quirement on a, namely that there should exist an
integer p such that
2ra(m-sq)-¥=25am =v 25p
Algebra reveals the fact that this condition on a
is precisely that of rationality":
a=pla
‘We now proceed, making full use of this somewhat
bizarre ansatz. (Presently, we will consider the
ease when a is irrational.) The product of q suc~
cessive A matrices will be called “@.” The con-
dition of physicality is now transferred from the
gto the matrix Q. It ean be shown without trou
ble that the correct condition on @ is that its two
‘eigenvalues be of unit magntiude. ‘That condition
cean then be shown to be equivalent to requiring its
trace to be less than or equal to 2, in absolute
value. Hence, a concise test for the boundedness
of the 2's is the following:
[tr@(6;]=2.
‘Trace conditions of this type have been found by
other authors.*"# ‘This one was discovered by
Professor G. Obermair, and extensively used by
the author. Now it can be shown that the only way
that v affects the value of Tr@ is additively, 1.
that as v changes, the shape of the graph of Tr@,
plotted against «, is unchanged —it merely moves
as a whole, up and down. (A proof of essentially
this fact can be found in Ref. 2.) Therefore
‘TrO(6jv)=TrQ(e)+2/(v), where flv) is a periodic
function of unit amplitude, and @(e) is defined as
Q(G 1/24). We are interested in all values «
which, for some v, yield bounded g's. (Such val-
ues will be called “eigenvalues” of the difference
equation.) Therefore, we want to form the union
of all eigenvalues ¢, as v varies. Since 2/(v)
ranges between +2 and -2, the condition on the
trace can be rewritten as follow:
Ineo|<4.
The trace of @ is always a polynomial of degree “ ENERGY LEVELS AND WAVE FUNCTIONS OF BLOCH 224
4; hence one might expect the above condition to
be satisfied in roughly q distinct regions of the
€ axis (one region centered on each root). This
is indeed the case, and is the basis for a very
striking (and at first disturbing) fact about this
problem: when a=p/q, the Bloch band always
breaks up into ¢zecisely q distinct energy bands.
Since small variations in the magnitude of a ean
produce enormous fluctuations in the value of the
denominator q, one is apparently faced with an
unacceptable physical prediction, However, nature
| ingenious enough to find a way out of this ap-
parent anomaly. Before we go into the resolution,
however, let us mention certain facts about the
spectrum belonging to any value of a. Most can
be proven trivially: (i) Spectrum(a) and spectrum
(a+N) are identical. (ii) Spectrum(a) and spec-
trum(-a) are identical. (il) ¢ belongs to spec-
trum(a) if and only if ~¢ belongs to spectrum(a)..
(jv) If € belongs to spectrum (a) for any a, then
== €=44, ‘The last property is a little subtler
than the previous three; it can be proven in dif-
ferent ways. One proof has been published."*
From properties (i) and (iv), it follows that a
graph of the spectrum need only include values of
€ between +4 and ~4, and values of a in any unit
Interval. We shall look at the interval [0, 1]. Fur-
thermore, as a consequence of properties, the
graph inside the above-defined rectangular region
‘must have two axes of reflection, namely the hor-
‘igontal line a=4, and the vertical line €=0. A
plot of spectrum(a), with a along the vertical axis,
appears in Fig. 1. (Only rational values of a with
denominator less than 50 are shown.)
1V. RECURSIVE STRUCTURE OF THE GRAPH
This graph has some very unusual properties.
‘The large gaps form a very striking pattern some-
‘what resembling a butterfly; perhaps equally strik-
ing are the delicacy and beauty of the fine-grained
structure. ‘These are due to a very intricate
scheme, by which bands cluster into groups, which
themselves may cluster into larger groups, and
so on. The exact rules of formation of these hier
archically organized clustering patterns (IVs) are
what we now Wish to cover. Our description of 1's
will be based on three statements, each of which
describes some aspect of the structure of the
scraph. All of these statements are based on ex-
tremely close examination of the numerical data,
and are to be taken as “empirically proven” theo
rems of mathematics. It would be preferable to
have a rigorous proof but that has so far eluded
capture. Before we present the three statements,
Jet us first adopt some nomenclature. “unit
celt” is any portion of the graph located between
successive integers’ and.N + 1—in fact we will
call that unt cell the Nth unit cell, Every unit cett
has a “local variable” 8, which runs from 0 to 1;
in particular, 8 is defined to be the fractional part
of a, usually denoted as {a}. At 8=0 and 8=1,
there Is one band which streiches across the fl
width ofthe cell, separating it from its upper and
lower neighbors; this band is therefore called a
ll wall.” It turns out that certain rational val-
‘ues of 6 play a very important role in the descrip-
ton of the structure of a unt cell; these are the
“pure cases”
FIG. 1. Spectrum inside
unit cell. € ls the hori
ontal variable, ranging
able, ranging from 0 to 1. 202 DOUGLAS R. HOFSTADTER 4
WN and 1-1/8 (v= 2};
and the “special cases"
N/QN+1) and (+1)/2N+1) (= 2)
(of the special cases, those with numerator N are
the “lower” special cases, and those with numera~
torN+1 are the “upper”). The spectra belonging
to these rational values form a “skeleton” on which
the rest of the graph is hung. Figure 2 shows that
skeleton; in it are shown the bands belonging to
pure cases (up toN=97); in addition, one out of
the 2.V'+1 bands per special case is included, the
centermost (i.e., the N+ sf, counting from either
end). The rest of the graph can be built up from
this skeleton by a recursive process. Roughly,
that process amounts to compressing the skeleton
down to a small fraction of its size, distorting its
vertical and horizontal scale in the process, and
inserting this shrunken skeleton in between neigh-
boring “ribs” of the large skeleton. When appro-
priately shrunken skeletons have been inserted
between each pair of ribs, then the process is
reiterated on the next level down; and this must
continue indefinitely. Our goal is to turn this pic
turesque deseription into a precise description,
and then to extract physical consequences from
this weird structure. For this, we need the three
statements;
Slatement I. At the height inside any cell where
its local variable equals a pure case 1/N or 1
-1/N, there are N bands between the left and
right borders of the cell. (in unit cells, when N
is even, there seem to be only N-1 bands, be~
shown with a “skeleton”: the spectra
as the ceater band belonging to special cases = N/2N
‘F1) and 6=(V+1)/2N+1), The L chain, C chain, and
R chain are shown, formed by
Joining bands in the
‘The labeling scheme for subcells in the three chains is
Indicated,
‘cause the two centermost bands touch in the mid-
dle, where ¢=0.) As N goes to infinity, the ratio
of Band size to gap size goes to zero (in other
words, the bands become negligibly thin, com-
pared to the gaps). Furthermore, the pure-case
bands are distributed in such a way that the entire
length of each cell roof and cell floor is ap-
proached, in the limit that WV goes to infinity.
Moreover, the number of pure-case bands per
unit energy interval is a slowly varying and rough-
ly constant function; that is, there is no cluster-
ing of the bands belonging to a pure case.
‘At heights where the local variable equals a spe-
lal case, there is a set of bands, of which only
the centermost is of interest here. The width of
these center bands approaches zero as.N goes to
infinity. When upper special cases are considered,
these bands approach a limit point, which is the
inner edge of one of the two bands at 4; when lower
special cases are considered, the limit-point is the
inner edge of the other band at
‘The next two statements involve the concept of
“eubeelis,” which are at the core of the recursive
description of the graph’s structure; but the con-
pt of subcells can only be defined after the
;keleton” has been introduced (statement 1). This.
is the reason that the following definition has been
‘sandwiched between statements, It is best under-
stood with the help of Fig. 2
Rules for Subcell Construction. The L and R
subcells of any cell are formed as follows: Con-
nect the edges of the outermost bands of neighbor-
ing pure cases by straight lines. The trapezoidal
boxes thus ereated form the “L chain” and the “R
chain” (on the left- and right-hand sides of the
cell),
‘The C subcells of any cell are formed as follows:
Connect the outer edges of the next-to-outermost
bands of pure cases with N'>2 by straight lines.
‘This will produce two large boxes whose sides are
unions of infinitely many straight-line segments.
‘The remaining C subeells are formed by joining
the centermost bands of neighboring special cases
by straight lines. All the C subcells taken together
form the “C chain.” Each subeell has a unique
label; the labeling scheme is shown in Fig. 2. We
now affirm the existence of large empty swaths
crossing the graph.
Statement IT. "The regions of a cell outside its
subcells are gaps (contain no bands or portions of
bands).
Finally, statement 11f contains the essence of the
recursive nature of this graph.
Statement Ii. Each subcell of any cell ean be
given its own local variable, defined in terms of
the local variable of the parent cell. (See below.)
Each subcell, when indexed by its own local vari- 4
ENERGY LEVELS AND WAVE
able, Is a cell in its own right, in that it satisfies
statements 1, 11, and II.
The subcell’s local variable is defined as follows:
Let f be the “outer” local variable (i.e., that of
the parent cell), and " be the “inner” local vari-
able. Assume first that @<4. Then let N be de-
fined by
ve[1/a]
(Note: The notation [x] stands for the greatest
Integer less than or equal to x; it follows that 8
is the denominator of the pure case just above #.)
If the subeell is of L type or R type, then the
equation relating f and 3" is
B= (Near
{Note how this forces 2 to lie between 0 and 1.)
Let us denote the function of which yields this
value of 8! by “A(3).”
If the subeell is of C type, then the relation be-
tween inner and outer local variables is
Bestar,
(a’} (fractional part of a)
Let us denote this function of f by “T(a).”
Finally, if g is between $ and 1, then 6" is equal
to the value of a’ which belongs to 1-2.
‘The statements are a little startling; they need
evidence. In Figs. 3 and 4 are plotted two “rectan-
gularized” subcells of a unit cell, namely L, and
C,. A “rectangularized” cell is made from the
cell itself by a family of one-dimensional linear
transformations. There is a linear stretching at
each height, which makes the effective width of
the cell be the same at every height (Like a unit
and the bands as stretched in that way are
then plotted using the cell's own local variable,
rather than that of its parent cell, as the vertical
axis. The characteristic butterfly pattern of the
large gaps is very obvious in the rectangularized
FIG. 8, Reotangularlzation of Ly.
FUNCTIONS OF BLOCH a3
FIG. 4. Rectangularization of C,, The number of bends
calculated was much smaller, which explains why 20
litle detail ie visible, All the bands shown belong to the
Dure-case part of tho skeleton ofthis subeell. (Compare
Fig. 2.)
graphs. Note, however, that pure cases with even
denominators inside the L. cell do not possess the
“degeneracy” property (of having two bands which
“kiss” at the center).
‘The recursive structure as here presented con-
firms in the main (but differs in detail with) the
important but extremely difficult article by Az
bel’, which states that the spectrum is entirely
determined by the continued fraction of a. The
‘connection is through the A function. If the local
variable function A is iterated, one obtains the
following representation for
1
est
Xo
i.
‘which is unique, and will terminate for any ration-
ala. Azbel’ predicts that spectrum(a) will con-
sist of N, bands, each of which breaks up into N,
subbands, each of which breaks up into N, sub-
bands, and so on. This is approximately the same
a8 our result, when all of the N's are large. Our
prediction is that the L and R cells will each con-
tain N, bands, but the number inside the C cell is
not given by this expansion. As the nesting contin-
ues, N, subbands are indeed found in the Land &
ssubeells of each of the L and R cells, but in the
C subcells, once again there is no simple predic
tion based on the continued fraction expansion.
Qualitatively, though, Azbel's prediction contains
the essence of the structure, and is very intuitively
appealing.
From this recursive breakdown of the graph there 28 DOUGLAS R. HOFSTADTER 4
follow a number of theorems, most of which in-
volve somewhat tedious topological reasoning (the
proots in complete detail are worked out in the
author's thesis"). First of all i is important to
be able to pinpoint any particular cell, no matter
hhow deeply itis nested inside other cells. A sim-
ple notation will do this for us:_ the outermost cell
iswrittentirst, followedby successively shrinking
cells inside it... For example, “U,LCyRsl,”
stands for a cell-in-a-cell-in-a-cell-in-a-cell-in~
coll, The subscripts are tobe interpreted as
shown in Fig. 2. ('U," stands for the unit cell
where [a]=N. However, the notation for the unit
cell is usually omitted, since all unit celis are
identical.)
‘A result which is quite difficult to establish is
the simple fact that all cells are (nearly) homeo
morphic to each other. (Homeomorphisms are the
topological version of isomorphisms: a homeo-
morphism is a one-to-one continuous mapping be
tween two manifolds whose inverse is also con-
tinuous.) The “nearly” has to be included since
there is a feature which could not be preserved
under a continuous mapping, and that is the “de-
generacy” at rationals with even denominators
which exists in unit cells, but not in L or R cells,
‘This means that there is a “branch cut” across
which the homeomorphism does not carry. To be
precise, each cell can be cut into two pieces — a
Jeft and'a right hall. For unit cells, the dividing
line is merely the vertical line at €=0; for other
cells, the dividing line can be defined In terms of
the center bands of rationals with odd denomina-
tors. ‘The leit and right halves of any cell, as de-
termined by its dividing line, are homeomorphie
to each other and to the halves of every other cell
as well. However, the homeomorphism ean only
bbe extended over the line in ease both cells are of
the same type, in the sense that they share the
property of degeneracy, or share the property of
its absence.
\V. HOW THE BANDS ARE CLUSTERED
‘We now can make a precise definition ofthe
cluster patterns. Suppose we wish to describe the
distribution of bands at the value a=p/a. Let 6
={a}, s0 that p is the local variable for the unit
cell to which a belongs. ‘The recursive decompo-
sition telis us that the spectrum at 8 consists of
three parts, which must be separated by gaps: one
inside an L subcell, one inside a C subeell, and
‘one inside an R subcell. Furthermore, the L and
R subcells contain bands at that height with a It
belonging to A(), and the C subcell contains bands
at the height with a 1 belonging to T(s). tn other
words, the Il at a consists of three I's, {rom right
to left, belonging to A(S), T(), Alp), respective-
ly. Let us take the example of the value a=
Us spectrum is shown below:
A suggestive symbolic representation for the
cluster pattern is
(2-1-2)-(2-8-2)-(2-1-2)
‘The five bands on either side are located inside
the L and R chains; the central seven are located
inside the C chain. The reason the breakdown is
5-1-5 is explained recursively as follows
For the L and R subcelis, the local variable is
siven by
Be Wweey
so that §”=A(a)=2. ‘The demoninator is 5, hence
wo oxpect to see 6 bands inside L, and R,,
For the C subcell, the oral variable is given by
rears @ea/ er,
which yields
arta'h= (eb
‘The analysis then “predicts” that the spectrum at
a1= 2 will consist of a set of five bands belonging
to the local variable 3; then a gap; then a set of
seven bands belonging to the loeal variable #5 then
another gap; then another set of five bands belong
ing to the local variable {. But the analysis can
be carried further, because the very same opera~
tions ean be earried out inside the subeels, start
ing with thetr local variables and deriving local
variables which are even more local. For 3 and
4, this gives
15)= (2) 1(0)11(8)
ng)=m@n@n0).
If is useful to adopt the notation “N” as shorthand
for “M(1/N),” because, according to statement 1,
the bands belonging to 1/N are smoothly spread
out across the cell to which they belong, vith no
clustering. And 10) is denoted "I" because at
=0 there is only one band, With this shortoand,
then, we can write
m1@)=2-1-2,
m3)
‘And these IPs can then be stuffed back into the
original I for $:
M()=(2-1-2)-(2-3-2)-(2-1-2).
‘This coincides with what our eye told us. There is
a guarantee that this recursive analysis of I's will
come to an end, because the two operations which 4 ENERGY LEVELS AND WAVE FUNCTIONS OF BLOCH aus
‘produce new local variables always reduce the
numerator or denominator of the input fraction,
In the end, one must eventually wind up with pure
cases, or zero. The number of levels which one
must descend before this process terminates is,
however, rather difficult to predict.
VI. SPECTRA BELONGING TO IRRATIONAL FIELDS
‘The only case in whieh it is easy to predict what
will happen is if you begin with an irrational value
of a. Inthat case, the two operations yield new
irrational values, which in turn yield irrational
values, etc., ad infinitum, This leads to the very
Interesting question, “What is left — if anything —
fin the spectrum of an irrational field, according
to this process?” Readers who are familiar with
‘the pathology of point sets may already be antici-
pating the answer: there is indeed something left,
and it is homeomorphie to the Cantor set. (The
Cantor set is an uncountable yet measure-zero set
of reals in an interval; see Ref. 18 for a detailed
‘exposition of ite fundamental properties.)
‘To demonstrate this starting from the three
statements, one looks at the sequences of nested
cells which are created by the repeated recursion
{in statement (li). That Js, given the original ir-
rational a, one knows that its spectrum is con-
fined to some particular unit cell. Statement (i
‘says that the confinement can be further specified,
as being inside three particular subcells of that
cell. Reapplication of statement (ili) creates more
deeply nested confining cells; for rational a the
process terminates, but for irrational a the end
product is uncountably many different infinite se-
quences of nested cells, It is a well-known theo-
rrem of topology that any nested sequence of closed
Intervals whose lengths tend to zero contains a
‘unique limit point; its two-dimensional generaliza-
tion to closed sets whose maximum dimension
‘shrinks to zero is immediate, and that theorem is
‘what tells us that the spectrum belonging to any
irrational value of a consists of uncountably many
points, between any pair of which there is a finite
gap. Rigorous topological analysis establishes
that the spectrum is indeed homeomorphie to the
Cantor set.
It is Legitimate to question whether these values
of € are actually eigenvalues of the difference
equation, i.e., whether, in fact, the wave function
e(m) does remain bounded as m goes to infinity.
‘Numerical work suggests that the answer is yes:
such values really are the eigenvalues. It would
be highly interesting to see a rigorous proof of
this fact, or a refutation. Until proven wrong,
however, we shall adopt this construction via re-
ccursion as the definition of the spectrum belonging
‘to an irrational field value,
VIL, MAGNETIC FIELD FLUCTUATIONS CREATE A
ILURRED GRAPH
When this is done, we have finally achieved an
Important result: we have found a spectrum for
every single value of a. Now the crucial question
is, “How physical is this spectrum?” After all,
it stil remains true that the spectrum at a rational
/q consists of q bands, and q is still 2 highly
‘yctuating function of p/4.. One can still feet sus
picious of the graph. Despite the intellectual mis
givings, though, the eye sees something rather
Continuous. ‘There is something to this visual in-
sight, and it can be stated formally in the follow
ing continuity theorem, which has been proven in
the author’s thesis: For any a, as a’ approaches
a, then al points of spectrum(a) are approached
by points belonging to spectrum(a’); furthermore,
‘only the points of spectrum(a) are so approached.
This theorem confirms the eye's assessment,
‘that vertical motion along the graph is “‘contin
“uous,” in some sense; yet there Is something dis
continuous about vertical mation as well. Define
AM(q) to be the Lebesgue measure of spectrum(a).
For all rational a, M(a) is positive, since every
rational has bands of positive length. But for all
irrational a, M(a) is zero. Therefore M has very
peculiar behavior: at rational values, M is dis-
continuous, since there are irrationals arbitrarily
near any rational; yet at irrational values, M is
continuous. [The proof of this latter statement can
bbe found in the author's thesis; it depends on a
careful examination of how spectrum(a’) is deter
rined by sequences of nested cells, when a” is
taken to be arbitrarily close to irrational values
of a.] The function Mf is continuous at all irxa~
tionals, discontinuous at all rationals. This is @
direct consequence of our recursive picture, and
once again makes one wonder whether the graph
is physically meaningful, oF not
Fortunately, there is a very simple resolution
to this problem, consisting in the observation that
every physical parameter has an experimental un-
certainly init, which smears it over some inter
val. Thus, the magnetic field, no matter how
carefully controlled, has some fluctuations, which
ray be terribly small. This suggests the following
concept: form a union of the spectra of all @ with-
in a"window” of height Aa. ‘This ean be thought
of as a blurred version of the graph, created by
rapid up-down jiggling, where the amplitude of
the jiggling is given by $a. A blurred graph is
shown in Fig. 5, using Aa= 2. As you can see,
the result of the smearing-process yields a graph
with a radically simplified appearance. It ean be
proven that the number of bands in any smeared
‘graph is bounded by tho constant 1/4a1+ 1, for all
«a, and that the band edges change smoothly with 206 DOUGLAS R. HOFSTADTER
FIG. 5, One quadrant of the smeared graph created by
using 30,
‘a. This establishes a totally continuous behavior
for all magnetic field values, as a consequence of
the imprecision of the field value. This also gets,
rid, of course, of the measure anomaly. As Aa
approaches zero, the fine structure of the graph
4s bit by bit recovered; the infinitely fine-grained
detail never returns (for positive 4a), but more
and more of it is revealed by decreasing the un-
certainty a. Of course, at the unphysical value
Aa=0, the entire graph returns.
VII, CORRESPONDENCE WITH RAUH'S LANDAU
LEVEL APPROACH
‘One unexpected feature of the recursive nature
of the graph is how it corroborates a picture set
forth by Rauh concerning the broadening of Landau
levels when a periodic potential is “turned on.”
In Rauh’s work, the simplest possible two-dimen
sionally periodic potential, V (x, »)=2¥,(coske
+cosky), is chosen as a perturbing potential act
ing on an electron in an initially pure Landau state.
‘The same difference equation arises, with a totally
different interpretation: z(m) represents the am-
plitude of a Landau state of fixed principal quantum
number, whose center of localization along the axis
of square integrability is ma/a (with a as we have
defined it), and ¢ is proportional to the energy
splitting. More interesting is the interpretation of
a. Instead of measuring the flux in flux quanta, it
measures the reciprocal of that quantity. There-
fore, a large a in Rauh’s equation means a small
eld. One can use the equations linking inner and
outer variables to establish a link between Ravh's
conclusions and our graph. This is done as fol
lows. Observe the way the L chain turns into a
very thin line as it approaches the Bloch band it
is so thin that it resembles a single level, split by
a perturbation, Therefore, we choose to identify
the leftmost band with a Landau level, perturbed
by the periodic potential of the exystal. The split-
ting of the band is present in our picture, since in
reality the line ls composed of very thin L cells.
[At height 2 (assuming <4), the structure inside
each of those cells is given by spectrum (x), where
B= (N42)
Here, Vis integral, and « is between 0 and 1, Now
by the first symmetry property of the graph,
pectrum(N +x)
spectrum(1/A).
Notice that this says that the split-up of the lowest-
lying Landau level is given by the same eigenvalue
equation, but with parameter 1/9 instead of 6.
‘This is completely consistent with Rauh’s work.
Moreover, one can identify other chains in the
‘graph with Landau levels, and under this identifi-
‘eation, it turns out that each one of them splits up
Ina pattern given by spectrum (1/g). ‘The natural
candidate for the 2nd-lowest Landau level is the
L chain located inside C,; the 3rd lowest Is the
L chain inside C,Cy, and so on. ‘The number of
such levels is essentially 1/3; half of them are
chains inside nested C cells, and the other half
are their symmetric counterparts: R chains in-
side nested C cells. To determine how any one
of them is split, we must iterate the formation of
local variables. In particular, to derive the split-
ting of the nth Landau level, we must begin with 8,
apply the F function n -1 times to it, and finish
by taking A of the result, As before, let
Nest,
‘with integral, and x between 0 and 1. Further,
assume N is at least 4, Then by definition, (a)
=x. Simple calculation shows also that
T(p)={(N -2)+s}*.
From this expression, we can directly read off
AE): itis also x. Now if N-2 is also at least
4, then we can immediately get
TIT@)=[(N- +21",
fand A of this is, once again, x. So it will go, with
2 being subtracted from the integer in the denom!~
nator over and over again, as long as that integer
stays 4 or more. When 2 is small (i.e., when N
is big), then the number of I's which can be iter-
ated before the integer ceases to satisfy that con-
dition is roughly 4N. This implies that there are
roughly £N Landau levels to the loft of center, and
symmetrically, 4 to the right of center, all NV
of which are split according to the pattern of spec~
trum(x) —but as before, spectrum(x) and spec-
‘trum(1/A) are identical.” Therefore all the Landau
levels do split in a similar way. And we have
shown that their number is roughly N, which is to
spectrum (x) 1“ ENERGY LEVELS AND WAVE FUNCTIONS OF BLOCH a7
say 1/B, Therefore the separation between Lan-
‘dau levels is roughly 88E,. This corresponds to
the spacing which one can calculate using an ef-
fective-mass approximation at the edges of the
band. Altogether, the Landau-level-based theory
and the Bloch-band-based theory achieve in this.
way a satisfying harmony.
1X, WAVE FUNCTIONS AND IRRATIONAL FIELDS.
In certain other approaches to this problem,
notably those based on the magnetic translation
group," the rationality of a is forced if one seeks
representations of the magnetic translation group
by the Frobenius method, which involves finding
an invariant subgroup. For some subgroup of mag-
netic translations to be invariant, all of its mem-
bers must commute, and this in turn forces cer-
tain phase factors, invotving the flux through the
parallelogram defined by the two translations, to
be unity. ‘The end result is that one must choose
a rational value #/q for a, and the invariant sub-
‘group consists of “superlattice” translations,
‘where the superlattice consists of lattice points
separated from each other by q lattice spacings in
both x and y directions. That way, the amount of
flux is always an integer, the phase factors are
always unity, and the subgroup of magnetic trans
lations is indeed invariant. The problem with this
‘whole approach is that such superlattices can only
bbe defined in the case of rational fields, and there
seems to be no obvious way to extend the results.
to irrational fields.
‘An alternative type of “superlattice” ean be
formulated, however, which comes up naturally
in the context of our difference equation. One be-
‘gins with the observation that there are solutions
of the Bloch-Floquet type to the difference equa-
tion —that is, solutions with the property that
alm enP)=e!? (mn),
where» is any integer, P is a constant, and 1s
wave number. One's first guess might well be
that P must be an integer, corresponding to mov-
ing through an integral number of lattice spacings.
‘This assumption is erroneous, however; P need
not be integral. Indeed, the correct minimal
period P is 1/a, which may be any real whatso-
fever, rational or irrational. This is proven in
exactly the same way as for the Mathieu equation,
of which our difference equation is, in some
senses, a discrete counterpart.
‘The crucial fact in the proof is that the coeffi
cients in the difference equation are themselves
periodic in the variable m, with period P=1/a.
‘Therefore when m is replaced by m+P in the dif-
ference equation, gm) becomes g(m+P) but the
coefficients are unaltered, which says that if (m)
4s a solution, then so is (m+). Now there are
two linearly independent solutions to a difference
‘equation of second order (which ours is}; let them
bbe gon) and ga(m)- Then g(m+P) and ga0n-+P)
are also solutions; but since \(m) and g,(n) form
abasis, there must be numbers C,, such that
(eee ): es &) (2).
gains P)) \cay Ca) \eaton
‘Now we can find a linear transformation which will
diagonalize the 22 matrix; this transformation
will mix g, and g, to produce new functions 4 and
«gg with the property that
sdlons P)=caiton) (=1,2),
where ¢ is an elgenvalve of the C matrix. ‘This
proves the Bloch-Floquet theorem for the differ-
‘ence equation, with period 1/a. A corollary is
‘that any solution (m) can be expressed as
glm=e™*G(m),
where
and where G(m) is a periodic function of period P.
When c is 1 (which happens, as in the Mathieu
‘equation, at one edge of each band), then g(m)
=Gim), 80 thatthe difference equation has a pure—
ly periodic solution of period P. In any case, the
distance 1/a plays the role of a fundamental per-
fod associated with the difference equation.
The difference equation per se allows us only to
determine G(m) when m is an integer. But the
perlodieity of G(r) allows us to interpolate be
foveen integers, and to determine G there also.
‘This comes about because the period i/o is
(in general) not an integer. Suppose, for instance,
that a= Then the period is of length . Now
G(0), GA), G(2), and G(S) all fal within one per-
sod, ‘but G(a) is beyond GC), and hence equals
G(G). similarly, 6(8)=6(H), G(6)=6(4), and so
‘on. Finally, G(H7)=G(0) and the whole eycle starts
over again, Therefore, we ean plot the values G(0)
‘through 6(17) inside one period of length P; they
Will appear in some rearranged order. The (wo
orders and thelr relation are shown below. Integer
order:
0123456789100 2314151617
Ree
IstP 2nd P 3rdP ath P sth P
In the figure below, the five complete periods
shown above are superimposed, to give the rear—
ranged order (note that the scale of the two figures
1s different): 228 DOUGLAS R. HOFSTADTER “4
071441118155 1229166133 100
‘The sequence of integers in the rearranged order
Js the successive multiples of 7, taken modulo 17.
‘This is because 7 occurs exactly ;. beyond the
period boundary in the upper figure, and gy is the
‘minimum distance possible, The general Tule for
the rearranged order when a= p/1 is to take the
multiples of j (modulo q), where p is defined by
the congruence
pb. (modulo 4)
So far we have concentrated on what happens when
a is rational; but the same process of folding back
all values of G(m) into one period of length P ean
be carried out. In the irrational case, however,
the reordering will create a dense distribution of
points inside the whole period. This is one place
‘where irrational fields seem to make more physi-
cal sense than rational fields, in that one can de-
termine the values of their wave functions on a
dense set, rather than at just a discrete set.
Hf one takes a sequence of rational values a,
‘which approach an irrational value (and whose de-
nominators therefore must go to infinity), the var-
ous periods 1/, are all approximately the same,
and it is therefore possible to compare the reor-
dered wave functions of these rationals, to see if,
some trend emerges, pointing the way to the re-
ordered wave function at the irrational field. One
must also be sure to choose eigenvalues which are
very close to each other; that this can be done is
consequence of the continuity theorem stated
above. Such a process of comparison was carried
out numerically for the following sequence of frac-
tions (shown with their continued-fraction expan-
sions) and their largest eigenvalues:
9864) ,
3.028 50),
023 983 268).
‘The wave functions in rearranged order are shown
Jn Fig, 6, It appears that an overall shape is es~
tablished by the fraction with a low denominator
(in this case #), and details of the shape are de
termined by fractions with higher and higher de-
nominators. Note how these fractions, which are
close in value but which have very different denom-
Inators, have magnetic periods P of very nearly
the same length, which allows the direct compari-
son of their wave functions. This figure is strong.
. 8
FIG. 6, Values of the wave function inside one mag
notie period P=1/a, shown for three values of «(and
their largest eigenvalues): triangles: a= 4 (€=2.2664);
circles: a= 3 (€=3.02850); dota: aH (€=3.025 989 268)
‘The x axis represents, in each ease, a physteal distance
of P(=1/a) lattice spacings. ‘The vertical seale s such
‘hat the highest dot represents the value 1.
evidence for the idea that the Limiting ease —
namely, the wave-function for an irrational a — is
‘continuous function which can be obtained from
the discrete points supplied by the difference
‘equation by translating them all into a single mag-
netic period of length 1/a.
In this connection, it is also interesting to point
‘oat that the one-dimensional “superlattice” of per-
iod 1/a ean be related to the magnetic translation
‘group, in our model. It can be verified easily that
all Landau gauge magnetic translation operators!
TyylF)= et dee 1008,
where
B=H(~y,0,0)
commutes with he effective Hamiltonian defined
earlier. (This is not the case with the true Ham-
Atonian for a crystal electron in a magnetic field.)
However, magnetic translation operators do not
in general commute with other magnetic transla
tion operators. The condition of commutation is
that the parallelogram which they define should
intercept an integral number of flux quanta. Now
since the Landau gauge leads naturally to a one-
‘dimensional mathematical treatment in which all
the interesting phenomena happen along the x axis,
it would seem natural to look for a commuting set
‘of magnetic translation operators whose y spacing
is “trivial” (i.e., is based on the lattice spacing),
and whose x spacing contains information about the
field. If we allow any magnetic translation in the
y direction as long as it is through an integral
‘umber of lattice spacings, then the commutation
‘condition quantizes the allowed magnetic transla-
tions along x; and the condition is precisely that
they must be through 1/e lattice spacings. The
reason for this is that a rectangle of dimensions
a/a1 along x, and a along y intercepts precisely
fone flux quantum,
‘This observation suggests that the best choice
of unit cell for a “magnetic superlattice” may not
be a square of q lattice [Link] a side (whieh
only can be done for rational fields), but rather,
4 rectangle with one side equal to the lattice spac-
ing, and the other side such that exactly one flux 4 ENERGY LEVELS AND WAVE FUNCTIONS OF BLOC, 2009
quantum is intercepted. And that is the super-
lattice defined by the period P for the difference
equation,
‘Somewhat related to these ideas is the article by
‘Chambers,"* in which orbits, “hyperorbits” (and
50 on) are discussed. In particular there is a dis-
eussionof the correspondence between simple or-
bits and bands, “hyperorbits” and subbands, and
X. POSSIBLE EXPERIMENTAL TEST
Finally, I would like to comment on the possibil-
ity of looking for the features predicted by this
‘model experimentally. At first glance, the idea
seems totally out of the range of possibility, since
a value of a=1 in a crystal with the rather gener-
ous lattice spacing of a=2 A demands a magnetic
field of roughly 10° G. It has been suggested, how-
ever (by Lowndes among others), that one could
manufacture a synthetic two-dimensional lattice
of considerably greater spacing than that which
characterizes real crystals. ‘The technique in-
volves applying an electric field across a field
effect transistor (without leads). ‘The effect of
such a field is to drive electrons (or holes) to one
side of the device, where they will crowd together
ina thin layer, essentially creating a two-dimen-
sional gas of charged particles. Now if the device
is prepared in advance with a dielectric layer
which is nonuniform, and which in fact is periodic
in each of its two dimensions, then the two-
