Where Do Bands Come From? — From Molecular Orbitals to Crystal Bands

— From the 2×2 matrix of H2 to Bloch’s theorem, the density of states, the origin of the gap, and finally s, p, d, and f orbitals on a two-dimensional square lattice (for second- and third-year undergraduates)

The energies of electrons in a solid are not discrete levels but form continuous bands. In lectures this fact tends to be handed to you together with Bloch’s theorem, as something to accept without derivation. In fact, simply extending the story of the bonding and antibonding orbitals of the H2 molecule is enough to derive the band, Bloch’s theorem, and the gap, all by hand. The mathematics we use is only quadratic equations and the addition theorems of trigonometric functions (§11 and §12, which treat d and f, add a counting of orbital symmetries to this).

The conclusion first Lining up N atoms gives exactly N levels. All of them fit inside the window α±2|β|, so for N → ∞ they become a continuous band of width 4|β|. The width is determined only by the overlap between neighboring orbitals (the resonance integral β). And the moment a “second kind” appears among the atoms or the bonds, a gap Eg=2(Δ/2)2+(β1−β2)2 opens at the zone boundary, and a metal turns into an insulator.

0. Setup — The Model We Use (Tight-Binding / Hückel Model)

Place N atoms in a row at equal spacing a, and put one s orbital φj (j = 1, …, N) on each atom. Write the crystal wave function as a superposition of atomic orbitals (LCAO),

(1) ψ= ∑j=1N cjφj

We keep only the following two matrix elements of the Hamiltonian.

(2) ⟨φj|H|φj⟩ =α, ⟨φj|H|φj±1⟩ =β(<0)
SymbolMeaningIn a word
αEnergy of the atomic orbital itself (Coulomb integral)The level of the isolated atom. May be taken as the energy zero
βHopping to the neighboring atom (resonance integral, hopping)Negative for s orbitals. The larger the orbital overlap, the larger |β|
NNumber of atomsOn the order of 107/cm in a crystal
aSpacing between atoms (lattice constant)Typically 2–4 Å

These are the only approximations: hopping only between nearest neighbors (we discard the matrix elements between atoms two sites apart, i.e., every other atom, such as ⟨φj|H|φj±2⟩ and so on), and the overlap integral between orbitals on different atoms is taken to be 0 (⟨φi|φj⟩=δij). In chemistry this is called the “Hückel method,” and in physics the “tight-binding model.” Writing the Schrödinger equation Hψ=εψ in components, the j-th row reads

(3) αcj+ β(cj−1+cj+1) =εcj

This single equation (a three-term recurrence relation) is the star of this page, and everything below is just solving this equation.

1. First, the Intuition (The 30-Second Version)

Below, we put this into equations properly.

2. Two Atoms — Bonding and Antibonding in H2

For N = 2, Eq. (3) gives two equations, for j = 1, 2 (c0=c3=0). The condition for a nontrivial solution is the secular determinant

(4) | α−εβ βα−ε | =(α−ε)2−β2=0

and the solutions are

(5) ε±=α±β , ψ±= 12 (φ1±φ2)
Watch the sign Since β < 0, the lower level is α+β (the in-phase φ1+φ2: the bonding orbital, 0 nodes), and the higher level is α−β (the out-of-phase φ1−φ2: the antibonding orbital, 1 node). The spacing between the two levels is 2|β|.

3. Three Atoms — A Hint of a Law

For N = 3, the secular determinant is a 3×3 tridiagonal matrix, and expanding along the first row gives

(6) | α−εβ0 βα−εβ 0βα−ε | = (α−ε) [(α−ε)2−2β2] =0

so there are three solutions:

(7) ε=α+2β ,α ,α−2β

The coefficients (before normalization) are, from the bottom up,

(8) (1,2,1) , (1,0,−1) , (1,−2,1)

The numbers of nodes are 0, 1, and 2 — just as intuition says. In the middle level the coefficient of the central atom is 0, and its energy stays at α (a nonbonding orbital). Now look closely at Eq. (7): 2=2cos⁡(π/4), 0=2cos⁡(2π/4), so all three levels can be written as ε=α+2βcos⁡(mπ/4). Eq. (5) for N = 2 takes the same form, because 2cos⁡(π/3)=1. There is a hint of a general law here.

4. General N — Solving with a Single Recurrence Relation [The Core]

We rewrite Eq. (3) in the form of a recurrence relation.

(9) βcj+1 +(α−ε)cj +βcj−1 =0

Only the end atoms (j = 1 and j = N) have a neighbor on just one side, so the form of the equation seems bound to differ, but by setting the coefficients of the nonexistent atoms to 0 we can bring every j under the same Eq. (9):

(10) c0=0 , cN+1=0

These boundary conditions are just like those of a vibrating string (amplitude 0 just outside both ends). So, as for the string, we make a sinusoidal ansatz.

(11) cj=sin⁡(jθ)

This automatically satisfies c0=0. To substitute into Eq. (9), we need just one identity that follows from the addition theorem:

(12) sin⁡((j+1)θ) + sin⁡((j−1)θ) =2sin⁡(jθ)cos⁡θ

Using this, Eq. (9) becomes

[2βcos⁡θ +(α−ε)] sin⁡(jθ)=0

and the condition for this to hold for every j is that the bracket vanish (= 0), that is,

(13) ε=α+2βcos⁡θ

What remains is the other boundary condition cN+1=sin⁡((N+1)θ)=0. This quantizes θ:

(14) θm= mπN+1 , m=1,2,…,N

To summarize, the exact solution for the eigenvalues and eigenvectors of the N-atom chain is:

(15) εm=α+2β cos⁡(mπN+1) , cj(m) = 2N+1 sin⁡(mjπN+1)

m = 0 gives all coefficients 0 (not a solution), and m = N + 1 also gives sin⁡(jπ)=0, so it is not a solution either; m > N + 1 only repeats solutions already found, so there are exactly N independent solutions. As a check, putting in N = 2, we find from 2cos⁡(π/3)=1 that ε=α±β (Eq. (5) ✓), and for N = 3, ε=α±2β,α (Eq. (7) ✓).

Checked numerically as well Eq. (15) has been checked against numerical diagonalization of the tridiagonal matrix for N = 2–13, and both the eigenvalues and the eigenvectors agree to machine precision (10−12 or smaller). Tab ① of the simulator is a direct plot of this equation.

5. N → ∞ — The Birth of the Band

From Eq. (15), we can read off all the properties of the band.

(a) Every level is inside the window. Since cos only takes values between −1 and 1, every level satisfies

(16) α−2|β| <εm< α+2|β|

The spread of the levels (highest minus lowest) is 4|β|cos⁡(π/(N+1)) and its limit as N → ∞ is the width 4|β| — increasing N does not increase the width.

(b) The spacing shrinks as 1/N. Estimating the spacing between neighboring levels near the center of the band (differentiating Eq. (15) with respect to m), we get

(17) Δε≈ 2π|β|N+1

For a chain with a = 3 Å and a length of 1 cm, N ≈ 3×107, so with |β| = 1 eV we get Δε ≈ 2×10−7 eV. This is far smaller than the thermal energy at room temperature (about 0.026 eV), so the levels form a practical continuum — and that is a band.

Fig. 1. The number of atoms N and the energy levels (Eq. (15), |β| = 1 eV). Starting from the bonding/antibonding pair at N = 2, the N levels crowd together inside the window α ± 2|β| (dashed lines). The width of the window does not depend on N.

Only β determines the width of the band N does not appear in the width 4|β|. Since |β| is set by the orbital overlap, spatially extended valence orbitals → large |β| → a wide band, core orbitals contracted close to the nucleus → |β| ≈ 0 → almost the atomic level, unchanged. The fact that “even in a solid the core levels are almost the same as in the atom, and only the valence electrons form bands” follows from this single line.

(c) Filling in the electrons gives a metal. For monovalent atoms (such as H or Na) there are N electrons. Filling the levels from the bottom with spin pairs, N/2 levels fill exactly half of the band. Right above, empty levels wait at an interval of Δε ≈ 10−7 eV, so an electric field can accelerate the electrons without limit — this is a metal. Please try tracing the continuous deformation from the “molecule” at N = 2, whose HOMO–LUMO gap was as large as 2|β|, to the “metal” with zero gap, with simulator ① for yourself.

When N is odd If an odd number of monovalent atoms are lined up, the number of electrons is also odd, so the highest occupied level holds only one electron, leaving one level without a spin partner. The property that becomes a “half-filled metal” as N → ∞ is already showing up in the finite chain. In simulator ①, this level is drawn as a line that is only half colored in.

6. Solving It Again as a Ring — Bloch’s Theorem

Solid-state physics textbooks use not an open chain but periodic boundary conditions. That is, the two ends of the chain are joined to form a ring, so that

(18) cj+N=cj

Because there are no ends any more, the solutions can be written not as standing waves but as traveling waves:

(19) cj= eikja

This is the one-dimensional version of Bloch’s theorem: when we move forward by one lattice site (j → j + 1), the wave function just acquires the phase eika and its shape does not change. Substituting into the recurrence relation (9), this time using eika+e−ika=2cos⁡ka we get

(20) E(k)= α+2βcos⁡ka

— this is the band dispersion. From the periodic boundary condition (18), eikNa=1, that is, the allowed values of k are

(21) k= 2πmNa , m=−N2+1,…,N2

(exactly N values), and the range −π/a<k≤π/a is the first Brillouin zone (for a k outside this range, k+2π/a gives the same cj, so it is not a new state).

How the open chain and the ring are connected The standing wave of §4, sin⁡(kja)=(eikja−e−ikja)/2i is, in other words, an equal superposition of the +k and −k traveling waves. The energy formulas (15) and (20) also have the same cos; the only difference is the spacing of k (π/(N+1)a or 2π/Na). Bulk properties (bandwidth, density of states, and so on) agree for N → ∞ with either boundary condition. That is why textbooks use the periodic boundary condition, which makes the calculation easier.

From the dispersion (20) we can also read off the dynamics. The velocity of a wave packet (the group velocity) is

(22) vg= 1ℏ dEdk =− 2βaℏ sin⁡ka

which is 0 at the bottom and top of the band (k = 0, ±π/a). Expanding cos to second order near the bottom gives E≈Ebottom+|β|a2k2 and, comparing with the free-electron ℏ2k2/2m*, we find

(23) m*= ℏ22|β|a2

— the effective mass. Putting in a = 3 Å and |β| = 1 eV gives m* ≈ 0.42 me. The wider the band (the larger |β|), the “lighter” the electrons — the relation used every day in semiconductor physics appears here. How the allowed k points fall onto the curve (20) can be seen in simulator ②.

7. Density of States — van Hove Singularities

Within the band the levels are not evenly spaced. We count “how many levels there are per unit energy.” By Eq. (21), the k values are equally spaced with spacing 2π/(Na), so the number of states in a width dk is Na/(2π)·dk. Since two values, +k and −k, correspond to the same energy,

(24) g(E) =2⋅ Na2π |dEdk|−1

Here |dE/dk|=2|β|a|sin⁡ka|; rewriting the sin in terms of E using (20) gives

(25) g(E)= Nπ 4β2− (E−α)2

The density of states diverges at the band edges E=α±2|β| — these are the van Hove singularities. Eq. (24) itself tells us why: at the edges dE/dk = 0, so the energy hardly changes when k is moved, and a great many states pile up at the same energy.

As a check, integrating over all energies gives ∫gdE=(N/π)⋅π=N — the total number of states is exactly N. At the center of the band, g(α)=N/(2π|β|).

The look changes with the dimension A divergence like 1/√ is a feature of one dimension only. In three dimensions, g(E)∝E−Eedge, so it rises from 0 at the band edge, and the divergence remains as a “kink in the slope.” The sharp peaks lined up in the optical absorption spectra of quasi-one-dimensional materials such as carbon nanotubes are these one-dimensional van Hove singularities actually being seen. In simulator ③, you can watch a histogram, made by actually counting a finite number N of levels, cling ever closer to Eq. (25).

8. The Origin of the Gap — The Moment Two Atoms Form a Pair

So far the chain has had identical atoms and identical bonds, so the unit cell was one atom. In this section we write the interatomic spacing as a0. Now we introduce a “second kind.” There are two ways to do it.

8.1 What happens first is a “doubling of the unit cell”

Whichever way we do it, one thing happens in common. It takes two atoms for the same pattern to repeat — that is,

(26) a=2a0 (unit cell doubled) ⟹ πa= π2a0 (zone halved)

When the period in real space gets longer, the zone in reciprocal space shrinks by the inverse factor. Then, of the states that were in the original zone |k|≤π/a0, the outer half sticks out of the new zone. The states that have lost their place are shifted by the reciprocal lattice vector 2π/a=π/a0 and folded back into the new zone. This is band folding (the folding up of the bands).

The equation of the folded-back branch is easy to write. By Eq. (20), the original band is E(k)=α+2βcos⁡ka0, so if we replace k→k+π/a0, we obtain

(27) E(k+πa0) =α+2β cos⁡(ka0+π) =α−2βcos⁡ka0

— this is the original branch turned upside down about E=α. At the boundary of the new zone k=π/2a0 we have cos⁡(π/2)=0, so these two branches both have E = α — that is, they cross exactly. This one point, where two degenerate states meet, is exactly where the gap opens.

So far, nothing has happened yet Note that the folding itself has not changed any physics. If all the atoms are the same, using a unit cell of 2a0 is merely our own choice (a way of choosing a supercell); we have only redrawn the bands, and not a single energy level has moved. A crossing remains a crossing, with no gap. Only when a second kind is really introduced is the degeneracy at this crossing lifted. In simulator ④, leave Δ and δ at 0 and you can see this directly.
A half-filled chain wants to dimerize In a chain with one electron per atom, the electrons fill half of the band, so the Fermi wave number is kF=π/2a0. As we just saw, when the unit cell is doubled, the new zone boundary falls exactly at the position of this kF. The states that go down when the gap opens are those near the boundary on the filled side only, while those that go up are on the empty side. Distorting the lattice costs elastic energy (proportional to the square of the size of the displacement δ), but the electronic gain is larger by a logarithmic factor, so as long as δ is small, the electronic gain always wins. Since the total energy goes down on balance, a one-dimensional half-filled metal gains by dimerizing spontaneously. This is the essence of the Peierls transition, and “why is the period doubled?” is the same question as “why 2kF?”

8.2 The crossing splits — the 2×2 matrix

By Bloch’s theorem, if we assume that the two components in the unit cell (uk,vk) are multiplied by eikna, the problem reduces to a 2×2 matrix for each k:

(28) H(k)= ( α−Δ/2 β1+β2e−ika β1+β2eika α+Δ/2 )

Since the eigenvalues of a 2×2 matrix are “mean ± √((difference/2)² + |off-diagonal|²),”

(29) E±(k)= α± (Δ/2)2 +β12 +β22 +2β1β2 cos⁡ka

The band has split into two. The quantity under the square root is largest at k = 0 ((Δ/2)2+(β1+β2)2), and smallest at the zone boundary k = ±π/a ((Δ/2)2+(β1−β2)2), so the spacing between the two bands is always smallest at the zone boundary, and the gap there is

(30) Eg=2 (Δ/2)2 +(β1−β2)2

Let us check the three limiting cases.

CaseGap EgExample material
Δ = 0, β1 = β2 (all the same) 0 Just the band of §6 folded into half the zone. Remains a metal
Δ ≠ 0, β1 = β2 (ionic) Δ Ionic crystals such as NaCl. Electrons shift toward the more electronegative side
Δ = 0, β1 ≠ β2 (bond alternation) 2|β1 − β2| Polyacetylene (Peierls transition, dimerization)
Metal → insulator With one electron per atom, there are 2 electrons per unit cell. This is (counting spin) exactly the capacity of the lower band. The moment the gap opens, the “single half-filled band” splits into a “full valence band + an empty conduction band,” and the metal becomes an insulator (semiconductor). A uniformly spaced one-dimensional metallic chain dimerizes spontaneously, opens a gap, and thereby lowers its energy — this is the Peierls transition.
Why does it open at the “zone boundary”? As we saw in 8.1, that is the point where the original branch and the folded-back branch cross, that is, the place where the energies of two states are exactly equal. Two degenerate states are split at first order by even the weakest perturbation, as long as the matrix element connecting them is nonzero (where there is no degeneracy only second-order effects appear, so the levels hardly move). In real space, the wave at the zone boundary is a standing wave with a wavelength of exactly 2a0, and there are two possibilities: “put the antinodes on the A atoms” or “put the antinodes on the B atoms.” If all the atoms are the same, these two have the same energy, but the moment A and B differ, the energy splits depending on where the antinodes are put — and the size of that splitting is the gap. In simulator ④, move Δ and δ away from 0 and watch the moment the crossing comes undone.

9. Where to Go from Here — Three Dimensions and Real Materials

Everything so far carries over directly to three dimensions. For a simple cubic lattice with only nearest neighbors kept, the recurrence relation (9) just splits into the three directions x, y, and z, and we get

(31) E(k)= α+2β ( cos⁡kxa+ cos⁡kya+ cos⁡kza )

— this is a band of width 12|β|. In real materials, each atom has several orbitals (s, p, d, …), so the 2×2 matrix of Eq. (28) grows into a matrix of size “number of orbitals × number of atoms,” and there are also several bands. Sorting out which orbitals can mix with which is the job of group theory (the symmetry of the crystal), and the tool that does the actual calculation and plotting, including this part, is CrystOD, developed in our laboratory. The CrystOD documentation contains an English version of the same “molecule → crystal” toy model as on this page, and orbital diagrams of real molecules and crystals (NH3 and ScF3).

The overall picture of band theory (the Fermi surface, phonons, and applications to specific heat) is covered in the textbooks “Solid-State Physics” and “Computational Materials Science” among the Advanced Resources. In the next section, §10, we go back a step to the two-dimensional square lattice and see how the shape of the bands changes with the orientation of the orbitals (s and p).

10. The Two-Dimensional Square Lattice — s and p Orbitals

In §9 we wrote that “it is the same in three dimensions,” but the two-dimensional square lattice, one step before that, has an important feature that can never be seen in one dimension: the orientation of the orbitals starts to matter. This section corresponds to lecture notes (8), (9), and (11) (“LCAO Theory”).

10.1 Going to two dimensions — just multiply

In a square lattice of lattice constant a, denote the atomic orbital of the atom that is n-th along the x axis and m-th along the y axis by χnm. Applying Bloch’s theorem in both the x and y directions, the wave function is

(32) Ψkx,ky = ∑n=1N ∑m=1N einkxa eimkya χnm

This is just Eq. (19) of the one-dimensional case multiplied by a phase in the y direction. After that we take the expectation value with the same conventions as in §0 (α for the same atom, β for nearest neighbors, 0 otherwise). The normalization is ∬Ψ*Ψdxdy=N2, and the numerator is N2(α+2β(cos⁡kxa+cos⁡kya)), so

(33) E(kx,ky) =α+2β ( cos⁡kxa + cos⁡kya )

x and y have separated cleanly. Since the number of nearest neighbors has increased from 2 to 4, the bandwidth doubles as well:

(34) W=8|β| (4|β| in 1D, 12|β| in 3D)
Verified numerically as well For an N×N square lattice with periodic boundary conditions (N = 4, 6, 8), we numerically diagonalized the matrix of dimension N2, and confirmed that the set of eigenvalues agrees with Eq. (33) evaluated at kx,ky=2πj/Na to machine precision (10−15).

10.2 How is a band “diagram” drawn? — Γ–X–M–Γ

Here a difficulty arises. In one dimension E(k) could be drawn as a single curve, but in two dimensions, E(kx,ky) is a surface. Drawn on paper it looks three-dimensional and is hard to read.

So textbooks follow only a representative path inside the Brillouin zone and reduce it to a single line. For the square lattice, the points are

Symbol(kx, ky)MeaningValue of the s band
Γ(0, 0)Center of the zone. All atoms in phaseα + 4β
X(π/a, 0)Midpoint of an edge. Only the signs along x alternateα
M(π/a, π/a)Corner. Checkerboard patternα − 4β

By symmetry, −π/a≤k≤0 and 0≤k≤π/a carry the same content, so it is enough to look at the positive side. Going around Γ → X → M → Γ passes through the bottom, middle, and top of the band. In simulator ⑤, this path is drawn on the contour map, and moving along it with the slider changes the phase pattern in real space at the same time.

What real space looks like at Γ, X, and M

10.3 p orbitals — with orientation, the overlap splits into σ and π

Now the real business of two dimensions begins. An s orbital is spherically symmetric, so it overlapped with every neighbor in the same way. A p orbital, however, has an orientation. The overlap with a neighbor then divides into two kinds.

Following the convention of the lecture notes, we assume βσ>0 and βπ>0 (note that this sign convention is different from β < 0 in §2). With only nearest neighbors kept in the square lattice, px, py, and pz do not mix with one another, and the three bands are independently given by

(35) Epx (kx,ky) =α +2βσcos⁡kxa −2βπcos⁡kya (σ along x, π along y) Epy (kx,ky) =α −2βπcos⁡kxa +2βσcos⁡kya (π along x, σ along y) Epz (kx,ky) =α −2βπcos⁡kxa −2βπcos⁡kya (out of plane; π for both)

The form of the equations is very similar to (33); the difference is that the σ term enters with +, and the π term with −. Since pz sticks out of the plane, it overlaps side by side in both x and y, so only π appears.

Why do σ and π enter with opposite signs? As we saw in §2 for one dimension, the basic rule is that the bonding combination is lower. Consider px orbitals lined up along x. At k=0 (all in phase), the adjacent heads face each other as + and −, so the state is antibonding. So the σ term raises the energy at k=0 — this is the reason for the + (the same story as “β > 0 for p orbitals” in lecture notes (4)). For π, with the sides next to each other, it is the opposite: k=0 is bonding, so it enters with −.
Independently checked against Slater–Koster Let (l,m,n) be the direction cosines of the bond direction. From the Slater–Koster formula Exx=l2Vppσ+(1−l2)Vppπ we built the 3×3 Bloch matrix and diagonalized it numerically on a 17×17 grid of points over the whole BZ. If we identify βσ=Vppσ and βπ=−Vppπ, the results agree with Eq. (35) to machine precision. In the square lattice the off-diagonal elements (the px–py hybridization) do not arise for nearest neighbors, so the three bands are independent — this too we confirmed by calculation.

10.4 Reading the three bands

Using the same values as in the lecture notes, α=10, βσ=2, βπ=1 (in eV), we list the values at the high-symmetry points.

Γ (0,0)X (π/a,0)M (π/a,π/a)What to notice
px1248 Drops sharply from Γ to X. The direction in which the σ overlap acts
py12168 Just px rotated by 90°. Highest at X
pz61014 Only π, so the same shape as s. Γ is the bottom

Three things can be read from this.

“The bottom of py is at M” is a statement about the path In the band diagram py appears lowest at the M point, but this is true only along the path Γ–X–M–Γ. Over the whole zone, the minimum of py is at the point (0,π/a) with a value of 4 eV, the same as that of px at the X point. This point moves to the X point under the interchange kx↔ky (a 90° rotation of the square lattice), so it is “another X point” that is symmetrically equivalent to X. It is a good example of the fact that a band diagram is only a cross section cut out of a surface. If you switch simulator ⑤ to its contour map, you can see that the patterns of px and py are related by nothing more than a 90° rotation.

For a three-dimensional simple cubic lattice, adding just one term cos⁡kza to Eq. (33) takes us back to Eq. (31), and the high-symmetry points are Γ = α+6β, X = α+2β, M = α−2β, R = α−6β (the lower half of lecture notes (9)).

11. d Orbitals — Three Kinds: σ, π, and δ

So far we have had s and p. For d and f the procedure is exactly the same. The only difference is that the kinds of “how it overlaps with a neighbor” increase, and once that is sorted out, the equations that follow look just like those of §10.

11.1 The kind of overlap is determined by m about the bond axis

Take the line joining two atoms as the axis, and classify the orbitals by the magnetic quantum number m about that axis. m=0 is σ, |m|=1 is π, |m|=2 is δ, and |m|=3 is φ. An s orbital has only σ, p has σ and π, d has σ, π, and δ, and f goes up to σ, π, δ, and φ. Only this part grows — and that is all there is to d and f.

In pictures, σ is head to head, π is side by side, δ is an overlap in which four lobes face each other, and φ is an overlap in which six lobes touch alternately. The higher |m| is, the more lobes there are, and the parts that touch become thinner.

What is important is that the signs of the two-center integrals are determined from the start. Taking the bond axis as z′, the real spherical harmonic with (l,m) transforms under z′→−z′ with the parity (−1)l+m. If the parity is even, the lobes facing each other have the same sign, and when multiplied by the attractive potential the integral is negative. If it is odd, the reverse holds. That is,

(36) Vm= (−1) l+m+1 βm, βm>0
σπδφ
s (l = 0)−
p (l = 1)+−
d (l = 2)−+−
f (l = 3)+−+−

The rule for p that we fixed by hand in §10, “the σ term enters with +, the π term with −,” is exactly the second row of this table. The β<0 for s is the first row as well. A single rule settles everything from s to f at once, so the rest is mechanical.

11.2 Which orbitals mix? — Decided by parity

For p in §10 the three bands were independent. For d and f this is not the case. However, we can tell in advance, before any counting, which partners mix.

Consider a bond along x. Two atoms lined up along it are symmetric under y→−y and also under z→−z. Therefore orbitals whose parities differ under these two reflections never mix. In the same way, a bond along y is symmetric under x→−x and z→−z. Imposing both, in the end only orbitals whose three parities (x, y, z) all coincide mix.

d orbitalxyzPartner
dxyoddoddevenAlone
dyzevenoddoddAlone
dzxoddevenoddAlone
dx²−y²eveneveneven2×2 with dz²
dz²eveneveneven2×2 with dx²−y²

Of the 5 orbitals, 3 stand alone, and only the remaining 2 pair up in a 2×2 block — now we can see our way through. In general, the Bloch matrix has exactly the same form as in §10:

(37) H(k) =αI +2E(x^) cos⁡kxa +2E(y^) cos⁡kya

where (E(x^) is the matrix of the two-center integrals with the neighbor in the x direction; since E(−x^)=E(x^), the sum of eik·R becomes 2cos), and the table above says that this matrix is block diagonal.

11.3 The three that stand alone — “π along x, δ along y”

When dxy orbitals are lined up along x, with respect to the bond axis they are dz′x′ — that is, pure π. For dyz they are dx′y′ — pure δ. If the direction is changed to y, the roles are interchanged. This feature — “even for the same orbital, whether it is σ, π, or δ depends on the direction” — is what makes d interesting.

(38) Edxy =α+2βπS (π along both x and y) Edyz =α −2βδcos⁡kxa +2βπcos⁡kya (δ along x, π along y) Edzx =α +2βπcos⁡kxa −2βδcos⁡kya (π along x, δ along y)

Here we have defined the abbreviations S=cos⁡kxa+cos⁡kya and D=cos⁡kxa−cos⁡kya. Since βδ is small, dyz hardly disperses at all along x and disperses well only along y — this is what you see in simulator ⑤: select dyz and the energy contours turn into a striped pattern.

11.4 The 2×2 of dx²−y² and dz² (eg in cubic symmetry) — orbitals of the same symmetry mix

Now the remaining two. Taking the x direction as the axis again, both dx²−y² and dz² turn into a mixture of the σ type (d3x2−r2) and the δ type (dy2−z2). Finding the coefficients, we get

(39) dx2−y2 =32σ −12δ , dz2 =−12σ −32δ

That is, dx²−y² is 3/4 σ and 1/4 δ, and dz² is 1/4 σ and 3/4 δ (the squares of the coefficients). The lobes of dx²−y² point straight at the four neighbors, so it has a large σ component — just as the picture suggests. Putting this into (37) and diagonalizing the 2×2 gives

(40) E±=α −(βσ+βδ)S ± |βσ−βδ|2 S2+3D2

At D=0 (that is, kx=ky, as at Γ or M) the two do not mix and remain pure dx²−y² and dz². The farther from there, the more the 3D2 term acts and the more they mix.

11.5 Reading the high-symmetry points

For βσ=1.5, βπ=1, βδ=0.25 eV (Harrison’s ratio σ : π : δ ≈ 6 : 4 : 1), the table gives E−α.

Γ (0,0)X (π/a,0)M (π/a,π/a)What to notice
dxy+4.000−4.00 Only π, so the s band turned upside down
dyz+1.50+2.50−1.50 x is δ, so Γ→X is almost flat
dzx+1.50−2.50−1.50 Just dyz rotated by 90°
dx²−y², dz² pair (upper)−2.25+2.165+4.75 dz² at Γ, dx²−y² at M
dx²−y², dz² pair (lower)−4.75−2.165+2.25 dx²−y² at Γ (most stable)

Incidentally, the band diagram of simulator ⑤ is drawn with lines that are thicker and darker the more of the selected orbital a band contains. The “sharing of darkness” at the X point that we just saw is how a half-and-half eigenvector is drawn under this convention.

Why is the M point the Γ point turned upside down? — The bipartite lattice The right-hand side of (37) is linear in cos⁡kxa and cos⁡kya. Therefore, if we shift k by Q=(π/a,π/a), the two cosines both change sign, and the whole matrix becomes H−αI→−(H−αI). All the eigenvalues change sign — this is what lies behind the relation between Γ and M. This property holds because the square lattice is a bipartite lattice (it can be painted in two colors in a checkerboard pattern, so that neighbors always have different colors), and it always holds for any of s, p, d, and f as long as only nearest neighbors are kept. Incidentally, at k=(π/2a,π/2a) both cosines are 0, so all the bands gather at α (this is why all the lines meet at one point partway along M–Γ in the band diagram).

11.6 Seeing the real thing — obtaining the irreducible representations at the Γ point with CrystOD

The classification so far can be obtained without counting by hand. If you run the CrystOD command crystod-group with a space group and a k point, it returns how the basis functions decompose into irreducible representations under the little group of that k point. For the Γ point of the square lattice (the two-dimensional p4mm embedded in three dimensions as P4/mmm), the command is

$ crystod-group --generate-basis --space-group P4/mmm --kpoint 0 0 0

That is all. Extracting the output for degree 2 (d orbitals), we get

* Decomposition *
 2.0 [GM1+(1)] + 1.0 [GM5+(2)] + 1.0 [GM2+(1)] + 1.0 [GM4+(1)]

* Irreducible representations for basis functions *
  GM1+(1): [x^2 + y^2, z^2]
  GM5+(2): [yz, xz]
  GM2+(1): [-x^2 + y^2]
  GM4+(1): [xy]

The 6 dimensions appear because there are six monomials of degree 2, x2 through xz, which is the sum of l=2 (5 functions) and l=0 (the 1 function r2). Since r2 is totally symmetric, it takes up one GM1+, and subtracting that, the five d orbitals are GM1+ + GM2+ + GM4+ + GM5+ — in exact agreement with the four levels counted in §11.5 (one of which is a doublet).

The degree-3 part (f orbitals) of the same output is as follows.

* Decomposition *
 2.0 [GM3-(1)] + 3.0 [GM5-(2)] + 1.0 [GM2-(1)] + 1.0 [GM4-(1)]

* Irreducible representations for basis functions *
  GM3-(1): [z^3, x^2 z + y^2 z]
  GM5-(2): [x^3, y^3, x^2 y, xy^2, xz^2, yz^2]
  GM2-(1): [xyz]
  GM4-(1): [-x^2 z + y^2 z]

The 10 dimensions are the sum of l=3 (7 functions) and l=1 (3 functions such as r2x), and the latter is GM3- + GM5-. Subtracting it, the seven f orbitals are GM3- + 2×GM5- + GM2- + GM4- (1 + 4 + 1 + 1 = 7). The “three single levels and two doubly degenerate pairs” that we will see in §12 is exactly this arrangement.

CrystODMullikenBasis functionsCorresponding orbitals on this page
Γ1+A1gx²+y², z² s, dz²
Γ2+B1gx²−y² dx²−y²
Γ3+A2g(does not appear at degree 2) —
Γ4+B2gxy dxy
Γ5+Egyz, xz dyz, dzx
Γ2−B1uxyz fxyz
Γ3−A2uz, z³ pz, fz³
Γ4−B2uz(x²−y²) fz(x²−y²)
Γ5−Eux, y px, py, and the remaining 4 f orbitals

In the band diagram of simulator ⑤, this is what appears at the right edge as, for example, x²−y² (Γ2+) (these are labels at the Γ point, so at the X and M points other symbols appear).

About the numbering The numbering of Γ depends on the order in which the one-dimensional irreducible representations are listed, which differs among references (some conventions swap the order of A2 and B1). This page and simulator ⑤ are aligned with the output of CrystOD. On the other hand, the basis functions and the Mulliken symbols do not depend on the convention. When comparing with other tables, it is safest to match using the right two columns of the table above.

12. f Orbitals — Even φ Appears

Finally, f. There is not a single new idea. Simply running through the procedure of §11 as it is gives the seven bands.

12.1 The seven split into “one single plus three 2×2 blocks”

We make the parity table of §11.2 for f.

f orbitalxyzPartner
fz³evenevenodd 2×2 block of these two
fz(x²−y²)evenevenodd
fxz²oddeveneven 2×2 block of these two
fx(x²−3y²)oddeveneven
fyz²evenoddeven 2×2 block of these two
fy(3x²−y²)evenoddeven
fxyzoddoddoddAlone

Only fxyz stands alone; the rest form three pairs of two. There is no need to solve a 7×7 matrix anywhere; solving a 2×2 three times finishes the job.

12.2 The single orbital, and the pair that contains z

fxyz is pure δ in both the x and y directions (about the bond axis it becomes fx′y′z′). The pair fz³ and fz(x²−y²) is a mixture of π and φ in either direction, with weights 3/8 and 5/8 (swapped between the two orbitals). The result is

(41) fxyz: E=α+2βδS fz3, fz(x2−y2): E±=α −(βπ+βφ)S ± |βπ−βφ|4 S2+15D2

This has exactly the same form as (40). The only differences are that (σ,δ) is replaced by (π,φ), the 2 in the denominator becomes 4, and the 3 inside the √ becomes 15. As |m| goes up one step at a time, the way the hybridization acts also changes regularly.

12.3 The two pairs involving x and y — σ and π interchange with direction

The pair fxz² and fx(x²−3y²) behaves a little differently. Along x it is a combination of σ and δ, along y a combination of π and φ, so the kinds of overlap themselves are interchanged depending on the direction. The weights are 3/8 and 5/8 for the x direction, and 1/16 and 15/16 for the y direction. Diagonalizing, we get

(42) E±=α +(βσ+βδ) cos⁡kxa −(βπ+βφ) cos⁡kya ± u2 −uv2 +v2 u=(βδ−βσ) cos⁡kxa , v=(βπ−βφ) cos⁡kya

The pair fyz² and fy(3x²−y²) is obtained from this by interchanging x and y (as expected, since they map onto each other under a 90° rotation of the square lattice). Since u2−uv/2+v2 is a positive-definite quadratic form with a negative discriminant, the quantity inside the √ can never be negative.

12.4 The high-symmetry points, and why the f bands are narrow

For βσ=1.0, βπ=0.6, βδ=0.25, βφ=0.10 eV, the values of E−α are as follows.

Pair (2×2)Γ (0,0)X (π/a,0)M (π/a,π/a)Main component at Γ
fz³, fz(x²−y²) −1.15 / −1.65+0.968 / −0.968+1.65 / +1.15 The upper one is fz³ (pure)
fxz², fx(x²−3y²) +1.55 / −0.45−1.159 / −2.741+0.45 / −1.55 Mixed
fyz², fy(3x²−y²) +1.55 / −0.45+2.741 / +1.159+0.45 / −1.55 Mixed
fxyz (alone)+1.000−1.00 —

Each cell shows E+ / E− (the upper and lower signs of ± in Eqs. (41) and (42)). Which branch belongs to which orbital changes with k — for example, fz³ is the upper branch at Γ and the lower branch at M.

Why are the f bands narrow? The total width of the values above is 5.48 eV. Built by the same procedure, d had 9.50 eV and p had 12.0 eV (these were calculations in which βσ was chosen separately as 1.0, 1.5, and 2.0 eV for f, d, and p, so please look at the trend rather than the numbers themselves. If p is recalculated with the same βσ=1.0 and βπ=0.6 as f, it comes to 6.4 eV). For an overlap with larger |m|, the two orbitals touch only in a “grazing” way. For φ, the six lobes merely touch alternately, and βφ is only about a tenth of βσ. The larger l is, the more kinds of overlap are available, but since most of them are the small kinds, the band conversely becomes narrower. For real 4f electrons, in addition, the orbital itself is contracted close to the nucleus, so the band is narrower still — this is why the 4f electrons of the rare earths look “almost atomic” and keep their localized magnetic moments. Conversely, 5f (actinide) orbitals are more extended and sit on the borderline between forming bands and being localized.
Checked three different ways The equations of §11 and §12 were all solved by hand and then checked against the following three. (1) Independent reconstruction of the two-center matrices: numerical integration of the real spherical harmonics over the sphere gives the rotation matrix to the bond-axis direction Diμ, and with it we set E=Ddiag(V|m|)DT to assemble the matrices anew. (2) Numerical diagonalization of the H(k) built from it on a 21×21 grid of points over the whole BZ. (3) Direct diagonalization, just as it is, of a real-space N×N supercell with periodic boundary conditions (for f, 6×6×7 = 252 levels). All of them agreed with the equations in the text to within 1×10−14. Incidentally, calculating s and p in the same framework gives exactly the same values, digit for digit, as the original equations of simulator ⑤ — a confirmation that the sign rule of (36) is consistent with §10.

Mochizuki Research Group, Tokyo University of Science. For the interactive version, see band-simulator (① From molecules to crystals / ② Bloch’s theorem and E(k) / ③ Density of states / ④ Origin of the band gap / ⑤ 2D square lattice (s, p, d, f)). §10 corresponds to lecture notes (8), (9), and (11) (“LCAO Theory”), and §11 and §12 extend it to d and f, with the irreducible representations at the Γ point matched to the output of crystod-group --generate-basis (§11.6). Eqs. (15), (20), (25), (29), (30), (33), (35), (38), (40), (41), and (42) have all been verified against numerical diagonalization (see the box at the end of §12).