Non-Orthogonal Basis

While the perpendicular grid of a Cartesian coordinate system offers a simple and intuitive way to describe space, the real world rarely adheres to such perfect order. From the atomic lattices of crystals to the overlapping orbitals that form chemical bonds, nature's most fundamental structures are often skewed and non-orthogonal. Adopting a basis that mirrors this inherent geometry simplifies the physics, but it demands a departure from familiar mathematical rules. This departure, however, is not a complication but a gateway to a more profound understanding of geometry and its connection to physical law. This article navigates this richer mathematical landscape. The first chapter, ​​Principles and Mechanisms​​, will deconstruct the mathematical machinery that governs non-orthogonal spaces, introducing the essential concepts of the metric tensor, the dual basis, and the distinction between covariant and contravariant vector components. Following this, the chapter on ​​Applications and Interdisciplinary Connections​​ will demonstrate how these principles are not just abstract theories but indispensable tools used across crystallography, materials science, and quantum chemistry to solve real-world problems.

Imagine you're drawing on a perfect sheet of graph paper. The lines are straight, perpendicular, and evenly spaced. This is the world of the ​​orthonormal basis​​—the familiar Cartesian coordinates $(x, y, z)$. In this world, life is simple. The distance to a point is given by Pythagoras's theorem, $d^2 = x^2 + y^2 + z^2$. The dot product between two vectors is a simple sum of the products of their components. Every direction is independent and plays by the same rules. It’s a beautifully ordered, but ultimately artificial, universe.

The real world, from the sprawling lattices of crystals to the intricate architecture of a molecule, rarely lays itself out on a perfect square grid. The most natural way to describe a honeycomb lattice, a diamond crystal, or the bonds emanating from a carbon atom isn't with perpendicular axes, but with axes that follow the material's own inherent structure. These natural axes are often skewed, stretched, or both. They form a ​​non-orthogonal basis​​.

By choosing a basis that respects the natural symmetry of the problem, we often simplify the physics. But this convenience comes at a price: we must update our mathematical rulebook. The simple, familiar formulas of graph-paper geometry no longer apply directly. This is not a failure; it is a doorway to a deeper, more general, and far more beautiful understanding of space itself.

Let's say we have two vectors, $\vec{G}$ and $\vec{F}$, in a skewed, two-dimensional world. Our basis vectors, $\vec{e}_1$ and $\vec{e}_2$, are no longer perpendicular. We know the vectors in terms of how many "steps" we take along each basis vector: $\vec{G} = G^1\vec{e}_1 + G^2\vec{e}_2$ and $\vec{F} = F^1\vec{e}_1 + F^2\vec{e}_2$. How do we compute their dot product, $\vec{G} \cdot \vec{F}$?

We can't just multiply the components, $G^1F^1 + G^2F^2$. That formula implicitly assumes that $\vec{e}_1 \cdot \vec{e}_1 = 1$, $\vec{e}_2 \cdot \vec{e}_2 = 1$, and $\vec{e}_1 \cdot \vec{e}_2 = 0$. In our skewed world, this is not true. We must go back to first principles:

Look at what appeared! The calculation requires a complete description of the geometry of our basis: all the dot products between the basis vectors themselves. We assemble these values into a matrix called the ​​metric tensor​​, denoted $g_{ij}$:

In our 2D case, this is the matrix $g = \begin{pmatrix} \vec{e}_1 \cdot \vec{e}_1 \vec{e}_1 \cdot \vec{e}_2 \\ \vec{e}_2 \cdot \vec{e}_1 \vec{e}_2 \cdot \vec{e}_2 \end{pmatrix}$. Using this, the dot product formula becomes compact and elegant:

This is the universal formula for the dot product in any basis. In a standard orthonormal basis, $g_{ij}$ is just the identity matrix, and we recover the familiar formula as a special case. The metric tensor is the keeper of the geometry, a Rosetta Stone that translates component-space algebra into real-world geometric facts.

This tensor is not just an abstract computational tool. It has a beautiful geometric meaning. For a 2D basis, the determinant of the metric tensor, $\det(g)$, is equal to the square of the area of the parallelogram formed by the basis vectors. For a 3D basis, it's the square of the volume of the parallelepiped. It literally tells you the size of the "unit cell" of your coordinate system.

In our comfortable Cartesian world, a vector's components are unambiguous. The component $V_x$ is both the length of the vector's shadow projected onto the x-axis and the number of steps you must take along the x-axis to get to the vector's tip. Projection and linear combination are one and the same.

In a non-orthogonal world, these two ideas diverge, giving rise to two distinct but equally valid types of components for the very same vector.

1. ​​Contravariant Components ($V^i$):​​ These are the "step-counting" or "ladder" components. They are the coefficients of the linear combination that builds the vector from its basis vectors: $\vec{V} = V^1\vec{e}_1 + V^2\vec{e}_2 + \dots$. They tell you "how much" of each basis vector you need.

2. ​​Covariant Components ($V_i$):​​ These are the "projection" or "shadow" components. They are found by taking the dot product of the vector with each basis vector: $V_i = \vec{V} \cdot \vec{e}_i$. They measure the projection of $\vec{V}$ along the direction of each $\vec{e}_i$.

In materials science, this distinction is critical. Imagine describing the traction (force) on an anisotropic crystal. The most natural basis vectors, $\mathbf{a}_i$, follow the crystal axes. If you measure the physical force projected onto each axis direction, you are measuring something related to the covariant components. But if you want to express the force vector as a sum of forces directed purely along those axes, you are asking for the contravariant components. These two sets of numbers will be different, and you can only convert between them using the metric tensor via the "index-raising and lowering" rules: $V_i = g_{ij}V^j$ and $V^i = g^{ij}V_j$, where $g^{ij}$ is the inverse of the metric tensor.

This duality of components begs a question: is there a more elegant way to think about this? Is there a structure that unifies these two "faces" of a vector? The answer is a resounding yes, and it is the concept of the ​​dual basis​​ (or ​​reciprocal basis​​).

For any given basis $\{\vec{e}_i\}$, there exists a unique partner basis, the dual basis $\{\vec{e}^j\}$, defined by one simple, powerful relationship:

where $\delta^j_i$ is the Kronecker delta (1 if $i=j$, 0 otherwise). Think of the dual basis as a set of perfect "interrogation" tools. The vector $\vec{e}^1$ is constructed to be perfectly orthogonal to $\vec{e}_2, \vec{e}_3, \dots$ and scaled such that its dot product with $\vec{e}_1$ is exactly 1.

With this tool, finding components becomes beautifully symmetric:

• To find the ​​contravariant​​ component $V^i$, you dot the vector with the corresponding dual basis vector: $V^i = \vec{V} \cdot \vec{e}^i$.
• To find the ​​covariant​​ component $V_i$, you dot the vector with the corresponding original basis vector: $V_i = \vec{V} \cdot \vec{e}_i$.

This reveals the deep connection: the covariant components of a vector are just its components in the dual basis, and the contravariant components are its components in the original basis. What seemed like two different kinds of components are just the same idea viewed from the perspective of two different, intimately related bases.

Nowhere is the concept of a non-orthogonal basis more crucial than in quantum mechanics, particularly in the study of molecules. A molecule is built from atoms, and we describe its electrons using molecular orbitals (MOs). A powerful idea, the Linear Combination of Atomic Orbitals (LCAO), is to build these MOs from the simpler atomic orbitals (AOs) of the constituent atoms.

So, we use the atomic orbitals centered on each nucleus—a 1s orbital on this hydrogen, a 2p orbital on that carbon—as our basis set $\{\chi_{\mu}\}$. But are these AOs orthogonal? An atomic orbital is a wavefunction that, while decaying rapidly, extends throughout all of space. The tail of an orbital on one atom inevitably extends into the region of an orbital on a neighboring atom. Their spatial ​​overlap​​ is unavoidable.

Therefore, their inner product (the integral of their product over all space) is non-zero: $\langle \chi_{\mu} | \chi_{\nu} \rangle = S_{\mu\nu} \neq 0$. The ​​overlap matrix​​ $S$ is nothing other than the metric tensor for our chosen basis of atomic orbitals!.

This has profound consequences. The inner product between two molecular orbitals $|\psi\rangle = \sum c_\mu |\chi_\mu\rangle$ and $|\varphi\rangle = \sum d_\nu |\chi_\nu\rangle$ is not a simple sum of coefficient products, but is given by $\langle\psi|\varphi\rangle = \mathbf{c}^\dagger S \mathbf{d}$. The total number of electrons is not the trace of the density matrix $\rho$, but $\mathrm{Tr}(\rho S)$. Most strikingly, the time-independent Schrödinger equation, $\hat{H}|\psi\rangle = E|\psi\rangle$, when written in this non-orthogonal basis, transforms from a standard eigenvalue problem into a ​​generalized eigenvalue problem​​:

The very fabric of the theory's central equation is altered. The simple $E$ is replaced by $ES$, weaving the geometry of the basis into the dynamics of the system.

We are left with this more complicated generalized eigenvalue problem. While solvable, our best numerical algorithms are designed for the simpler standard form, $A\mathbf{x} = \lambda\mathbf{x}$. Can we transform our problem back to this familiar territory?

Yes, by changing the basis. We can construct a transformation matrix, $X$, that takes us from our non-orthogonal atomic orbitals to a new set of orthonormal basis functions. A particularly elegant way to do this is ​​Löwdin symmetric orthogonalization​​, where we define $X = S^{-1/2}$.

Applying this transformation converts the generalized problem $H \mathbf{c} = E S \mathbf{c}$ into an equivalent standard eigenvalue problem $H' \mathbf{c}' = E \mathbf{c}'$, where $H' = X^\dagger H X$ is the transformed Hamiltonian in the new orthonormal basis. Now we can unleash our powerful numerical solvers.

It is absolutely crucial to understand what is and is not happening here. This orthogonalization is a ​​change of representation​​. It's a computational trick. We are not changing the physics. The molecular orbitals, the orbital energies, the total energy, the electron density, and whether the molecule is magnetic or not—all physical observables remain identical. We have simply changed the language we use to describe them, choosing a more computationally convenient one.

However, this powerful tool comes with a warning. If our initial basis set contains functions that are very similar to each other—a condition called near-linear dependency—the overlap matrix $S$ will have some eigenvalues that are very close to zero. The matrix $S^{-1/2}$ will then have enormous eigenvalues, and the transformation can become numerically unstable, wildly amplifying tiny errors and potentially destroying the calculation. This practical danger highlights the delicate dance between theoretical elegance and computational reality.

From the geometry of crystals to the quantum reality of molecules, non-orthogonal bases are not a complication to be avoided, but a tool to be embraced. They force us to distinguish between a vector and its components, to discover the hidden symmetry of dual spaces, and to appreciate the deep unity between the metric tensor of geometry and the overlap matrix of quantum mechanics. They reveal a richer, more general, and ultimately more truthful picture of the world.

Having explored the principles of non-orthogonal bases, a natural question arises regarding their practical application. The utility of a mathematical framework is ultimately measured by its ability to describe real-world phenomena. In this context, non-orthogonal systems are not a mathematical luxury but a practical necessity for describing the world as it is. The moment we step away from the idealized chalkboard grid of perpendicular axes and look at the real world, we find things are slanted, skewed, and overlapping.

In this chapter, we will see how these ideas blossom into powerful tools across an astonishing range of disciplines, from the geometry of crystals to the very foundations of quantum field theory. The story is one of unification, where a single mathematical concept—the metric tensor—becomes the key to unlocking diverse physical phenomena.

Let's start with the most intuitive domain: geometry. In a standard Cartesian system, calculating lengths, angles, and volumes is second nature. The Pythagorean theorem, the simple dot product $\vec{a} \cdot \vec{b} = a_x b_x + a_y b_y + a_z b_z$, these are the comfortable tools of our trade. But what happens when our basis vectors—our fundamental "rulers"—are not mutually orthogonal?

Imagine mapping the structure of a crystal. The atoms don't sit on a neat cubic grid; they form a lattice described by primitive vectors $\vec{a}_1, \vec{a}_2, \vec{a}_3$ that can point at any angle to one another. This is a living, breathing non-orthogonal basis. If we have the coordinates of two atoms in this basis, say $(u_1, u_2, u_3)$ and $(v_1, v_2, v_3)$, the simple dot product formula fails spectacularly. To find the real-world distance, we need to know the geometry of the basis itself. This information is entirely captured by the ​​Gram matrix​​, $G_{ij} = \vec{a}_i \cdot \vec{a}_j$, which we can think of as the metric tensor for our particular coordinate system. The true inner product between two vectors with coordinate representations $\mathbf{u}$ and $\mathbf{v}$ becomes $\mathbf{u}^T G \mathbf{v}$. Every geometric quantity, from the distance between a point and a plane to the angle between two chemical bonds, must be re-expressed using this metric tensor.

This has profound consequences. For instance, what is the volume of the primitive unit cell of our crystal? In an orthogonal system, it's just the product of the lengths of the basis vectors. In a non-orthogonal system, the volume is given by the scalar triple product, $|\vec{a}_1 \cdot (\vec{a}_2 \times \vec{a}_3)|$. Remarkably, the square of this volume is exactly equal to the determinant of the Gram matrix, $\det(G)$. The geometry of the space is encoded in a single number! This allows us to calculate how the volume of a material changes under deformation, a crucial aspect of materials science and engineering.

The same challenge appears in a completely different context: computer simulations of periodic systems like liquids or solids. To avoid edge effects, physicists use Periodic Boundary Conditions (PBC), where the simulation box is replicated infinitely in all directions. When calculating the force between two particles, we must use the "Minimum Image Convention" (MIC) to find the shortest distance between them, accounting for all their periodic images. If the simulation box is a rectangular prism (an orthogonal basis), the algorithm is simple: for each coordinate $x, y, z$, you just find the shortest distance component-wise. But many important crystal structures, like the hexagonal close-packed lattice, have non-orthogonal primitive cells. In this case, the simple component-wise wrapping algorithm fails. The true shortest distance corresponds to finding the displacement vector that lies within a special shape called the ​​Wigner-Seitz cell​​. Correctly implementing the MIC in a general non-orthogonal lattice requires an algorithm that explicitly uses the metric tensor to minimize the true Euclidean distance, a beautiful and practical link between computational physics and pure geometry.

The real fun begins when we enter the quantum world. In quantum chemistry, we often try to describe the complex behavior of electrons in molecules by building molecular orbitals (MOs) from simpler, atom-centered atomic orbitals (AOs). This is the famous LCAO (Linear Combination of Atomic Orbitals) approximation. The problem is, when two atoms are close enough to form a bond, their atomic orbitals inevitably overlap. The basis of atomic orbitals {$\phi_i$} is inherently non-orthogonal.

This isn't just a minor inconvenience; it strikes at the heart of the Schrödinger equation. In an orthonormal basis, the time-independent Schrödinger equation $\hat{H}\Psi = E\Psi$ becomes a standard matrix eigenvalue problem, $H \mathbf{c} = E \mathbf{c}$, where $H$ is the Hamiltonian matrix and $\mathbf{c}$ is the vector of coefficients. But when the basis is non-orthogonal, a new player enters the game: the overlap matrix, $S_{ij} = \langle \phi_i | \phi_j \rangle$. The variational principle leads us to the ​​generalized eigenvalue problem​​:

This equation is one of the most important in all of quantum chemistry. It tells us that the energies and shapes of molecular orbitals depend not just on the Hamiltonian ($H$), which describes the physics of energy, but also on the overlap matrix ($S$), which describes the geometry of the basis. You cannot understand one without the other. The off-diagonal elements of the Fock matrix (a sophisticated version of $H$) represent the coupling or mixing between atomic orbitals, but their physical interpretation is inextricably tied to the corresponding overlap elements in $S$. This same structure appears not just for single electrons, but also when we build many-electron wavefunctions from a basis of non-orthogonal configuration state functions (CSFs), as is common in Valence Bond theory.

Faced with the generalized eigenvalue problem, one common strategy is to first transform our non-orthogonal basis into an orthonormal one. The Löwdin symmetric orthogonalization is a particularly elegant way to do this, creating a new set of orbitals {$\chi_i$} that are orthonormal while being "as close as possible" to the original atomic orbitals. But this mathematical convenience comes at a physical price. When you force orbitals that naturally overlap to become orthogonal, you change their character. This "orthogonalization penalty" can be quantified. For instance, the new orthogonal orbitals have a higher total kinetic energy than the original ones, because forcing them apart introduces more curvature, or "wiggles," into their wavefunctions.

Even more strikingly, consider the simplest chemical bond in the hydrogen molecule, H$_2$. A simple model describes the covalent bond as a state where one electron is on proton A and the other is on B, and vice versa. But what if we build this model using Löwdin-orthogonalized orbitals instead of the true, overlapping atomic ones? When we translate the resulting wavefunction back into the language of the original, physical orbitals, we find that it is no longer purely covalent. It now contains a significant admixture of "ionic" states ($\text{H}^+\text{H}^-$), where both electrons are on the same proton. The amount of this spurious ionic character is directly proportional to the overlap integral, $S$. This is a profound lesson: non-orthogonality is not a bug, it's a feature! It is intimately connected to the nature of chemical bonding.

Despite these subtleties, working in an orthogonalized basis has immense practical advantages. For example, when trying to calculate the partial electric charge on an atom in a molecule, different methods give different answers. Methods like Mulliken analysis, which work directly in the non-orthogonal basis, are notoriously unstable; adding more basis functions can cause the calculated charges to swing wildly. In contrast, Löwdin analysis, which is performed in the symmetrically orthogonalized basis, is far more robust. This is because it effectively partitions the electron density based on the subspaces spanned by each atom's orbitals, a property that is much less sensitive to the inclusion of nearly redundant functions in the basis set.

The influence of non-orthogonality extends to the deepest and most elegant parts of physics. Consider the role of symmetry. Group theory is a powerful tool that simplifies quantum problems by exploiting the symmetry of a molecule or crystal. The standard theory is built on unitary representations, which works perfectly in an orthonormal basis. But what happens in our non-orthogonal world? The matrices that represent the symmetry operations (like rotations or reflections) are no longer unitary in the standard sense. Instead, they satisfy a modified condition, $R(g)^\dagger S R(g) = S$, making them "$S$-unitary." Consequently, the entire machinery of symmetry analysis, including the powerful projection operators used to construct symmetry-adapted linear combinations (SALCs), must be reformulated to explicitly include the overlap matrix $S$ at every step.

Perhaps the most breathtaking consequence arises when we consider the very foundations of many-body quantum theory. In second quantization, we describe a system of many particles by defining field operators, which are expanded in a basis of single-particle states. These fields are governed by canonical (anti-)commutation relations, like $\{\psi(\mathbf{r}), \psi^\dagger(\mathbf{r}')\} = \delta(\mathbf{r}-\mathbf{r}')$ for fermions. This is the bedrock of quantum field theory. We've always assumed, usually without saying so, that the underlying single-particle basis is orthonormal. What if it isn't?

If we expand our field operator $\psi(\mathbf{r})$ in a non-orthogonal basis $\{\phi_i\}$ with creation/annihilation operators $\{c_i^\dagger, c_i\}$, and demand that the field itself still obeys the canonical relations, we find something astonishing. The operators $\{c_i\}$ can no longer obey their own simple canonical relations. The familiar anticommutator $\{c_i, c_j^\dagger\} = \delta_{ij}$ must be replaced by:

The inverse of the overlap matrix appears directly in the fundamental algebra of our quantum operators! To recover the simple textbook algebra, one must define new operators that are linear combinations of the old ones, using the matrix $S^{1/2}$ as the transformation key. The geometry of our chosen basis dictates the very structure of the quantum mechanical algebra.

From calculating distances in a skewed crystal to rewriting the commutation relations of quantum fields, the thread that connects these disparate domains is the metric tensor, $S$ or $G$. It is the dictionary that translates our descriptions from an arbitrary, convenient basis to the invariant, physical reality of the world. Far from being a mere mathematical complication, the study of non-orthogonal systems reveals the deep and beautiful unity between the geometry of space and the laws of physics.