The tight-binding model is a description of quantum states in which a system is represented by localized orbitals and the couplings between them. Originally developed to explain electrons in crystals, it describes how states associated with individual atoms combine into extended electronic bands. Its characteristic parameters are on-site energies, which specify the energies of localized states, and hopping amplitudes, which describe their coupling. These ingredients turn a continuous Schrödinger equation into a matrix eigenvalue problem.
Tight binding is a central method of condensed-matter physics because it connects atomic structure and chemical bonding with electrical, optical, and magnetic properties. Simple versions explain band formation and the distinction between metals and insulators; more elaborate versions describe semiconductors, surfaces, nanostructures, correlated electrons, and topological phases. Closely related lattice models also describe light in coupled waveguides, vibrations in mechanical arrays, and ultracold atoms in optical lattices.
The term has two related meanings. In its original sense, it is an approximation appropriate when electrons remain strongly associated with atomic orbitals and interatomic coupling is comparatively weak. In modern usage, it also denotes a general localized-basis Hamiltonian, whose parameters may be fitted to experiments or derived from first-principles calculations. A modern tight-binding representation can therefore remain accurate even when its localized states are not weakly perturbed orbitals of isolated atoms.
Physical foundations and historical development
An isolated atom has discrete electronic energy levels. When many identical atoms are brought together, states of similar energy interact, and their degeneracy is lifted. For a crystal containing equivalent atoms, one relevant orbital per atom produces spatial eigenstates. In a macroscopic sample, the energy differences between neighboring eigenstates are usually so small that the levels form an effectively continuous band. If spin is independent and twofold degenerate, that band can accommodate electrons.
The underlying mechanism is quantum-mechanical coupling, not classical motion between fixed atomic trajectories. A state initially concentrated near one atom can evolve into a superposition involving neighboring atoms because the Hamiltonian operator has nonzero matrix elements between their orbitals. These matrix elements are conventionally called hopping integrals. Their magnitudes depend on orbital shapes, interatomic distance, bond orientation, and the potential created by the surrounding material. The same physics is closely connected to molecular orbital theory: bonding and antibonding levels in a small molecule become bands when the construction is extended periodically.
Felix Bloch’s 1928 work on electrons in periodic potentials established the symmetry framework for crystalline electronic states. Tight-binding reasoning became one of the principal ways to construct states satisfying that framework. Gregory Wannier introduced the localized functions now bearing his name in 1937, providing a systematic connection between localized orbitals and Bloch eigenstates. In 1954, John Slater and George Koster developed a practical scheme for expressing interatomic matrix elements through bond-direction factors and a relatively small set of parameters. Their method made multiorbital tight-binding calculations considerably more tractable.
The traditional contrasting approximation is the nearly free electron model. That approach begins with extended plane waves and treats the periodic crystal potential as a perturbation. Tight binding instead begins with localized states and introduces coupling between sites. These are different starting points rather than mutually exclusive descriptions: both can produce Bloch bands, energy gaps, and effective carrier masses. Their usefulness depends on which basis represents the relevant states economically.
Historically, tight binding was particularly natural for narrow bands derived from transition-metal orbitals, molecular crystals, and covalently bonded solids. Its later development broadened this scope. Localized orbitals constructed from accurate band calculations can span strongly hybridized states, while lattice Hamiltonians can isolate general mechanisms without reproducing every chemical detail. Consequently, the method functions both as a quantitative electronic-structure technique and as a framework for deliberately simplified theoretical models.
Localized orbitals, hopping, and the matrix problem
Let denote an orbital of type centered on site . A single-particle wave function is expanded as
The Hamiltonian and overlap matrix elements are
Substitution into the Schrödinger equation gives
This is a generalized eigenvalue problem. If the basis is orthonormal, is the identity matrix and the equation reduces to . Atomic orbitals centered on different atoms are generally not exactly orthogonal, so neglecting overlap is an additional approximation rather than part of the definition of tight binding.
Diagonal matrix elements are on-site energies. Off-diagonal elements describe coupling between different orbitals, whether on the same atom or on different atoms. A common one-orbital convention writes the Hamiltonian as
where each neighboring pair is included once. Here the actual off-diagonal matrix element is . Other conventions call the matrix element itself , without extracting a minus sign. Band formulas must therefore be read together with the chosen Hamiltonian convention.
Hermiticity requires , ensuring real eigenvalues. Individual hopping signs and phases can change when the phases of basis orbitals are redefined. Observable quantities do not depend on this arbitrary basis choice, although phase combinations around closed loops can have physical significance. An isolated hopping amplitude is therefore not always an independently measurable property.
The phrase “nearest-neighbor model” means that only couplings between the closest sites are retained. Second-neighbor and longer-range terms can be added systematically. Localized orbitals often make these matrix elements decay with separation, producing a sparse matrix, but their decay need not be identical for every orbital or bonding environment.
Orthonormalization can also change the apparent range of the model. Symmetric orthogonalization transforms the Hamiltonian to
provided the overlap matrix is positive definite. Even if the original matrices contain only short-range elements, the transformed Hamiltonian can acquire longer-range couplings. A short-range nonorthogonal model and a short-range orthogonal model are consequently not generally equivalent term by term.
Periodic crystals and Bloch Hamiltonians
In a crystal, sites are labeled by a lattice vector and an orbital index within a unit cell. Translational symmetry implies that a hopping matrix element depends on the separation between cells rather than their absolute positions. Bloch’s theorem then converts a problem containing a macroscopic number of orbitals into a collection of small problems labeled by crystal wave vector .
Using Bloch sums with phases , the Hamiltonian takes the form
An analogous expression defines for a nonorthogonal basis. At each , the energies follow from
Other Bloch-sum conventions include the positions of orbitals within the unit cell in their phase factors. They produce different-looking matrix elements but the same energy spectrum when applied consistently.
Wave vectors differing by a reciprocal lattice vector describe equivalent translational behavior. It is therefore sufficient to calculate states within a Brillouin zone. For an orthonormal basis containing independent orbitals per unit cell, the Hamiltonian has eigenvalues at each wave vector, before any additional spin multiplicity is counted. Hybridization between orbitals produces avoided crossings when symmetry permits them; symmetry can also protect degeneracies.
Band energies determine several important properties. The group velocity of a wave packet in band is
where is the reduced Planck constant. Near a nondegenerate band extremum, the inverse effective-mass tensor is
These quantities connect hopping parameters to carrier motion: stronger hopping typically produces greater dispersion, although the detailed curvature depends on geometry and interference.
Electronic occupation follows Fermi–Dirac statistics. At zero temperature, states are occupied up to the Fermi level. A partially filled band ordinarily permits metallic behavior within an independent-electron treatment, whereas completely filled bands separated from empty bands by a band gap give a band insulator. Band touching, disorder, and electron interactions introduce important qualifications.
The eigenvectors contain information that the energies alone do not provide. They determine orbital character, sublattice weights, and matrix elements relevant to optical transitions and geometric responses. Two models with similar energy bands can therefore predict different observables if their eigenvectors or representations of physical operators differ.
Solvable chains and the origin of band features
A uniform one-dimensional chain illustrates the basic construction. Suppose each site has one orthonormal orbital, lattice spacing , on-site energy , and nearest-neighbor hopping , with . Periodic boundary conditions give
The bandwidth is . The lowest-energy state has neighboring coefficients in phase, while the highest-energy state alternates in sign between neighboring sites. The band therefore expresses the continuous progression from bonding to antibonding combinations.
Near the minimum at ,
Comparison with a parabolic dispersion gives
A larger hopping amplitude yields a smaller effective mass in this particular band-edge limit. If , all sites are decoupled and the band collapses to the atomic energy . For a finite open chain of sites, the allowed standing-wave energies instead are
The density of states shows how many levels occur in a given energy interval. For the infinite uniform chain, normalized to one state per site and per spin,
inside the band and zero outside it. The divergences at the band edges arise because the dispersion becomes locally flat. They are examples of van Hove singularities and demonstrate that a band’s state distribution is not determined by its width alone.
A chain with two sites per unit cell gives a richer example. In the Su–Schrieffer–Heeger model, introduced in 1979 in work on polyacetylene, alternating bonds have hopping magnitudes and . With equal on-site energies set to zero,
where is now the two-site unit-cell period. Its bands are
For nonnegative hoppings, the gap is .
An open alternating chain can support states localized near its ends. Their existence depends on the pattern of strong and weak bonds, the termination, and the symmetries retained by the Hamiltonian. In a finite chain, opposite-end states can hybridize and split slightly in energy. Thus the bulk dispersion, boundary geometry, and symmetry must be considered together; a two-band formula alone does not completely specify the spectrum of a finite sample.
Multiorbital models and parameter construction
Real materials usually require several orbitals on each atom. An orbital is approximately isotropic, whereas and orbitals have directional structures that strongly affect hopping. A multiorbital model must therefore describe both chemical identity and bond orientation. The Slater–Koster method separates these ingredients into radial coupling parameters and geometrical factors constructed from the direction cosines of an interatomic bond.
For example, let be the components of the unit vector along a bond. In the two-center approximation, a hopping element between two orbitals has the form
while
The parameters and distinguish orbital overlap along and transverse to the bond axis. This construction enforces the appropriate angular behavior without fitting a separate parameter for every bond orientation. More elaborate tables include , , and orbitals and different atomic species.
Parameter values can be obtained empirically by fitting measured gaps, effective masses, optical transitions, or photoemission dispersions. They can also be fitted to calculations based on density functional theory. The fitting target matters: reproducing bands close to the Fermi energy does not guarantee accurate high-energy states, and reproducing eigenvalues does not necessarily guarantee accurate wave functions. Some semiconductor models include excited -like or -like basis states to improve conduction-band dispersions even when those orbitals do not dominate the simplest valence-bond picture.
A systematic alternative uses Wannier functions, which are localized functions constructed from selected Bloch states. For a chosen set of bands, their Hamiltonian matrix elements are Fourier transforms of an appropriately transformed Bloch Hamiltonian. The resulting hoppings provide a real-space representation of the selected electronic subspace. If the target bands overlap in energy with others, a disentanglement procedure may be needed to define that subspace smoothly across wave-vector space.
Wannier functions are not unique: wave-vector-dependent unitary transformations within the selected subspace change their shapes and the numerical hopping elements without changing the represented physics. Localization procedures select useful representatives, but an individual hopping value is not independent of this choice.
Reducing a model further also requires care. Eliminating high-energy orbitals generally produces energy-dependent effective couplings. For low-energy orbitals coupled to high-energy orbitals , the exact projected eigenvalue equation involves
An energy-independent short-range model is then an approximation to this expression. Virtual transitions through eliminated orbitals can generate longer-range hopping and alter on-site energies, explaining why effective parameters need not resemble simple isolated-atom integrals.
Graphene, semiconductors, and orbital chemistry
Graphene is a particularly transparent application of tight binding. Its honeycomb lattice has two carbon atoms per primitive cell, conventionally labeled and . The in-plane , , and orbitals participate primarily in strong bonding, while one orbital per atom supplies the low-energy bands. Retaining only nearest-neighbor coupling between orbitals gives
where are the three vectors from an site to adjacent sites.
The two bands have energies
At the inequivalent corners and of the Brillouin zone, the three hopping contributions cancel and . Expanding about either point gives approximately conical bands,
with , where is the carbon–carbon bond length. With about nanometres and of order electronvolts, the velocity is of order metres per second. The two sublattice amplitudes form a pseudospin, and the low-energy Hamiltonian resembles a massless Dirac equation.
This resemblance does not mean that the electrons move at the speed of light or that sublattice pseudospin is their actual spin. It is an emergent description of a particular dispersion and wave-function structure. Different on-site energies on the two sublattices produce a mass-like term and open a gap. Same-sublattice hopping breaks the simplest model’s electron–hole spectral symmetry, while edges and defects introduce states absent from the infinite pristine sheet.
Covalent semiconductors require a different orbital treatment. In silicon and diamond-structure materials, tetrahedral bonding mixes and character through orbital hybridization. An basis captures the main valence-bond structure, but quantitatively reliable conduction bands often require additional basis functions or longer-range terms. Silicon’s indirect gap, with valence- and conduction-band extrema at different wave vectors, is an important test of model quality.
Orbital chemistry also controls transition-metal compounds. Crystal environments split -orbital energies, while metal–ligand hybridization produces effective couplings that can differ greatly among orbital channels. A model containing only metal orbitals may encode ligand-mediated motion in its effective hopping parameters; a model retaining both metal and ligand states resolves those processes explicitly. The appropriate choice depends on whether the target is a broad band dispersion, optical excitation, magnetic exchange, or low-energy interacting behavior.
Magnetic fields, spin, and band topology
A magnetic field changes the phases accumulated by charged particles moving between sites. In a localized basis, its orbital effect is often represented by the Peierls substitution. For a matrix element that transfers a particle of charge from site to site ,
where is the vector potential. For an electron, . The product of these phases around a closed loop depends on the enclosed magnetic flux, rather than on the gauge-dependent vector potential at individual points.
This approximation is especially useful when the field varies slowly on the orbital scale and field-induced changes in the orbital shapes are small. It does not automatically include all atomic magnetic effects or couplings between bands. In a periodic two-dimensional lattice, rational magnetic flux per elementary plaquette can enlarge the magnetic unit cell and split bands into magnetic subbands. Douglas Hofstadter’s 1976 square-lattice calculation displayed the intricate flux-dependent spectrum now known as the Hofstadter butterfly.
Spin adds further matrix structure. A Zeeman term couples spin to the magnetic field, while spin–orbit coupling couples spin and orbital motion. Atomic models often include a term proportional to ; effective lattice models may instead use spin-dependent hopping matrices constrained by crystal symmetry. A model with spatial orbitals generally becomes a -dimensional Bloch Hamiltonian when spin is included explicitly.
The eigenvectors of a Bloch Hamiltonian also carry geometric information. Their evolution across wave-vector space defines the Berry phase and Berry curvature. For a suitable isolated occupied-band subspace in two dimensions, integrating the curvature over the Brillouin zone yields a Chern number. A nonzero total occupied-band Chern number produces a quantized Hall response when the Fermi level lies in a bulk gap.
F. Duncan Haldane’s 1988 honeycomb-lattice model demonstrated that such a phase can arise from complex further-neighbor hopping with zero net magnetic flux through a unit cell. The Kane–Mele model, proposed in 2005, combined honeycomb geometry with spin–orbit coupling to describe a time-reversal-invariant quantum spin Hall phase. These models helped establish tight binding as a principal language for topological insulators.
Topological conclusions require the appropriate symmetries, filling, and bulk gap. They also depend on the global structure of occupied states, not merely on whether two energy bands appear to invert near one point. Boundary states provide important signatures, but their detailed dispersion depends on the termination and on perturbations allowed at the edge.
Electron interactions and disorder
The simplest tight-binding Hamiltonian is a single-particle model. Filling its orbitals with electrons accounts for Fermi statistics but does not, by itself, describe correlations caused by electron–electron repulsion. Interactions can be introduced using creation and annihilation operators. A standard example is the Hubbard model, developed in 1963 in work by John Hubbard and, independently, Martin Gutzwiller and Junjiro Kanamori:
Here creates an electron of spin on site , and is the energy cost of opposite-spin double occupation.
The kinetic term favors delocalization, whereas repulsion can suppress charge fluctuations. At appropriate integer filling, sufficiently strong correlations can produce a Mott insulator even when an independent-electron description predicts a partly filled band. The ratio , where is a characteristic bandwidth, is useful for judging the competition, but dimensionality, orbital degeneracy, lattice geometry, and magnetic order also matter. There is no universal critical ratio applicable to all Hubbard systems.
At half filling and in the strong-coupling limit of the nearest-neighbor single-orbital model, virtual hopping between singly occupied sites generates antiferromagnetic exchange with scale
Actual charge motion is energetically costly in this limit, but virtual double occupation lowers some spin configurations more than others. This gives a direct connection between tight-binding hopping parameters and effective magnetic interactions. More complicated orbital structures can produce different signs and anisotropies of exchange.
Disorder provides a separate route to nonmetallic behavior. Random on-site energies represent spatially varying local potentials, while random hoppings can model bond disorder or structural irregularity. Philip Anderson’s 1958 analysis showed that sufficiently strong disorder can localize quantum states through interference, a phenomenon called Anderson localization. In conventional one-dimensional noninteracting systems with uncorrelated disorder, generic states are localized even for weak disorder, although the localization length may become very large.
Interaction-driven and disorder-driven insulation are not interchangeable. A Mott insulator involves correlation effects tied to interaction and filling; an Anderson insulator can occur in a noninteracting model because eigenstates fail to extend through the system. Both mechanisms may coexist in real materials.
Interactions also change the computational task fundamentally. A one-particle Hamiltonian with orbitals is an matrix, but the corresponding many-electron Hilbert space grows combinatorially. Mean-field approximations, exact diagonalization, quantum Monte Carlo, and tensor-network methods address different regimes with different limitations. The simplicity of the hopping part does not imply that the interacting model is easy to solve.
Numerical methods, transport, and engineered lattices
For a periodic material with a modest number of orbitals per unit cell, numerical tight-binding calculations typically diagonalize on a wave-vector mesh. Band plots use selected paths through the Brillouin zone, whereas charge densities, total occupations, and response functions require suitable integration over its volume. A band plot alone can miss extrema away from the chosen path and is insufficient to establish a small gap reliably.
Finite systems are treated directly in real space. Surfaces, interfaces, vacancies, and nanoparticles are represented by changing on-site energies, deleting sites, or modifying couplings. Because local orbitals often give sparse matrices, iterative eigensolvers can obtain selected states without computing the entire spectrum. Polynomial-expansion methods can estimate densities of states for much larger systems. Computational cost depends not only on atom count but also on the number of orbitals per atom, hopping range, and requested observables.
Open-system transport is commonly formulated using a Green’s function. A finite device region connected to left and right leads has retarded Green’s function
where is the central Hamiltonian and the self-energies and encode the effect of the semi-infinite leads. With
the coherent transmission is
Within the corresponding elastic, noninteracting or effective single-particle description, the zero-temperature linear conductance follows the Landauer formula,
when all spin channels are included in . An explicit factor of two is used instead when the transmission is defined for one of two identical spin channels. These conventions must not be mixed. Interactions and inelastic scattering require extensions beyond this elementary expression.
Tight-binding matrices also describe engineered wave systems. In arrays of optical waveguides, amplitudes on neighboring guides exchange energy through evanescent coupling, and propagation distance can play the mathematical role of time. Coupled resonators and mechanical arrays similarly realize site energies and intersite couplings, although their natural dynamical variables differ from electronic wave functions.
Ultracold atoms in deep optical lattices provide another realization: localized states form near potential minima, and barrier penetration supplies hopping. Appropriate conditions lead to lattice models with controllable interactions and geometry. The common tight-binding structure makes these systems useful for studying localization, interference, and topology, but electronic occupation statistics and electrical response do not transfer automatically to classical-wave realizations.
Accuracy, transferability, and interpretive limits
The accuracy of a tight-binding model is determined by its basis, retained couplings, parameter construction, and intended energy range. A minimal model may correctly explain why a crossing occurs while giving inaccurate velocities or optical intensities. A larger model may reproduce a reference calculation closely but offer less transparent physical interpretation. Neither simplicity nor parameter count alone establishes reliability.
Transferability is a major issue. Parameters fitted at one lattice constant may fail under compression, strain, or changes in coordination. Bond lengths alter orbital overlap, while local charge redistribution changes on-site energies and screening. Distance-dependent hopping laws can improve structural transferability, but they are not universally accurate substitutes for recalculating the electronic environment. Defects and surfaces are especially demanding because their coordination and charge states differ from those of the bulk training or fitting structures.
Self-consistent extensions address some of these effects by allowing effective potentials to depend on orbital occupations or atomic charges. Density-functional tight-binding methods approximate a density-functional energy through expansions around a reference density and combine electronic terms with additional energy contributions. Such methods are related to ordinary band-structure tight binding but should not be identified with an arbitrary fitted hopping Hamiltonian.
Electronic eigenvalues alone also do not define a complete total-energy model. Structural relaxation and molecular dynamics require forces derived from a consistent energy expression. In nonorthogonal bases, derivatives of both the Hamiltonian and overlap matrix enter. Additional repulsive or double-counting terms may be necessary, depending on how the model was derived. Simply summing occupied band energies generally does not yield the full binding energy.
Response properties require similar care. In a localized basis, the position operator can contain intracell and interorbital matrix elements. Consequently, differentiating an arbitrarily chosen Bloch matrix with respect to wave vector does not automatically supply every physical current or optical matrix element unless basis conventions and operator representations are handled consistently. A band-only fit can therefore be insufficient for predicting absorption strengths, polarization, or nonlinear responses.
There are also mathematical restrictions on localization. For an isolated band subspace with nonzero Chern number, a complete orthonormal set of exponentially localized Wannier functions cannot be chosen while retaining the usual lattice-translation structure for that subspace alone. This does not prevent a local-orbital tight-binding model from exhibiting Chern bands: the larger orbital space can contain both occupied and unoccupied bands, while the obstruction applies to the topologically nontrivial subspace selected from it. The distinction is important when constructing reduced models intended to preserve both locality and topology.