Doras F516 ‘W PENNYLANE ux / Understanding The Haar Measure Understanding the Ha e Olivia Di Matteo Published March 22, 2021. Last updated January 1, 2024, If you've ever dug into the literature about random quantum ¢ anything related to the structure and properties of unitary op statement like the following: “Assume that U is sampled uni measure’. In this demo, we're going to unravel this cryptic st, it means. You'll gain an understanding of the general concep special properties, and you'll learn how to sample from it usi scientific computing frameworks. By the end of this demo, statement in your own work with confidence! Note To get the most out of this demo, itis helpful if you are far dimensional functions, the Bloch sphere, and the concept factorizations of unitary matrices (see, e.g,, 4.5.1 and 4.5 Measure Measure theory is a branch of mathematics that studies thin area, or volume, but generalized to mathematical spaces an¢ measure tells you about how “stuff” is distributed and conce An intuitive way to understand the measure is to think about can be parametrized by three numbers—depending on what} coordinates (z, y, z), or it may be more convenient to use s} Suppose you wanted to compute the volume of a solid sphel integrating over the three coordinates p, ¢, and 8. Your first each parameter over its full range, like so: ve[ [" [avase htipsifpennylane aifgniidemosttutoral_hear_measure! Understanding the Haar measure | PennyLane Demos Contents Measure The Haar measure v Fun (and not-so-fun) fact Conclusion References About the author Downloads @ Download Python script F Download Notebook © veewon cit FS copy Bex citation 190Doras F516 Understanding the Haar measure | PennyLane Demos ‘W PENNYLANE ‘with respect to the parameters. For example, consider twos} the same difference in @ and ¢, but at different values of 0: Even though the differences d@ and d¢ themselves are the s equator of the sphere than there is near the poles. We must computing the integral! Specifically, we multiply by the functi function mean that the most weight will occur around the eq weight near the poles where @ = 0 and @ =m. Similar care must be taken for p. The contribution to volume more than for a small p—we should expect the contribution surface area of a sphere of radius r is 4r?. On the other hand, for a fixed p and @, the length of the dd is these facts together, we find that the actual expression for th rete ae v-[ [ [ p'sind dpa 0 Jo Io These extra terms that we had to add to the integral, p? sing weights portions of the sphere differently depending on whe know the measure to properly integrate over the sphere, kno) means to perform another important task, that of sampling We can't simply sample each parameter from the uniform dit experienced already, this doesn't take into account how the s measure describes the distribution of each parameter and gi obtain something properly uniform, The Haar measure htipsifpennylane.afgrilidemosttutoral_hear_measure! Contents Measure Conclusion References About the author Downloads Download Python seript & Download Notebook © veewon cit 220Doras F516 Understanding the Haar measure | PennyLane Demos ‘W PENNYLANE yenerar on parameters like so, yeryuun wucau eH cos(9/2) eN/2 sin(9/2) U(9, 8,0) ( For every dimension N, the unitary matrices of size N x N can perform operations on elements of this group, such as a them, or sample uniformly over them, just as we can do to p¢ with respect to the sphere, we have to add the measure in or regions of space. The Haar measure provides the analogous unitary group. For an N-dimensional system, the Haar measure, often denc elements of U(IN). For example, suppose f is a function the would like to take its integral over the group. We must write t measure, like so: Fron £V tant evn) As with the measure term of the sphere, dz itself can be b| on individual parameters. While the Haar measure can be de mathematical form gets quite hairy for larger dimensions—ir requires at least N? — 1 parameters, which is a lot to keep t of a single qubit (N’ = 2), then show how things generalize. Single-qubit Haar measure The single-qubit case provides a particularly nice entry point comparison to spheres by visualizing single-qubit states on measure provides a recipe for sampling elements of the unit given the structure of the group. One useful consequence of sample quantum states uniformly at random—we simply ger them to a fixed basis state such as |). We'll see how this works in good time. First, we'll take a look measure and do things wrong. Suppose we sample quantum the parametrization above, but sample the angles w, ¢, and between [0, 27) (fun fact: there is a measure implicit in this constant value, because each point is equally likely to be sar htipsifpennylane aifgniidemosttutoral_hear_measure! Contents Measure Conclusion References About the author Downloads @ Download Python script F Download Notebook © veewon cit 3720Doras F516 Understanding the Haar measure | PennyLane Demos ‘W PENNYLANE import numpy as np import matplotlib.pyplot as plt # set the random seed np.random.seed(42) # Use the mixed state simulator to save some step! dev = qnl.device('default.mixed', wires=1) @qnl .qnode (dev) def not_a_haar_random_unitary(): # Sample all parameters from their flat unifol phi, theta, omega = 2 * np.pi * np.random.uni qnl.Rot(phi, theta, omega, wires=0) return qnl.state() num_samples = 2021 not_haar_samples = [not_a_haar_random_unitary() fi In order to plot these on the Bloch sphere, we'll need to do of states into Bloch vectors. X = np.array([[@, 1], [1, @]]) Y = np.array([{[@, -19], [13, @]]) Z = np.array({(1, @], [@, -1]]) # Used the mixed state simulator so we could have def convert_to_bloch_vector(rho): “""convert a density matrix to a Bloch vector ax = np.trace(np.dot (rho, X)).real ay = np.trace(np.dot (rho, Y)).real np.trace(np.dot(rho, Z)).real return [ax, ay, az] not_haar_bloch_vectors = np.array([convert_to_blo} With this done, let's find out where our “uniformly random" st htipsifpennylane aifgniidemosttutoral_hear_measure! Contents Measure Conclusion References About the author Downloads @ Download Python script F Download Notebook © veewon cit 4720Doras F516 ‘W PENNYLANE " Helper function to plot vectors on a sphel fig = plt.figure(figsize=(6, 6)) ax = fig-add_subplot(111, projection="3d") fig.subplots_adjust(left=0, right=1, bottom=9 ax.grid(False) ax.set_axis_of f() ax.view_init(30, 45) ax.dist = 7 # Draw the axes (source: https://github.com/m X, Y, 2 = np.array([[-1.5,0,0], [@,-1-5,0], [ U,V, W= mp.array( [13,20], [2,30], [9,83] ax.quiver(x, y, 2) Us Vv, Wy arrow_length_ratiy ax.text(2, ®, 1.7, r"|@)", color="black", font ax.text(®, ®, -1.9, r"[1)", color="black", for ax.text(1.9, @, 8, r"[+)", colors"black", font ax.text(-1.7, 0, 0, r”|-)", color="black", for ax.text(®, 1.7, 0, r"[i#)", color="black", for ax.text(0,-1.9, 8, r"|i-)", color="black", for ax.scatter( bloch_vectors[:,0], bloch_vectors[:,1], 6 plot_bloch_sphere(not_haar_bloch_vectors) htipsifpennylane aifgniidemosttutoral_hear_measure! Understanding the Haar measure | PennyLane Demos Contents Measure Conclusion References About the author Downloads Download Python seript & Download Notebook © veewon cit 520Doras F516 Understanding the Haar measure | PennyLane Demos ‘W PENNYLANE {oy Out: /one/runner /work/qml/qml/demonstrations/t ax.dist = 7 You can see from this plot that even though our parameters) there is a noticeable amount of clustering around the poles ¢ parameters being uniform, the output is very much not unifo) measure is larger near the equator, and if we just sample uni area as much. To take that into account we will need to sam} weight the different parameters appropriately. For a single qubit, the Haar measure looks much like the cas component. Intuitively, all qubit state vectors have length 1, ¢ role here. The parameter that we will have to weight differen! measure is identical to that we had to do with the polar axis) sample the @ uniformly at random in this context, we must si Pr(6) = sin @, We can accomplish this by setting up a cust rv_continuous in scipy. htipsifpennylane aifgniidemosttutoral_hear_measure! Contents Measure Conclusion References About the author Downloads Download Python seript & Download Notebook © veewon cit 620Doras F516 Understanding the Haar measure | PennyLane Demos ‘W PENNYLANE class sin_prob_dist(rv_continuous) Contents def _pdf(self, theta # The 0.5 is so that the distribution is) return 0.5 * np.sin(theta) Measure # Samples of theta should be drawn from between @ sin_sampler = sin_prob_dist(a=0, b=np.pi) @qml . qnode(dev) Conclusion def haar_random_unitary(): phi, omega = 2 * np.pi * np.random.uniform(si; petorences ‘theta = sin_sampler.rvs(siz ) # Sample thet, qnl.Rot(phi, theta, omega, wires=0) About the author return qnl.state() haar_samples = [haar_random_unitary() for _ in ra) DOwnloads haar_bloch_vectors = np.array([eonvert_t0_81°ch My psntoas python sept plot_bloch_sphere(haar_bloch_vectors) + Download Notebook © veewon cit htipsifpennylane aifgniidemosttutoral_hear_measure! 720Doras F516 Understanding the Haar measure | PennyLane Demos ‘W PENNYLANE Out: /home/runner/work/qml/qm1/demonstrations/1 ax.dist = 7 a 7 We see that when we use the correct measure, our qubit stat the sphere. Putting this information together, we can now wr Haar measure: dyrz = sin 848 - doo Show me more math! While we can easily visualize the single-qubit case, this is number of qubits. Regardless, we can still obtain a mathe arbitrary dimensions. in the previous section, we expressed 1 htipsifpennylane aifgniidemosttutoral_hear_measure! Contents Measure Conclusion References About the author Downloads ® Download Python script & Download Notebook ©) View on GitHubDoras F516 Understanding the Haar measure | PennyLane Demos ‘W PENNYLANE HU @ Set UF par Many of these parametrization come to us from the study ¢ broken down into elementary operations involving only a few parameters of the physical apparatus used to implement the Rather than qubits, such operations act on modes, or qumog the N-dimensional special unitary group. This group, written consisting of all N’ x N unitary operations with determinan potential global phase) Note Elements of SU(N) and U(N) can still be considered as where NV is a power of 2, but they must be translated from qubit operations. (In PennyLane, this can be done by feedi operation directly. Alternatively, one can use quantum con sequence of elementary gates such as Pauli rotations and Tip If you haven't had many opportunities to work in terms of | documentation is a good starting point. For example, we saw already above that for V = 2, we can} e092 co5(6/2) = U(4,8,0) = (co sin(0/2) ‘This unitary can be factorized as follows: U(4, 0,0) = ( 0 ) (ee 2 oe?) \sin(o/2) 5 The middle operation is a beamsplitter; the other two operat! for N = 2, dtp = sin @d0dwdd—note how the parameter i measure in a different way than those of the phase shifts. As there are multiple ways to decompose the unitary. Such dece acting on N modes as a sequence of operations acting only phase shifts. Shown below are three examples [2], [3], [4]: htipsifpennylane aifgniidemosttutoral_hear_measure! Contents Measure Conclusion References About the author Downloads @ Download Python script F Download Notebook © veewon cit 920Doras F516 Understanding the Haar measure | PennyLane Demos ‘W PENNYLANE In these graphics, every wire is a different mode. Every box rt modes, and the number in the box indicates the number of p simply phase shifts on individual modes. The blocks contain parameters, such as the U(¢, 0, w) above. Those containing modes with 2 parameters, similar to the 3-parameter ones bt Although the decompositions all produce the same set of op parametrization may have consequences in practice. The firs that leads to a recursive definition of the Haar measure. The such that an SU(N) operation can be implemented by san¢ between two SU(N — 1) transformations, like so: The Haar measure is then constructed recursively as a prod the parameters in the first SU(N — 1) transformation; the ¢ lone SU(2) transformation; and the third term depends ont transformation. SU (2) is the "base case” of the recursion—we simply havet htipsifpennylane aifgniidemosttutoral_hear_measure! Contents Measure Conclusion References About the author Downloads @ Download Python script F Download Notebook © veewon cit 10720Doras F516 Understanding the Haar measure | PennyLane Demos ‘W PENNYLANE dy = sinOd9 did Contents Measure Moving on up, we can write elements of SU(3) as a sequen Haar measure dis then consists of two copies of dy12, with account the middle transformation. Conclusion = References Su(2) 18, 6 About the author Og = SvnQ, d8,du4 dg, Sik8 Downloads @ Download Python script For SU(4) and upwards, the form changes slightly, but still! % — Download Notebook dyiy—1 with a term in between. © Viewon ittiub 2] d poy For larger systems, however, the recursive composition allow transformations on the lower modes to be grouped. We can some of the parameters: htipsifpennylane aifgniidemosttutoral_hear_measure! 11202o2aieint F518 Understanding the Haar measure | PenryLane Demos ‘W PENNYLANE soca) jena suey om tel This leads to one copy of djzy-1, which well denote as dyih, set of terms (as detailed in [2], this is called a coset measure Putting everything together, we have that Oy duy = dply_, x sinOy, sin") ( =a The middle portion depends on the value of IV, and the paral (N — 1)'th SU(N) transformation. This is thus a convenie} dimensional Haar measure for the unitary group. As a finaln parametrized in different ways, the underlying Haar measure being an invariant measure, as will be shown later. Haar-random matrices from the QF Nice-looking math aside, sometimes you just need to genera Haar-random matrices. It would be very cumbersome to samt so many parameters; furthermore, the measure above requir fixed way. There is a much quicker way to perform the samp} decomposition of complex-valued matrices. This algorithm i following steps: hitpsipennylane allgmi/demostutorial haar_measure! Contents Measure Conclusion References About the author Downloads @ Download Python script F Download Notebook ©) View on GitHub 2120Doras F516 Understanding the Haar measure | PennyLane Demos W PENNYLANE ornorererreonovsy . 2. Compute a QR decomposition Z = QR. 3. Compute the diagonal matrix A = diag(Rii/|Rii|)- 4, Compute Q' = QA, which will be Haar-random. from numpy.linalg import qr def qr_haar(N): "Generate a Haar-random matrix using the QR # Step 1 A, B = np.random.normal(size=(N, N)), np.randh Z=Asaj*B # Step 2 QR = ar(z) # Step 3 Lambda np.diag([R[i, i] / np.abs(R[i, i]) fi # Step 4 return np.dot(Q, Lambda) Let's check that this method actually generates Haar-randon plotting on the Bloch sphere @qm1. qnode (dev) def qr_haar_random_unitary(): qml.Qubitunitary(qr_haar(2), wires=2) return qnl.state() qr_haar_samples = [qr_haar_random_unitary() for _ gr_haar_bloch_vectors = np.array([convert_to_bloc| plot_bloch_sphere(qr_haar_bloch_vectors) ‘ a htipsifpennylane aifgniidemosttutoral_hear_measure! Contents Measure Conclusion References About the author Downloads @ Download Python script F Download Notebook © veewon cit 13720Doras F516 Understanding the Haar measure | PennyLane Demos ‘W PENNYLANE {oy Out: /one/runner /work/qml/qml/demonstrations/t ax.dist = 7 As expected, we find our qubit states are distributed uniform is what's implemented in packages like scipy’s unitary_grou Now, it's clear that this method works, but itis also importan fairly straightforward—the base of our samples is a matrix fu typical distribution. This isn’t enough by itself, since unitary r and columns must be orthonormal. These constraints are wi generic QR decomposition consists of an orthonormal matri Since our original matrix was complex-valued, we end up wit why not stop there? Why do we then perform steps 3 and 4? Steps 3 and 4 are needed because, while the QR decomposit is properly Haar-random. In [5], it is explained that a uniform also yield a uniform distribution over the eigenvalues of thos htipsifpennylane aifgniidemosttutoral_hear_measure! Contents Measure Conclusion References About the author Downloads Download Python seript & Download Notebook © veewon cit 14720Doras F516 Understanding the Haar measure | PennyLane Demos ‘W PENNYLANE OM qUET ye Cann eaNe en UNIAN CEgOTErTT QR = QAMR = Q'R’ Step 3 removes this redundancy k to a unique value of Q’ = QA, and a uniform distribution of Try it! Use the qr_haar function above to generate random unita eigenvalues. Then, comment out the lines for steps 3 and distribution is no longer uniform. Check out reference [5] f Fun (and not-so-fun) facts We've now learned what the Haar measure is, and both an ar sampling quantum states and unitary operations uniformly a many neat properties that play a role in quantum computing, Invariance of the Haar measure Earlier, we showed that the Haar measure is used when inte¢ Fon £Vtent VeU(N) One of the defining features of the Haar measure is that itis, transformations. That is, Fup LOW MAY = HWY, Vevtn) Veu(N) This holds true for anyother N x N unitary W! A conseque Haar-random, then so is V7, V1, and any product of anothe may be taken on either side) Another consequence of this invariance has to do with the s} must all come from the same distribution. This is because tt permutations, since permutations are unitary— the whole thi how the entries are ordered, so all distributions must be the! numbers a + bi where both a and b has mean 0 and variane ensemble we used in the QR decomposition above, but with to orthonormality), htipsifpennylane aifgniidemosttutoral_hear_measure! Contents Measure Conclusion References About the author Downloads @ Download Python script F Download Notebook © veewon cit 1520Doras F516 Understanding the Haar measure | PennyLane Demos ‘W PENNYLANE ‘An unfortunate (although interesting) property of the Haar fl phenomenon of concentration of measure. Most of the “stuf certain area, and this gets worse as the size of the system in by looking at the sphere. For the 3-dimensional sphere, we si around the equator, and how the measure takes that into acc This property becomes increasingly prominent for higher-din Important The concentration described here is not referring to what) sampled quantum states (points on the Bloch sphere) inet around the poles. However, that is not unrelated. Concent measure itself is concentrated, and which parts of the spa For the case of the sphere, itis the equatorial area, and wk ‘that concentration into account, we obtained an uneven di Let's consider an V-dimensional unit sphere. Points on thes parametrized by N — 1 real coordinates. Suppose we have + sphere to real numbers. Sample a point 2 on that sphere frot value of f(«). How close do you think the result will be to th the entire sphere? A result called Levy's lemma [7], [8] expresses how likely itis from the mean. It states that, for an x selected uniformly at! from Ef] by some amount ¢ is bounded by: Pr(|f(z) — Elf]| > €) < 2ex A constraint on the function f is that it must be Lipschitz co! constant of the function. The important aspect here is the lik mean by an amount decreases exponentially with c. Furthe makes the deviation exponentially less likely. Now, this result seems unrelated to quantum states—it cone, However, recall that a quantum state vector is a complex vec similar to vectors on a sphere. If you “unroll" a quantum stati stacking its real and complex parts, you end with a vector of like a unit vector on the sphere in this dimension. Given that quantum state vectors can be converted to vectors on spher will also demonstrate concentration. This is bad news! To do useful things in quantum computing qubits we have, the more our randomly sampled states will will concentrate around the maximally entangled state [8]). T htipsifpennylane aifgniidemosttutoral_hear_measure! Contents Measure Conclusion References About the author Downloads @ + 90 Download Python script Download Notebook View on GitHub 1620Doras F516 Understanding the Haar measure | PennyLane Demos ‘W PENNYLANE Haar measure and barren plateaus Suppose you are venturing out to solve a new problem using quantum eigensolver. A critical component of such methods Having now learned a bit about the properties of the Haar mi sense to use this for the parametrization. Variational ansaet: circuits, so why not choose an ansatz that corresponds direc unitaries? The initial parameter selection will give you a state Then, since this ansatz spans the entire Hilbert space, you're target ground state with your ansatz, and it should be able te Unfortunately, while such an ansatz is extremely expressive) possible state), these ansaetze actually suffer the most from Barren plateaus are regions in the cost landscape of a paran and its variance approach 0, leading the optimizer to get stu recently in the work of [10], wherein closeness to the Haar m. expressivity. The closer things are to the Haar measure, the more prone to exhibiting barren plateaus. 8) Unitary Group Lesion © ‘Accessible space U Image source: [10]. A highly expressive ansatz that can} unitaries (i.e,, an ansatz capable of producing unitaries manner) is very likely to have flat cost landscapes and suf Itturns out that the types of ansaetze know as hardware-effl problem if they are “random enough’ (this notion will be form [9] that this is a consequence of the concentration of measu values of gradients and variances can be computed for clas¢ with respect to the Haar measure, and it is shown that these number of qubits, and thus huge swaths of the cost landsca| htipsifpennylane aifgniidemosttutoral_hear_measure! Contents Measure Conclusion References About the author Downloads Download Python seript & Download Notebook © veewon cit 17120Doras F516 Understanding the Haar measure | PennyLane Demos ‘W PENNYLANE The Haar measure plays an important role in quantum comp with sampling random circuits, or averaging over all possible with respect to the Haar measure. There are two important aspects of this that we have yet tot is efficient to sample from the Haar measure—given that the exponential in the number of qubits, certainly not. But a mort always sample from the full Haar measure? The answer to th Depending on the task at hand, you may be able to take a shi design. In an upcoming demo, we will explore the amazing w applications! References [1] M.A. Nielsen, and I. L. Chuang (2000) “Quantum Compu Cambridge University Press. 2(1,2,3) H. de Guise, 0. Di Matteo, and L. L. Sanchez-Soto. (2 transformations", Phys. Rev. A 97 022328. (arXiv) [3] W.R. Clements, P. C. Humphreys, B. J. Metcalf, W. S. Kol “Optimal design for universal multiport interferometers" [4] M. Reck, A. Zeilinger, H. J. Bernstein, and P. Bertani (199 discrete unitary operator”, Phys. Rev. Lett.73, 58-61 5(1,2,3) F. Mezzadri (2006) “How to generate random matric (arXiv) [6] E.Meckes (2019) "The Random Matrix Theory of the Cle University Press. [7] M. Gerken (2013) “Measure concentration: Levy's Lemm 8(1,2) P. Hayden, D. W. Leung, and A. Winter (2006) “Aspects Phys. Vol. 265, No. 1, pp. 95-117. (arXiv) 9(1,2) J.R. McClean, S. Boixo, V. N. Smelyanskiy, R. Babbust quantum neural network training landscapes”, Nature 10(1,2,3) Z. Holmes, K. Sharma, M. Cerezo, and P. J. Coles (} gradient magnitudes and barren plateaus". (arXiv) About the author htipsifpennylane aifgniidemosttutoral_hear_measure! Contents Measure Conclusion References About the author Downloads @ Download Python script F Download Notebook © veewon cit 18202o2aieint F518 Understanding the Haar measure | PenryLane Demos ‘W PENNYLANE Olivia Di Matteo Olivia is an Assistant Professor in the Department of Electri¢ University of British Columbia. Her research interests are qui algorithms, Total running Ask a question on the f Share demo v Related Dem Unitary designs Que Vv hitpsipennylane allgmi/demostutorial haar_measure! Contents Measure Conclusion References About the author Downloads @ Download Python script F Download Notebook ©) View on GitHub 19120Doras F516 ‘W PENNYLANE For researchers Features. Demos Datasets Performance Learn Videos Documentation Education yQOQ ino @ # ‘Stay updated with our newsletter © Copyright 2024 | Xanadu | All rights reserved Understanding the Haar measure | PennyLane Demos For learners Contents Learn Codebook Education Videos Challenges Demos Measure Glossary Conclusion References About the author Downloads @ Download Python script TensorFlow, the TensorFlow logo and any related marks are tradem, Privacy Policy | TermsofService | htipsifpennylane aifgniidemosttutoral_hear_measure! CookiePolicy | ¥ Download Notebook © veewon cit 20720