On Fri, Jul 14, 2023, 9:03 PM James Bowery <[email protected]> wrote:

> https://www.mdpi.com/1099-4300/25/5/763
>

It's an interesting paper that helped me understand quantum computing a
little better. Recall that the state of a quantum computer with n qubits is
described by a unit vector in a 2^n dimensional space (a Hilbert space).
The only allowed operations are rotations (unitary), which are time
reversible. Copying, AND, and OR are not allowed, but swap, NOT, and
controlled NOT are allowed as long as the number of inputs and outputs of
each gate are the same. You can read the state only once. The result is one
of the possible 2^n bit strings with probability equal to the square of the
magnitude of the unit vector in that dimension, after which the state is
destroyed.

For example, the 4 dimensions of a 2 qubit system would be labeled |00>,
|01>, |10>, |11>. A NOT operation flipping the first bit would actually be
a 90 degree rotation in the |00>, |10> plane from (1,0,0,0) to (0,0,1,0).
You can have partial rotations like a 45 degree Hadamard gate resulting in
(.707,0,.707,0), which would read as a coin flip on the first qubit.

What the authors did was define quantum algorithmic complexity of a bit
string as the minimum number of quantum operations needed to compute the
string. Unlike ordinary algorithmic complexity, this is computable because
quantum circuits, a subset of Boolean circuits, always halt. My
interpretation of what they found empirically is that a random bit string
has an average complexity of about 6 2-input gates or 4 3-input gates drawn
from a minimum universal set of gates.

A gate set is universal if you can construct any function from it. For
example, in Boolean logic, 2 input NAND is universal. In quantum computing,
the 2 input set {NOT, CNOT} and 3 input set {CCX} (or Toffoli) gate are
both universal. A CNOT gate flips one qubit if the other is 1. A CCX gate
flips one qubit if the other two are both 1.

Their result implies that quantum complexity is not useful for induction.
The Kolmogorov complexity of a string of n identical bits is log n plus a
language dependent constant. But the Boolean and quantum complexities are
both O(n), one gate for each output bit. The quantum complexity of a random
string increases from n to O(n log n) since it takes log n bits to specify
each of the ~12n gate inputs. Occam's Razor depends on input being
compressible to predict future inputs. Boolean logic, and by implication,
quantum computing, cannot compress.

The result is helpful by giving one more reason why quantum computing is
not the path to AGI. Quantum computing is often misunderstood as computing
an exponential set of paths in parallel, when for most purposes it is
actually weaker than classical computing.  Leveraging the power of quantum
computing depends on arranging the computation so that the components of
the unit vector reinforce for a subset of the qubits, like in computing the
quantum Fourier transform for Shor's or Grover's algorithms for cracking
crypto. But it doesn't solve NP complete problems in polynomial time or
anything like that. The brain is not a quantum computer. Neural networks,
and learning in general, are not time reversible.


------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/Tb2574dcac5560d73-M638314d8d4ae2d29c924c5dc
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to