Ul ak hing Desi hie rs wi imipitdKne Pn Ree 9, (106108, ‘Melko, Machine learning phases of matter. Nature Physics, 19(6)491, [10| Peter Broecker, Juan Carrasquilla, Roger G Melko, and Simon ‘Trebst. Machine learning quantum phases of matter beyond the fermion sign problem. Scientific reports, 7(1):8823, 2017. [11] Kelvin Ci’ng, Juan Carrasquilla, Roger G Melko, and Ehsan Khatami. Machine leaning phases of strougly correlated fermions. Physical Review X, 7(3)-031038, 2017. [12] Evert PL Van Nieawonburg, YeHua Liu, and Sebastian D Huber. Learning phase transitions by conf sion. Nature Physics, 13(5):425, 2017. [13] Lonis-Frangois Arsenanit, Alejandro Lopez-Bezanilla, © Anatole von Lilienfeld, and Andrew J Millis. Machine leaming for many-body physics: the case of the anderson impurity model. Physical Review B, 90(15):155136, 2014, [14] Aaron Gilad Kusne, Tieren Gao, Apurva Mehta, Ligin Ke, Manh Cuong Nguyen, Kal-Ming Ho, Vindinuir Antropov, Cal-Zhuang Wang, Matthew J Kramer, Christian Long, et al. Ou-the-fly machine-learning for high-throughput experimeats: search for rare-carth-free permanent magnets. Sciculific reports, 4.6967, 2014. (15) Phil De Luna, Jennifer Wei, Yoshua Bengio, Alin Aspuni-Guzik, and Kaward Sargent. Use machine learning to find energy materials. Nature, 552(7683):23-25, 2017, [16] Jennifer N Wei, David Duvenaud, and Alin Aspuru-Gi chemistry reactiors. ACS central science, 2(10):725-7: k. Neural networks for the prediction of organic , 2016. [17] Jacob Bismonto, Peter Wittek, Nicola Pancotti, Patrick Rebentrost, Nathan Quantum machine learning. Nature, 549(7671):195, 2017, be, and Seth Lloyd. [18] Seth Lloyd, Masoud Mohseni, and Patrick Rebentrost. Quantum algorithms for supervised and unsu- pervised machine learning. 2013 Preprint at https: //arxiv.org/abs/1307.0411v2 [19] Patrick Rebentrost, Masoud Mohsent, and Seth Lloyd. Quant classification. Physical review letters, 113(18):190503, 2014. a Support vector machine for big data [20] Hartmnt Neven, Geordie Rose, and William G Macready. Image recognition with an adiabatic quantum, computer i, mapping to quadratic unconstrained binary optimization. 2008. Preprint at https: //arxiv org/abs/0804.4457, (21) Hartmut Neven, Vasil $ Denchey, Geordie Rose, and William @ Macready. ‘Training a binary classifier ‘with the quantum adiabatic algorithm. 2008. Preprint at https: //arxiv.org/abs/0811 0416. [22] Hartmut Neven, Vasil § Deachev, Geordie Rose, and William G Macready. Training a large seale classifier ‘with the quantum adiabatic algorithm. 2009. Preprint at https: //arxiv.org/abs/0912..0779. [23] Nathan Wiebe, Ashish Kapoor, and Krysta M Svore. Quantum deep learning. Quantum Information 6 Computation, 16(7-8):541-587, 2016. [24] Giuseppe Carleo and Matthias ‘Troyer. Solving the quantum many-body problem with artific networks. Selence, 355(6325):602 606, 2017. neural [25] Gincomo Torlai, Gugliclmo Mazola, Juan Carrasquilla, Matthias Troyer, Roger Melko, and Giuseppe Carleo, Nenral-network quantum state tomography. Nature Physics, 14(5):447, 2018, [26] Sabre Kais. Introduction to quantum information and computation for chemistry. Quantum Injormation ‘and Computation for Chemistry, pages 1-38, 2014.arrangement of gates (a fermionic swap network) it is possible to simulate electronic structure Hamiltonian with linear depth and connectivity. ‘These results present significant improvement on the cast of quantum simulation for both variational and phase estimation based quantum chemistry simulation metbods. Recently, ‘Troyer and coworkers proposed using a restricted Boltzmann machine to solve quantum many-body problems, for both stationary states and time evolution of the quantum Ising and Heisenberg models [24] However, this simple approach haz to be modified for cases where the wave function's phase is required for accurate caleniations [25] Herein, we propose a three-layered RISM structure that includes the visible and hidden layers, plus a new layer correction for the signs of coefficients for basis functions of the wave function. We will show that this ‘model has the potential to solve complex quantum many-body problems and to obtain very accurate results for simple molecules as compared with the results calculated by a finite minimal basis set, STO-3G. We also employed a quansum algerithm to help the optimization of training procedure. Results Three-layers restricted Boltzmann machine. We will begin by briefly outlining the original RBM struc- ture as described by [24]. For a given Hamiltonian, #7, and a trial state, }6) = Y7, o(a)[z), the expectation value ean be written as[24]: _ (olHtld) _ Ao leat le")x"d) _ Y» Ola) lal l2")o(a") (ole) LlAe rib) Ylo@)P (219) will be used throughout this letter to express the overlap of the complete wave function funetion [z), 6(2) is the complex conjugate of $(2) Wo can map the above to a RBM model with visible layer units of, of... and hidden layer units ha, fa. Pe with oF, hy © {-1,1}. We uso visible unite oF to reprosent the spin stato of a qubit i up or down. The cof m qubits is represeatal by the basis [:) — Jojo3...03). oz) — y/P(@) where P(e) i ained by the RDM. "The probability of a specific set (#) io where 6(¢ with the basi ry for x from the distribution det 2) Within the above a, and 6; are trainable weights for units of and hj. wij are trainable weights describing the connections hetween @7 and hy (see Figure 1.) By setting (/1) as the objective function of this RBM, we can use the standard gradient decent method to update parameters, effectively minimizing (I) to obtain the ground state energy. However, previous prescriptions considering the use of REMSs for electronic structure problems have found difficulty as (e,) can only be non-negative values. We have thus appeuded an additional layer to the neural network architecture to compensate for the lack of sign features specific to electronie structure problem We propose an RBM with three layers. ‘The first layer, o*, describes the parameters building the wave function. ‘The h’s within the second layer are parameters for the coefficients for the wave functions and the third layer s, represents the signs associated |): (2) sloiogu.o3) ~ tank? daa + 0) @Figure 4: Constructions of rsstricted Boltzmann machine. az the original mstricled Boltzmann machine (IBM) structure with visible o* and hidden ft layers. b: Improved RBM structure with three layers, visible, hidden and sign, i uld be: ponitive or Decauise tre hie added another function for the cocfficients, the distribution isnot solely decided by REM. We also nocd to add our sign Function into the distribution. Within this scheme, ¢ is a regulation and dj are weights for af. (e00 Figuro 1). Our final objective function, now with [d) ~ 53, #(2)a(2)|2), booomen Lea HENZ) a1 la po(z')s(2") Velo After setting the objective function, the learning procedure is performed by sampling to get the distribution of 6(z) and calculating to get s(x). We then proceed to calculate the joint distribution determined by o() and s(z). ‘The gradients are determined by the joint distribution and we use gradient decent method to optimize (I) (sce Supplementary Note 1). Calculating the the joint distribation is efficient because s(2) is only related to 2. (my - @ Electronic Structure Hamiltonian Preparation. ‘The clsctronie stricture is represented by NV single. particle orbitals which ean be empty or occupied by a spinles electron{33 A= J hayalas + 5 YO hyutelogay 6 rere fy ad fy ae one and two-electron integrals. In this study we use the minimal basis (STO-AG) to calculate them. a! and a; are creation and annihilation operators for the orbital j Equation (8) is then transformed to Pali matrices representation, which is achieved by the Jordan-Wigner transformation[3l]. The final electronic structure Hamiltonian tales the general form with af, € {40,2 42,1} where 07.,44,2. are Pauli matrices and I is the identity matrix[35]: H=D ho. + YD wigcseh+ np eno} + (6) in shat tikesSeapets pretani ope sie jo ondoqsaap pides on tea Tendo Hf woreRNRAP ya YR ere soxpHOsy Shecre| copay perme Jo Saqgen of Spee ears Hide p Semmes my toi, Gaoeeeete aoe (8), sor yo Syeoqdnno omy “osry “sponqHe tc: jo sqKaN oy HH deRONEOd mADAR SAR "¢.°8-OFH PUP HET TH Jo Aton apes panos Buyyeo}W JOSIUSAL 2p OANEHL Pp 2 ow m q e oe gf ERENT on eae ae ae ee ES so scoresults for electronic structure, quantum machine learning techniques are set to become powerful tools to perform electronic structure calculations and assist in designing new materials for specific applications. Methods Preparation of the Hamiltonian of Hs, Li and HO. We treat Hz molecule with 2olectrons minimal basis STO.2G and use the Jordan-Wigner traneformation[34). ‘The final Hamiltonian is of 4 qubits, We trent Lil molecule with 4electrons in a minimal basis STO.4G and use the Jordan-Wignee [24]. We armumed the first two lowest orbitale are occupied by electrons and the the final ‘of 4 qubits. We treat H0 molecule with I0-electrons in a minimal hacis STO.2G, we use Jonan-Wigner transform 1s ate oceupied hy electron and first two highest energy orbitals are not occupied all time. We also use the spin symmetry in [12,43] to redhice another two qubits. With the reduction of the number of qubits, finally we have 6 qubits Flamiltonian [35, 44]. All calculations of integrals in second quantization and transformations of electronic structure are done by OpenFermion| 45] and Psit[4) Gradient estimation. The two fimetions g(2) and s(x) are both real function. Thus, the gradient for parameter py can be estimated as 2((HincDp,) — (Hee](Dpy)) Where ige(z) = {itt isso called local energy, Dpx(2) = SESE. (...) represents the expectation value of joint distribution determined by 0(2) and s(r) (details in Supplementary Note 1), transforma Hamilton [24]. We assume the first four lowest energy 01 Implementation Details. In our simulation we choose small constant learning rate 0.01 to avoid trapping im local minimum. All parameter are initialized as a random number between (002,002). ‘The range of initial random parameter is to avoid gradient vanishing of tanh. For each calculation we just need 1 reusing ancilla qubit all the time. ‘Thus, in the simulation, the number of required qubits is m-+n-+1. All calculations do not consider the no'se and system error (details in Supplementary Note 5). Data availability. ‘The data and codes t! corresponding author upon reasonable request. support the findings of this study are available from the References [1] Gooffrcy E Hinton ond Rashn R Salakhutdinor. Reducing the dimen rnotworks. Selence, 313(5786):504 507, 2006. ality of data with neural 2] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):496, 2015, [S| Florian Hase, Stéphanie Valleau, Kdward Pyzer-Knapp, and Alin Aspuru-Guzik. Machine learning exciton dynatnies. Chemical Science, 7(8):5139 5147, 2016. [i] Florian Hise, Christoph Kreisbeck, and Alin Aspura Guzik Machine learning for quantum dynamics: deep learning of excitation energy transfer properties. Chemical science, 8(12):8419-8126, 2017. [5] Grégoire Montavon, Matthias Rupp, Vivekanand Gobre, Alvaro Vazquez-Mayagoitia, Katja Hansen, Alexandre Tkatchenko, Klans-Robert Milller, and O Anatole Von Lilienfeld. Machine learning of molec ular electronic properties in chemical compound space. New Journal of Pkysies, 15(9):095003, 2013 [6] Zachary W Ulissi, Andrew J Medford, Thomas Bligaard, and Jens K Norskov. To address surface y isa reaction network complexity using scaling relations machine learning aud dft cal munications, B:14621, 2017. ations. Nature com- [7] Felix Brockherde, Leslie Vogt, Li Li, Mark E Tuckerman, Kieron Burke, and Klaus-Robert Miller. Bypassing the kohn-sham equations with machine learning. Nature communications, 8(1):872, 2017.[27] Ammar Daskin and Sabre Kais. Direct application of the phase estimation algorithm to find the eigen- values of the hamiltonians. Chemical Physics, 2018. BS) Alin Asparu-Guzik, Anthony D Dutoi, Peter J Love, and Martin Mead-Gordon. Simulated quantum op , “ a ‘computation of molecular encrgis. Seience, SU9(STAL}174 1707, 2006 [29] PJJ O'Malley, Ryan Babbush, ID Kivlichan, Jonathan Romero, JR. McClean, Rami Pedram Roashan, Andrew Tranter, Nan Ding, et al. Sealable quantum Simm energies. Physical Review X, 6(3):031007, 2016 [80] Tan Kassal, James D Whitfield, Alejandro Perdomo-Ortiz, Man-Hoag Yung, aud Alin Aspuru-Guzik ‘Simulating chemistry using quautum computers. Annual review of physical chemistry, 62185 207, 2011, [1] Ryan Babbush, Peter J Love, and Alin Aspura-Guzik Adiabatic quantum simul chemistry. Scientific reports, 4:6608, 2014 [82] lan D Kivlichan, Jarrod McClean, Nathan Wiebe, Craig Gidney, Alin Aspuru-Guzik, Gamet Kin-Lie Chan, and Ryan Babbush, Quantum simulation of electronic structure with linear depth and connectivity. Physical review letters, 120(11):110501, 2018. of quantum [23] Benjamin P Lanyon, James D Whitfield, Geoff G Gillett, Michael E Goggin, Marcelo P Almeida, Ivan Kassal, Jacob D Biamonte, Masoud Mohseni, Ben J Powell, Mateo Barbieri, et al. ‘Towards quantum chemistry on a quantum computer. Nature chemistry, 2(2)-106, 2010. [84] Eduardo Fradkin. Jordan-wigner transformation for quantum-spin systems in two dimensions and frac- tional statistics. Physical review leiters, 63(3):322, 1989. [25] Rongain Xia, Tong Bian, and Sabro Kais, Electronic structure calculations and the ising hamiltonian, ‘The Journal of Physical Cheméatry B, 122(12):2384 3305, 2017. (36 Christopher M Dobson. Chemical space and biology. Nature, 432(7019):824, 2004. [87] Lorenz C Blum and Jean-Louls Reymond, 970 million druglike siaall molecules for virtual sereeuing in the chemical universe database gdb-13. Journal of the American Chemical Society, 1$1(25):8782 8733, 2008. (38) Daniel P Tabor, Loic M Roch, Semion K Saikin, Christoph Kreisbeck, Dennis Sheberla, Joseph H Montoya, Shyam Dwaraknath, Muratahan Aykol, Carlos Ortiz, Hermann ‘Tribukait, et al. Accelerating the discovery of materials for clean energy in the era of smart automation. Nat. Ren. Mater., 2018. [39] Daniel P Tabor, Loic M Roch, Semion K Saikin, Christoph Kreisbeck, Dennis Sheberla, Joseph H Montoya, Shyam Dwaraknath, Muratahan Aykol, Carlos Ortiz, Hermann Tribukalt, et al. Accelerating the discovery of materials for clean energy in the era of smart automation Nat. lev. Mater., 3:8 20, amis, [40] © Anatole von Liienfeld. Quantam machine learning in chemical compound space. Angemandte Chemie International Edition, 57(16) 4164-4169, 2018 (41) Raghunathan Ramakrishnan, Pavlo O Dral, Matthias Rupp, and Q Anatole Von Lilienfeld. Quantum ‘chemistry structures and properties of 134 Kilo molecules. Scientific data, 1:140022, 2014 [#2] Abhinav Kandala, Antonio Mezzacapo, Kristan Temme, Maika ‘Takita, Markus Brink, Jerry M Chow, and Jay M Gambetta. Hardware-effcient variational quantum eigensolver for small molecules and quan tum magnets. Nature, 549(7671):242, 2017. [43] Sergey Bravyi, Jay M Gambetta, Antonio Mezzacapo, and Kristan ‘Temme. ‘Tapering off qubits to simulate fermionic hamiltonians. 2017. Preprint at https: //arxiv.org/abs/1701 .08213, 44] ‘Teng Dian, Daniel Murphy, Rongxin Xin, Ammar Daskin, and Sabre Kais. Comparison study of quantum, computing methods for simulating the hamiltonian of the water molecule. 2018. Preprint at betpe //azxiv .org/abs/1804.05453.Supplementary Note 2 Sequential applications of controlled-rotation algorithm: ‘The probability for each combination y—{o%, h} can be written as: Sai Fy FE BED, woe hy Ply) (14) However, we do not directly cnleulate the P(y) but we do some modification on P(y) to increase the succesful EATS MSE, miele probability of our algorithm. We calenlate Q(y) = ee srboce' hte lage umber to increase the successful probability of our measurements. First wo use Ry gnte to achiove a superposi qubit are initialiond at state [0). / ensk esl oo Ry 2aresin iy eae) IO) 8 RyBarenin Vey =a) PsN) ~ E VOIP) (15) Sie Eon whee OD = See re tt) = bf ahh) ofall possible a and h. The system qubite and the ancilla The next step is to calculate each term of &+s""", which is achieved hy controlled rotations gates. ‘The idea is, for each time controlled rotation, we calculate two angles Bj. —~ aresin| Ve®/Fe~ Fl) and 65 — 2aresin(Vew™o7Fe-TPoTE), We use controlled-totation CRy(0kja) and CR, (Gj) which are Controlled by combination of a7, hy as working qubits to do rotation on the ancilla qubit. ‘The controlled rotation is to check the working qubits and then do the corresponding rotation @j1 OF Biz. All controlled rotation gates can be expressed as below: CRa,.1 = Crt hy @ Ry(aresin( Vers Ee eal/E)) + (Dye, + Eogshy + Faz hz) OL Dat in, @ Hey (2aresin( Vem" /Fe MI) + (Cop.n, + Eat hy + Fat ay) OF Fossa, @ Ry Qaresin( Velo Wal) + (Cesny + Dezany + Foz, 1 Chey : (16) OR. Raa a = Fai a, @ Ry(2aresin( Vews/Fe-besl/E)) + (Coz, + Daz ny + Lot hs) OT where Cez,s, = Boz ® Bry» Des.h, = Az ® Ba, Eas pj = Boz ® Any, For n, = Aes @ An, and _ flo}, [o 9 heb Between the calculation of two ws, we need to doa measurement onthe ania qubit to make sure the state of system qubits collapse to the wanted state. Measuring ancilla.qubit in |!) means the state of system qubits collapse to the wanted state as we initialize the ancilla qubit in [0). We then do the meamroment, if and only if the axcilla qubit is in |L) we continuo with a now ancilla qq initialized in 0), othorwise we start from boginning. The probability of succsas is very large sineo wo choos ke an a largo number.our quantum algorithm when k is large, which requires large number of sampling and may not be efficient. In the Supplementary Figure 3, we can sce that at the final procedure of optimization, the fluctuation is very large due to large k, which can be decreased by increasing the number of sampling. Because we investigated small molecule system Hz, LiH and HzO, k is not very large and the quantum algorithm is eificient. For large &, our quantum algorithm may require large sampling.Supplementary Note 1 Derivation of the gradient For an electronic structure Hamiltonian prepared by second quantization and Jordan-Wigner traneformation{24}, 17, aol a tral wave function, [é) ~ S1,4(2)«(z)}2), the expectation value ean be written as{47} (AHO) — Dae HENS) LI" o(2’)s(2") “Tel o(z)s(a)P ris acombination of {7,03...03) and |x) o{03...03). (H} = (9) If we set (x) = Hf2)s(2), because (7) and s(7) are all real valne functions, then the gradient ean be calculated as[47, 24} apt) ~ Tansee) Ep PCe) (lL 0) F(x 2)) (2) + 0(2)2,.0(2)) ug) IP) Ys POE If we set Eige(2) = SUM?) and D,,(2) = “eS, the gradiont can be written as{47}: sam ie() Dyula 40(0)E le)= ao Fb Maclt) SO le) + Dp (DIE i) I SPOOF ‘tac! D yy) ~ (Fc) Doe) where (..) represent the expectation value of distribution determined by ®(x). (r|/1\g) = (6l2f|2) for that H's a real symmetric matrix due to Jordan-Wigner transformation. px is the parameters a,,b;,1 1 for ky iterations, ‘Thus we havel 67} D(a) = hot Fo )nnae Das(a) = Ztanh,) — 5j)anne (12) Dasa) = 5tank(0, oz — Sothys, Da) Dy (2) = shere 8; —S2,wyo8 +b. (Jenne represents the distrbntion determined solely by RBM. We co not need to calenlate the second term of Da, Ds, and Dw, for that they will be cancelled when calculating the gradient d,, (I). We use the gradient decent method to optimize our RBM, yielding the global minirmm corresponding to the ground energy. Psa ~ Pk ~ 04d, i) (13) Where 04 is the learning rate for ki iteration, controlling the conver tant we reach the maximum number of iternifona, ‘The gradient is cot by sampling. ce rate. We can continue iterating sted by the distribution calculated[45] Jarrod R. McClean, lan D Kivlichan, Damian $ Steiger, Yudong Cao, E Schuyler Fried, Craig Gidney, ‘Thomas Hiner, Vojtiéch Havligek, Zhang Jiang, Matthew Necley, et al. Openfermion: the electronic structure package for quantum computers. 2017. Preprint at https://arziv.org/abs/1710.07629. ]46] Robert M Parrish, Lori A Burns, Daniel GA Smith, Andrew C Simmonett, A Eugene DePrince IIL, Edward G Hohenstein, Ugur Bezkaya, Alexander Yu Sokolov, Roberto Di Remigio, Ryan M Richard, otal. Psid 1.1: An open-source clectronic stmacture program emphasizing automation, advanced libraries, and interoperability. Journal of chemical theory and computation, 13(7):2185 3197, 2017. 147| Giuseppe Carleo, Matthias Troyer, Giacomo Torlai, Roger Melko, Juan Carrasquilla, and Guglielmo Mazzola, Neural-network quantum states. Bulletin of the American Physical Socicty, 2018, 48) Michael A Nielsen and Isaac L Chuang. Quantum computation and quantum information. Cambridge University Press, Cambridge, 2000. Acknowledgement We would thank to Dr. Ross Hoehn, Dr. Ziuan Hu and Teng Bian for critical rea S.K and R.X are grateful for the support from Integrated Data Science Ini ing and useful discussions. sive Grants, Purdue University Author Contribution S.K designed the research, R.X performed the calculations. Both discussed the results and wrote the paper. Additional Information Supplementary Information. $ :pplementary Materials are available. Competing Interests. The authors declare no competing interestad Hh in 7s asa SE TOOTH TOL TEHOT UOT STOO STOTT « » -ns| 0 r -ns|\ Supplementary Figure 3: ‘The energis calculated by ISM during the optimization. a Optimization procedure for Hz. b Optimization procedure for Lill, ¢ Optimization pmeesture for HO. ‘The distribution we want to sampling for the quantum algorithm QW) (as) In our controlled rotation algorithm, wo use a k ax mgulation to incroaes the probability of success as the proof in the Supplementary Note 3, the lower bound of probability of success would become 1 Fe, (19) ‘Thus, if no regulation (k = 1), the probability of success would become which means we need exponential number of measurements to get enough successfil sampling, ‘making no speedup in quantum algorithm. If we add a regulation of k, In simulation we use & = constant number of measure 3D, luis, the probability becomes leuts Lo get enough successful Sampling. which needs Afier we get the distribution, we ned to caleulate all distribution to the power of & and norma get the wanted distribution. kis a large number, which is around 5 at final for H, Lill and 1,0 in our simulation, "To decrease the errors in calculating power of k, we have to increase the number of sampling for 15After we finish all measurements, the distribution should be Q(j). We just measure the first n +m qubits of the system register to obtain the probability distribution. After we get the distribution, we calculate all probabilities to the power of & and normalize to get the Gibbs distribution Supplementary Note 3 Lower bound of successful sampling The suc essful probability P can be written as: = F. fonwr yh P HOD AO, Fed If-we choose k= O(S>, ; lwij|), we have P > z2ivy which means the lower bound of probability of sampling is a constant’ In the simulation, we choose k ~ maz(3 3°, , wish 1) because larger larger sampling errors. This particular choice of k gives us lower bound of success as e~*. But in numerical simulation, the probability is much larger than e~*, see Supplementary Note 5. Supplementary Note 4 Complexity and error of the algorithin For a CU) conditioned by |11), it can be decomposed as the below[45}: v vt VE Supplementary Figure 1: ‘The decomposition of the C(U) gate where V? =U. In our algorithm, U ~ fy(0), ths we can choose V = fy(0/2) to achieve the decomposition. C>(U) conditioned by |00), |10) or OL) Can be achieved by adding X gates on controling qubits. Fur each wij, We have 4 C2(U/) whieh meaus the gates complesity seales «9 (ran) aud the musnber of qubits for our algorithms scales to O(0nn), which can be reduced 1 O(n +n) if considering qublt reuse. Because the lower bound of probability of sucemsful sampling i constant, ifthe aumber of succesful saripling is Nq, the complexity for each iteration is O(Ngmin). The only error comes fora the error of sampling if not considering noise in the quantum computeSupplementary Note 5 Implementation details for H:, Li and H,0 Here we present the probabilitias of successful sampling whon calculating H, LiH and H,0 for bond length equals to 1.75 Angstrom. 2 My fog LK fos| po = Treat erator a b 4:0 Supplementary Figure 2: ‘The probability of sucvesful sunpling during (he optimization. a Optimization proce- dure for Hy. b Oplimizatioa procedure for Lill. e Oplimization procedure for HO. Here we present the the changes of energy during optimization when calcalating Hz, LiH and HzO for bond length equals to 1.75 Angstrom.