On Ternary Coding and Three-Valued Logic

Subhash Kak

Abstract.
Mathematically, ternary coding is more efficient than binary coding. It is little used
in computation because technology for binary processing is already established and
the implementation of ternary coding is more complicated, but remains relevant in
algorithms that use decision trees and in communications. In this paper we present
a new comparison of binary and ternary coding and their relative efficiencies are
computed both for number representation and decision trees. The implications of
our inability to use optimal representation through mathematics or logic are
examined. Apart from considerations of representation efficiency, ternary coding
appears preferable to binary coding in classification of many real-world problems of
artificial intelligence (AI) and medicine. We examine the problem of identifying
appropriate three classes for domain-specific applications.

Keywords: optimal coding, decision trees, ternary logic, artificial intelligence

Introduction
The problem of optimal coding of numbers has been examined by many scholars
(e.g. [1][2]) and it is known that ternary coding is more efficient than binary coding
[3]. The reason why it is not used extensively in hardware is because technology for
binary processing is already established and the implementation of ternary coding is
more complicated in computation, although it is used in digital communications
where error correction coding is employed [4]. Ternary coding remains relevant in
algorithms that use decision trees.

Let’s consider coding of numbers to an arbitrary base b. Given that the probability
of the use of the b symbols can be taken to be the same and equal to 1/𝑏𝑏, the
information associated with each symbol is log 𝑏𝑏. The efficiency of the coding
scheme per symbol is

ln 𝑏𝑏
𝐸𝐸(𝑏𝑏) = 𝑏𝑏
(1)

To find the value of b for which it is a maximum we take its derivative with respect
to b and equate that to zero. This yields the condition that ln 𝑏𝑏 = 1, from which we
conclude that the optimal base is e, with E(e)= 0.368 nats or 0.531 bits.

1
Although e cannot be used for a rational representation of numbers, we can use it
as a measure to mark the performance of other values of b as in Table 1.

Table 1. Efficiency of coding for certain bases

b 2 e 3 4 5 8 10 12 20 100
E(b) nats 0.347 0.368 0.366 0.347 0.322 0.260 0.230 0.207 0.150 0.046
E(b) bits 0.500 0.531 0.528 0.500 0.465 0.375 0.331 0.298 0.216 0.066

The efficiency is quite close to the maximum for b=3 (a value of 0.366 as compared
to the optimal value for e which is 0.368), with the next best value coming at the
bases 2 and 4. The efficiency at b=3 is off from the optimal by 0.5 percent, whereas
it is off by 5.7% at b=2. After this the values decline monotonically as shown in
Figure 1 in which the most interesting region is for the base values that lie between
2 and 4.

Figure 1. Efficiency of different number bases

If we consider only the capacity of a code to carry information, then binary comes
up considerably short compared to the optimal value. For binary, the value is
𝑒𝑒
ln 2 = 0.693 which is short by ln 2 = 0.307 from the optimal. (In bits, the
corresponding numbers are 1.0 and 0.443, respectively.) For ternary, the coding is
𝑒𝑒
vastly more efficient since ln 3 = 1.099 and ln 3 = −0.099. As we see, the carrying
capacity differences are much more significant compared to those based on
efficiency measure given in equation (1). For example, for b = 2, the carrying
capacity is less by 30.7%.

2
In this paper we present a new comparison of binary and ternary coding and their
relative efficiencies are computed both for number representation and decision
trees. The implications of our inability to use optimal representation through
mathematics or logic are investigated. From a practical point of view, we note the
importance of the use of ternary logic in medicine and artificial intelligence (AI) and
the expectation that it will lead to better performance in certain applications.

Philosophical issues
Since coding has aspects that go beyond representation of data, the issue of
optimal coding leads to many philosophical questions. Since binary is not optimal,
there ought to be a logic that has more than two classes but it is not clear how such
a logic manifests itself in natural processes or in cognition.

Absolute optimality appears to be unachievable as it corresponds to a non-integer,

irrational value. We may conclude that mathematics is not an optimal or complete
way to represent reality and this shortcoming may be added to the known
limitations of mathematics [5].

Although interactions with Nature are normally seen as responses to binary

questions in terms of “yes” or “no” [6][7], this is incorrect if a third category of
“maybe” is included.

This third “indeterminate” category may be seen as the origin of non-classical

aspects of reality. A relevant parallel from quantum mechanics is that the
observation cannot reveal what the state of the system was prior to this
observation [8][9]. This indeterminacy creates unique challenges if one wishes to
build a quantum system for computation [10][11].

Sometimes physicists search for mathematical arguments for the ontological basis
of reality and why one has symmetries of different kind (as in [12]). Speaking in a
similar vein, one may say that the universe is 3-dimensional for that gives the
optimal coding in integer values. Likewise, one may add that the brain performs a
2.5 perception [13] for it makes us come close to the optimal value on the “screen”
of our mind.

3
Comparison for bases 2 and 3
The two bases of most interest are 2 and 3. Consider the count of all numbers over
the range (0, 𝑁𝑁 = 2𝑛𝑛 − 1), and let the number of bits we come across when the
numbers are written in the most compact form (for example, 001 is written as 1,
and so on) be represented by S. We will use the product 𝑅𝑅 = 𝑆𝑆 × 𝑏𝑏 = 𝑆𝑆𝑆𝑆 to be the
range associated with the representation so that this range can be meaningfully
compared across different bases.

In general, the value of S to any base is as follows:

𝑗𝑗=𝑛𝑛
𝑆𝑆𝑏𝑏 (𝑁𝑁 = 𝑏𝑏 𝑛𝑛 − 1) = ∑𝑗𝑗=1 𝑗𝑗(𝑏𝑏 𝑗𝑗 − 𝑏𝑏 𝑗𝑗−1 ) (2)

This can be simplified as given below:

1
𝑆𝑆𝑏𝑏 (𝑁𝑁 = 𝑏𝑏 𝑛𝑛 − 1) = 𝑏𝑏−1 [𝑛𝑛(𝑏𝑏 𝑛𝑛+1 − 𝑏𝑏 𝑛𝑛 ) − 𝑏𝑏 𝑛𝑛 + 1] (3)

For the special case of b=2, this simplifies to:

𝑆𝑆2 (𝑁𝑁 = 2𝑛𝑛 − 1) = (𝑛𝑛 − 1)2𝑛𝑛 + 1 (4)

Thus when n=3, that is N=7, we get the numbers 1,10,11,100,101,110,111 for a
total of 17 locations that equals 17 bits. The formula (2) gives us 2 × 23 + 1 = 17,
and the corresponding S22 is 34.

For N=8, for which n=2, the numbers are 1, 2, 10, 11, 12, 20, 21, 22 for a total of 14
symbols which is the same S3(8)= 1/2[2(33 − 32 ) − 32 + 1]. The S3b count of the
range is 14 × 3 = 42.

For a comparison for N=8 for binary (b=2) coding we get to add 8 more (for 8=1000)
to the previous count of 34 to get the same range count of 42 as for base 3. If we
throw in b=4 for another point of comparison, the numbers are 1, 2, 3, 10, 11, 12,
13 for S4(7)=11 and the corresponding range of 44.

Table 2 provides the total range count for the bases 2, 3, and 4 for different N.

4
 Table 2. Range counts
N 1 2 3 4 7 8 9 15 16 26 27 31 63 64 80 100
Sb, b=2 2 6 10 16 34 42 50 98 108 208 218 258 642 656 880 1160
Sb, b=3 3 6 12 18 36 42 51 105 114 204 216 264 648 660 852 1152
Sb, b=4 4 8 12 20 44 52 60 108 120 220 232 280 684 700 956 1276

We observe that the best R=Nb values keep on switching between the bases 2 and
3, depending on how close one is to the transitions at the powers of these bases. At
2i-1, the base 2 counts are a bit lower (the differences are 2, 2, 7, 6, 6), but at 3i-1,
the counts of base 3 are lower by a much more substantial margin (0, 4, 8); and at
N=100, it remains 8.

A comparison of bases 2 and 3 on absolute count (Figure 2) shows that their

difference is relatively small.

Figure 2. Comparison of bases 2 and 3

Ternary trees
Trees are ubiquitous in data structures [14]. Every node (excluding a root) in a tree
is connected by a directed edge from exactly one other node, which is the parent
node. Alhough binary trees are used most frequently, there are situations where
ternary trees offer advantages.

Excellent trees for data values associated with different probabilities are given by
Huffman coding or arithmetic coding [15]. Huffman coding is optimal if all symbol
probabilities are integral powers of the inverse of the coding alphabet. This is hardly
ever going to be true of a real-world situation and in those cases Huffman coding
can take up to one extra bit per symbol.

5
By dispensing with the restriction that each symbol must translate into an integral
number of bits, arithmetic coding achieves greater efficiency. But since its
implementation is more complex, we shall limit our discussion to Huffman coding.

The coding of different symbols (or messages) of unknown probabilities by efficient

search trees is equivalent to representation by sequences as shown in Figure 3 for 9
symbols.

We will first assume that the probabilities of the symbols are the same, then a
modification of a straight enumeration of the symbols should provide the most
efficient coding.

Figure 3. Binary and ternary trees

In the case of the example of Figure 3, the average length of a symbol in the binary
case is 29/9=3.22 bits, whereas ternary coding shown on the right gives us 2 trits.
Since 1 trit equals 𝑙𝑙𝑙𝑙𝑙𝑙2 3 = 1.585 bits, ternary coding is more efficient for this
example.

In the enumeration mapping, which must satisfy the prefix condition, we run
through all the elements of one size and then replace the largest number by the
next sized number. The counting will be as shown here for both binary and ternary
(Table 3).
Table 3. Coding for different size symbols
Binary Ternary
3: 0, 10, 11 3: 0, 1, 2
4: 00, 01, 10, 11 4: 0, 1, 20, 21
5: 00, 01, 10, 110, 111 5: 0, 1, 20, 21, 22
6: 00, 01, 100, 101, 110, 111 6: 0, 10, 11, 20, 21, 22
7: 00, 010, 011, 100, 101, 110, 111 7: 0, 10, 11, 12, 20, 21, 22
8:000, 001, 010, 011, 100, 101, 110, 111 8: 00, 01, 10, 11, 12, 20, 21, 22
9: 000, 001, 010, 011, 100, 101, 110, 1110, 1111 9: 00, 01, 02, 10, 11, 12, 20, 21, 22

6
A comparison between binary and ternary search trees for unknown symbol
frequencies is given in Table 4.

Table 4. Comparison of binary and ternary coding for different

number of symbols
Symbols 3 4 5 6 7 8 9 10
Binary (bits) 1.66 2 2.4 2.7 2.86 3 3.2 3.4
Ternary (trits) 1 1.5 1.6 1.83 1.86 2 2 2.2
Ternary (bits) 1.59 2.38 2.54 2.91 2.93 3.17 3.17 3.48

We observe that ternary is superior to binary where the number of symbols is in

powers of 3; and binary is superior otherwise.

It is quite clear from the underlying combinatorics that as it is a larger set, ternary
coding is more efficient than binary only for specific values as shown in Table 4.
When considering non-uniform probabilities, the smaller binary code alphabet
increases the advantage even further.

Now assume that the non-uniform probabilities are according to the first digit
1
phenomenon [16], in which the probabilities are 𝑃𝑃(𝑑𝑑) = 𝑙𝑙𝑙𝑙𝑙𝑙𝑟𝑟 (1 + 𝑑𝑑), where r is
the base and d= 1,… r-1. We present this as an excurses to bring to the reader this
property of natural random numbers, which might be relevant in some classification
problems, although as we show below it makes no difference for efficiency of
coding here.

Because of the logarithmic structure of the probabilities, there are various relations
amongst them. In all of these, we have (if warranted by the presence of the
relevant numbers) Prob [1] = Prob [2+3], Prob [2] = Prob [4+5], and Prob [3] = Prob
[6+7]. For base 10, we also have:

Prob [1] = Prob [5+6+7+8+9]

Prob [2] = Prob [5+8+9]
Prob [4] = Prob [8+9]

These values for the first digit frequencies for different bases are given in Table 5.

7
 Table 5. First digit frequencies in different bases (3 to 10)

We see that if we are considering first digits to any base, binary coding is more
efficient, as is to be expected. We mention it here to make a distinction between
the mathematical coding process and the more fundamental question of using a
ternary classification of data obtained from a natural process. If Nature uses the
most efficient ternary coding, it perhaps gets manifested at a level that is deeper
than the use of a ternary alphabet.

Ternary classification
It is possible that ternary classification, when optimized appropriately, will reveal
the deeper ontological basis corresponding with ternary coding but that can only be
a research program at this stage. Meanwhile, let me list a few applications where
ternary classification has been used. In urban analysis, a ternary representation
with the three classes as cyber (C), physical (P), and human (H) gives better results
[17].

Ternary classification has been used with the classes positive, negative, and neutral
in many problems related to response to drugs. In the problem of classifying
chemicals with biomedical properties, one may be considering agonistic,
antagonistic, or no receptor activities [18].

Real world applications are associated with multi-class classification problems with
unknown categories. This is true, for example, in automatic webpage annotation,
and disease diagnosis with unknown classes [19]. The current regime of two testing
classes (those being administered the drug and those in the control sample who are
given the placebo) could be replaced by three classes (related to psychological and
physical profiles together with the control class) and this should help find hitherto

8
undiscovered correlations between mind states and health. One could also classify
subjects in 3 categories, and it is significant that a ternary classification of healthy
patients is used in at least one traditional medicine system [20][21].

The search for best three-class representation of data in medicine is a question that
requires much analysis and research. Unsupervised clustering algorithms, such as
hierarchical clustering, model-based clustering, k-means clustering, and self-
ordering maps, may be used [22][23][24] and they offer improvement of
performance [25]. But they are dependent on the quality and the nature of the data
and they do not automatically obtain the best three-class model or reveal a
fundamental ternary ontological structure.

Ternary logics
Three-valued logics have truth values indicating true, false, and an indeterminate
third value. One can directly see the application of this to databases where some
data entries are missing. Indeed, the database structural query
language SQL implements ternary logic to handle comparisons with NULL field
content, which stands for the missing data. This may be necessary even when an
actual value exists, but it is not recorded in the database [26].

SQL uses AND, OR, and NOT tables where the intermediate value is intended to be
interpreted as UNKNOWN. Explicit comparisons with NULL, including that of
another NULL yields UNKNOWN.

If the truth values are represented numerically as -1, 0, 1 for false, unknown, and
true, respectively, then we can set up tables for the 𝑁𝑁𝑁𝑁𝑁𝑁(𝐴𝐴), 𝐴𝐴 ∩ 𝐵𝐵, and 𝐴𝐴 ∪ 𝐵𝐵
where 𝐴𝐴 ∪ 𝐵𝐵 is 𝑀𝑀𝑀𝑀𝑀𝑀(𝐴𝐴, 𝐵𝐵) and 𝐴𝐴 ∩ 𝐵𝐵 is 𝑀𝑀𝑀𝑀𝑀𝑀(𝐴𝐴, 𝐵𝐵) (Table 6).

Table 6. NOT, OR (max function), and AND (min function) tables for a three-
valued logic

9
These functions can be used for developing different logical operations on the
variables.

A similar truth table can be used to consider the specific states of photon
polarization, A, that are used in quantum computations: horizontal (H = -1),
entangled or diagonal (D = 0) and vertical (V = 1). The second variable, B, is the state
of the measurement apparatus or filter which can be aligned to horizontal (H = -1),
diagonal (D = 0) or vertical (V = +1) orientations. The groups of photons, all in the
same state, is measured correctly if the output intensity is the same as the input
intensity (the two coincide) or if it is zero (the output is the complement of the
input); the output state is 0, if the output intensity is half of the input intensity
[11][27]. This sets up the truth table of Table 8:

B
AUB -1 0 +1
-1 -1 0 +1
A 0 0 1 0
+1 +1 0 +1
Table 8. Truth table for a quantum state measurement problem

Conclusions
In this paper we presented a new comparison of binary and ternary coding, and
their relative efficiencies were computed both for number representation and
decision trees. The implications of our inability to use optimal representation
through mathematics or logic were examined. We proposed that interactions with
Nature, which are seen to be binary at the deepest level, need to be enhanced by a
third category of “maybe”, and this additional category is like the fundamental
indeterminacy of quantum mechanics. This approach seems to be another way of
looking at the classical-quantum divide, which is normally viewed through the lens
of differently operating probability theory [28][29].

The most important question raised by our consideration of the near-optimality of

ternary classification, and superiority over binary, is how these classes should be
identified in any application. The objective is not to go to unconstrained
unsupervised learning with its many classes, but rather to identify a triplicate-logic

10
structure that offers superior understanding compared to binary classification and
provides better performance in AI applications.

References

1. Frieder, G., Fong, A. and Chow, C.Y. (1973). A balanced ternary computer.
Conference Record of the 1973 International Symposium on Multiple-valued Logic,
pp. 68–88.
2. Hurst, S.L. (1984). Multiple-valued logic - Its status and its future. IEEE Trans.
Computers, C-33, 1160–1179.
3. Hayes, B. (2001). Third base. American Scientist 89: 490-494
4. Perrins, E. and Rice, M. (2008). PAM representation of ternary CPM. IEEE
Transactions on Communications 56, 2020-2024.
5. Kak, S. (2004). The Architecture of Knowledge. Motilal Banarsidass.
6. Brassard, G. (2005). Is information the key? Nat. Phys. 1: 2–4.
7. Chiribella, G., Perinotti, P., D’Ariano, G.M. (2011). Informational derivation of
quantum theory. Phys. Rev. A, 84, 012311:1–012311:47.
8. Penrose, R. (1989). The Emperor's New Mind. Oxford University Press.
9. Kak, S. (2018). Epistemic view of quantum communication. In Quantum
Foundations, Probability and Information, A. Khrennikov and B. Toni (editors).
Springer-Verlag International, pp. 119-128.
10. Kak, S. (1999). The initialization problem in quantum computing. Foundations of
Physics 29: 267-279
11. Kak, S. (2002). Quantum information and entropy. Int. Journal of Theoretical
Physics 46: 860-876
12. Vaidman, L. (2016). Many-worlds interpretation of quantum mechanics. The
Stanford Encyclopedia of Philosophy, E.N. Zalta (ed.)
13. Marr, D. and Vaina, L. (1982). Representation and recognition of the movements of
shapes. Proceedings of the Royal Society of London B, 214:501-524.
14. Knuth, D. (1997). The Art of Computer Programming: Fundamental Algorithms.
Addison-Wesley.
15. MacKay, D.J.C. (2003). Information Theory, Inference, and Learning Algorithms.
Cambridge University Press.
16. Kak, S. (2018). Variations on the Newcomb-Benford law. arXiv:1806.06695
17. Pan, Y. et al. (2016). Urban big data and the development of city intelligence.
Engineering 2: 171-178.
18. Zhang, Q. e al. (2017). A ternary classification using machine learning methods of
distinct estrogen receptor activities within a large collection of environmental
chemicals. Science of The Total Environment 580: 1268-1275

11
19. Guo, H, Wang, W. (2015). An active learning-based SVM multi-class classification
model. Pattern Recognition 48: 1577-1597.
20. Kak, S. (2006). Ayurveda. In Encyclopedia of India, Stanley Wolpert (ed.). Thomson
Gale.
21. Keswani, N. H. (ed.) (1974). The Science of Medicine and Physiological Concepts in
Ancient and Medieval India. All India Institute of Medical Sciences.
22. Yeung K.Y., Fraley, C., Murua, A., Raftery, A.E., Ruzzo, W.L. (2001). Model-based
clustering and data transformations for gene expression data. Bioinformatics 17:
977–87.
23. Fraley C. (2003). Enhanced software for model-based clustering, discriminant
analysis, and density estimation: MCLUST. Journal of Classification. 20: 263–86.
24. Franceschi, P., Wehrens, R. (2014) Self-organizing maps: a versatile tool for the
automatic analysis of untargeted imaging datasets. Proteomics. 14: 853–61.
25. Lynch, C.M., van Berkel, V.H., Frieboes, H.B. (2017) Application of unsupervised
analysis techniques to lung cancer patient data. PLoS ONE 12(9): e0184370.
26. Bergmann, M. (2008). An Introduction to Many-Valued and Fuzzy Logic: Semantics,
Algebras, and Derivation Systems. Cambridge University Press.
27. Kak, S. (2013). Probability constraints and the classical/quantum divide.
NeuroQuantology 11: 600-606.
28. Kak, S. (2014). Observability and computability in physics. Quantum Matter 3: 172-
176.
29. Goyal, P., Knuth, K.H. (2011). Quantum theory and probability theory: Their
relationship and origin in symmetry. Symmetry, 3, 171–206.

12