Computing
Science
NovemberDecember,
2001
Third
Base
Brian
Hayes
People
count by tens and machines count by twos—that pretty much sums up
the way we do arithmetic on this planet. But there are countless other
ways to count. Here I want to offer three cheers for base 3, the ternary
system. The numerals in this sequence—beginning 0, 1, 2, 10, 11, 12,
20, 21, 22, 100, 101—are not as widely known or widely used as their
decimal and binary cousins, but they have charms all their own. They
are the Goldilocks choice among numbering systems: When base 2 is
too small and base 10 is too big, base 3 is just right.
Cheaper
by the Threesome
Under the skin, numbering systems are all alike. Numerals in various
bases may well look different, but the numbers they represent are
the same. In decimal notation, the numeral 19 is shorthand for this
expression:
1 x 10^{1 }+ 9 x 10^{0}.
Likewise the binary numeral 10011 is understood to mean:
1 x 2^{4 }+ 0 x 2^{3 }+ 0 x 2^{2 }+ 1 x 2^{1}
+ 1 x 2^{0},
which adds up to the same value. So does the ternary version, 201:
2 x 3^{2} + 0 x 3^{1 }+ 1 x 3^{0}.
The general formula for a numeral in any positional notation goes
something like this:
…
d_{3}r^{3} + d_{2}r^{2
}+ d_{1}r^{1 }+ d_{0}r^{0}.…
Here r is the base, or radix, and the coefficients d_{i}
are the digits of the number. Usually, r is a positive integer
and the digits are integers in the range from 0 to r–1, but
neither of these restrictions is strictly necessary. (You can build
perfectly good numbers on a negative or an irrational base, and below
we’ll meet numbers with negative digits.)
To say that all bases represent the same numbers, however, is not
to say that all numeric representations are equally good for all purposes.
Base 10 is famously well suited to those of us who count on our fingers.
Base 2 dominates computing technology because binary devices are simple
and reliable, with just two stable states—on or off, full or empty.
Computer circuitry also exploits a coincidence between binary arithmetic
and binary logic: The same signal can represent either a numeric value
(1 or 0) or a logical value (true or false).
The cultural preference for base 10 and the engineering advantages
of base 2 have nothing to do with any intrinsic properties of the
decimal and binary numbering systems. Base 3, on the other hand, does
have a genuine mathematical distinction in its favor. By one plausible
measure, it is the most efficient of all integer bases; it offers
the most economical way of representing numbers.
How do you measure the cost of a numeric representation? If you simply
count digits, then the biggest base will always win; for example,
base 1,000,000 can represent any number between 0 and decimal 999,999
in a single digit. The trouble is, that single digit can be any of
a million different symbols, all of which you must somehow recognize.
At the opposite pole are unary, or base1, numbers. The unary representation
of decimal 1,000,000 needs only one type of symbol, but that symbol
is repeated a million times. (Unary notation is in a category apart
from other bases—it’s not really a positional number system—but in
the present context it serves as a useful limiting case.)
Among all possible ways of writing the numbers up to a million, neither
base 1,000,000 nor base 1 seems ideal; as a matter of fact, you could
hardly do worse than either of these choices. Minimizing the number
of digits causes an explosion in the alphabet of symbols, and vice
versa; when you squish down one factor, the other squirts out. Evidently
we need to optimize some joint measure of a number’s width (how many
digits it has) and its depth (how many different symbols can occupy
each digit position). An obvious strategy is to minimize the product
of these two quantities. In other words, if r is the radix
and w is the width in digits, we want to minimize rw
while holding r^{w} constant.
Curiously, this problem is easier to solve if r and w
are treated as continuous rather than integer variables—that is, if
we allow a fractional base and a fractional number of digits. Then
it turns out (see Figure 1) that the optimum radix is e,
the base of the natural logarithms, with a numerical value of about
2.718. Because 3 is the integer closest to e, it is almost
always the most economical integer radix (see Figure 2).
Consider again the task of representing all numbers from 0 through
decimal 999,999. In base 10 this obviously requires a width of six
digits, so that rw=60. Binary does better: 20 binary digits
suffice to cover the same range of numbers, for rw=40. But
ternary is better still: The ternary representation has a width
of 13 digits, so that rw=39. (If base e were a practical
choice, the width would be 14 digits, yielding rw=38.056.)
Figure
1
Trit
by Trit by Trit
This special property of base 3 attracted the notice of early computer
designers. On the hypothesis that a computer’s component count would
be roughly proportional both to the width and to the depth of the
numbers being processed, they suggested that rw might be a
good predictor of hardware cost, and so ternary notation would make
the most efficient use of hardware resources. The earliest published
discussion of this idea I’ve been able to find appears in the 1950
book Highspeed Computing Devices, a survey of computer technologies
compiled on behalf of the U.S. Navy by the staff of Engineering Research
Associates.
At about the same time as the ERA survey, Herbert R. J. Grosch proposed
a ternary architecture for the Whirlwind computer project at MIT.
Whirlwind evolved into the control system for a military radar network,
which stood vigil over North American airspace through 30 years of
the Cold War. Whirlwind was also the proving ground for several novel
computer technologies—including magnetic core memory—but ternary arithmetic
was not among the innovations tested; Whirlwind and its successors
were binary machines.
As it happens, the first working ternary computer was built on the
other side of the Iron Curtain. The machine was designed by Nikolai
P. Brusentsov and his colleagues at Moscow State University and was
named Setun, for a river that flows near the university campus. Some
50 machines were built between 1958 and 1965. Setun operated on numbers
composed of 18 ternary digits, or trits, giving the machine
a numerical range of 387,420,489. A binary computer would need 29
bits to reach this capacity; in terms of rw, the ternary design
wins 54 to 58.
Unfortunately, Setun did not realize the potential of base 3 to reduce
component counts. Each trit was stored in a pair of magnetic cores,
wired in tandem so that they had three stable states. A pair of cores
could have held two binary bits, which amounts to more information
than a single trit, and so the ternary advantage was squandered.
Along with ternary arithmetic, a computer built of base3 hardware
can also exploit ternary logic. Consider the task of comparing two
numbers. In a machine based on binary logic, comparison is often a
twostage process. First you ask, “Is x less than y?”;
depending on the answer, you may then have to ask a second question,
such as “Is x equal to y?” Ternary logic simplifies
the process: A single comparison can yield any of three possible outcomes:
“less,” “equal” and “greater.”
Ternary computers were a fad that faded, though not quickly. In the
1960s there were several more projects to build ternary logic gates
and memory cells, and to assemble these units into larger components
such as adders. In 1973 Gideon Frieder and his colleagues at the State
University of New York at Buffalo designed a complete base3 machine
they called ternac, and created a software emulator of it. Since then
the idea of ternary computing has had occasional revivals, but you’re
not going to find a ternary minitower in stock at CompUSA.
Why did base 3 fail to catch on? One easy guess is that reliable threestate
devices just didn’t exist or were too hard to develop. And once binary
technology became established, the tremendous investment in methods
for fabricating binary chips would have overwhelmed any small theoretical
advantage of other bases. Furthermore, it’s only a hypothesis that
such an advantage exists. Everything hinges on the assumption that
rw is a proper measure of hardware complexity, or in other
words that the incremental cost of increasing the radix
is the same as the incremental cost of increasing the number of digits.
Figure
2
But
even if ternary circuits don’t find a home in computer hardware, the
Goldilocks argument favoring base 3 may apply in other contexts. Suppose
you are creating one of those dreadful telephone menu systems—Press
1 to be inconvenienced, Press 2 to be condescended to, and so forth.
If there are many choices, what is the best way to organize them?
Should you build a deep hierarchy with lots of little menus that each
offer just a few options? Or is it better to flatten the structure
into a few long menus? In this situation a reasonable goal is to minimize
the number of options that the wretched caller must listen to before
finally reaching his or her destination. The problem is analogous
to that of representing an integer in positional notation: The number
of items per menu corresponds to the radix r, and the number
of menus is analogous to the width w. The average number of
choices to be endured is minimized when there are three items per
menu.
Turning
to Ternary Dust
Although numbers are the same in all bases, some properties of numbers
show through most clearly in certain representations. For example,
you can see at a glance whether a binary number is even or odd: Just
look at the last digit. Ternary also distinguishes between even and
odd, but the signal is subtler: A ternary numeral represents an even
number if the numeral has an even number of 1s. (The reason is easy
to see when you count powers of 3, which are invariably odd.)
More than 20 years ago, Paul Erdös and Ronald L. Graham published
a conjecture about the ternary representation of powers of 2. They
observed that 2^{2} and 2^{8} can be written in ternary
without any 2s (the ternary numerals are 11 and 100111 respectively).
But every other positive power of 2 seems to have at least one 2 in
its ternary expansion; in other words, no other power
of 2 is a simple sum of powers of 3. Ilan Vardi of the Institut des
hautes études scientifiques has searched up to 2^{6973568802}
without finding a counterexample, but the conjecture remains open.
Figure
3
The digits of ternary numerals can also help illuminate a peculiar
mathematical object called the Cantor set, or Cantor’s dust. To construct
this set, draw a line segment and erase the middle third; then turn
to each of the resulting shorter segments and remove the middle third
of those also, and continue in the same way. After infinitely many
middle thirds have been erased, does anything remain? One way to answer
this question is to label the points of the original line as ternary
numbers between 0 and 0.222.… (The repeating ternary fraction 0.222…
is exactly equal to 1.0.) Given this labeling, the first middle third
to be erased consists of those points with coordinates between 0.1
and 0.122…, or in other words all coordinates with a 1 in the first
position after the radix point. Likewise the second round of erasures
eliminates all points with a 1 in the second position after the radix
point. The pattern continues, and the limiting set consists of points
that have no 1s anywhere in their ternary representation. In the end,
almost all the points have been wiped out, and yet an infinity of
points remain. No two points are connected by a continuous line, but
every point has neighbors arbitrarily close at hand. It’s hard to
form a mental image of such an infinitely perforated object, but the
ternary description is straightforward.
The
Jewel in the Triple Crown
“Perhaps
the prettiest number system of all,” writes Donald E. Knuth in The
Art of Computer Programming, “is the balanced ternary notation.”
As in ordinary ternary numbers, the digits of a balanced ternary numeral
are coefficients of powers of 3, but instead of coming from the set
{0, 1, 2}, the digits are –1, 0 and 1. They are “balanced” because
they are arranged symmetrically about zero. For notational convenience
the negative digits are usually written with a vinculum, or overbar,
instead of a prefixed minus sign, but here the vinculum is shown as
an overstrike, thus: 1.
As an example, the decimal number 19 is written 1101
in balanced ternary, and this numeral is interpreted as follows:
1 x 3^{3 }– 1 x 3^{2} + 0 x 3^{1} + 1 x 3^{0},
or in other words 27–9+0+1. Every number, both positive and negative,
can be represented in this scheme, and each number has only one such
representation. The balanced ternary counting sequence begins: 0,
1, 11, 10, 11, 111,
110, 111. Going in the opposite
direction, the first few negative numbers are 1,
11, 10, 11,
111, 110, 111.
Note that negative values are easy to recognize because the leading
trit is always negative.
The
idea of balanced number systems has quite a tangled history. Both
the Setun machine and the Frieder emulator were based on balanced
ternary, and so was Grosch’s proposal for the Whirlwind project. In
1950, Claude E. Shannon published an account of symmetrical signeddigit
systems, including ternary and other bases. But none of these 20thcentury
inventors was the first. In 1840, Augustin Cauchy discussed signeddigit
numbers in various bases, and Léon Lalanne immediately followed up
with a discourse on the special virtues of balanced ternary. Twenty
years earlier, John Leslie’s remarkable Philosophy of Arithmetic
had set forth methods of calculating in any base with either signed
or unsigned digits. Leslie in turn was anticipated a century earlier
by John Colson’s brief essay on “negativoaffirmative arithmetick.”
Earlier still, Johannes Kepler used a balancedternary scheme modeled
on Roman numerals. There is even a suggestion that signeddigit arithmetic
was already implicit in the Hindu Vedas, which would make the idea
very old indeed!
What makes balanced ternary so pretty? It is a notation in which everything
seems easy. Positive and negative numbers are united in one system,
without the bother of separate sign bits. Arithmetic is nearly as
simple as it is with binary numbers; in particular, the multiplication
table is trivial. Addition and subtraction are essentially the same
operation: Just negate one number and then add. Negation itself is
also effortless: Change every 1 into a 1, and vice
versa. Rounding is mere truncation: Setting the leastsignificant
trits to 0 automatically rounds to the closest power of 3.
The bestknown application of balanced ternary notation is in mathematical
puzzles that have to do with weighing. Given a twopan balance, you
are asked to weigh a coin known to have some integral weight between
1 gram and 40 grams. How many measuring weights do you need? A hasty
answer would be six weights of 1, 2, 4, 8, 16 and 32 grams. If the
coin must go in one pan and all the measuring weights in the other,
you can’t do better than such a powersof2 solution. If the weights
can go in either pan, however, there’s a ternary trick that works
with just four weights: 1, 3, 9 and 27 grams. For instance, a coin
of 35 grams—1101 in signed ternary—will balance on
the scale when weights of 27 grams and 9 grams are placed in the pan
opposite the coin and a weight of 1 gram lies in the same pan as the
coin. Every coin up to 40 grams can be weighed in this way. (So can
all helium balloons weighing no less than –40 grams.)
James Allwright, who maintains a Web site promoting balanced ternary
notation, suggests a monetary system based on the same principle.
If both a merchant and a customer have just one bill or coin in each
powerof3 denomination, they can make exact change for any transaction.
Martha
Stewart’s File Cabinet
Some weeks ago, rooting around in files of old clippings and correspondence,
I made a discovery of astonishing obviousness and triviality. What
I found had nothing to do with the content of the files; it was about
their arrangement in the drawer.
Imagine a fastidious office worker—a Martha Stewart of filing—who
insists that no file folder lurk in the shadow of another. The protruding
tabs on the folders must be arranged so that adjacent folders always
have tabs in different positions. Achieving this staggered
arrangement is easy if you’re setting up a new file, but it gets messy
when folders are added or deleted at random.
Figure
4
A drawer filled with “halfcut” folders, which have just two tab positions,
might initially alternate leftrightleftright. The pattern
is spoiled, however, as soon as you insert a folder in the middle
of the drawer. No matter which type of folder you choose and no matter
where you put it (except at the very ends of the sequence), every
such insertion generates a conflict. Removing a folder has the same
effect. Translated into a binary numeral with left=0 and right=1,
the pristine file is the alternating sequence …0101010101.… An insertion
or deletion creates either a 00 or a 11—a flaw much like a dislocation
in a crystal. Although in principle the flaw could be repaired—either
by introducing a second flaw of the opposite polarity or by flipping
all the bits between the site of the flaw and the end of the sequence—even
the most maniacally tidy recordkeeper is unlikely to adopt such practices
in a real file drawer.
In my own files I use thirdcut rather than halfcut folders; the
tabs appear in three positions, left, middle and right.
Nevertheless, I had long thought—or rather I had assumed without bothering
to think—that a similar analysis would apply, and that I couldn’t
be sure of avoiding conflicts between adjacent folders unless I was
willing to shift files to new folders after every insertion. Then
came my Epiphany of the File Cabinet a few weeks ago: Suddenly I understood
that going from halfcut to thirdcut folders makes all the difference.
It’s easy to see why; just interpret the drawerful of thirdcut folders
as a sequence of ternary digits. At any position in any such sequence,
you can always insert a new digit that differs from both of its neighbors.
Base 3 is the smallest base that has this property. Moreover, if you
build up a ternary sequence by consistently inserting digits that
avoid conflicts, then the choice of which symbol to insert is always
a forced one; you never have to make an arbitrary selection among
two or more legal possibilities. Thus, as a file drawer fills up,
it is not only possible to maintain perfect Martha Stewart order;
it’s actually quite easy.
Deletions, regrettably, are more troublesome than insertions. There
is no way to remove arbitrary elements from either a binary or a ternary
sequence with a guarantee that two identical digits won’t be brought
together. (On the other hand, if you’re fussy enough to fret about
the positions of tabs on file folders, you probably never throw anything
away anyhow.)
The protocol for avoiding conflicts between thirdcut file folders
is so obvious that I assume it must be known to file clerks everywhere.
But in half a dozen textbooks on filing—admittedly a small sample
of a surprisingly extensive literature—I found no clear statement
of the principle.
Strangely enough, my trifling observation about arranging folders
in file drawers leads to some mathematics of wider interest. Suppose
you seek an arrangement of folders in which you not only avoid putting
any two identical tabs next to each other, but you also avoid repeating
any longer patterns. This would rule out not only 00 and 11 but also
0101 and 021021. Sequences that have no adjacent repeated patterns
of any length are said to be “square free,” by analogy to numbers
that have no duplicated prime factors.
In binary notation, the onedigit sequences 0 and 1 are obviously
square free, and so are 01 and 10 (but not 00 or 11); then among sequences
three bits long there are 010 and 101, but none of the other six possibilities
is square free. If you now try to create a fourdigit squarefree
binary sequence, you’ll find that you’re stuck. No such sequences
exist.
What about squarefree ternary sequences? Try to grow one digit by
digit, and you’re likely to find your path blocked at some point.
For example, you might stumble onto the sequence 0102010, which is
square free but cannot be extended without creating a square. Many
other ternary sequences also lead to such dead ends. Nevertheless,
the Norwegian mathematician Axel Thue proved almost a century ago
that unbounded squarefree ternary sequences exist, and he gave a
method for constructing one. The heart of the algorithm is a set of
digit replacement rules: 0>12, 1>102, 2>0. At each
stage in the construction of the sequence, the appropriate rule is
applied to each digit, and the result becomes the starting point for
the next stage. Figure 4 shows a few iterations of this process. Thue
showed that if you start with a squarefree sequence and keep applying
the rules, the sequence will grow without bound and will never contain
a square.
More recently, attention has turned to the question of how many ternary
sequences are square free. Doron Zeilberger of Rutgers University,
in a paper coauthored with his computer Shalosh B. Ekhad, established
that among the 3^{n} ndigit ternary sequences
at least 2^{n}^{/17} are square free. Uwe Grimm
of the Universiteit van Amsterdam has tightened this lower bound somewhat;
he has also found an upper bound and has counted all the ndigit
sequences up to n=110. It turns out there are 50,499,301,907,904
ways of arranging 110 ternary digits that avoid all repeated patterns.
I’ll have to choose one of them when I set up my squarefree file
drawer.
