In the last post, we have looked at the Deutsch-Jozsa algorithm that is considered to be the first example of a quantum algorithm that is structurally more efficient than any classical algorithm can probably be. However, the problem solved by the algorithm is rather special. This does, of course, raise the question whether a similar speed-up can be achieved for problems that are more relevant to practical applications.
In this post, we will discuss an algorithm of this type – Grover’s algorithm. Even though the speed-up provided by this algorithm is rather limited (which is of a certain theoretical interest in its own right), the algorithm is interesting due to its very general nature. Roughly speaking, the algorithm is concerned with an unstructured search. We are given a set of N = 2n elements, labeled by the numbers 0 to 2n-1, exactly one of which having a property denoted by P. We can model this property as a binary valued function P on the set that is zero on all but one elements. The task is to locate the element x0 for which P(x0) is true.
Grover’s algorithm
Grover’s algorithm presented in [1] proceeds as follows to locate this element. First, we again apply the Hadamard-Walsh operator W to the state of an n-qubit system to obtain a superposition of all basis states. Then, we iteratively apply the following sequence of operations.
- Apply a conditional phase shift S, i.e. apply the unique unitary transformation that maps
to
.
- Apply the unitary transformation D called diffusion that we will describe below
Finally, after a defined number of outcomes, we perform a measurement which will collaps the system into one of the states . We claim – and will see why this is true below – that for the right number of iterations, this value x will, with a high likelihood, be the solution to our problem, i.e. equal to x0.
Before we can proceed, we need to define the matrix D. This matrix is along the diagonal, with N = 2n, and
away from the diagonal. In terms of basis vectors, the mapping is given by
Consequently, we see that
where is the average across the amplitudes
. Thus geometrically, the operation D performs an inversion around the average. Grover shows that this operation can be written as minus a Hadamard-Walsh operation followed by the operation that flips the sign for
, followed by a second Hadamard-Walsh transformation.
For the sake of completeness, let us also briefly discuss the first transformation employed by the algorithm, the conditional phase shift. We have already seen a similar transformation while studying the Deutsch-Jozsa algorithm. In fact, we have shown in the respective blog post that the circuit displayed below (with the notation slightly changed)
performs the required operation
Let us now see how why Grover’s algorithm works. Instead of going through the careful analysis in [1], we will use bar charts to visualize the quantum states (exploiting that all involved matrices are actually real valued).
It is not difficult to simulate the transformation in a simple Python notebook, at least for small values of N. This script performs several iterations of the algorithm and prints the result. The diagrams below show the outcome of this test.
Let us go through the diagrams one by one. The first diagram shows the initial state of the algorithm. I have used 3 qubits, i.e. n = 3 and N = 8. The initial state, after applying the Hadamard-Walsh transform to the zero state, is displayed in the first line. As expected, all amplitudes are equal to 1 over the square root of eight, which is approximately 0.35, i.e. we have a balanced superposition of all states.
We now apply one iteration of the algorithm. First, we apply the conditional phase flip. The element we are looking for is in this case located at x = 2. Thus, the phase flip will leave all basis vectors unchanged except for and it will change the amplitude of this vector to – 0.35. This will change the average amplitude to a value slightly below 0.35. If we now perform the inversion around the average, the amplitudes of all basis vectors different from
will actually decrease, whereas the amplitude of
will increase. The result is displayed in the second line of the diagram.
Thus, what really happens in this case is an amplitude amplification – we increase the amplitude of one component of the superposition while decreasing all the others.
The next few lines show the result of repeating these two steps. We see that after the second iteration, almost all of the amplitude is concentrated on the vector , which represents the solution we are looking for. If we now perform a measurement, the result will be 2 with a very high probability!
It is interesting to see that when we perform one more iteration, the difference between the amplitude of the solution and the amplitudes of all other components decreases again. Thus the correct choice for the number of iterations is critical to make the algorithm work. In the last line, we have plotted the difference between the amplitude of and the other amplitudes (more precisely, the ratio between the amplitude of
and the second largest amplitude) on the y-axis for the different number of iterations on the x-axis. We see that the optimal number of iterations is significantly below 10 (actually five iterations give the best result in this case), and more iterations decrease the likelihood of getting the correct result out of the measurement again. In fact, a careful analysis carried out in [2] shows that for large values of N, the best number of iterations is given by
, and that doubling the number of iterations does in general lead to a less optimal result.
Generalizations and amplitude amplification
In a later paper ([3]), Grover describes a more general setup which is helpful to understand the basic reason why the algorithm works – the amplitude amplification. In this paper, Grover argues that given any unitary transformation U and a target state (in our case, the state representing the solution to the search problem), the probability to meet the target state by applying U to a given initial state can be amplified by a sequence of operations very much to the one considered above. We will not go into details, but present a graphical representation of the algorithm.
So suppose that we are given an n-qubit quantum system and two basis vectors – the vector t representing the target state and an initial state s. In addition, assume we are given a unitary transformation U. The goal is to reach t from s by subsequently applying U itself and a small number of additional gates.
Grover considers the two-dimensional subspace spanned by the vectors s and U-1t. If, within this subspace, we ever manage to reach U-1t, then of course one more application of U will move us into the desired target state.
Now let us consider the transformation
where Ix denotes a conditional phase shift that flips the phase on the vector . Grover shows that this transformation does in fact leave our two-dimensional subspace invariant, as indicated in the diagram below.
He then proceeds to show that for sufficiently small values of the matrix element Uts, the action of Q on this subspace can approximately be described as a rotation by the angle . Applying the operation Q n times will then approximately result in a superposition of the form
Thus if we can make the first coefficient very small, i.e. if is close to a multiple of
, then one application of U will take our state to a state very close to t.
Let us link this description to the version of the Grover algorithm discussed above. In this version, the initial state s is the state . The transformation U is the Hadamard-Walsh transformation W. The target state is
where x0 is the solution to the search problem. Thus the operation It is the conditional phase shift that we have denoted by S earlier. In addition, Grover shows in [1] already that the diffusion operator D can be expressed as -W I0 W. Now suppose we apply the transformation Q n times to the initial state and then apply U = W once more. Then our state will be
This can be written as
Regrouping this and using the relation D = -W I0 W, we see that this is the same as
Thus the algorithm can equally well be described as applying W once to obtain a balanced superposition and then applying the sequence DS n times, which is the formulation of the algorithm used above. As in this case, we also recover the result that the optimal number of iterations is
for large N.
Applications
Grover’s algorithm is highly relevant for theoretical reasons – it applies to a very generic problem and (see the discussions in [1] and [2]) is optimal, in the sense that it provides a quadratic speedup compared to the best classical algorithm that requires O(N) operations, and that this cannot be improved further. Thus Grover’s algorithm provides an interesting example for a problem where a quantum algorithm delivers a significant speedup, but no exponential speedup as we will see it later for Shor’s algorithm.
However, as discussed in detail in section 9.6 of [4], the relevance of the algorithm for practical applications is limited. First, the algorithm applies to an unstructured search, i.e. a search over unstructured data. In most practical applications, we deal with databases that have some sort of structure, and then more efficient search algorithms are known. Second, we have seen that the algorithm requires applications of the transformation UP. Whether this is better than the classical algorithm does of course depend on the efficiency with which we can implement this with quantum gates. If applying UP requires O(N) operations, the advantage of the algorithm is lost. Thus the algorithm only provides a significant speedup if the operation UP can be implemented efficiently and there is no additional structure that a classical algorithm could exploit.
Examples of such problems are brute forces searches as they appear in some crypto-systems. Suppose for instance we are trying to break a message that is encrypted with a symmetric key K, and suppose that we know the first few characters of the original text. We could then try to use an unstructured search over the space of all keys to find a key which matches at least the few characters that we know.
In [5], a more detailed analysis of the complexity in terms of qubits and gates that a quantum computer would have to attack AES-256 is made, arriving at a size of a few thousand logical quantum bits. Given the current ambition level, this does not appear to be completely out of reach. It does, however, not render AES completely unsecure. In fact, as Grover’s algorithm essentially results in a quadratic speedup, a code with a key length of n bits in a pre-quantum world is essentially as secure as the same code with a key length of 2n in a post-quantum world, i.e. roughly speaking, doubling the key length compensates the advantage of quantum computing in this case. This is the reason why the NIST report on post-quantum cryptography still classifies AES as inherently secure assuming increased key sizes.
In addition, the feasibility of a quantum algorithm is not only determined by the number of qubits required, but also by other factors like the depth, i.e. the number of operations required, and the number of quantum gates – and for AES, the estimates in [5] are significant, for instance a depth of more than 2145 for AES-256, which is roughly 1043. Even if we assume a switching time of only 10-12 seconds, we still would require astronomical 1031 seconds, i.e. in the order of 1023 years, to run the algorithm.
Even a much less sophisticated analysis nicely demonstrates the problem behind these numbers – the number of iterations required. Suppose we are dealing with a key length of n bits. Then we know that the algorithm requires
iterations. Taking the decimal logarithm, we see that this is in the order of 100.15*n. Thus, for n = 256, we need in the order of 1038 iterations – a number that makes it obvious that AES-256 can still be considered secure for all practical purposes.
So overall, there is no reason to be overly concerned about serious attacks to AES with sufficiently large keys in the near future. For asymmetric keys, however, we will soon see that the situation is completely different – algorithms like RSA or Elliptic curve cryptography are once and for all broken as soon as large-scale usable quantum computer become reality. This is a consequence of Shor’s algorithm that we will study soon. But first, we need some more preliminaries that we will discuss in the next post, namely quantum Fourier transforms.
References
1. L.K. Grover, A fast quantum mechanical algorithm for database search, Proceedings, 28th Annual ACM Symposium on the Theory of Computing (STOC), May 1996, pages 212-219, available as arXiv:quant-ph/9605043v3
[2] M. Boyer, G. Brassard, P. Høyer, A. Tapp, Tight bounds on quantum searching, arXiv:quant-ph/9605034
3. L.K. Grover, A framework for fast quantum mechanical algorithms, arXiv:quant-ph/9711043
4. E. Rieffel, W. Polak, Quantum computing – a gentle introduction, MIT Press
5. M. Grassl, B. Langenberg, M. Roetteler, R. Steinwandt, Applying Grover’s algorithm to AES: quantum resource estimates, arXiv:1512.04965