1.1 Linear Algebra for Quantum Mechanics

1.1.1 Basic definitions and notations

We let 𝔽 denote either the real field ℝ or the complex one ℂ. For a complex number z = x+iy ∈ℂ, with x,y ∈ℝ, we write the conjugate z^∗ := x−iy. We let ℳ_m,n(𝔽) denote the space of matrices of dimension m × n with entries in 𝔽 and ℳ_n(𝔽) whenever m = n. For A := (a_ij)_{1≤i≤m; 1≤j≤n} ∈ℳ_m,n(𝔽), A^∗ := (a_ij^∗)_{1≤i≤m; 1≤j≤n} is the complex conjugate. If A ∈ℳ_n(𝔽), we write A^⊤ for its transpose and A^† := (A^∗)^⊤ for its Hermitian conjugate. We finally denote I the identity matrix and write I_n whenever we wish to emphasise the dimension, and 0_m,n the null matrix in ℳ_m,n(𝔽). Recall that a matrix A ∈ℳ_n(𝔽) is invertible (or non-singular) if there exists B ∈ℳ_n(𝔽) such that AB = BA = I_n. Given two matrices A ∈ℳ_p,m(𝔽) and B ∈ℳ_q,n(𝔽), we define their tensor product as

Since a vector is a particular case of a matrix, for u ∈𝔽^m and v ∈𝔽ⁿ, we can write

1.1.2 Inner products

A vector space V over the field 𝔽 is a set endowed with

• a commutative, associative addition operation,
• an operation of multiplication by a scalar.

The addition and the multiplication by a scalar have the following properties (for scalars α,β ∈𝔽 and vectors u,v ∈V):

• v + 0 = v;
• v + (−v) = 0;
• α(βv) = (αβ)v;
• (α + β)v = αv + βv;
• α(u + v) = αu + αv;
• 1 ⋅ v = v.

Armed with this, we can now define an inner product on V:

For example, the following spaces carry a natural inner product:

• The vector space ℂⁿ with the inner product ⟨u,v⟩ := u^†v = ∑ _i=1ⁿu_i^∗v_i;
• The space of complex-valued continuous functions on [0,1] with ⟨f,g⟩ := ∫ ₀¹f(t)^∗g(t)dt;
• If X,Y ∈ℳ_m,n(ℝ), then ⟨X,Y⟩ := Tr(X^⊤Y) = ∑ _i=1^m ∑ _j=1ⁿX_ijY_ij defines an inner product on the space of (real) matrices.

Projection matrices are particularly useful for geometric purposes:

In particular, if W is a vector subspace of 𝔽ⁿ with some orthonormal basis (w₁,…,w_d), it is then easy to check that the map 𝒫_W : 𝔽ⁿ →𝔽ⁿ onto W satisfying

defines an orthogonal projection.

1.1.3 From linear operators to matrices

Let V be a finite-dimensional vector space over 𝔽 and ⟨⋅,⋅⟩ a non-degenerate inner product on V. Given a linear operator 𝒜 : V →V, then, by the Riesz representation theorem  [309, Section III-6], there exists a unique linear operator 𝒜^† : V →V, called the adjoint operator, such that

Indeed, for any v ∈V, the map u ∈V⟨𝒜u,v⟩ is a linear functional, hence an element of the dual space V^† (the space of bounded linear functionals on V), therefore for each v ∈V, there exists v′∈V such that ⟨𝒜u,v⟩ = ⟨u,v′⟩. It is then easy to show that the map vv′ is linear, proving that the adjoint operator is uniquely defined. In the particular case where 𝒜 = 𝒜^†, the operator 𝒜 is called Hermitian, a key requirement in quantum mechanics:

For a Hermitian operator 𝒜, we then have, for any u ∈V,

by conjugate symmetry (Definition 1), and therefore ⟨𝒜u,u⟩ is real. Conversely, if ⟨𝒜u,u⟩ is real, then

Therefore, = 0; since this is true for all u ∈V, then 𝒜 = 𝒜^†.

The following property of operators shall be useful to ensure that systems driven by operators preserve distances, or norms:

Definition 4. The linear operator 𝒜 : V → V is called unitary if it is surjective and

Recall that a linear operator between two finite-dimensional normed spaces is bounded, and therefore continuous. For any u ∈V, this implies that ∥𝒜u∥ = ∥u∥, so that a unitary operator 𝒜 preserves the norm. In that case, 𝒜 is an isometry, therefore injective. Being also surjective, it is bijective and therefore its inverse exists. For a unitary operator 𝒜 and any u,v ∈V, we have

by definition of the adjoint, implying that

where ℐ is the identity operator.

Example (Real Matrices): If V = ℝⁿ with inner product ⟨u,v⟩ := u^⊤v for u,v ∈ℝⁿ, the linear operator 𝒜 can now be viewed as a matrix A in ℳ_n(ℝ). Its adjoint is nothing other than the transpose A^⊤, and therefore A is self-adjoint if and only if it is symmetric. In this case, if A is unitary (or orthogonal), then it is invertible with A⁻¹ = A^⊤. Rotation matrices in ℝ², which will play an important role later when constructing quantum circuits, are the only unitary maps of ℝ² onto itself and are of the form

for 𝜃 ∈ [0,2π) and δ ∈{−1,+1}.

Example (Complex Matrices): If V = ℂⁿ with inner product ⟨u,v⟩ := v^†u for u,v ∈ℂⁿ, the linear operator 𝒜 can now be viewed as a matrix in ℳ_n(ℂ). The adjoint of such a matrix is then the Hermitian conjugate A^† and A is called Hermitian if A = A^† and unitary if A^†A = I_n. We shall denote by 𝒰_n(ℂ) the set of unitary matrices in ℳ_n(ℂ). We will discuss Hermitian matrices over ℂ in more detail in Section 1.1.6.

1.1.4 Condition number

In order to manipulate matrices and measure them, we require matrix norms:

The condition number of a matrix is an important tool to understand the stability of linear equations of the form Ax = b, for A ∈ℳ_n(𝔽), b ∈𝔽ⁿ. Assuming A to be non-singular, the true solution is clearly x_∗ := A⁻¹b. Suppose, however, that the vector b is only known up to some (not necessarily quantum) measurement error, and one observes instead b + Δ_b. The solution is then A⁻¹(b + Δ_b) = x_∗ + Δ_x, with Δ_x := A⁻¹Δ_b. In particular, we can write, for any (sub-multiplicative) matrix norm ∥⋅∥,

From this inequality, we see that the quantity ∥A⁻¹∥∥A∥ bounds the relative error in the solution with respect to the relative error in the measurement of the input vector b. This leads to the following terminology:

Definition 6. Given a matrix A ∈ ℳ_n(𝔽) and a sub-multiplicative norm ∥⋅∥, we call

the condition number (with respect to the norm ∥⋅∥) of the matrix A (and assign to it infinite value if A is singular).

Remark: The definition of the condition number above holds for any matrix norm ∥⋅∥, but admits a more explicit representation in the particular case of the spectral norm ∥⋅∥₂, defined as

where ∥x∥₂ := is the L ₂ norm for vectors. If the matrix A is not singular, then

where λ_max(A) and λ_min(A) denote the largest and smallest eigenvalues of A.

1.1.5 Matrix decompositions and spectral theorem

Having defined essential properties of (complex) matrices, we now introduce several essential tools that allow us to gain a better understanding of their properties.

The Singular Value Decomposition is a key tool to analyse the properties and behaviours of matrices. It is ubiquitous in applied statistics and machine learning and allows us to reduce the explanatory dimension of a large matrix into a small number of meaningful components.

The numbers {σ₁,…,σ_p} are called the singular values of A and are uniquely defined. The columns of U and V are the left-singular and right-singular vectors of A, in the sense that, if σ ∈{σ₁,…,σ_p}, then there exist a column u of U and a column v of V such that Av = σu and A^†u = σv. Recall that the rank of a matrix is defined as the dimension of the span of its columns. As a corollary of the Singular Value Decomposition theorem, the rank of a matrix is therefore equal to the number of non-zero singular values. The Singular Value Decomposition is general in the sense that it holds for any matrix. In the particular case of square matrices, the Schur decomposition and the Spectral Theorem provide refinements.

The Spectral Theorem is a cornerstone result in the theory of linear operators, and in particular for (finite-dimensional) matrices. Recall that an operator 𝒜 : V →V is called normal if it commutes with its adjoint, namely if 𝒜𝒜^† = 𝒜^†𝒜. Self-adjoint (or Hermitian) operators are clearly normal, yet the converse is not true in general. Recall further that an eigenvector of 𝒜 is a non-zero vector u ∈V such that 𝒜u = λu for some λ ∈ℂ, and we denote by σ(𝒜) the set of eigenvalues of 𝒜.

The following result, which is more general than the subsequent spectral theorem, allows us to decompose any arbitrary complex square matrix.

Note that since U is unitary, then U⁻¹ = U^†. We call the matrix T the Schur transform of A and the identity in the theorem means that A and T are similar, so in particular, possess the same eigenvalues, all located on the diagonal of T. If A is a normal matrix, then so is T, and therefore T must be diagonal and we write T = D for clarity. In this case, we say that the matrix A is diagonalisable with A = UDU^†, where the diagonal entries of D are the eigenvalues of A and the column vectors of U are the orthonormal eigenvectors of A.

For each eigenvalue λ ∈ σ(𝒜), denote the corresponding eigenspace

Since the vector space V is the orthogonal direct sum of the eigenspaces (indexed by the eigenvalues of 𝒜), we can then write the spectral decomposition

where 𝒫_λ is the orthogonal projection operator onto 𝒱_λ. Note that such an operator is naturally self-adjoint  [309, Theorem 2, Section III-1].

1.1.6 Hermitian matrices

We introduced above Hermitian matrices as the set of matrices A over the complex field ℂ such that A = A^†. As fundamental building blocks of quantum computing, we investigate their properties further. Clearly, a real matrix is Hermitian if and only if it is symmetric, in which case A^⊤ = A.

The Singular Value Decomposition (Theorem 1) takes a particular flavour in the case of Hermitian matrices:

Proof. The Spectral Theorem shows that there exist a unitary matrix U ∈𝒰_n(ℂ) and a diagonal matrix Σ ∈ℳ_n(ℂ) such that A = UΣU^†, where the diagonal elements of Σ are the eigenvalues of A. Assuming (i), we can define B = UU^†∈ℳ_n(ℂ). Then clearly

since U is unitary. The equality A = B^†B is also obvious. The latter implies that

Finally, assume (iv) and let λ be an eigenvalue of A with eigenvector u. Then

Since the latter is strictly positive, then clearly λ ≥ 0.

The following property lies at the core of Hamiltonian simulation of quantum systems:

Recall that for a matrix A ∈ℳ_n(ℂ), its exponential is given by

In practice, though, given a Hermitian matrix A, finding the corresponding unitary matrix U is not easy. The Hamiltonian simulation problem is defined as follows.

Hamiltonian Problem: Given a Hermitian matrix A ∈ℳ_n(ℂ), a time t > 0, a tolerance level 𝜀 > 0, and some matrix norm ∥⋅∥, find a unitary matrix U such that ≤ 𝜀.

1.1.7 Rotation matrices

Rotation matrices, and later their quantum gate equivalents will play a key role in building quantum circuits. Let us start with the following lemma:

Lemma 1. If a matrix A ∈ℳ_n(ℂ) is such that A² = I, then for any 𝜃 ∈ℝ,

Proof. This follows directly from the series expansion

which has an infinite radius of convergence. □

Lemma 1 will prove essential for computational purposes. As simple examples, consider the following:

Exercise: Compute e^i𝜃A for A ∈{X,Y,Z} and 𝜃 ∈ℝ, where

For any α ∈ [0,2π), consider now the map ℛ_α : ℝ² →ℝ² such that

which is basically a rotation of angle α and does not affect the norm of the input vector. To the map ℛ_α, we can associate the (rotation) matrix R_α such that ℛ_α(u) = R_αu for any u ∈ℝ². It is easy (exercise) to show the following:

Lemma 2. The matrix R_α has the form

This representation is the general form of a rotation matrix in ℝ² (introduced in (1.1.3)).

Exercise: Write the matrices e^i𝜃A for A ∈{X,Y,Z} from the previous exercise as rotation matrices.

1.1.8 Polar coordinates

Recall that a point z = x + iy, with x,y ∈ℝ, lying on the unit circle can be written as z = e^i𝜃 with 𝜃 ∈ [0,2π). Indeed, simply let x = r cos(𝜃), y = r sin(𝜃) and add the constraint r = 1. Consider now a general vector u ∈ℂ² of the form

with α,β ∈ℂ such that |α|² + |β|² = 1. Here, (e₁,e₂) forms a basis of ℝ²:

In polar coordinates, we can then write

Note that arbitrary multiplication phases have no influence – a fact of key importance in quantum mechanics – because, for any γ ∈ℝ,

so that in fact, multiplying u by the global phase e^−i𝜃_α and letting 𝜃 := 𝜃_β −𝜃_α, we consider

Write temporarily r_βe^i𝜃 = x + iy. Insisting on u being on the unit sphere further imposes ∥u∥² = 1, namely

1	= ∥u∥2 = rαe1 + (x + iy)e2†rαe1 + (x + iy)e2
	= rαe1⊤ + (x − iy)e 2⊤rαe1 + (x + iy)e2
	= rα2 + x2 + y2,

since (e₁,e₂) is orthonormal. This is nothing more than the equation of the unit sphere. In polar coordinates, we can write

and clearly r = 1 since we are on the unit sphere. Therefore

u	= cos(𝜃)e1 + sin(𝜃)cos(ϕ) + isin(𝜃)sin(ϕ)e2
	= cos(𝜃)e1 + sin(𝜃)eiϕe 2.

1.1.9 Dirac notations

Given a vector v ∈ℂⁿ, Dirac’s ket and bra notations read

With these notations, the operation := ⟨u|v⟩ defines an inner product on ℂⁿ. The notation for the standard orthonormal basis in  ℂⁿ is ()_{i=0,…,n−1}, i.e.,

In coordinates, we can write, for any u,v ∈ℂⁿ,

and therefore,

1.1.10 Quantum operators

In the language of Dirac’s notations, we can define the outer product ⟨v| (for u ∈U and v ∈V) as a linear operator from V to U, two vector spaces, as

In particular, ⟨v| is the projection on the one-dimensional space generated by . Any linear operator can be expressed as a linear combination of outer products as

where and are the standard basis vectors (1.1.9).

Similarly to the linear algebra setting above, we can define an eigenvector of a linear operator 𝒜 : V →V as a non-zero vector such that

for some complex eigenvalue λ. Associated with any linear operator 𝒜, the adjoint operator 𝒜^† satisfies

Indeed, in the language of linear operators above, we have

by definition of the inner product (Definition 1).

1.1.11 Tensor product

Given two vector spaces U and V of dimensions m and n, the tensor product U ⊗V is a vector space of dimension mn. For u ∈U and v ∈V, we can form the vector := ⊗∈U ⊗V with the following properties:

• = + , for any u′∈U;
• = + , for any v′∈V;
• α = = , for any α ∈ℂ.

Given the linear operators 𝒜 : U →U and ℬ : V →V, we can then define their tensor product as an operator 𝒜⊗ℬ on U ⊗V:

which can be represented in matrix form as A ⊗ B ∈ℳ_mn,mn(ℂ).

This Dirac formalism, fully anchored in (classical) linear algebra, now opens the gates to a proper dive into the foundations of quantum mechanics.