uantum Chemistry Edition IRAN. LEVINE Chemistry Department Brooklyn College City University of New York Brooklyn, New York ZPA KBs PRENTICE HALL, Upper Saddle River, New Jersey 07458Contents PREFACE ix 1 THE SCHRODINGER EQUATION 1 1.1 Quantum Chemistry, 1 1.2 Historical Background of Quantum Mechanics, 2 13 The Uncertainty Principle, 5 14 The Time-Dependent Schrodinger Equation, 7 1.5 The Time-Independent Schrédinger Equation, 12 1.6 Probability, 14 1.7 Complex Numbers, 16 18 Units, 18 1.9 Summary, 18 THE PARTICLE INA BOX 21 2.1. Differential Equations, 21 2.2 Particle in a One-Dimensional Box, 22 2.3 ‘The Free Particle in One Dimension, 28 2.4 Particle in a Rectangular Well, 29 2.5 Tunneling, 31 2.6 Summary, 32 OPERATORS 35 3.1 Operators, 35 3.2 Eigenfunctions and Eigenvalues, 39 3.3 Operators and Quantum Mechanics, 40 ; 3.4 The Three-Dimensional Many-Particle Schrédinger Equation, 46 3.5 The Particle in a Three-Dimensional Box, 49 3.6 Degeneracy, 52 3.7 Average Values, 53 3.8 Requirements for an Acceptable Wave Function, 57 3.9 Summary, 58 THE HARMONIC OSCILLATOR 62 4.1 Power-Series Solution of Differential Equations, 62 4.2. The One-Dimensional Harmonic Oscillator, 65 4.3 Vibration of Molecules, 74 4.4 Numerical Solution of the One-Dimensional Time-Independent Schrédinger Equation, 78 4.5 Summary, 89Contents ANGULAR MOMENTUM 94 5.1 Simultaneous Specification of Several Properties, 94 5.2 Vectors, 97 5.3 Angular Momentum of a One-Particle System, 102 5A The Ladder-Operator Method for Angular Momentum, 115 5.5 Summary, 120 THE HYDROGEN ATOM 123 6.1 The One-Particle Central-Force Problem, 123 6.2 Noninteracting Particles and Separation of Variables, 125 63 Reduction of the Two-Particle Problem to Two One-Particle Problems, 127 6.4 The Two-Particle Rigid Rotor, 130 6.5 The Hydrogen Atom, 134 6.6 The Bound-State Hydrogen-Atom Wave Functions, 142 6.7 Hydrogenlike Orbitals, 150 68 The Zeeman Effect, 154 6.9 Numerical Solution of the Radial Schrédinger Equation, 157 6.10 Summary, 158 THEOREMS OF QUANTUM MECHANICS 163 7A Introduction, 163 7.2, Hermitian Operators, 164 7.3 Expansion in Terms of Figenfunctions, 170 7.4 Eigenfunctions of Commuting Operators, 175, 75 Parity, 178 7.6 Measurement and the Superposition of States, 182 7.7 Position Eigenfunctions, 187 7.8 The Postulates of Quantum Mechanics, 190 7.9 Measurement and the Interpretation of Quantum Mechanics, 194 7.10 Matrices, 198 711 Summary,201 THE VARIATION METHOD 208 8.1 The Variation Theorem, 208, 8.2 Extension of the Variation Method, 212 83 Determinants, 213 8.4 Simultaneous Linear Equations, 217 8.5 Linear Variation Functions, 220 8.6 Matrices, Eigenvalues, and Eigenvectors,228 8.7 Summary, 235 vvi Contents 9 10 W 12 PERTURBATION THEORY 245 9.1. Introduction, 245 9.2. Nondegenerate Perturbation Theory, 246 9.3 Perturbation Treatment of the Helium-Atom Ground State, 252 9.4 Variation Treatments of the Ground State of Helium, 256 9.5 Perturbation Theory for a Degenerate Energy Level, 259 9.6 Simplification of the Secular Equation, 263 9.7 Perturbation Treatment of the First Excited States of Helium, 265 9.8 Comparison of the Variation and Perturbation Methods, 272 9.9 Time-Dependent Perturbation Theory, 272 9.10 Interaction of Radiation and Matter, 275 9.411 Summary,277 ELECTRON SPIN AND THE PAULI PRINCIPLE 282 10.1 Electron Spin,282 10.2 Spin and the Hydrogen Atom, 285 10.3 The Pauli Principle, 285 10.4 The Helium Atom, 288 10.5 The Pauli Exclusion Principle, 290 10.6 Slater Determinants, 295 10.7 Perturbation Treatment of the Lithium Ground State, 297 10.8 Variation Treatments of the Lithium Ground State, 298 10.9 Spin Magnetic Moment, 299 10.10 Ladder Operators for Electron Spin, 300 10.11 Summary, 302 MANY-ELECTRON ATOMS 305 1L.1 The Hartree-Fock Seif-Consistent-Field Method, 305 11.2 Orbitals and the Periodic Table, 312 11.3 Electron Correlation, 315 11.4 Addition of Angular Momenta,318 11.5 Angular Momentum in Many-Electron Atoms, 323 11.6 Spin-Orbit Interaction, 335 11.7 The Atomic Hamiltonian, 337 118 The Condon-Slater Rules, 339 11.9 Summary, 342 MOLECULAR SYMMETRY 347 12.1 Symmetry Elements and Operations, 347 12.2 Symmetry Point Groups, 355 12.3 Summary, 362Contents vii 13 ELECTRONIC STRUCTURE OF DIATOMIC MOLECULES 366 13.1 The Born-Oppenheimer Approximation, 366 13.2 Nuclear Motion in Diatomic Molecules, 370 13.3 Atomic Units, 375 13.4 The Hydrogen Molecule Ion, 376 13.5 Approximate Treatments of the H3 Ground Electronic State, 381 13.6 Molecular Orbitals for H Excited States, 390 13.7 MO Configurations of Homonuclear Diatomic Molecules, 396 13.8 Electronic Terms of Diatomic Molecules, 402 13.9 The Hydrogen Molecule, 407 13.10 The Valence-Bond Treatment of H,, 410 13.11 Comparison of the MO and VB Theories, 414 13.12 MO and VB Wave Functions for Homonuclear Diatomic Molecules, 416 13.13 Excited States of Hy, 419 13.14 Electron Probability Density, 421 13.15 Dipole Moments, 423 13.16 The Hartree-Fock Method for Molecules, 426 13.17 SCF Wave Functions for Diatomic Molecules, 436 13.18 MO Treatment of Heteronuclear Diatomic Molecules, 439 13.19 VB Treatment of Heteronuclear Diatomic Molecules, 442 13.20 The Valence-Electron Approximation, 443 13.21 Cl Wave Functions, 444 13.22 Summary, 451 44 THE VIRIAL THEOREM AND THE HELLMANN-FEYNMAN THEOREM 459 14.1 The Virial Theorem, 459 14.2 The Virial Theorem and Chemical Bonding, 466 143 The Hellmann—Feynman Theorem, 469 14.4 The Electrostatic Theorem, 472 14.5 Summary, 478 15 AB INITIO AND DENSITY-FUNCTIONAL TREATMENTS OF MOLECULES 480 15.1 Ab Initio, Density-Functional, Semiempirical, and Molecular-Mechanics Methods, 480 15.2 Electronic Terms of Polyatomic Molecules, 481 153 The SCF MO Treatment of Polyatomic Molecules, 485, 15.4 Basis Functions, 486 15.5 Speeding Up Hartree-Fock Calculations, 494 15.6 The SCF MO Treatment of H,0, 498 15.7 Population Analysis, 505 15.8 The Molecular Electrostatic Potential and Atomic Charges, 508 15.9 Localized MOs, 511viii Contents 16 17 15.10 The SCF MO Treatment of Methane, Ethane, and Ethylene, 517 15.11 Molecular Geometry, $28 15.12 Conformational Searching, 539 15.13 Molecular Vibrational Frequencies, 545 15.14 Thermodynamic Properties, $48 15.15 Ab Initio Quantum Chemistry Programs, 550 15.16 Performing Ab Initio Calculations, 551 15.17 Configuration Interaction, 557 15.18 Moller-Plesset (MP) Perturbation Theory, 563 15.19 The Coupled-Ciuster Method, 568 15.20 Density-Functional Theory, 373 45.21 Composite Methods for Energy Calculations, 592 15.22 Solvent Effects, 593 15.23 Relativistic Effects, 602 15.24 Valence-Bond Treatment of Polyatomic Molecules, 604 15.25 The Generalized Valence-Bond Method, 612 15.26 Chemical Reactions, 613 SEMIEMPIRICAL AND. MOLECULAR-MECHANICS TREATMENTS OF MOLECULES 626 16.1 Semiempirical MO Treatments of Planar Conjugated Molecules, 626 16.2 The Free-Electron MO Method, 627 16.3 The Hiickel MO Method, 629 16.4 The Pariser-Parr—Pople Method, 650 16.5 General Semiempirical MO Methods, 652 16.6 The Molecular-Mechanics Method, 664 16.7 Empirical and Semiempirical Treatments of Solvent Effects, 680 16.8 Chemical Reactions, 684 COMPARISONS OF METHODS 693 17.1 Molecular Geometry, 693 17.2 Energy Changes, 696 17.3 Other Properties, 703 17.4 Hydrogen Bonding, 705 17.5 Conclusion, 707 17.6 The Future of Quantum Chemistry, 708 APPENDIX 710 BIBLIOGRAPHY 712 ANSWERS TO SELECTED PROBLEMS 715 INDEX 72115.1 AB INITIO, DENSITY-FUNCTIONAL, SEMIEMPIRICAL, 480 CHAPTER 15 Ab Initio and Density- Functional Treatments of Molecules AND MOLECULAR-MECHANICS METHODS For polyatomic molecules, the presence of several nuclei makes quantum-me calculations harder than for diatomic molecules. Moreover, the electronic tion of a diatomic molecule is a function of only one parameter—the inte tance. In contrast, the electronic wave function of a polyatomic molecule several parameters—the bond distances, bond angles, and dihedral angles about single bonds (these angles define the molecular conformation), A full cal treatment of a polyatomic molecule involves calculation of the elect function for a range of each of these parameters. The equilibrium bond distz angles are then found as those values that minimize the electronic energy + nuclear repulsion. The four main approaches to calculating molecular properties are methods, semiempirical methods, the density-functional method, and the mechanics method. Semiempirical molecular quantum-mechanical methods use a simp tonian than the correct molecular Hamiltonian and use parameters whose adjusted to fit experimental data or the results of ab initio calculations; an exa the Hiickel MO treatment of conjugated hydrocarbons (Section 16.3), wl one-electron Hamiltonian and takes the bond integrals as adjustable rather than quantities to be calculated theoretically. In contrast, an ab initio principles) calculation uses the correct Hamiltonian and does not use data other than the values of the fundamental physical constants, A Hartree calculation seeks the antisymmetrized product ® of one-electron functions mizes [ ©*/T@ dr, where A is the true Hamiltonian, and is thus an ab initio tion. (Ab initio is Latin for “from the beginning” and indicates a calculation fundamental principles.) The term ab initio should not be interpreted to mean correct.” An ab initio SCF MO calculation uses the approximation of taking antisymmetrized product of one-electron spin-orbitals and uses a finite (and incomplete) basis set.Section 15.2 Electronic Terms of Polyatomic Molecules 481 The density-functional method (Section 15.20) does not attempt to calculate the mo.ecular wave function but calculates the molecular electron probability density p and calculates the molecular electronic energy from p. ‘The molecular-mechanics method (Section 16.6) is not a quantum-mechanical method and does not use a molecular Hamiltonian operator or wave function. Instead, it views the molecule as a collection of atoms held together by bonds and expresses the molecular energy in terms of force constants for bond bending and stretching and other parameters. A free database of references for molecular ab initio and density-functional cal- culations is at qcldb.ims.acjp/index.html. ! ELECTRONIC TERMS OF POLYATOMIC MOLECULES For polyatomic molecules the operator 5? for the square of the total electronic spin angular momentum commutes with the electronic Hamiltonian, and, as for diatomic molecules, the electronic terms of polyatomic molecules are classified as singlets, dou- blets, triplets, and so on, according to the value of 25 + 1. (The commutation of $? and AZ, holds provided spin-orbit interaction is omitted from the Hamiltonian; for mole- cules containing heavy atoms, spin-orbit interaction is considerable, and S is not a good quantum number.) For linear polyatomic molecules, the operator F., for the axial component of the total electronic orbital angular momentum commutes with the electronic Hamiltonian, and the same term classifications are used as for diatomic molecules; we have such pos- sibilities as 'S*,!E-,°*, ‘I, and so on. For linear polyatomic molecules with a cen- ter of symmetry, the g, i classification is added. For nonlinear polyatomic molecules, no orbital angular-momentum operator commutes with the electronic Hamiltonian, and the angular-momentum classification of electronic terms cannot be used. Operators that do commute with the electronic Hamiltonian are the symmetry operators Og of the molecule (Section 12.1), and the electronic states of polyatomic molecules are classified according to the behavior of the electronic wave function on application of these operators. Consider H,O as an example. In its equilibrium configuration, the water molecule belongs to group 2, with the symmetry operations EB C) x2) G2) (asa) The standard convention [R. S. Mulliken, J. Chem. Phys, 23, 1997 (1955); 24, 1118 (1956)] takes the molecular plane as the yz plane (Fig. 15.1). We readily find that each of the symmetry operations commutes with the other three. Therefore, the electronic wave functions can be chosen as simultaneous eigenfunctions of all four symmetry operators. Since Oy is the unit operator, we have O sth. = Ya Each of the remaining symmetry operators satisfies 02 = 1, and so each has as its eigenvalues +1 and 71 [Eq. (7.55)]. Therefore, each electronic wave function of H,O is an eigenfunction of Oz: with eigenvalue +1 and an eigenfunction of each of the other three symmetry opera- tors with the eigenvalues + 1 or —1.Section 15.20 Density-Functional Theory 573 Method MP2 CISD MP3 CISDT MP4SDQ CCD CCSD QCISD MP4 AtR, 165 92 79 7A 43 38 30 27 2d AIR 278 2.0 229 167 3 128 Ta 64 59 Method MPS MP6 CCSD(T) CCSDT-1 QCISD(T) CCSDT CISDTQ CCSDTQ AIR 13 05 05 04 o8 03 02 0.01 AUR” S51 14 12 09 08 08 LL 0.03 ‘The very accurate CCSDTO, CCSDT, CISDTQ, CCSDT-1, and MP6 methods are much too computationally demanding to be used regularly. A review article states that “the CCSD(T) and QCISD(T) methods appear to be the most accurate, yet computationally tractable, schemes though recent experience on some challenging systems indicates that CCSD(T) is applicable for a wider range of problems [than QCISD(1)]” [K. Raghavachari and J. B. Anderson, J, Phys. Chern., 100, 12960 (1996)]. These two methods require similar amounts of computer time and give very accurate molecular geometries and vibrational frequencies in addition to accurate energies. (See also Chapter IN)A 1996 review article gave the maximum feasible molecular size for a CCSD(T) calcula- tion requiring 10 energy and gradient evaluations as 8 to 12 first-row nonhydrogen atoms for a DZP basis set [M.Head-Gordon, J. Phys. Chem., 100, 13213 (1996)]. DENSITY-FUNCTIONAL THEORY The electronic wave function of an n-electron molecule depends on 3n spatial and spin coordinates. Since the Hamiltonian operator (15.10) contains only one- and two-electron spatial terms, the molecular energy can be written in terms of integrals involving only six spatial coordinates (Problem 15.82). In a sense, the wave function of a many-electron molecule contains more information than is needed and is lacking in direct physical sig- nificance. This has prompted the search for functions that involve fewer variables than the wave function and that can be used to calculate the energy and other properties. ‘The Hohenberg-Kohn Theorem. In 1964, Pierre Hohenberg and Walter Kohn proved that for molecules with a nondegenerate ground state, the ground-state molecular energy, wave function, and all other molecular electronic properties are uniquely determined by the ground-state electron probability density p(x, yz), (Section 13.14), a function of only three variables [P. Hohenberg and W. Kohn, Phys. Rev, 136, B864 (1964)]. (The zero subscript indicates the ground state.) One says that the ground-state electronic energy Ep is a functional of py and writes Ey = Eo{po], where the square brackets denote a functional relation. Density-functional theory (DET) attempts to calculate E, and other ground-state molecular properties from the ground-state electron density po. What is a functional? Recall that a function f(x) is a rule that associates a number with each value of the variable x for which the function f is defined. For example, the function f(x) = x? + 1 associates the number 10 with the value 3 of x and associates a number with each other value of x. A functional F[f) is a rule that associates a num- ber with each function f. For example, the functional F[f] = J™_ /*(x)f(x) dx associ- ates a number, found by integration of |f|* over all space, with each quadratically574 Chapter 15 Ab Initio and Density-Functional Treatments of Molecules integrable function f(x). The variational integral W[4] = (6 |A| 6)Xl¢) isa & of the variation function # and gives a number for cach well-behaved # ‘The proof of the Hohenberg-Kohn theorem is as follows. The ground-st tronic wave function yy of an n-electron molecule is an eigenfunction of # electronic Hamiltonian of Eq. (13.5), which, in atomic units, is : =-tSvitduysEE4 Z. Tie The quantity v(r;), the potential energy of interaction between electron nuclei, depends on the coordinates x, y, z; of electron i and on the nuclear Since the electronic Schrédinger equation is solved for fixed locations of the: nuclear coordinates are not variables for the electronic Schrédinger equati v(t) in the electronic Schrédinger equation is a function of only x), Yq Zp Whiel cate by using the yector notation of Section 5.2. In DFT, v(r)) is called th potential acting on electron i, since it is produced by charges external to the electrons. Once the external potential v(r,) and the number of electrons n are 5 electronic wave functions and allowed energies of the molecule are deter: solutions of the electronic Schrédinger equation. Hohenberg and Kohn pro} systems with a nondegenerate ground state, the ground-state electron prol sity po(t) determines the external potential (except for an arbitrary additive and determines the number of electrons. Hence, the ground-state wave energy (and, for that matter, all excited-state wave functions and energies) a mined by the ground-state electron density. ‘To see that p,(r) determines the number of electrons, we integrate (13.1 all space and use the normalization of ¢ to get Jpo(r) dr To see that p,(r) determines the external potential v(F,), we suppose false and that there are two external potentials v, and vs (differing by mor: stant) that each give rise to the same ground-state electron density pg. Let H, an the n-electron Hamiltonians (15,106) corresponding to v,(r,) and v4(r,), wher are not necessarily given by (15.107); they can be any external potential. Let y and E), and Eq, be the normalized ground-state wave functions and energie Hamiltonians, [Note that even if A’, is a molecular electronic Hamiltonian by (15.107), 0,(1;) is not restricted to the form (15.107) but can be any function: and io, must be different functions, since they are eigenfunctions of Hamilto differ by more than an additive constant (Problem 15.61). If the ground state i generate, then there is only one normalized function, the exact ground-state we tion wo, that gives the exact ground state energy Ej when used as a trial ve function, and, according to the variation theorem, use of any normalized wel function that differs from y will make the variational integral greater than Ey (1 8.10): that is, (4|H1|4) > Ep if & # Wo and the ground state is nondegenerate. Th using Uo, as a trial function with the Hamiltonian A, givesSection 15.20 Density-Functional Theory 575 Exe < WoolAalos) = Wow Aa + As ~ Habbo) = Wools ~ Halton) + Woslldos) The Hamiltonians A, and A, differ only in their external potentials v, and v%, so A, ~ Ay = Xini[vr) — o(r,)] and we have Ena < (veal zl 7) + Eos ‘The quantities v,(r) and v,(r) are one-electron operators, and using Eq. (13.131), we get — u(t) Fe < [puaC [ese ~ ete + Ey where, since the integration is over Wo,, we get the electron density pp, corresponding 10 Woy. If we go through the same reasoning with a and b interchanged, we get Fou < {onde ~ extlldt + Bag By hypothesis, the two different _wave functions give the same electron densit: Poa 1 Putting 9. = Po, and adding the last two inequalities, the two integrals can- “cel and we get Eq, + Ey < Eqy + Epg: This result is false, so our initial assumption that two different external potentials could produce the same ground-state electron density must be false. Hence, the ground-state electron probability density pp determines the external potential (to within an additive constant that simply affects the zero level of energy) and also determines the number of electrons. Hence py determines the molec- ular electronic Hamiltonian and so determines the ground-state wave function, energy, and other properties. The ground-state electronic energy Ey is thus a functional of the function py(r), which we write as Ey = E,[po],where the v subscript emphasizes the dépéndence of Ey on the external potential u(r), which differs for different molecules. The purely electronic Hamiltonian (13.5) is the sum of electronic kinetic-energy terms, electron-nuclear attractions, and electron-clectron repulsions. Taking the aver- age of (13.5) for the ground state, we have E = T + Vy, + V.e where, for notational convenience, overbars instead of angular brackets have been used to denote averages. Each of the average values in this equation is a molecular property determined by the ground-state electronic wave function, which, in turn, is determined by p)(r). Therefore, each of these averages is a functional of py: Eq = Exlpo] = Tlpo] + VuelPo] + VeelPo| From (13.1), Vue = Dies v(t), where v(t) = — Da Z,/tiq in atomic units, so = (v] Dye where (13.131) was used, and where »(r) is the nuclear attraction potential-energy function for an electron located at point r. Thus Vy_{/p9] is known, but the functionals T(po] and V,.{ po] are unknown. We have rm) = | polt)o(r) de (15.108) Ey = Ex(po] Joxeree dr + T{po] + Veelpo] = ose) de + F[ po] (15.109)576 Chapter 15 Ab Initio and Density-Functional Treatments of Molecules where the functional F(po], defined by F[a)] = T[oo] + Vi.[po), is independent external potential. Equation (15.109) does not provide a practical way to from po, because the functional F[o] is unknown. The Hohenberg-Kohn Variational Theorem. To turn (15.109) from a relation to a practical tool, we need a second theorem proven by Hohenbe Kohn, and an approach developed by Kohn and Sham. Hohenberg and Kohn that for every trial density function p,(r) that satisfies {p,(x) dr = n and p,(e) = all r, the following inequality holds: Ey < Ey[p,), where F, is the energy funel (15.109). Since Z, = E,[po], where py is the true ground-state electron density, ground-state electron density minimizes the energy functional E,[p,,] Gust as t normalized ground-state wave function minimizes the variational integral). The proof of the Hohenberg-Kohn variational theorem is as follows. satisfy the above two conditions of integrating to n and being nonnegative. | Hohenberg-Kohn theorem, p,, determines the external potential v,, and this j determines the wave function y, that corresponds to the density p,. (Actug is only true if there exists an external potential 2, that will give rise to an metric wave function that corresponds to p,,. If this condition holds, p,. is ‘epresentable. It turns out that not all p,’s are v-representable. This has not “any practical difficulties in applications of DFT. Also, Levy has reformulate Hohenberg-Kohn theorems in a way that eliminates the need for v-represent: See Parr and Yang, Sections 3,3,3.4, and 7.3.) Let us use the wave function yf, responds to p,, as a trial variation function for the molecule with Hamiltonian # variation theorem gives (Welly) = (vet 40,4 Som) a We) = Fo= Fale] Using the fact that the average kinetic and potential energies are functiona electron density, and using (15.108) with ty replaced by dh, we have for (15.110) Foe) + Vado) + [pee de 2 Ep The functionals T and V,, are the same in (15.109) and (15.111) although the fut pp and p,, differ. The left side of (15.111) differs from the corresponding expres (15.109) only by having py replaced by py: use of (15.109) with py replaced (15.111) gives E,{p,)] > E,[q), which proves that any trial electron density ca a lower ground-state energy than the true ground-state electron density, Hohenberg and Kohn proved their theorems only for nondegenerate ¢ states Subsequently, Levy proved the theorems for degenerate ground states (: and Yang, Sec. 3.4). The Kohn-Sham Method. If we know the ground-state electron densi the Hohenberg-Kohn theorem tells us that it is possible in principle to calculate ground-state moiecular properties from py, without having to find the moleculaSection 15.20 Density-Functional Theory 577 function. [In the traditional quantum-mechanical approach, one first finds the wave function and then finds p by integration; Eq, (13.128).] The Hohenberg-Kohn theorem does not tell us how to calculate Ey from pp [since the functional F in (15.109) is unknown], nor does it tell us how to find pp without first finding the wave function. A key step toward these goals was taken in 1965 when Kohn and Sham devised a practical method for finding pp and for finding Ep from py (W. Kohn and L. J. Sham, Phys. Rev., 140, A1133 (1965)]. Their method is capable, in principle, of yielding exact results, but because the equations of the Kohn-Sham (KS) method contain an unknown functional that must be approximated, the KS formulation of DFT yields approximate results. Kohn and Sham considered a fictitious reference system (denoted by the sub- script s and often called the noninteracting system) of n noninteracting electrons that each experience the same external potential-energy function v,(r,), where 0,(r)) is such as to make the ground-state electron probability density p,(r) of the reference system equal to the exact ground-state electron density po(r) of the molecule we are interested in; p,(r) = po(t). Since Hohenberg and Kohn proved that the ground-state probability density function determines the external potential, once p,(r) is defined for the refer- ence system, the external potential »,(r,) in the reference system is uniquely determined, although we might not know how to actually find it. The electrons do not interact with one another in the reference system, so the Hamiltonian of the reference system is A= DWV? +o) = DAS where AS = - IV? + v(x) (15.112) ry ASS is the one-electron Kohn-Sham Hamiltonian. Use of a fictitious system of nonin- teracting electrons should not be too disturbing. Recall that we used a system of non- interacting electrons in the Section 9.3 perturbation treatment of the He atom. We can relate the fictitious Kohn-Sham reference system to the real molecule by writing the Hamiltonian H, = 7 + 3; v4(r,) + AV. where the parameter A ranges from 0 (no interelectronic repulsions, which is the reference system) to 1 (the real molecule), and 1 is defined as the external potential that will make the ground-state electron density of the system with Hamiltonian /, equal to that of the real molecule’s ground state. Since the reference system s consists of noninteracting particles, the results of Section 6.2 and the Pauli principle show that the ground-state wave function yf. of the reference system is the antisymmetrized product (Slater determinant) of the lowest- energy Kohn-Sham spin-orbitals Sof the reference system, where the spatial part S(r,) of each spin-orbital is an eigenfunction of the one-electron operator Eas that is, yp = luau, = OS(r)or (15.113) ASSGES = eKSgkS (15.114) where g; is a spin function (either a or B) and the e's are Kohn-Sham orbital energies.578 Chapter 15 Ab Initio and Density-Functional Treatments of Molecules For a closed-shell ground state, the electrons are paired in the orbitals, with two electrons of opposite spin having the same spatial orbital (as in the RHF method). Kohn and Sham rewrote the Hohenberg-Kohn equation (15.109) AT be defined by AT ip] = Tp] - Tle] where, for convenience, the zero subscript on p is omitted in this and equations. AT is the difference in the average ground-state electronic k between the molecule and the reference system of noninteracting ele: tron density equal to that in the molecule. Let a8 { tenon re na where 7 is the distance between points x, yy, 2, and x2, Yo, Ze $$ Se(ri)p(r2)ria! dr; dra is the classical expression (in atomic units) for interelectronic repulsion energy if the electrons were smeared out into distribution of charge with electron density p.The charge dQ, ina tiny dr, of such a distribution is dQ, = —ep(r,)dr, and the potential en sion between dQ, and the charge in the volume element dr, i er 4p(r;)plra)dr,dr>. Integration over dr, gives the repulsion energy bet the charge distribution; integration over dr, and multiplication by | gives sion energy, where the factor 3 is needed to prevent counting each repulsic as the repulsion between dQ, and dQ, and once as the repulsion betwee: With-the definitions (15.115) and (15.116), (15.109) becomes E[o} =[o (r)o(r) dr + Tp] + = | foe eleidets) dr, dr, + AT[p] + The functionals AT and AV,, are unknown. Defining the exchange-co1 functional E,.[p) by Ep] = AT(p] + AV.Lp] we have Ey = Eo] = [ern dr + Tio] +5 3 | ate seve) deydr + E. ‘The motivation for the definitions (15.115), (15.116), and (15.117) is to terms of three quantities, the first three terms on the right side of (15 easy to evaluate from p and that include the main contributions to the: energy, plus a fourth quantity £,., which, although not easy to evaluateSection 15.20 Density-Functional Theory 579 be a relatively small term. The key to accurate KS DFT erties is to get a good approximation to E, Before we can evaluate the terms in (15.118), we need to find the ground-state electron density. Recall that the fictitious system of noninteracting electrons is defined to have the same electron density as that in the ground state of the molecule: p, = po. It is readily proved (see Problem 15.67) that the electron probability density of an neparticle system whose wave function [Eq. (15.113)] is a Slater determinant of the spin-orbitals ukS = 68Sc; is given by ¥/_, oS. Therefore culation of molecular prop- p= => pe (15.119) a How do we evaluate the terms in (15.118)? Using (15.107), we have Soleu(r) dr = — Ey Ze fplt,)rid dei, which is easily evaluated if p(r) is known. The T, term in (15.118) is the kinetic energy of the system of noninteracting electrons with wave function y, in (15.113) equal to a Slater determinant of orthonormal Kohn-Sham spin-orbitals. We have T,[p] = —3}(, |=. V?|y,). The Slater~Condon rules [Table 11.3 and Eq. (11.78)] give T.[p] = —3D; (@SS(1)|¥/6S5(1)). Thus (15.118) becomes 4 vp Tateapts : Re -32, [an — FB ersenivaissay + $f [PRR aejdes + Bue] (15.120) We can therefore find Ey from p if we can find the KS orbitals @° and if we know what the’ functional £,, is. The electronic energy including nuclear repulsion is found by adding the internuclear repulsion Vy to (15.120). The Kohn-Sham orbitals are found as follows. The Hohenberg-Kohn variational theorem tells us that we can find the ground-state energy by varying p (subject to the constraint [p dr = n) so as to minimize the functional E,[p]. Equivalently, instead of varying p, we can vary the KS orbitals f°, which determine p by (15.119). (In doing so, we must constrain the 0's to be orthonormal, since orthonormality was assumed when we evaluated T,.) Just as one can show that the orthonormal orbitals that minimize the Hartree-Fock expression for the molecular energy satisfy the Fock equation (13.148), ‘one can show that the Kohn-Sham orbitals that minimize the expression (15.120) for the molecular ground-state energy satisfy (for a proof, see Parr and Yang, Section 7.2): —ty?— pze+ [2 @ Tie ne dry + v,.(1) |OS(1) = e051) (15.121) where the function 2¢(1) is defined by (15.124), From (15.114) and (15.112), alterna- tive ways to write (15.121) are (BV? + v(1)]08S(1) = efS0%S (a) (15.122) ANS (1)08S(1) = efS0¥S (1) (15.123)Chapter 15 Ab Initio and Density-Functional Treatments of Molecules The exchange-correlation potential v,. is found as the functional deriva exchange-correlation energy E,.: dE scl lt) Sle) The precise definition of the functional derivative need not concern us (see Yang, Appendix A). The following formula allows one to find the function tive of most functionals that occur in DFT. For a functional defined by Vae(F) = F[p] = [fas Ys Zs Ps Pas Py pz) dx dy dz where p is a function of x, y, and z that vanishes at the limits of the integral, Px = (8p/2x),.. etc, the functional derivative can be shown to be given by 5F/ 5p = ag/ap — (a/dx)(ag/ap,) ~ (0/ay)(4g/ép,) — (8/4z\ag/ap.) If E,<[p] is known, its functional derivative is readily found from (1 (15.125), and so %- is known. The functional E,«{p] in (15.120) is a numb tional derivative of E,.[p] is a function of p (see Problem 15.62 for an since pis a function of r, 2,- is a function of r, that is, of x, y, and z. Sometins write v,¢ aS v,,(p(r)). [In the Kohn-Sham eigenvalue equation (15.123), the taken as r; rather than as r.] ‘The one-electron Kohn-Sham operator A*S(1) in (15.123) is the same as operator (15.82) in the Hartree-Fock equations except that the exchange ~Xj.1k; in the Fock operator are replaced by 2,. (Problem 15.63), which effects of both exchange (antisymmetry) and electron correlation. ey There is only one problem in using the Kohn-Sham method to find one knows what the correct functional £,.[p] is. Therefore, both £,, in expression (15.120) and v,¢ in (15.121) and (15.124) are unknown. Various tions to E,, will be discussed shortly. The Kohn-Sham orbitals 6° are orbitals for the fictitious reference noninteracting electrons, so, strictly speaking, these orbitals have no physic cance other than in allowing the exact molecular ground-state p to be caleu (15.119). The density-functional (DF) molecular wave function is not a Sla nant of spin-orbitals. In fact, there is no DF molecular wave function. Hoy tice, one finds that the occupied Kohn-Sham orbitals resemble molecula calculated by the Hartree-Fock method, and the Kohn-Sham orbitals can as Hartree-Fock MOs are used) in qualitative MO discussions of molecula and reactivities (see E. J. Baerends and O. V. Gritsenko, J. Phys. Chem. (1997); Gritsenko et al.,/. Chem. Phys., 107, 5007 (1997)]. (Note that, strict! Hartree-Fock orbitals also have no physical reality, since they refer to « model system in which each electron experiences some sort of average & other electrons.) For a closed-shell molecule, each Hartree-Fock occupied-orbital ener approximation to the negative of the energy needed to remove an electro orbital (Koopmans’ theorem). However, this is not true for Kohn-Sham orSection 15.20 Density-Functional Theory 581 gies. The one exception is eS for the highest-occupied KS orbital, which can be proved to be equal to minus the molecular ionization energy. (With the currently used approx- imations t0 E,q ionization energies calculated from KS highest-occupied-orbital ener- gies agree poorly with experiment.) Various approximate functionals E,,[p] are used in molecular DF calculations. To study the accuracy of an approximate F,,[p], one uses it in DF calculations and compares calculated molecular properties with experimental ones. The lack of a sys- tematic procedure for improving E,,[p] and hence improving calculated molecular properties is the main drawback of the DF method. In a “true” density-functional theory, one would deal with only the electron den- sity (a function of three variables) and not with orbitals, and would search directly for the density that minimizes E,[p]. Because the functional E, is unknowa, one instead uses the Kohn-Sham method, which calculates an orbital for each electron. Thus, the KS method represents something of a compromise with the original goals of DFT. The exchange-correlation energy E,, in (15.117) contains the following com- ponents: the kinetic correlation energy {the AT term in (15.117), which is the difference in T for the real molecule and the reference system of noninteracting clectrons; Eq. (15.115)], the exchange energy (which arises from the antisymmetry requirement), the Coulombic correlation energy (which is associated with interelectronic repulsions), and a self-interaction correction (SIC). The SIC arises from the fact that the classical charge-cloud electrostatic-repulsion expression 3 ffo(t,)p(ts)riz dr dr, in (15.116) erroneously allows the portion of p in dr, that comes from the smeared-out part of a particular clectron to interact with the charge contributions of that same electron to p throughout space. In actuality, an electron does not interact with itself. (Note that for a one-electron molecule, there is no interelectronic repulsion, but the expression Lffolr)o(rritdr dr, erroneously gives an interelectronic repulsion.) The kinetic energy T, of the reference system turns out to be close to T of the real molecule, and AT/T is small. However, the contribution of AT to £,- in (15.117) is not negligible. ‘The Local-Density Approximation (LDA). Hohenberg and Kohn showed that if p varies extremely slowly with position, then £,,{p] is accurately given by EPA(p) = [plelautn) a (15.126) where the integral is over alll space, dr stands for dx dy dz, and e,,(p) is the exchange plus correlation energy per electron in a homogeneous electron gas with electron density p. Jellium is a hypothetical electrically neutral, infinite-volume system consist- ing of an infinite number of interacting electrons moving in a space throughout which positive charge is continuously and uniformly distributed; the number of electrons per unit volume has a nonzero constant value p. The electrons in the jellium constitute a homogeneous (or uniform) electron gas. Taking the functional derivative of E:P*, one finds [Eqs. (15.124) and (15.125)] See exlole)) + o(e) A as.27) LDA yldA = xe 3p ap582 Chapter15 Ab Initio and Density-Functional Treatments of Molecules Kohn and Sham suggested the use of (15.126) and (15.127) as approxim and »,. in (15.120) and (15.121), a procedure that is called the local der mation (LDA). One can show that e,, can be written as the sum of exch relation parts: £,c(p) = €,(e) + &() where i edo) = ~3(2)"oony® The correlation part e,(p) has been calculated and the results have been ex; very complicated function e¥¥% of p by Vosko, Wilk, and Nusair (VWN); Yang, Appendix E; S. H. Vosko, L. Wilk, and M. Nusair, Can. J. Phys., 58, ‘Thus ep) = e¢*N(p) where eW*™ is a known function, From (15.124), (15.125), (15.126), (15 (15.129), we get (Problem 15.64) DADA = oO 4 DEA, EPA = — [(3/ar)p(n}]2, ofA = Ey [pozde = ~3(3)" [ioer*ar ‘The method for finding the LDA quantities e, and ¢, is as follows (for Parr and Yang, Appendix E). Consider a uniform electron gas (UEG) ¥ where k is some constant value. As noted after (15.125), se = Vee(P(8)), a constant for a particular UEG, 0,. is a constant for a particular UEG. (Of cos have different values for two UEGs with different electron densities) In th equations (15.121) for the reference system that corresponds to the UEG, the « can be omitted without affecting the eigenfunctions (Problem 4.49). Also, second term in brackets in (15.121) (the external potential) must be rep attraction between electron 1 and the uniformly distributed positive charge EG is electrically neutral, the positive charge density equals the electron de second and third terms in brackets in (15.121) cancel. Thus, the term —} surviving term in ASS for the UEG. The UEG KS orbitals can thus be ta dimensional free-particle wave functions Aes"**)**=) [recall (2.30), ¥ of A is chosen to give the desired electron density in (15.119). Because the Ul trically neutral in cach region of space, the sum of the electrostatic repulsia the smeared-out electrons {the third term on the right side of (15.120)], the between the smeared-out electrons and the continuous positive charge dis first term on the right of (15.120) modified to correspond to continuous p. and the repulsions between parts of the positive charge distribution [ana internuclear repulsion term to be added to (15.120)] adds to zero. This leav side of the energy expression (15.120) only the E,. term and the kineti T., which is readily evaluated from the known KS orbitals. Breaking E,, inte E, and E, (Eq, (15.133)], one evaluates E, from (15.134) and the KS orbitSection 15.20 Density-Functional Theory 583 result shown in (15.132). This leaves only E, as unknown, One then does an accurate nu- merical solution of the UEG Schrédinger equation to find the energy for the particular density p = k, Combining this energy with the already calculated KS energy terms gives the unknown £, for that p. Repetition of the entire procedure for many density values gives the UEG E, as a function of p. From E, and £,,we find &, and &,. ‘The Functionals E, and E,. As an aid to developing approximate functionals for use in KS DFT, the functional £,, is written as the sum of an exchange-energy ‘functional F,, and a correlation-energy functional E,: Ege = Ex + Ee (15.133) E, is defined by the same formula used for the exchange energy in Hartree-Fock the- ory, except that the Hartree-Fock orbitals are replaced by the Kohn-Sham orbitals. ‘The Hartree-Fock exchange energy of a closed-shell molecule is given by the terms in (13.145) that involve the exchange integrals Ky. Replacing the Hartree-Fock orbitals by the Kohn-Sham orbitals in (13.147), we have for a closed-shell molecule £,= FS S@saesali/ndosmase) (15.134) mf where the factor of } comes from the fact that in (13.145) we are summing over the orbitals, whereas in (15.134) we are summing over the electrons, which gives four times as many terms in the double sum in (15.134) (Problem 15.65). Since, in practice, KS orbitals are found to rather closely resemble Hartree-Fock orbitals, the DFT exchange energy is close to the Hartree-Fock exchange energy. Having defined E,, the correlation-energy functional £, is defined as the difference between E,, and E,; that is E, = E,, ~ E,,and (15.133) follows. When E, is evaluated using the definition (15.134) and E, is evaluated by one of the currently available models (such as the LDA), one obtains poor results for molecular properties. Thus, in practice it is best to model both E, and F,, because this leads to cancellation of errors and better results. One therefore uses the LDA (or one of its improved versions discussed later) to find both E, and E. Both E, and E, are negative, with |E,| being much larger than |E,|. The definition of E,in DFT differs from the definition (11,16) of the correlation energy in Hartree-Fock theory, but analysis and calculations show that these two quantities are nearly equal [E. K.U. Gross et al. in B. B. Laird et al. (eds.) Chemical Applications of Density Functional Theory, American Chemical Society, 1996, Chapter 3]. Starting with accurate electron densities from MRCI wave functions for Li,, N>, and F;, Gritsenko and co-workers used an iterative procedure to calculate KS orbitals and the KS quantities E, and E, for each of these molecules at three internuclear distances [O. V. Gritsenko et al., J. Chem. Phys., 107, 5007 (1997)]. The KS E, and E, values were found to be very close to the corre- sponding Hartree-Fock (HF) values ES"and E!Fat the equilibrium internuclear distances, but agreement between these quantities decreased as the internuclear distances increased. For N, at R,, values in hartrees for these quantities and for quantities in (15.115) are: E, = 13.114, EY? = 13.008; E, = —0.475, EG* = —0.469; AT = 0.329, T = 109.399. The correlation kinetic energy AT is a significant part of E, ‘The Xa Method. The following approximation for E,-gives the Xa method (the X stands for exchange). Here, the correlation contribution to E,, is omitted (it is584. Chapter 15 Ab Initio and Density-Functional Treatments of Molecules substantially smaller in magnitude than the exchange contribution) and the & contribution is taken as eae = -2(2)"elipeitar where a is an adjustable parameter; a values from § to 1 have been used. Fus differentiation of (15.135) [Eqs. (15.124) and (15.125)] gives the Ka exchange p as 08" = ~(3a/2)(3p/m)". Note that with a = § the Xa expression (15.1 becomes equal to the exchange part (15.132) of the LDA E,, expressio method gives rather erratic results in molecular calculations and has been s by better approximations to E,,. The Xa method was developed by Slater pri work of Hohenberg, Kohn, and Sham, and was viewed by Slater as an appro to the Hartree-Fock method. It is sometimes called the Hartree-Fock-Slater = However, the Xa method is best viewed as a special case of DFT. Performing Kohn-Sham Density-Functional Calculations, How does molecular density-functional calculation with E4P4 (or some other function starts with an initial guess for p, which is usually found by superposing « clectron densities of the individual atoms at the chosen molecular geometry. initial guess for p(r), an initial estimate of v,.(r) is found from (15.127) and (I this initial v,,(r) is used in the Kohn-Sham equations (15.121), which are sol initial estimate of the KS orbitals. In solving (15.121), the 0s are usually exp terms of a set of basis functions y, (08° = ¥/_1¢,;x,) to yield equations that = the Hartree-Fock—Roothaan equations (13.157) and (13.179), except that matrix elements F,, = (y,|F| x,) are replaced by the Kohn-Sham matrix € AKS = (y, [A y,), where A¥S is in (15.122) and(15.123). Thus, instead of (13.1: DFT with a basis-set expansion of the orbitals, one solves the equations " ® SeGe-sg=0, .F= 1,2. The basis functions most commonly used in molecular KS DFT caler contracted Gaussians, but some DF programs use STOs or still other basis ‘The KS equations can also be solved numerically, without using a ba: expansion of the orbitals, but this choice is not often used. ‘The initially found 6's are used in (15.119) to get an improved electron which is then used to find an improved »,., which is then used in the KS (15.121) to find improved KS orbitals, and so on. The iterations continue no further significant change in the density and the KS orbitals. KS DFT cal involve iterations until self-donsistency between the exchange-correlation Y,< and the KS orbitals 6 in (15.121) has been reached. They are thus a Ki calculation. Once the calculation has converged, the ground-state energy E, in found from the converged p and E\PThe dipole moment can be calcul: using (13.144), and other one-electron properties can be found from (13.131 gradients of the energy have been developed for KS DFT calculations, so the rium geometry is readily found using one of the methods of Section 15.11.Section 15.20 Density-Functional Theory 585 second derivatives of the KS DFT energy are available, and DF vibrational frequen- cies are readily calculated. One significant difference between KS DFT calculations and Hartree-Fock cal- culations arises from the fact that v£P4(r) and versions of v,, more accurate than the LDA are very complicated functions of the coordinates. This makes it impossible to analytically evaluate the integrals (y, |vxd x,), which occur in AES. Instead, (y, ved x,) is evaluated numerically by evaluating the integrand at each point of a grid of points in the molecule and performing a summation. [An alternative approach is to expand v,.(r) using an auxiliary set of basis functions (not the same set used to expand the orbitals), where the expansion coefficients are chosen to give a good least-squares fit to values of v,.(r) evaluated at a grid of points.] DF calculations that use a basis-set expansion of the KS orbitals must deal with the same Coulomb matrix elements J,, [Eq. (15.13)] that occur in Hartree-Fock calcu- lations. Therefore, such DF calculations can be speeded up and linear scaling obtained for large molecules by the same techniques used to speed up Hartree-Fock calcula- tions, namely, direct and semidirect methods, neglect of integrals smaller than a thresh- old value, the continuous fast multipole and quantum-chemical tree-code methods, the J-matrix engine. For details, see the references in Section 15.5. Also the conjugate- gradient density-matrix search method can be used to avoid diagonalizing the Kohn— Sham matrix. Deviations of KS DF calculated results from the true values are due to use of approximate E,. and vq, expressions, and to basis-set inadequacies. See Chapter 17 for detailed comparisons. The programs Gaussian, Q-Chem, Jaguar, ACES II, Turbomole, MOLPRO, CADPAC, and SPARTAN (Section 15.15) all contain KS DFT modules. Some pro- grams that only do KS DF calculations are ADF (wwwscm.com/), deMon (www. cerca.umontreal.ca/deMon/), DeFT (www.chem.uottawa.ca/DeFT.html), and Dgauss (www.oxmol.co.uk/prods/unichem/cap/Dgauss.html). ‘The technology of DF calculations is not as mature as that of HF calculations, and because of the variety of possible procedures used, DF calculations done with two different programs using the same E,, and the same basis set may not give quite the same results. ‘The Local-Spin-Density Approximation (LSDA). For open-shell molecules and molecular geometries near dissociation, the Jocal-spin-density approximation (LSDA) gives better results than the LDA. Whereas in the LDA, electrons with opposite spins paired with each other have the same spatial KS orbital, the LSDA allows such electrons to have different spatial KS orbitals 6X5 and 64S; the LSDA is thus analogous to the UHF method (Section 15.3), which allows different spatial Hartree-Fock orbitals for electrons with different spins. The theorems of Hohenberg, Kohn, and Sham do not require using different orbitals for electrons with different spins (unless an external magnetic field is present), and if the exact functional E,,[p] were known, one would not do so. With the approximate £,, functionals that are used in KS DFT calculations, it is advantageous to allow the possibility of different orbitals for electrons with different spins, so as to improve calculated properties of open-shell species and species with geometries near dissociation.586 Chapter 15 —_Ab Initio and Density-Functional Treatments of Molecules ‘The generalization of density-functional theory that allows different electrons with different spins is called spin-density-functional theory (Parr Chapter 8). In spin-DFT, one deals separately with the electron density p%(r) spin-e electrons and the density p*(r) of the spin-f electrons, and functionals E,¢ become functionals of these two quantities: E,. = E,.[p%, p®]. One deals w= rate Kohn-Sham eigenvalue equations for the spin-a orbitals and the spin-S where these equations have v%. = SE xeo*, o°]/8p* and similarly for vé.. For like CH; or the O, triplet ground state, the number n® of a electrons differs number of electrons, so here p + p*, and spin-DFT will give different electrons with different spins. Gunnarsson and Lundqvist did an LSDA spin-DFT calculation of the cule, expanding the occupied KS orbitals using the 1s, and 1s, AOs as basis For internuclear separations up to 3.2 bohrs, they found the lowest energy KS to be ofS = ofS = N(Ls, + 1s,). However, for internuclear separations 3.2 bohrs, they found the lowest-energy KS orbitals to be 65 = N(1s, = OSS = N(cls, + 15,), where c <1 and c decreased to zero as the internuch tion increased to infinity. Having the spin-a electron’s KS orbital differ ti spin-f electron allowed the H, molecular energy versus internuclear separ. to show the proper dissociation behavior, corresponding to dissociation to gen atom with a spin-a electron and a second H atom with a spin-6 electron | that the RHF wave function of H; dissociates improperly. The UHF H, wave & shows proper dissociation (Szabo and Ostlund, Section 3.8.7).] As in the UHF method, allowing differing KS orbitals for electrons with spins can produce a wave function for the reference system s that is not an tion of $2, but this spin-contamination is less of a problem in KS DFT than in. method, - For species with all electrons paired and molecular geometries in the the equilibrium geometry, we can expect that p* = p®, and spin-DFT will re ordinary form of DFT. Despite the fact that p in a molecule is not a slowly varying function of! the LSDA works surprisingly well for calculating molecular equilibrium vibrational frequencies, and dipole moments, even for transition-metal where Hartree-Fock calculations often give poor results. (For detailed Chapter 17.) However, calculated LSDA molecular atomization energies inaccurate. Accurate dissociation energies require functionals that go beyond Gradient-Corrected and Hybrid Functionals. The LDA and LSDA are the uniform-electron-gas model, which is appropriate for a system where slowly with position. The integrand in the expression (15.126) for ESP is a only p,and the integrand in EES? is a function of only p* and p*. Functios beyond the LSDA aim to correct the LSDA for the variation of electron position. They do this by including the gradients [Eq. (5.31)] of p% and = integrand. Thus BAP =| 10°. VOTHeee ee ee ee Section 15.20 Density-Functional Theory 587 where f is some function of the spin densities and their gradients. The letters GGA stand for generalized-gradient approximation. The term gradient-corrected func- tional is also used. (Gradient-corrected functionals are often called “nonlocal” functionals but, strictly speaking, this is a misuse of the mathematical meaning of non- local.) ES is usually split into exchange and correlation parts, which are modeled separately: E96A = EGA 4 G68 (15.138) Approximate gradient-corrected exchange and correlation energy functionals are developed using theoretical considerations such as the known behavior of the true (but unknown) functionals E, and £, in various limiting situations as a guide, with, per- haps some empiricism thrown in. Some commonly used gradient-corrected exchange functionals E, are Perdew and Wang’s 1986 functional (which contains no empirical parameters), designated PW86 or PWx86, Becke’s 1988 functional, denoted B88, Bx88, Becke88, or B, and Perdew and Wang’s 1991 exchange functional PWx91. The explicit form of the B88 exchange functional is EBS = ELPA gS | (15.139) omaa 1 + 6bx, sil where x, = |Vo"|/(o"), sinh”! x = In[x + (x? + 1)!7], 0 is an empirical parameter whose value 0.0042 atomic units was determined by fitting known Hartree-Fock exchange energies (which are close to KS exchange energies) of several atoms, and [see (15.132) and Problem 15.66] 3/(6\? geen =~ 2(8)" [ne + erat (15.140) The PWx86 functional (which has no empirical parameters) and the B88 functional work about equally well in predicting molecular properties. Commonly used gradient-corrected correlation functionals £, include the Lee— Yang—Parr (LYP) functional, the Perdew 1986 correlation functional (P86 or Pc86), the Perdew—Wang 1991 parameter-free correlation functional (PW91 or PWc91), and the Becke correlation functional called Bc95 or B96. ‘The Perdew-Burke-Ernzerhof (PBE) exchange and correlation functional has no empirical parameters [Phys. Rev. Lett, 77,3865 (1996)]. Some E, and E, values in hartrees for the Ar atom are (A. D. Becke in Yarkony, Chapter 15): EHF = —30.19, ELS>* = —27.86, E8** = —30.15; EMF = —0.722, E}S°* = —1.431, E®¥°! = 0,768, where the Hartree-Fock (HF) values should be good esti- mates of the KS DFT values. Any exchange functional can be combined with any correlation functional. For example, the notation BLY P/6-31G* denotes a DF calculation done with the Becke 1988 exchange functional and the Lee-Yang-Parr correlation functional, with the KS orbitals expanded in a 6-31G* basis set. The letter $ (which acknowledges Slater’s Xa method) denotes the LSDA exchange functional (15.140). VWN denotes the Vosko-Wilk-Nusair expression for the LSDA correlation functional (actually, these workers gave two differ-588 Chapter15 Ab Initio and Density-Functional Treatments of Molecules ent expressions for E{S>*, which are sometimes referred to as VWN3 and an LSDA calculation can be denoted by the letters LSDA or by SVWN. Hybrid exchange-correlation functionals are widely used. A hybrid mixes together the formula (15.134) for E, with gradient-corrected E, and E, For example, the popular B3LYP (or Becke3LYP) hybrid functional (where the 3 cates a three-parameter functional) is defined by EY? = (1 — aq — a,)EMSPA + age + a, EBS + (1 — a JEYS + a EO? where ES" (which is sometimes denoted £'", since it uses a Hartree-Fock de of E,) is given by (15.134), and where the parameter values ay = 0.20, a, = 0.72. 2m 0.81 were chosen to give good fits to experimental molecular atomization ener B3PW91 hybrid functional replaces £1? in (15.141) with E°™", and uses values, Becke’s one-parameter hybrid functional 196 (also called B1B95) is EBS . £8 4 q(Ze3* — E®), where the empirical-parameter value a found by fitting atomization energies. As an improvement on the B3LYP, B3PW91, and B1B96 hybrid funet Becke [A. D. Becke, J. Chem. Phys., 107, 8554 (1997); H. L. Schmider and Chem. Phys., 108, 9624 (1998)] proposed the hybrid functional Ey = EQS + cE + EGOA where ¢, is a parameter and E9C* and £9 are certain GGA functionals that. three and six parameters, respectively. The values of the 10 parameters in determined as the set that gave the best fit to experimental energy data in the C set (Section 15.21). Using a numerical method to solve the Kohn-Sham equati as to avoid basis-set truncation error), Becke found that the functional (15.14 mean absolute error (MAE) of only 1.8 kcal/mol for 55 atomization energies. a: cant improvement over the functional B3PW91, which had an MAE of 2.4 kea these energies. However, Becke concluded that the great flexibility of the (15.142) “implies that the limits of accuracy of the GGA/exact-exchange fra have been reached... . It appears that [a further increase in accuracy] gained by continued experimentation with particular GGAs, but will require ne insights and perhaps higher-order density derivatives.” Several other functionals containing parameters fit to experimental molect have been proposed. The 2l-parameter Van Voorhis-Scuseria exchange-co functional VSXC has no mixing of exact exchange and performs slightly bets B3LYP for calculating atomization energies but slightly worse for bond lengths Voorhis and G. E. Scuseria, J. Chem. Phys., 109, 400 (1998)]. The nine param EDF! (empirical density functional version 1) were optimized specifically for us the rather small 6-31G* basis set [R. D. Adamson, P. M. W. Gill, and J. A. Pople Phys. Lett., 284, 6 (1998)]. EDF performs rather well for calculating atomizati gies and was not significantly improved by mixing in exact exchange. Katfafi devised the hybrid K2-BVWN (standing for Kafafi 2-parameter Vosko-Wilk-Nusair) exchange-correlation functional [S. A. Kafafi, J. Phys. G 102, 10404 (1998)]. K2-BVWN/6-311+G(2df) calculations of AH729g of 350 an MAE of only 1.4 kcal/mol, comparable to results from the computation more demanding G2 method (Section 15.21).Section 15.20 Density-Functional Theory 589 Going beyond the GGA, Becke introduced a 10-parameter exchange-correlation functional where f in (15.137) is a function of not only the densities and their gradients but also of V2p,,V ps, and the gradients of the KS orbitals [A. D. Becke, J. Chem. Phys., 109, 2092 (1998); H. L. Schmider and A. D. Becke, J. Chem. Phys., 109, 8188 (1998)]. This functional had a MAE of 1.54 keal/mol for 55 atomization energies as compared with 1.79 kcal/mol for Becke’s 10 parameter GGA functional, but the use of higher derivatives of p incteases the computational time required and Becke concluded that “it remains to be seen if higher-order DFT will offer sufficient advantages to supplant the GGA.” Gradient-corrected functionals and hybrid functionals give not only good equi- librium geometries, vibrational frequencies, and dipole moments, but also generally accurate molecular atomization energies. For example, BLYP/6-311+G(2d,p) and B3LYP/6-311+G(2d,p) calculations on the G2 data set gave MAEs of 3.9 and 3.1 kcal/mol, respectively (see Foresman and Frisch, Chapter 7). Overall, the hybrid func- tionals seem to give the best performance. See Chapter 17 for details. In doing DF calculations with functionals that go beyond the LSDA, quantum chemists sometimes cheat a bit by solving the equations (15.121) for the KS orbitals using only the LSDA form of »,,. They then calculate the energy (and the energy gradi- ent if geometry optimization is being done) using (15.120) with the LSDA KS orbitals and electron density and the GGA or hybrid E,,. This procedure (called perturbative, since it resembles the perturbation-theory calculation of E® + E using the true Hamiltonian with the zeroth-order wave function) saves considerable computational time at the cost of rather smail errors in the DFT energy and molecular properties. However, it is preferable that the KS orbitals be found using the same functional that is used to calculate the energy. When this is done, one says the calculation has been performed self-consistenily. The Past and Future of DFT. Hohenberg, Kohn, and Sham’s work was published in 1964 and 1965. [For a personal account, see P. C. Hohenberg, W. Kohn, and L. J. Sham, Adv. Quantum Chem., 21,7 (1990).] Quickly thereafter, physicists applied KS DFT using the LSDA to study solids with considerable success, and DFT became the dominant method of doing quantum-mechanical calculations on solids. Chemists were rather slow to apply DFT to molecules, because of numerical difficulties in doing reliable DFT molecular calculations, the lack of widely available molecular DFT computational programs, and perhaps partly in reaction to disappointment with the Xa method, which had been overpraised by some of its practitioners. The numerical difficulties were largely solved around 1980, and LSDA DFT molecular calculations in the 1980s achieved good results for molecular geometries but failed to give accurate dissociation energies. The next major advance in DFT was the introduction of gradient-corrected func- tionals in the mid-1980s, which Becke found to give accurate dissociation energics. ‘Also in 1988, analytic gradients were implemented in DFT, greatly facilitating calcula- tion of equilibrium geometries. The 1989 DFT book by Parr and Yang helped draw the attention of quantum chemists to DFT. In 1993, provision for DF calculations was added to the popular program Gaussian. In the mid-1990s molecular DFT calculations experienced explosive growth. DFT has the advantage of allowing for correlation effects to be included in a cal-590 Chapter 15 Ab Initio and Density-Functional Treatments of Molecules culation that takes roughly the same time as a Hartree-Fock calculation, wh not include correlation. A 1996 review article [M. Head-Gordon, J. Phys. Che 13220 (1996)] gave the following rough estimates of the maximum number of nonhydrogen atoms in a molecule with no symmetry for which 10 energy and evaluations could be done on a high-end workstation using a DZP basis set 2 ous methods: FCI CCSD(T) CCSD MP2 HF 2 8to12 10 to 15 25 to 50 50 to 200 Whether KS DFT should be classified as an ab initio method is a debate. If the true E,, were known and used, then KS DFT would be an method. However, the true E,, is unknown and must be replaced by a model E, as the LSDA or the LSDA with gradient corrections. Some people would consi use of ELS? disqualifies KS DFT as being an ab initio method, but others v Some of the gradient-corrected functionals contain empirical parameters, 2 hybrid functionals, the mixing constant(s) are determined empirically. Use of als with empirically determined parameters clearly disqualifies a method as be initio, but the number of parameters used in these versions of DFT is far fe the number used in common semiempirical theories such as AM1 or PM3 16.5), which use different parameters for each kind of atom, which is not true # The KS DFT method is usually considered in a category by itself, distinct from methods such as HF, CI, MP, and CC. Despite its successes, DFT is not a panacea. Some drawbacks and failin; are the following. ‘ The Hohenberg-Kohn-Sham theory is basically a ground-state theory. of KS DFT applicable to excited states have been developed [see Parr a Section 9.2; K. Burke and E. K. U. Gross in D, Joubert (ed.), Density Funes Springer, 1998), but the thcory has not reached the point where it allows acc a tical calculations to be readily done on individual molecular excited states. (Ox use DFT to calculate the lowest state of each symmetry; for example, one can ca the lowest singlet state and the lowest triplet state.) Because approximate functionals are used, KS DFT is not variational, yield an energy below the true ground-state energy. For example, a B3P% geometry optimization of HO gives an energy of 76.60 hartrees, compared true nonrelativistic energy of —76.44 hartrees (Table 15.2). Calculations with corrected functionals are size-consistent. The true £,, contains a self-intcraction correction that exactly cancels t interaction energy in }ffp(r;)p(r,)ry'dr, dr, but currently used functionals completely free of self-interaction. Because of the self-interaction error, available functionals give very incorrect U(R) curves at large internuclear dis symmetrical radical ions such as H}, He}, and F3 and overestimate the interme interaction in some charge-transfer complexes [Y. Zhang and W. Yang, J. Che 109, 2604 (1998)]. ‘The currently available KS DFT functionals often do not give good res activation energies of reactions.Section 15.20 Density-Functional Theory 591 Although KS DFT yields good results for most molecular properties, with the presently available functionals, KS DFT cannot match the accuracy that methods like CCSD(T) and QCISD(T) can achieve. Of course, CCSD(T) and OCISD(T) are lim- ited to dealing with small molecules, whereas DFT can handle rather large molecules. With methods such as CC, CI, and MP, the way to achieve more accurate results is clear. One uses larger basis sets and goes to higher orders of correlation (CCSD, CCSDT,...; CISD, CISDT,...; MP2, MP3,...), although how far one can go for a given size molecule is limited by currently available computing power. In KS DFT, there is no clear way to construct more-accurate £,, functionals; one must try out new functionals one by one to see whether they will give improved results. Many of the currently used £,, functionals fail for van der Waals molecules. For example, the BLYP, B3LYP, and BPW91 functionals predict no binding in He, and Ne; However, the PBE functional works fairly well here [Y. Zhang et al., J. Chem. Phys., 107, 7921 (1997)], as does the K2-BVWN functional. Also, Adams and Barone modified the parameters in the PW91 exchange functional to give the modified Perdew-Wang (mPW) exchange functional and found that the hybrid functional Bye = O.TSER! + 0.25E25% + EE*! (called mPW1PW) performs rather well for He, and Ne, and works well for bond lengths, atomization energies, and vibrational frequencies of ordinary molecules [C. Adamo and V. Barone, J. Chem. Phys., 108, 664 (1998)]. Although gradient-corrected and hybrid functionals usually give good results for molecular properties, currently available forms of these functionals are known to be significantly in error. DFT shows that the exact E,, E,, v,, and v, must satisfy cer- tain conditions, and all currently available functionals violate at least some of these conditions [see C. Filipi et al., in J. M. Seminario (ed.), Recent Developments and Applications of Modern Density Functional Theory, Elsevier, 1996, p. 295]. The Hohenberg-Kohn theorem applied to the reference system of noninteracting electrons assures us that the true ground-state electron density determines the external potential v, in (15.122). Iterative methods have been devised that take a very accurate ground- state molecular electron density found from a high-level calculation (for example, CI) and use it to calculate v, for the corresponding reference system. From »,, we can use (15.121) and (15.122) to find v,- (r). The accurate »,, found from an accurate p for a particular atom or molecule can then be compared with the v,.’s found from currently used E,,’s. The results show that currently used »,,’s are substantially in error (M. E. Mura, PJ. Knowles, and C. A. Reynolds, J. Chem. Phys., 106, 9659 (1997); E. J. Baerends et al., in B. B. Laird et al. (eds.) Chemical Applications of Density-Functional Theory, American Chemical Society, 1996, Chapter 2]. Much current work is being done on the difficult job of developing improved Ey. functionals. Kohn and Sham took £,, as a functional of the density p. An alternative procedure, the optimized effective potential (OEP) method, takes E,, as a functional of the occupied KS orbitals, in the hope that this will make it easier to develop accurate functionals: The OEP method leads to equations that are hard to solve. Kreiger, Li and Iafrate (KLI) developed an accurate approximation to the OEP equations, thereby making them easier to deal with, and the KLI method has given good results in DF calculations on atoms [J. B. Krieger, Y. Li, and G. I. Iafrate in E. K. U, Gross and R. M. Dreizler (eds.), Density Functional Theory, Plenum, 1995, pp. 191-216}592 Chapter 15 Ab Initio and Density-Functional Treatments of Molecules 15.21 Some advocates of DFT believe that DFT will displace the method and Hartree-Fock based correlation methods (MP. CC, CI) and be dominant way of doing quantum-chemistry calculations and the main way cally interpreting chemical concepts. [DFT has been used to provide quantit: nitions of such chemical concepts as electronegativity, hardness and s reactivity; see Parr and Yang, Chapters 5 and 10 and W. Kohn, A. D. Becke Part, J. Phys. Chern., 100, 12974 (1996).] COMPOSITE METHODS FOR ENERGY CALCULATIONS A desirable goal is to compute a thermodynamic energy such as the molec ization energy or the enthalpy of formation, with chemical accuracy, which accuracy of 1 kcal/mol. Currently available functionais in DFT cannot do level methods such as CCSD(T), QCISD(T), CISDTO, and MP6 with large’ can do this but are much too costly to be feasible except for quite small mole aim of the compound methods G3 and CBS discussed in this section is to a kcal/mol accuracy with a computational time that allows calculations on = containing several nonhydrogen atoms ‘These compound methods use a series of ab initio calculations plus emp rections. The Gaussian-3 (G3) method (so-called because it is an improveme G1 and G2 methods) is designed to give a result close to what would be ob’ QCISD(T)/G3large calculation in much less computer time than required b calculation [L. A. Curtiss, K. Raghavachari, P. C. Redfern, V. Rassolov, and J. Chem. Phys., 109, 7764 (1998)]; G3large is an improved version of the 6-311 basis set. In the G3 method, the zero-point energy is found by scaling the fre found in a HF/6-31G* frequency calculation with the factor 0.893. All subseque culations are done at the optimized geometry found in an MP2/6-31G* ca One then computes a base energy £** from an MP4/6-31G(d) calculation. » corrections to £*** are then found as differences between £°** and energies from MP4/6-31+G(d), MP4/6-31G(2dfp), QCISD(T)/6-31G(d), and MP2 single-point calculations. These corrections are added to £°**, and an (“higher-level”) correction of —Artysic, — BYyapairea hartrees is added to con basis-set incompleteness. Here, npsie, is the number of valence electron pairs molecule and nyapaiea is the number of unpaired electrons; A and B are parameters with A = 0,0006386, B = 0.0002977 for molecules and A = B = 0.0001185 for atoms. The various corrections to E™ allow for the effeet cluding diffuse basis functions, polarization basis functions, and higher levels tron correlation. G2 [L. A. Curtiss et al., J. Chem. Phys., 94, 7221 (1991)], the predecessor uses larger basis sets in most of the single-point calculations, and so is slower # G3 uses a full MP2/G3large calculation, whereas the corresponding calculation. a frozen-core MP2 calculation. G3 includes spin-orbit corrections for atomic e: which G2 did not, G2 uses the same higher-level correction for atoms as for 1 The MP4 calculations are the most time-consuming steps in the G2 and G3 ods and limit their applicability to rather small molecules. The G2(MP2) and