★ See also the **PDF version of this chapter** (better formatting/references) ★

See any bugs/typos/confusing explanations? Open a GitHub issue. You can also comment below

- Understand one of the most important concepts in computing: duality
between code and data.
- Build up comfort in moving between different representations of
programs.
- Follow the construction of a “universal NAND program” that can
evaluate other NAND programs given their representation.
- See and understand the proof of a major result that compliments the
result last chapter: some functions require an
*exponential*number of NAND lines to compute. - Understand the
*physical extended Church-Turing thesis*that NAND programs capture*all*feasible computation in the physical world, and its physical and philosophical implications.

“The term code script is, of course, too narrow. The chromosomal structures are at the same time instrumental in bringing about the development they foreshadow. They are law-code and executive power - or, to use another simile, they are architect’s plan and builder’s craft - in one.”, Erwin Schrödinger, 1944.

“A mathematician would hardly call a correspondence between the set of 64 triples of four units and a set of twenty other units,”universal“, while such correspondence is, probably, the most fundamental general feature of life on Earth”, Misha Gromov, 2013

A NAND program can be thought of as simply a sequence of symbols, each
of which can be encoded with zeros and ones using (for example) the
ASCII standard. Thus we can represent every NAND program as a binary
string. This statement seems obvious but it is actually quite profound.
It means that we can treat a NAND program both as instructions to
carrying computation and also as *data* that could potentially be input
to other computations.

This correspondence between *code* and *data* is one of the most
fundamental aspects of computing. It underlies the notion of *general
purpose* computers, that are not pre-wired to compute only one task, and
it is also the basis of our hope for obtaining *general* artificial
intelligence. This concept finds immense use in all areas of computing,
from scripting languages to machine learning, but it is fair to say that
we haven’t yet fully mastered it. Indeed many security exploits involve
cases such as “buffer overflows” when attackers manage to inject code
where the system expected only “passive” data (see
XKCDmomexploitsfig). The idea of code as data reaches beyond
the realm of electronic computers. For example, DNA can be thought of as
both a program and data (in the words of Schrödinger, who wrote before
DNA’s discovery a book that inspired Watson and Crick, it is both
“architect’s plan and builder’s craft”).

For every NAND program \(P\), we can represent \(P\) as a binary string. In particular, this means that for any choice of such representation, the following is a well defined mathematical function \(EVAL:\{0,1\}^* \times \{0,1\}^* \rightarrow \{0,1\}^*\)

\[ EVAL(P,x) = \begin{cases} P(x) & |x|= \text{no. of $P$'s inputs} \\ 0 & \text{otherwise} \end{cases} \] where we denote by \(P(x)\) the output of the program represented by the string \(P\) on the input \(x\).

The above is one of those observations that are simultaneously both
simple and profound. Please make sure that you understand **(1)** how
for every fixed choice of representing programs as strings, the function
\(EVAL\) above is well defined, and **(2)** what this function actually
does.

\(EVAL\) takes strings arbitrarily of length, and hence cannot be computed by a NAND program, that has a fixed length of inputs. However, one of the most interesting consequences of the fact that we can represent programs as strings is the following theorem:

For every \(s,n,m \in \N\) there is a NAND program that computes the
function \[
EVAL_{s,n,m}:\{0,1\}^{S+n} \rightarrow \{0,1\}^m
\] defined as follows. We let \(S\) be the number of bits that are needed
to represents programs of \(s\) lines. For every string \((P,x)\) where
\(P \in \{0,1\}^S\) and \(x\in\{0,1\}^n\), if \(P\) describes an \(s\) line NAND
program with \(n\) input bits and \(m\) outputs bits, then
\(EVAL_{s,n,m}(P,x)\) is the output of this program on input \(x\).

Of course to fully specify \(EVAL_{s,n,m}\), we need to fix a precise
representation scheme for NAND programs as binary strings. We can simply
use the ASCII representation, though below we will choose a more
convenient representation. But regardless of the choice of
representation, bounded-univ is an immediate corollary of
NAND-univ-thm, which states that *every* finite function, and
so in particular the function \(EVAL_{S,n,m}\) above, can be computed by
*some* NAND program.

Once again, bounded-univ is subtle but important. Make sure you understand what this thorem means, and why it is a corollary of NAND-univ-thm.

bounded-univ can be thought of as providing a “NAND
interpreter in NAND”. That is, for a particular size bound, we give a
*single* NAND program that can evaluate all NAND programs of that size.
We call this NAND program \(U\) that computes \(EVAL_{s,n,m}\) a *bounded
universal program*. “Universal” stands for the fact that this is a
*single program* that can evaluate *arbitrary* code, where “bounded”
stands for the fact that \(U\) only evaluates programs of bounded size. Of
course this limitation is inherent for the NAND programming language
where an \(N\)-line program can never compute a function with more than
\(N\) inputs. (We will later on introduce the concept of *loops*, that
allows to escape this limitation.)

It turns out that we don’t even need to pay that much of an overhead for universality

For every \(s,n,m \in \N\) there is a NAND program of at most \(O(s^2 \log s)\) lines that computes the function \(EVAL_{S,n,m}:\{0,1\}^{S+n} \rightarrow \{0,1\}^m\) defined above.

Unlike bounded-univ, eff-bounded-univ is not a trivial corollary of the fact that every function can be computed, and takes much more effort to prove. It requires us to present a concrete NAND program for the \(EVAL_{s,n,m}\) function. We will do so in several stages.

- First, we will spell out precisely how to represent NAND programs as strings. We can prove eff-bounded-univ using the ASCII representation, but a “cleaner” representation will be more convenient for us.
- Then, we will show how we can write a program to compute
\(EVAL_{s,n,m}\) in
*Python*.We will not use much about Python, and a reader that has familiarity with programming in any language should be able to follow along. - Finally, we will show how we can transform this Python program into a NAND program.

A NAND program is simply a sequence of lines of the form

There is of course nothing special about these particular identifiers. Hence to represent a NAND program mathematically, we can simply identify the variables with natural numbers, and think of each line as a triple \((i,j,k)\) which corresponds to saying that we assign to the \(i\)-th variable the NAND of the values of the \(j\)-th and \(k\)-th variables. We will use the set \([t]= \{0,1,\ldots,t-1\}\) as our set of variables, and for concreteness we will let the input variables be the first \(n\) numbers, and the output variables be the last \(m\) numbers (i.e., the numbers \((t-m,\dots,t-1)\)). This motivates the following definition:

Let \(P\) be a NAND program of \(n\) inputs, \(m\) outputs, and \(s\) lines, and
let \(t\) be the number of distinct variables used by \(P\). The *list of
tuples representation of \(P\)* is the triple \((n,m,L)\) where \(L\) is a
list of triples of the form \((i,j,k)\) for \(i,j,k \in [t]\).

For every variable of \(P\), we assign a number in \([t]\) as follows:

- For every \(i\in [n]\), the variable
`X[`

\(i\)`]`

is assigned the number \(i\). - For every \(j\in [m]\), the variable
`Y[`

\(j\)`]`

is assigned the number \(t-m+j\). - Every other variable is assigned a number in \(\{n,n+1,\ldots,t-m-1\}\) in the order of which it appears.

The list of tuples representation will be our default choice for
representing NAND programs, and since “list of tuples representation” is
a bit of a mouthful, we will often call this simply the *representation*
for a program \(P\).

Our favorite NAND program, the XOR program:

`u = NAND(X[0],X[1])`

`v = NAND(X[0],u)`

`w = NAND(X[1],u)`

`Y[0] = NAND(v,w)`

Is represented as the tuple \((2,1,L)\) where
\(L=((2, 0, 1), (3, 0, 2), (4, 1, 2), (5, 3, 4))\). That is, the variables
`X[0]`

and `X[1]`

are given the indices \(0\) and \(1\) respectively, the
variables `u`

,`v`

,`w`

are given the indices \(2,3,4\) respectively, and
the variable `Y[0]`

is given the index \(5\).

Transforming a NAND program from its representation as code to the
representation as a list of tuples is a fairly straightforward
programming exercise, and in particular can be done in a few lines of
*Python*.

To obtain a representation that we can use as input to a NAND program, we need to take a step further and map the triple \((n,m,L)\) to a binary string. Here there are many different choices, but let us fix one of them. If the list \(L\) has \(s\) triples in it, we will represent it as simply the string \(str(L)\) which will be the concatenation of the \(3s\) numbers in the binary basis, which can be encoded as a string of length \(3s\ell\) where \(\ell = \ceil{\log 3s}\) is a number of bits that is guaranteed to be sufficient to represent numbers in \([t]\) (since \(t \leq 3s\)). We will represent the program \((n,m,L)\) as the string \(\expr{n}\expr{m}\expr{s}str(L)\) where \(\expr{n}\) and \(\expr{m}\) are some prefix free representations of \(n\), \(m\) and \(s\) (see prefixfreesec). Hence an \(s\) line program will be represented by a string of length \(O(s \log s)\). In the context of computing \(EVAL_{s,n,m}\) the number of lines, inputs, and outputs, is fixed, and so we can drop \(n,m,s\) and simply think of it as a function that maps \(\{0,1\}^{3s\ell + n}\) to \(\{0,1\}^m\), where \(\ell = \ceil{\log 3s}\).

To prove eff-bounded-univ it suffices to give a NAND program
of \(O(s^2 \log s) \leq O((s\log s)^2)\) lines that can evaluate NAND
programs of \(s\) lines. Let us start by thinking how we would evaluate
such programs if we weren’t restricted to the NAND operations. That is,
let us describe informally an *algorithm* that on input \(n,m,s\), a list
of triples \(L\), and a string \(x\in \{0,1\}^n\), evaluates the program
represented by \((n,m,L)\) on the string \(x\).

It would be highly worthwhile for you to stop here and try to solve this
problem yourself. For example, you can try thinking how you would write
a program `NANDEVAL(n,m,s,L,x)`

that computes this function in the
programming language of your choice.

Here is a description of such an algorithm:

**Input:** Numbers \(n,m\) and a list \(L\) of \(s\) triples of numbers in
\([t]\) for some \(t\leq 3s\), as well as a string \(x\in \{0,1\}^n\).

**Goal:** Evaluate the program represented by \((n,m,L)\) on the input
\(x\in \{0,1\}^n\).

**Operation:**

- We will create a
*dictionary*data structure`Vartable`

that for every \(i \in [t]\) stores a bit. We will assume we have the operations`GET(Vartable,i)`

which restore the bit corresponding to`i`

, and the operation`UPDATE(Vartable,i,b)`

which update the bit corresponding to`i`

with the value`b`

. (More concretely, we will write this as`Vartable = UPDATE(Vartable,i,b)`

to emphasize the fact that the state of the data structure changes, and to keep our convention of using functions free of “side effects”.) - We will initialize the table by setting the \(i\)-th value of
`Vartable`

to \(x_i\) for every \(i\in [n]\). - We will go over the list \(L\) in order, and for every triple
\((i,j,k)\) in \(L\), we let \(a\) be
`GET(Vartable,`

\(j\)`)`

, \(b\) be`GET(Vartable`

\(k\)`)`

, and then set the value corresponding to \(i\) to the NAND of \(a\) and \(b\). That is, let`Vartable = UPDATE(Vartable,`

\(i\),`NAND(`

\(a\)`,`

\(b\)`))`

. - Finally, we output the value
`GET(Vartable,`

\(t-m+j\)`)`

for every \(j\in [m]\).

Please make sure you understand this algorithm and why it does produce the right value.

To make things more concrete, let us see how we implement the above
algorithm in the *Python* programming language. We will construct a
function `NANDEVAL`

that on input \(n,m,L,x\) will output the result of
evaluating the program represented by \((n,m,L)\) on \(x\).

Accessing an element of the array `Vartable`

at a given index takes a
constant number of basic operations. Hence (since \(n,m \leq s\) and
\(t \leq 3s\)), the program above will use \(O(s)\) basic operations.

We now turn to describing the proof of eff-bounded-univ. To do this, it is of course not enough to give a Python program. Rather, we need to show how we compute the function \(EVAL_{s,n,m}\) by a NAND program. In other words, our job is to transform, for every \(s,n,m\), the Python code above to a NAND program \(U_{s,n,m}\) that computes the function \(EVAL_{s,n,m}\).

Before reading further, try to think how *you* could give a
“constructive proof” of eff-bounded-univ. That is, think of
how you would write, in the programming language of your choice, a
function `universal(s,n,m)`

that on input \(s,n,m\) outputs the code for
the NAND program \(U_{s,n,m}\) such that \(U_{s,n,m}\) computes
\(EVAL_{s,n,m}\). Note that there is a subtle but crucial difference
between this function and the Python `NANDEVAL`

program described above.
Rather than actually evaluating a given program \(P\) on some input \(w\),
the function `universal`

should output the *code* of a NAND program that
computes the map \((P,x) \mapsto P(x)\).

Our construction will follow very closely the Python implementation of
`EVAL`

above. We will use variables
`Vartable[`

\(0\)`]`

,\(\ldots\),`Vartable[`

\(2^\ell-1\)`]`

, where
\(\ell = \ceil{\log 3s}\) to store our variables. However, NAND doesn’t
have integer-valued variables, so we cannot write code such as
`Vartable[i]`

for some variable `i`

. However, we *can* implement the
function `GET(Vartable,i)`

that outputs the `i`

-th bit of the array
`Vartable`

. Indeed, this is nothing by the function `LOOKUP`

that we
have seen in lookup-thm!

Please make sure that you understand why `GET`

and `LOOKUP`

are the same
function.

We saw that we can compute `LOOKUP`

on arrays of size \(2^\ell\) in time
\(O(2^\ell)\), which will be \(O(s)\) for our choice of \(\ell\).

To compute the `update`

function on input `V`

,`i`

,`b`

, we need to scan
the array `V`

, and for \(j \in [2^\ell]\), have our \(j\)-th output be
`V[`

\(j\)`]`

unless \(j\) is equal to `i`

, in which case the \(j\)-th output
is `b`

. We can do this as follows:

- For every \(j\in [2^\ell]\), there is an \(O(\ell)\) line NAND program to compute the function \(EQUALS_j: \{0,1\}^\ell \rightarrow \{0,1\}\) that on input \(i\) outputs \(1\) if and only if \(i\) is equal to (the binary representation of) \(j\). (We leave verifying this as equals and equalstwo.)
- We have seen that we can compute the function \(IF:\{0,1\}^3 \rightarrow \{0,1\}\) such that \(IF(a,b,c)\) equals \(b\) if \(a=1\) and \(c\) if \(a=0\).

Together, this means that we can compute `UPDATE`

as follows:

Once we can compute `GET`

and `UPDATE`

, the rest of the implementation
amounts to “book keeping” that needs to be done carefully, but is not
too insightful. Hence we omit the details from this chapter. See the
appendix for the full details of how to compute the universal NAND
evaluator in NAND.

Since the loop over `j`

in `UPDATE`

is run \(2^\ell\) times, and computing
`EQUALS_j`

takes \(O(\ell)\) lines, the total number of lines to compute
`UPDATE`

is \(O(2^\ell \cdot \ell) = O(s \log s)\). Since we run this
function \(s\) times, the total number of lines for computing
\(EVAL_{s,n,m}\) is \(O(s^2 \log s)\). This completes (up to the omitted
details) the proof of eff-bounded-univ.

The NAND program above is less efficient that its Python counterpart,
since NAND does not offer arrays with efficient random access. Hence for
example the `LOOKUP`

operation on an array of \(s\) bits takes \(\Omega(s)\)
lines in NAND even though it takes \(O(1)\) steps (or maybe \(O(\log s)\)
steps, depending how we count) in *Python*.

It turns out that it is possible to improve the bound of
eff-bounded-univ, and evaluate \(s\) line NAND programs using a
NAND program of \(O(s \log s)\) lines. The key is to consider the
description of NAND programs as circuits, and in particular as directed
acyclic graphs (DAGs) of bounded in degree. A universal NAND program
\(U_s\) for \(s\) line programs will correspond to a *universal graph* \(H_s\)
for such \(s\) vertex DAGs. We can think of such as graph \(U_s\) as fixed
“wiring” for communication network, that should be able to accommodate
any arbitrary pattern of communication between \(s\) vertices (where this
pattern corresponds to an \(s\) line NAND program). It turns out that
there exist such efficient routing networks
exist that allow embedding any \(s\) vertex circuit inside a universal
graph of size \(O(s \log s)\), see this recent
paper for more on this issue.

To prove eff-bounded-univ we essentially translated every line
of the Python program for `EVAL`

into an equivalent NAND snippet. It
turns out that none of our reasoning was specific to the particular
function \(EVAL\). It is possible to translate *every* Python program into
an equivalent `NAND`

program of comparable efficiency.`C`

program. We can combine this with a C compiler to transform a Python
program to various flavors of “machine language”.

So, to transform a Python program into an equivalent NAND program, it is
enough to show how to transform a machine language program into an
equivalent NAND program. One minimalistic (and hence convenient) family
of machine languages is known as the *ARM architecture* which powers a
great many mobile devices including essentially all Android devices.`NAND2C`

function in the appendix.

Going one by one over the instruction sets of such computers and translating them to NAND snippets is no fun, but it is a feasible thing to do. In fact, ultimately this is very similar to the transformation that takes place in converting our high level code to actual silicon gates that are not so different from the operations of a NAND program. Indeed, tools such as MyHDL that transform “Python to Silicon” can be used to convert a Python program to a NAND program.

The NAND programming language is just a teaching tool, and by no means
do I suggest that writing NAND programs, or compilers to NAND, is a
practical, useful, or even enjoyable activity. What I do want is to make
sure you understand why it *can* be done, and to have the confidence
that if your life (or at least your grade in this course) depended on
it, then you would be able to do this. Understanding how programs in
high level languages such as Python are eventually transformed into
concrete low-level representation such as NAND is fundamental to
computer science.

The astute reader might notice that the above paragraphs only outlined
why it should be possible to find for every *particular*
Python-computable function \(F\), a *particular* comparably efficient NAND
program \(P\) that computes \(F\). But this still seems to fall short of our
goal of writing a “Python interpreter in NAND” which would mean that for
every parameter \(n\), we come up with a *single* NAND program \(UNIV_n\)
such that given a description of a Python program \(P\), a particular
input \(x\), and a bound \(T\) on the number of operations (where the length
of \(P\), \(x\) and the magnitude of \(T\) are all at most \(n\)) would return
the result of executing \(P\) on \(x\) for at most \(T\) steps. After all, the
transformation above would transform every Python program into a
different NAND program, but would not yield “one NAND program to rule
them all” that can evaluate every Python program up to some given
complexity. However, it turns out that it is enough to show such a
transformation for a single Python program. The reason is that we can
write a Python interpreter *in Python*: a Python program \(U\) that takes
a bit string, interprets it as Python code, and then runs that code.
Hence, we only need to show a NAND program \(U^*\) that computes the same
function as the particular Python program \(U\), and this will give us a
way to evaluate *all* Python programs.

What we are seeing time and again is the notion of *universality* or
*self reference* of computation, which is the sense that all reasonably
rich models of computation are expressive enough that they can “simulate
themselves”. The importance of this phenomena to both the theory and
practice of computing, as well as far beyond it, including the
foundations of mathematics and basic questions in science, cannot be
overstated.

One of the consequences of our representation is the following:

\[|Size(s)| \leq 2^{O(s \log s)}.\] That is, there are at most \(2^{O(s\log s)}\) functions computed by NAND programs of at most \(s\) lines.

Moreover, the implicit constant in the \(O(\cdot)\) notation in
program-count is at most \(10\).

The idea behind the proof is that we can represent every \(s\) line program by a binary string of \(O(s \log s)\) bits. Therefore the number of functions computed by \(s\)-line programs cannot be larger than the number of such strings, which is \(2^{O(s \log s)}\). In the actual proof, given below, we count the number of representations a little more carefully, talking directly about triples rather than binary strings, although the idea remains the same.

Every NAND program \(P\) with \(s\) lines has at most \(3s\) variables. Hence, using our canonical representation, \(P\) can be represented by the numbers \(n,m\) of \(P\)’s inputs and outputs, as well as by the list \(L\) of \(s\) triples of natural numbers, each of which is smaller or equal to \(3s\).

If two programs compute distinct functions then they have distinct representations. So we will simply count the number of such representations: for every \(s' \leq s\), the number of \(s'\)-long lists of triples of numbers in \([3s]\) is \((3s)^{3s'}\), which in particular is smaller than \((3s)^{3s}\). So, for every \(s' \leq s\) and \(n,m\), the total number of representations of \(s'\)-line programs with \(n\) inputs and \(m\) outputs is smaller than \((3s)^{3s}\).

Since a program of at most \(s\) lines has at most \(s\) inputs and outputs, the total number of representations of all programs of at most \(s\) lines is smaller than \[ s\times s \times s \times (3s)^{3s} = (3s)^{3s+3} \label{eqcountbound} \] (the factor \(s\times s\ times s\) arises from taking all of the at most \(s\) options for the number of inputs \(n\), all of the at most \(s\) options for the number of outputs \(m\), and all of the at most \(s\) options for the number of lines \(s'\)). We claim that for \(s\) large enough, the righthand side of \eqref{eqcountbound} (and hence the total number of representations of programs of at most \(s\) lines) is smaller than \(2^{4 s \log s}\). Indeed, we can write \(3s = 2^{\log(3s)}=2^{\log 3 + \log s} \leq 2^{2+\log s}\), and hence the righthand side of \eqref{eqcountbound} is at most \(\left(2^{2+ \log s}\right)^{3s+3} = 2^{(2+\log s)(3s+3)} \leq 2^{4s\log s}\) for \(s\) large enough.

For every function \(F \in Size(s)\) there is a program \(P\) of at most \(s\) lines that computes it, and we can map \(F\) to its representation as a tuple \((n,m,L)\). If \(F \neq F'\) then a program \(P\) that computes \(F\) must have an input on which it disagrees with any program \(P'\) that computes \(F'\), and hence in particular \(P\) and \(P'\) have distinct representations. Thus we see that the map of \(Size(s)\) to its representation is one to one, and so in particular \(|Size(s)|\) is at most the number of distinct representations which is it at most \(2^{4s\log s}\).

We can also establish program-count directly from the ASCII
representation of the source code. Since an \(s\)-line NAND program has at
most \(3s\) distinct variables, we can change all the non input/output
variables of such a program to have the form `Temp[`

\(i\)`]`

for \(i\)
between \(0\) and \(3s-1\) without changing the function that it computes.
This means that after removing extra whitespaces, every line of such a
program (which will be of the form form `var = NAND(var',var'')`

for
variable identifiers which will be either `X[###]`

,`Y[###]`

or
`Temp[###]`

where `###`

is some number smaller than \(3s\)) will require
at most, say, \(20 + 3\log_{10} (3s) \leq O(\log s)\) characters. Since
each one of those characters can be encoded using seven bits in the
ASCII representation, we see that the number of functions computed by
\(s\)-line NAND programs is at most \(2^{O(s \log s)}\).

A function mapping \(\{0,1\}^2\) to \(\{0,1\}\) can be identified with the
table of its four values on the inputs \(00,01,10,11\); a function mapping
\(\{0,1\}^3\) to \(\{0,1\}\) can be identified with the table of its eight
values on the inputs \(000,001,010,011,100,101,110,111\). More generally,
every function \(F:\{0,1\}^n \rightarrow \{0,1\}\) can be identified with
the table of its \(2^n\) values on the inputs \(\{0,1\}^n\). Hence the
number of functions mapping \(\{0,1\}^n\) to \(\{0,1\}\) is equal to the
number of such tables which (since we can choose either \(0\) or \(1\) for
every row) is exactly \(2^{2^n}\). Note that this is *double exponential*
in \(n\), and hence even for small values of \(n\) (e.g., \(n=10\)) the number
of functions from \(\{0,1\}^n\) to \(\{0,1\}\) is truly astronomical.

There is a function \(F:\{0,1\}^n\rightarrow \{0,1\}\) such that the shortest NAND program to compute \(F\) requires \(2^n/(100n)\) lines.

Suppose, towards the sake of contradiction, that every function \(F:\{0,1\}^n\rightarrow\{0,1\}\) can be computed by a NAND program of at most \(s=2^n/(100n)\) lines. Then the by program-count the total number of such functions would be at most \(2^{10s\log s} \leq 2^{10 \log s \cdot 2^n/(100 n)}\). Since \(\log s = n - \log (100 n) \leq n\) this means that the total number of such functions would be at most \(2^{2^n/10}\), contradicting the fact that there are \(2^{2^n}\) of them.

We have seen before that *every* function mapping \(\{0,1\}^n\) to
\(\{0,1\}\) can be computed by an \(O(2^n /n)\) line program. We now see
that this is tight in the sense that some functions do require such an
astronomical number of lines to compute. In fact, as we explore in the
exercises below, this is the case for *most* functions. Hence functions
that can be computed in a small number of lines (such as addition,
multiplication, finding short paths in graphs, or even the \(EVAL\)
function) are the exception, rather than the rule.

The list of triples is not the shortest representation for NAND programs. We have seen that every NAND program of \(s\) lines and \(n\) inputs can be represented by a directed graph of \(s+n\) vertices, of which \(n\) have in-degree zero, and the \(s\) others have in-degree at most two. Using the adjacency list representation, such a graph can be represented using roughly \(2s\log(s+n) \leq 2s (\log s + O(1))\) bits. Using this representation we can reduce the implicit constant in program-count arbitrarily close to \(2\).

We’ve seen that NAND gates can be implemented using very different systems in the physical world. What about the reverse direction? Can NAND programs simulate any physical computer?

We can take a leap of faith and stipulate that NAND programs do actually
encapsulate *every* computation that we can think of. Such a statement
(in the realm of infinite functions, which we’ll encounter in
chaploops) is typically attributed to Alonzo Church and Alan
Turing, and in that context is known as the *Church Turing Thesis*. As
we will discuss in future lectures, the Church-Turing Thesis is not a
mathematical theorem or conjecture. Rather, like theories in physics,
the Church-Turing Thesis is about mathematically modelling the real
world. In the context of finite functions, we can make the following
informal hypothesis or prediction:

If a function \(F:\{0,1\}^n \rightarrow \{0,1\}^m\) can be computed in the physical world using \(s\) amount of “physical resources” then it can be computed by a NAND program of roughly \(s\) lines.

We call this hypothesis the **“Physical Extended Church-Turing Thesis”**
or *PECTT* for short. A priori it might seem rather extreme to
hypothesize that our meager NAND model captures all possible physical
computation. But yet, in more than a century of computing technologies,
no one has yet built any scalable computing device that challenges this
hypothesis.

We now discuss the “fine print” of the PECTT in more detail, as well as the (so far unsuccessful) challenges that have been raised against it. There is no single universally-agreed-upon formalization of “roughly \(s\) physical resources”, but we can approximate this notion by considering the size of any physical computing device and the time it takes to compute the output, and ask that any such device can be simulated by a NAND program with a number of lines that is a polynomial (with not too large exponent) in the size of the system and the time it takes it to operate.

In other words, we can phrase the PECTT as stipulating that any function
that can be computed by a device of volume \(V\) and time \(t\), must be
computable by a NAND program that has at most \(\alpha(Vt)^\beta\) lines
for some constants \(\alpha,\beta\). The exact values for \(\alpha,\beta\)
are not so clear, but it is generally accepted that if
\(F:\{0,1\}^n \rightarrow \{0,1\}\) is an *exponentially hard* function,
in the sense that it has no NAND program of fewer than, say, \(2^{n/2}\)
lines, then a demonstration of a physical device that can compute \(F\)
for moderate input lengths (e.g., \(n=500\)) would be a violation of the
PECTT.

We can attempt at a more exact phrasing of the PECTT as follows. Suppose
that \(Z\) is a physical system that accepts \(n\) binary stimuli and has a
binary output, and can be enclosed in a sphere of volume \(V\). We say
that the system \(Z\) *computes* a function
\(F:\{0,1\}^n \rightarrow \{0,1\}\) within \(t\) seconds if whenever we set
the stimuli to some value \(x\in \{0,1\}^n\), if we measure the output
after \(t\) seconds then we obtain \(F(x)\).

We can phrase the PECTT as stipulating that if there exists such a
system \(Z\) that computes \(F\) within \(t\) seconds, then there exists a
NAND program that computes \(F\) and has at most \(\alpha(Vt)^2\) lines,
where \(\alpha\) is some normalization constant.*exponentially* in \(n\), it is not hard to set parameters so that even
for moderately large values of \(n\), such a system could not fit in our
universe.

To fully make the PECTT concrete, we need to decide on the units for
measuring time and volume, and the normalization constant \(\alpha\). One
conservative choice is to assume that we could squeeze computation to
the absolute physical limits (which are many orders of magnitude beyond
current technology). This corresponds to setting \(\alpha=1\) and using
the Planck units for volume and time. The
*Planck length* \(\ell_P\) (which is, roughly speaking, the shortest
distance that can theoretically be measured) is roughly \(2^{-120}\)
meters. The *Planck time* \(t_P\) (which is the time it takes for light to
travel one Planck length) is about \(2^{-150}\) seconds. In the above
setting, if a function \(F\) takes, say, 1KB of input (e.g., roughly
\(10^4\) bits, which can encode a \(100\) by \(100\) bitmap image), and
requires at least \(2^{0.8 n}= 2^{0.8 \cdot 10^4}\) NAND lines to compute,
then any physical system that computes it would require either volume of
\(2^{0.2\cdot 10^4}\) Planck length cubed, which is more than \(2^{1500}\)
meters cubed or take at least \(2^{0.2 \cdot 10^4}\) Planck Time units,
which is larger than \(2^{1500}\) seconds. To get a sense of how big that
number is, note that the universe is only about \(2^{60}\) seconds old,
and its observable radius is only roughly \(2^{90}\) meters. The above
discussion suggests that it is possible to *empirically falsify* the
PECTT by presenting a smaller-than-universe-size system that computes
such a function.*interactive proofs* and *program
checking* that we might encounter later in this book. Another,
perhaps more salient problem, is that while we know many hard
functions exist, at the moment there is *no single explicit
function* \(F:\{0,1\}^n \rightarrow \{0,1\}\) for which we can *prove*
an \(\omega(n)\) (let alone \(\Omega(2^n/n)\)) lower bound on the number
of lines that a NAND program needs to compute it.

One of the admirable traits of mankind is the refusal to accept limitations. In the best case this is manifested by people achieving longstanding “impossible” challenges such as heavier-than-air flight, putting a person on the moon, circumnavigating the globe, or even resolving Fermat’s Last Theorem. In the worst case it is manifested by people continually following the footsteps of previous failures to try to do proven-impossible tasks such as build a perpetual motion machine, trisect an angle with a compass and straightedge, or refute Bell’s inequality. The Physical Extended Church Turing thesis (in its various forms) has attracted both types of people. Here are some physical devices that have been speculated to achieve computational tasks that cannot be done by not-too-large NAND programs:

**Spaghetti sort:**One of the first lower bounds that Computer Science students encounter is that sorting \(n\) numbers requires making \(\Omega(n \log n)\) comparisons. The “spaghetti sort” is a description of a proposed “mechanical computer” that would do this faster. The idea is that to sort \(n\) numbers \(x_1,\ldots,x_n\), we could cut \(n\) spaghetti noodles into lengths \(x_1,\ldots,x_n\), and then if we simply hold them together in our hand and bring them down to a flat surface, they will emerge in sorted order. There are a great many reasons why this is not truly a challenge to the PECTT hypothesis, and I will not ruin the reader’s fun in finding them out by her or himself.**Soap bubbles:**One function \(F:\{0,1\}^n \rightarrow \{0,1\}\) that is conjectured to require a large number of NAND lines to solve is the*Euclidean Steiner Tree*problem. This is the problem where one is given \(m\) points in the plane \((x_1,y_1),\ldots,(x_m,y_m)\) (say with integer coordinates ranging from \(1\) till \(m\), and hence the list can be represented as a string of \(n=O(m \log m)\) size) and some number \(K\). The goal is to figure out whether it is possible to connect all the points by line segments of total length at most \(K\). This function is conjectured to be hard because it is*NP complete*- a concept that we’ll encounter later in this course - and it is in fact reasonable to conjecture that as \(m\) grows, the number of NAND lines required to compute this function grows*exponentially*in \(m\), meaning that the PECTT would predict that if \(m\) is sufficiently large (such as few hundreds or so) then no physical device could compute \(F\). Yet, some people claimed that there is in fact a very simple physical device that could solve this problem, that can be constructed using some wooden pegs and soap. The idea is that if we take two glass plates, and put \(m\) wooden pegs between them in the locations \((x_1,y_1),\ldots,(x_m,y_m)\) then bubbles will form whose edges touch those pegs in the way that will minimize the total energy which turns out to be a function of the total length of the line segments. The problem with this device of course is that nature, just like people, often gets stuck in “local optima”. That is, the resulting configuration will not be one that achieves the absolute minimum of the total energy but rather one that can’t be improved with local changes. Aaronson has carried out actual experiments (see aaronsonsoapfig), and saw that while this device often is successful for three or four pegs, it starts yielding suboptimal results once the number of pegs grows beyond that.

**DNA computing.**People have suggested using the properties of DNA to do hard computational problems. The main advantage of DNA is the ability to potentially encode a lot of information in relatively small physical space, as well as compute on this information in a highly parallel manner. At the time of this writing, it was demonstrated that one can use DNA to store about \(10^{16}\) bits of information in a region of radius about milimiter, as opposed to about \(10^{10}\) bits with the best known hard disk technology. This does not posit a real challenge to the PECTT but does suggest that one should be conservative about the choice of constant and not assume that current hard disk + silicon technologies are the absolute best possible.We were extremely conservative in the suggested parameters for the PECTT, having assumed that as many as \(\ell_P^{-2}10^{-6} \sim 10^{61}\) bits could potentially be stored in a milimeter radius region. **Continuous/real computers.**The physical world is often described using continuous quantities such as time and space, and people have suggested that analog devices might have direct access to computing with real-valued quantities and would be inherently more powerful than discrete models such as NAND machines. Whether the “true” physical world is continuous or discrete is an open question. In fact, we do not even know how to precisely*phrase*this question, let alone answer it. Yet, regardless of the answer, it seems clear that the effort to measure a continuous quantity grows with the level of accuracy desired, and so there is no “free lunch” or way to bypass the PECTT using such machines (see also this paper). Related to that are proposals known as “hypercomputing” or “Zeno’s computers” which attempt to use the continuity of time by doing the first operation in one second, the second one in half a second, the third operation in a quarter second and so on.. These fail for a similar reason to the one guaranteeing that Achilles will eventually catch the tortoise despite the original Zeno’s paradox.**Relativity computer and time travel.**The formulation above assumed the notion of time, but under the theory of relativity time is in the eye of the observer. One approach to solve hard problems is to leave the computer to run for a lot of time from*his*perspective, but to ensure that this is actually a short while from*our*perspective. One approach to do so is for the user to start the computer and then go for a quick jog at close to the speed of light before checking on its status. Depending on how fast one goes, few seconds from the point of view of the user might correspond to centuries in computer time (it might even finish updating its Windows operating system!). Of course the catch here is that the energy required from the user is proportional to how close one needs to get to the speed of light. A more interesting proposal is to use time travel via*closed timelike curves (CTCs)*. In this case we could run an arbitrarily long computation by doing some calculations, remembering the current state, and the travelling back in time to continue where we left off. Indeed, if CTCs exist then we’d probably have to revise the PECTT (though in this case I will simply travel back in time and edit these notes, so I can claim I never conjectured it in the first place…)**Humans.**Another computing system that has been proposed as a counterexample to the PECTT is a 3 pound computer of about 0.1m radius, namely the human brain. Humans can walk around, talk, feel, and do others things that are not commonly done by NAND programs, but can they compute partial functions that NAND programs cannot? There are certainly computational tasks that*at the moment*humans do better than computers (e.g., play some video games, at the moment), but based on our current understanding of the brain, humans (or other animals) have no*inherent*computational advantage over computers. The brain has about \(10^{11}\) neurons, each operating in a speed of about \(1000\) operations per seconds. Hence a rough first approximation is that a NAND program of about \(10^{14}\) lines could simulate one second of a brain’s activity.This is a very rough approximation that could be wrong to a few orders of magnitude in either direction. For one, there are other structures in the brain apart from neurons that one might need to simulate, hence requiring higher overhead. On ther other hand, it is by no mean clear that we need to fully clone the brain in order to achieve the same computational tasks that it does. Note that the fact that such a NAND program (likely) exists does not mean it is easy to*find*it. After all, constructing this program took evolution billions of years. Much of the recent efforts in artificial intelligence research is focused on finding programs that replicate some of the brain’s capabilities and they take massive computational effort to discover, these programs often turn out to be much smaller than the pessimistic estimates above. For example, at the time of this writing, Google’s neural network for machine translation has about \(10^4\) nodes (and can be simulated by a NAND program of comparable size). Philosophers, priests and many others have since time immemorial argued that there is something about humans that cannot be captured by mechanical devices such as computers; whether or not that is the case, the evidence is thin that humans can perform computational tasks that are inherently impossible to achieve by computers of similar complexity.There are some well known scientists that have advocated that humans have inherent computational advantages over computers. See also this. **Quantum computation.**The most compelling attack on the Physical Extended Church Turing Thesis comes from the notion of*quantum computing*. The idea was initiated by the observation that systems with strong quantum effects are very hard to simulate on a computer. Turning this observation on its head, people have proposed using such systems to perform computations that we do not know how to do otherwise. At the time of this writing, Scalable quantum computers have not yet been built, but it is a fascinating possibility, and one that does not seem to contradict any known law of nature. We will discuss quantum computing in much more detail later in this course. Modeling it will essentially involve extending the NAND programming language to the “QNAND” programming language that has one more (very special) operation. However, the main take away is that while quantum computing does suggest we need to amend the PECTT, it does*not*require a complete revision of our worldview. Indeed, almost all of the content of this course remains the same whether the underlying computational model is the “classical” model of NAND programs or the quantum model of QNAND programs (also known as*quantum circuits*).

While even the precise phrasing of the PECTT, let alone understanding
its correctness, is still a subject of research, some variant of it is
already implicitly assumed in practice. A statement such as “this
cryptosystem provides 128 bits of security” really means that **(a)** it
is conjectured that there is no Boolean circuit (or, equivalently, a
NAND gate) of size much smaller than \(2^{128}\) that can break the
system,*prove* such a statement for any
cryptosystem. This is related to the \(\mathbf{P}\) vs \(\mathbf{NP}\)
question we will discuss in future chapters.**(b)** we assume that no other physical mechanism can
do better, and hence it would take roughly a \(2^{128}\) amount of
“resources” to break the system.

- We can think of programs both as describing a
*process*, as well as simply a list of symbols that can be considered as*data*that can be fed as input to other programs. - We can write a NAND program that evaluates arbitrary NAND programs. Moreover, the efficiency loss in doing so is not too large.
- We can even write a NAND program that evaluates programs in other programming languages such as Python, C, Lisp, Java, Go, etc.
- By a leap of faith, we could hypothesize that the number of lines in
the smallest NAND program for a function \(F\) captures roughly the
amount of physical resources required to compute \(F\). This statement
is known as the
*Physical Extended Church-Turing Thesis (PECTT)*. - NAND programs capture a surprisingly wide array of computational
models. The strongest currently known challenge to the PECTT comes
from the potential for using quantum mechanical effects to speed-up
computation, a model known as
*quantum computers*.

Most of the exercises have been written in the summer of 2018 and haven’t yet been fully debugged. While I would prefer people do not post online solutions to the exercises, I would greatly appreciate if you let me know of any bugs. You can do so by posting a GitHub issue about the exercise, and optionally complement this with an email to me with more details about the attempted solution.

Which one of the following statements is false:

a. There is an \(O(s^3)\) line NAND program that given as input program
\(P\) of \(s\) lines in the list-of-tuples representation computes the
output of \(P\) when all its input are equal to \(1\).

b. There is an \(O(s^3)\) line NAND program that given as input program
\(P\) of \(s\) characters encoded as a string of \(7s\) bits using the ASCII
encoding, computes the output of \(P\) when all its input are equal to
\(1\).

c. There is an \(O(\sqrt{s})\) line NAND program that given as input
program \(P\) of \(s\) lines in the list-of-tuples representation computes
the output of \(P\) when all its input are equal to \(1\).

For every \(k \in \N\), show that there is an \(O(k)\) line NAND program that computes the function \(EQUALS_k:\{0,1\}^{2k} \rightarrow \{0,1\}\) where \(EQUALS(x,x')=1\) if and only if \(x=x'\).

For every \(k \in \N\) and \(x' \in \{0,1\}^k\), show that there is an \(O(k)\) line NAND program that computes the function \(EQUALS_{x'} : \{0,1\}^k \rightarrow \{0,1\}\) that on input \(x\in \{0,1\}^k\) outputs \(1\) if and only if \(x=x'\).

Suppose \(n>1000\) and that we choose a function
\(F:\{0,1\}^n \rightarrow \{0,1\}\) at random, choosing for every
\(x\in \{0,1\}^n\) the value \(F(x)\) to be the result of tossing an
independent unbiased coin. Prove that the probability that there is a
\(2^n/(1000n)\) line program that computes \(F\) is at most \(2^{-100}\).**Hint:** An equivalent way to say this is that you need to prove
that the set of functions that can be computed using at most
\(2^n/(1000n)\) has fewer than \(2^{-100}2^{2^n}\) elements. Can you see
why?

Prove that there is a constant \(c\) such that for every \(n\), there is
some function \(F:\{0,1\}^n \rightarrow \{0,1\}\) s.t. **(1)** \(F\) *can*
be computed by a NAND program of at most \(c n^5\) lines, but **(2)** \(F\)
can *not* be computed by a NAND program of at most \(n^4 /c\) lines.**Hint:** Find an approriate value of \(t\) and a function
\(G:\{0,1\}^t \rightarrow \{0,1\}\) that can be computed in \(O(2^t/t)\)
lines but *can’t* be computed in \(\Omega(2^t/t)\) lines, and then
extend this to a function mapping \(\{0,1\}^n\) to \(\{0,1\}\).

*Circuit Evaluation* typically. More
references regarding oblivious RAM etc..

Scott Aaronson’s blog post on how information is physical is a good discussion on issues related to the physical extended Church-Turing Physics. Aaronson’s survey on NP complete problems and physical reality is also a great source for some of these issues, though might be easier to read after we reach cooklevinchap on \(\mathbf{NP}\) and \(\mathbf{NP}\)-completeness.

Some topics related to this chapter that might be accessible to advanced students include:

- Lower bounds. While we’ve seen the “most” functions mapping \(n\) bits
to one bit require NAND programs of exponential size
\(\Omega(2^n/n)\), we actually do not know of any
*explicit*function for which we can*prove*that it requires, say, at least \(n^{100}\) or even \(100n\) size. At the moment, strongest such lower bound we know is that there are quite simple and explicit \(n\)-variable functions that require at least \((5-o(1))n\) lines to compute, see this paper of Iwama et al as well as this more recent work of Kulikov et al. Proving lower bounds for restricted models of straightline programs (more often described as*circuits*) is an extremely interesting research area, for which Jukna’s book provides very good introduction and overview.

Copyright 2018, Boaz Barak.

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

HTML version is produced using the Distill template, Copyright 2018, The Distill Template Authors.

## Comments

Comments are posted on the GitHub repository using the utteranc.es app. A GitHub login is required to comment. If you don't want to authorize the app to post on your behalf, you can also comment directly on the GitHub issue for this page.