Pretty Good State Transfer in Discrete-Time Quantum Walks Arxiv

Pretty good state transfer in discrete-time quantum walks

Ada Chan and Hanmeng Zhan

Department of Mathematics and Statistics, York University, Toronto, ON, Canada {ssachan, h3zhan}@yorku.ca

Abstract We establish the theory for pretty good state transfer in discrete- time quantum walks. For a class of walks, we show that pretty good state transfer is characterized by the spectrum of certain Hermitian adjacency matrix of the graph; more speciﬁcally, the vertices involved in pretty good state transfer must be m-strongly cospectral relative to this matrix, and the arccosines of its eigenvalues must satisfy some number theoretic conditions. Using normalized adjacency matrices, cyclic covers, and the theory on linear relations between geodetic angles, we construct several inﬁnite families of walks that exhibits this phenomenon.

1 Introduction arXiv:2105.03762v1 [math.CO] 8 May 2021 A quantum walk with nice transport properties is desirable in quantum computation. For example, Grover’s search algorithm [12] is a quantum walk on the looped complete graph, which sends the all-ones vector to a vector that almost “concentrate on” a vertex. In this paper, we study a slightly stronger notion of state transfer, where the target state can be approximated with arbitrary precision. This is called pretty good state transfer. Pretty good state transfer has been extensively studied in continuous-time quantum walks, especially on the paths [9, 22, 1, 5, 21]. However, not much

1 is known for the discrete-time analogues. The biggest diﬀerence between these two models is that in a continuous-time quantum walk, the evolution is completely determined by the adjacency or Laplacian matrix of the graph, while in a discrete-time quantum walk, the transition matrix depends on more than just the graph - usually, it is a product of two unitary matrices:

U = SC, where S permutes the arcs of the graph, and C, called the coin matrix, send each arc to a linear combination of the outgoing arcs of the same vertex. Thus, the relation between the walk and the graph can be rather complicated. In this paper, we consider a discrete-time quantum walk where the spectrum of U is determined by certain Hermitian adjacency matrix H of the graph. A generalization of this walk, called the twisted Szegedy walk, was proposed in [14], and applied in [19, 16]. We show that pretty good state transfer in this model reduces to m-strong cospectrality relative to H and some number theoretic conditions on the arccosines of its eigenvalues. Using normalized adjacency matrices, cyclic covers, and the theory on linear relations between geodetic angles, we construct several inﬁnite families of walks that exhibits this phenomenon.

2 A discrete-time quantum walk

Let X be a connected undirected graph. Let W be a complex adjacency matrix of X, that is, W = (wab)ab where wab 6= 0 if and only if a is adjacent to b. We say W is normalized if the entrywise product W ◦W is row-stochastic, that is, (W ◦ W )1 = 1. We construct a quantum walk with respect to a given pair (X,W ) of graph and normalized complex adjacency matrix. To each edge {a, b} of X, we associate a pair of opposite arcs, (a, b) and (b, a); we call a and b the tail and the head of the arc (a, b), respectively. A quantum state is a complex- valued function on the arcs of X. Over time, it evolves according to a unitary transition matrix U, which we deﬁne now. Let R be the permutation matrix that sends each arc (a, b) to its opposite direction (b, a). Let Nt be the

2 weighted tail-arc incidence matrix, where ( wab, u = a, (Nt)u,(a,b) = 0, u 6= a.

∗ ∗ Since W is normalized, NtNt = I and Nt Nt is a projection. The transition matrix U is then given by ∗ U := R(2Nt Nt − I). After t steps, the quantum walk will be in the state U tx, if it started with the state x.

3 Spectral decomposition of U

The spectral decomposition of a product of two involutions has been studied in [20], and widely applied to quantum walks whose transition matrices satisfy this property [15, 20, 14, 25, 16]. In this section, we consider the quantum walk with respect to a pair (X,W ) of graph and normalized complex adjacency matrix, and let U be the associated transition matrix: ∗ U = R(2Nt Nt − I). We establish a correspondence between the spectrum of U and that of the Hermitian adjacency matrix H := W ◦ W ∗. 3.1 Lemma. We have ∗ Nt RNt = H. ∗ Proof. Pick any two vertices a and b. The ab-entry of NtRNt is ! ! T ∗ X T X ea NtRNt eb = waue(a,u) R wbve(b,v) u∼a v∼b X X = wauwbvR(a,u),(b,v) u∼a v∼b ( w w , a ∼ b, = ab ba 0, a 6∼ b, where the last equality follows as R is the permutation matrix that reverses each arc.

3 Note that 2 ∗ 2 R = (2Nt Nt − I) = I, and so U is indeed a product of two involutions. As every involution equals twice a projection minus the identity, the following result turns out to be useful in determining the eigenspaces of U.

3.2 Lemma. [23, Ch 2] Let K and L be matrices with orthonormal columns. Let P = KK∗, Q = LL∗ and S = L∗K. Let

U = (2P − I)(2Q − I).

Then the eigenvalues of SS∗ lie in [0, 1]. Moreover, we have the following spectral correspondence between S and U.

(i) The 1-eigenspace of U is the direct sum

(col(P ) ∩ col(Q)) ⊕ (ker(P ) ∩ ker(Q)).

Moreover, the map z 7→ Lz is an isomorphism from the 1-eigenspace of SS∗ to col(P ) ∩ col(Q).

(ii) The (−1)-eigenspace of U is the direct sum

(col(P ) ∩ ker(Q)) ⊕ (ker(P ) ∩ col(Q)).

Moreover, z 7→ Kz is an isomorphism from ker(S) to col(P ) ∩ ker(Q), and z 7→ Lz is an isomorphism from ker(S∗) to ker(P ) ∩ col(Q).

(iii) The remaining eigenspaces of U are completely determined by the eigenspaces of SS∗. To be more speciﬁc, let µ ∈ (0, 1) be an eigenvalue of SS∗ and let θ ∈ R be such that cos(θ) = 2µ − 1. The map z 7→ ((cos(θ) + 1)I − (eiθ + 1)P )Lz

is an isomorphism from the µ-eigenspace of SS∗ to the eiθ-eigenspace of U, and the map

z 7→ ((cos(θ) + 1)I − (e−iθ + 1)P )Lz

is an isomorphism from the µ-eigenspace of SS∗ to the e−iθ-eigenspace of U.

4 We now apply the above lemma to our walk, with the observation that ∗ ∗ ∗ L = Nt and 2SS − I = Nt RNt. 3.3 Theorem. Let U be the transition matrix with respect to (X,W ), and let H = W ◦ W ∗. For each eigenvalue eiθ of U, there is an eigenvalue λ of H iθ such that λ = cos(θ). Moreover, if Fθ denotes the e -eigenprojection, and Eλ denotes the λ-eigenprojection, then Fθ and Eλ are related in the following way. (i) If θ = 0, then λ = 1 and

∗ NtF0Nt = E1.

(ii) If θ = π, then λ = −1 and

∗ NtFπNt = E−1.

(iii) If θ ∈ (−π, 0) ∪ (0, π), then 1 N F N ∗ = E . t θ t 2 λ

∗ Proof. Note that with L = Nt in Lemma 3.2, we have ∗ ker(Q) = ker(Nt Nt) = ker(Nt). Suppose θ = 0. By Lemma 3.2 (i), the projection onto col(P ) ∩ col(Q) is given by ∗ F0 = Nt E1Nt, ∗ and so NtF0Nt = E1. Now suppose θ = π. By Lemma 3.2, the map ∗ ∗ z 7→ Nt z deﬁnes an isomorphism from ker(S ) to ker(P ) ∩ col(Q). Since ker(S∗) = ker(SS∗) and 2SS∗ − I = H, we see that

∗ F1 = Nt E−1Nt. Finally, suppose θ ∈ (−π, 0) ∪ (0, π). Lemma 3.2 (iii) and a bit of calculation show that 1 F = (N ∗ − eiθRN ∗)E (N − e−iθN R). θ 2 sin2(θ) t t λ t t Therefore, 1 1 N F N ∗ = (I − eiθH)E (I − e−iθH) = E . t θ t 2 sin2(θ) λ 2 λ

5 We are interested in transferring a state that “concentrate” on some vertex a, that is, a state whose support is a subset of the outgoing arcs of a. ∗ One natural candidate for such states is Nt ea, where ea is a the characteristic vector of the vertex a. In [24], it is shown that for certain family of graphs, there are vertices a and b, an integer time t and a complex number γ such that t ∗ ∗ U (Nt ea) = γ(Nt eb). This is called perfect state transfer at time t. Figure 1 illustrates one such example.

1 • W = 2 A(X)

∗ 1 • Nt ea = 2 (e(0,1) + e(0,2) + e(0,4) + e(0,5))

∗ 1 • Nt eb = 2 (e(3,1) + e(3,2) + e(3,4) + e(3,5)) • t = 6

Figure 1: A graph with perfect state transfer

Unfortunately, perfect state transfer is rare. The main goal of this paper is to characterize a relaxation called pretty good state transfer, which occurs if there is a unimodular complex number γ such that for any > 0, there is an integer time t with

t ∗ ∗ U Nt ea − γNt eb < . The following result comes in handy when we analyze both perfect state ∗ ∗ transfer and pretty good state transfer from Nt ea to Nt eb. iθ 3.4 Lemma. Let e be an eigenvalue of U with eigenprojection Fθ. Let λ = cos(θ). Let Eλ be projection onto the λ-eigenspace of H. Then

∗ ∗ FθNt ea = αFθNt eb if and only if Eλea = αEλeb. Proof. The identity ∗ ∗ FθNt ea = αFθNt eb

6 is equivalent to ∗ FθNt (ea − αeb) = 0. (1) ∗ Since Fθ is a projection, NtFθNt is positive semideﬁnite. Thus, (1) holds if and only if ∗ NtFθNt (ea − αeb) = 0. The result now follows from Theorem 3.3.

4 Perfect and pretty good state transfer in a generic quantum walk

We develop the theory for perfect and pretty good state transfer in a generic quantum walk. Let U be a unitary matrix acting on Cn. Consider two states x and y in Cn. We say there is perfect state transfer from x to y, if there exists t ∈ Z and γ ∈ C such that U tx = γy.

Clearly, the above equation implies |hU tx, yi| = 1; on the other hand, by Cauchy-Schwarz inequality,

t t hU x, yi ≤ U x kyk = 1, where equality holds if and only if U tx = γy for some γ ∈ C. Thus, an equivalent deﬁnition of perfect state transfer is that for some t ∈ Z, t hU x, yi = 1. The same generic walk is said to admit pretty good state transfer from x to y, if there exists a unimodular γ ∈ C such that, for any > 0, there is a time t ∈ Z with t U x − γy < . Similar to perfect state transfer, pretty good state transfer can be deﬁned using |hU tx, yi|. 4.1 Lemma. Pretty good state transfer occurs from x to y if and only if for any > 0, there is t ∈ Z such that t hU x, yi > 1 − .

7 Proof. Suppose ﬁrst that pretty good state transfer occurs. Then there exists γ ∈ C such that for any > 0, there is a time t ∈ Z with

t U x − γy < , that is, U tx = γy + z, where z is a vector with norm less than . Thus,

t hU x, yi = |γhy, yi + hy, zi| ≥ 1 − |hy, zi| > 1 − .

We now prove the backward direction. Given δ > 0, let tδ ∈ Z be a time such that t 2 hU δ x, yi > 1 − δ /2. Deﬁne |hy, U tδ xi| γδ := . hy, U tδ xi

Then |γδ| = 1. Moreover,

2 tδ tδ tδ U x − γδy = hU x − γδy, U x − γδyi tδ = 2 − 2Re(γδhy, U xi) t = 2 − 2 hy, U δ xi < δ2.

Hence, there is a sequence of integer times {tk} and a sequence of unimodular complex numbers {γk} such that

tk lim U x − γky = 0. k→∞

Since the unit circle is compact, {γk} has a limit point γ with |γ| = 1. Therefore, for any > 0, there exists k such that tk U x − γky < , kγk − γk < , 2 2 and so tk tk U x − γy ≤ U x − γky + kγky − γyk < . We can say more about these two phenomena using the spectral information of U. Since U is unitary, its eigenvalues are of the form eiθr for some

8 iθr θr ∈ R. Let Fr denote the orthogonal projection onto the e -eigenspace. Then U has spectral decomposition

X iθr U = e Fr. r This allows us to derive three inequalities for |hU tx, yi|. 4.2 Lemma. Let U be a unitary matrix with spectral decomposition

X iθr U = e Fr. r For any time t ∈ Z and any two states x and y, we have t X hU x, yi ≤ |hFrx, Fryi| (2) r X ≤ kFrxkkFryk (3) r s s X 2 X 2 ≤ kFrxk kFryk (4) r r = 1 (5)

Moreover, (3) is tight if and only if Frx and Fry are parallel for each r, and (4) is tight if and only if kFrxk = kFryk for each r. Proof. First, note that * + t X itθr X itθr U x, y = e Frx, y = e hFrx, Fryi . r r Now take the absolute value on both sides. Inequality (2) follows from the triangle inequaity, Inequalities (3) and (4) follow from Cauchy Schwarz inequality, and Equation (5) follows from the fact that x and y are unit vectors. From Lemma 4.2, we see that (3) and (4) are simultaneously tight if and only if Frx = µrFry, |µr| = 1, ∀r (6) Following the notion in continuous-time quantum walks, we say two states x and y are strongly cospectral if they satisfy (6). Clearly, strong cospctrality is a necessary condition for perfect state transfer. The next lemma shows that it is also necessary for pretty good state transfer.

9 4.3 Lemma. If pretty good state transfer occurs from x to y, then x and y are strongly cospectral.

Proof. Suppose x and y are not strongly cospectral. Then at least one of (3) and (4) in Lemma 4.2 is a strict inequality. Thus, X |hFrx, Fryi| < 1, r and there exists > 0 such that X < 1 − |hFrx, Fryi|. r

Now for any t ∈ Z, t hU x, yi < 1 − . It follows from Lemma 4.1 that pretty good state transfer does not occur from x to y. We give a spectral characterization for perfect and pretty good state transfer. To start, we deﬁne

Θx := {r : Frx 6= 0} and call it the eigenvalue support of the state x. It is clear that strongly cospectral states have the same eigenvalue support.

4.4 Theorem. Let U be a unitary matrix with spectral decomposition

X iθr U = e Fr. r Perfect state transfer occurs from x to y if and only if both of the following hold.

(i) x and y are strongly cospectral, that is, for each r, there is a unimodular µr ∈ C such that Frx = µrFry.

(ii) There is t ∈ Z such that for all r, s ∈ Θx,

itθr itθs e µr = e µs.

10 Proof. Note that U tx = γy if and only if for each r, itθr e Frx = γFry. Thus, perfect state transfer implies (i) and (ii). Conversely, suppose (i) and itθr (ii) holds. Then we may let γ = e µr for some r ∈ Θx. It follows that U tx = γy.

4.5 Theorem. Let U be a unitary matrix with spectral decomposition

X iθr U = e Fr. r Pretty good state transfer occurs from x to y if and only if both of the following hold. (i) x and y are strongly cospectral, that is, for each r, there is a unimodular µr ∈ C such that Frx = µrFry.

(ii) There is a unimodular γ ∈ C such that, for any > 0, there exists t ∈ Z such that for each r ∈ Θx,

itθr e µr − γ < .

Proof. It suﬃces to show that given (i), pretty good state transfer is equivalent to (ii). Indeed,

t X itθr U x − γy = (e µr − γ)Fry,

r∈Θx and so

t 2 X itθr X itθr U x − γy = h (e µr − γ)Fry, (e µr − γ)Fryi

r∈Θx r∈Θx 2 X itθr 2 = e µr − γ kFryk .

r∈Θx

Therefore, (ii) holds if and only if for any > 0, there is t ∈ Z such that t U x − γy < .

11 5 A special form of strong cospectrality

In this section, we study pretty good state transfer with a spectral form of strong cospectrality, that is,

Frx = µrFry, ∀r ∈ Θx, where µr is some root of unity. For shortness, we say x and y are m-strongly cospectral if for some positive integer m,

2πiσr/m Frx = e Fry, ∀r ∈ Θx, where σr ∈ Zm and at least one of them is coprime to m. The following version of Kronecker’s Theorem turns out to be useful.

5.1 Theorem (Kronecker). Given α1, α2, . . . , αn ∈ R and β1, β2, . . . , βn ∈ R, the following are equivalent. (i) For any > 0, the system

|qαr − βr − pr| < , r = 1, 2, ··· , n n+1 has a solution {q, p1, p2, . . . , pn} ∈ Z .

(ii) For any set {`1, `2, . . . , `n} of integers such that

`1α1 + ··· + `nαn ∈ Z, we have `1β1 + ··· + `nβn ∈ Z. We now give a number theoretic characterization for pretty good state transfer with m-strongly cospectral states. 5.2 Theorem. Let U be a unitary matrix with spectral decomposition

X iθr U = e Fr. r Let x and y be two states that are m-strongly cospectral. Pretty good state transfer occurs from x to y if and only if for any set of integers {`r : r ∈ Θx} such that X X `rθr ≡ 0 (mod 2π), `r = 0,

r∈Θx r∈Θx we have X `rσr ≡ 0 (mod m)

r∈Θx

12 Proof. By Theorem 4.5, pretty good state transfer occurs if and only if there exists γ ∈ C such that, for any > 0, there is a time t ∈ Z with

i(tθ +2πσ /m) e r r − γ < , ∀r; this is equivalent to the existence of δ ∈ R such that for any > 0, there is t ∈ with Z θr δ σr t − − − pr < . 2π 2π m Now apply Theorem 5.1 with θ δ σ α = r , β = − r . r 2π r 2π m We see that pretty good state transfer occurs if and only if the following holds:

(a) There exists δ ∈ R such that, for any set of integers {`r : r ∈ Θx} such that X `rθr ≡ 0 (mod 2π),

r∈Θx it holds that X 2π X δ ` − ` σ ≡ 0 (mod 2π). (7) r m r r r∈Θx r∈Θx

We claim that (a) is equivalent to (b):

(b) For any set of integers {`r} such that X `rθr ≡ 0 (mod 2π), (8)

r∈Θx

if P ` = 0, then r∈Θx r X `rσr ≡ 0 (mod m). (9)

r∈Θx

The direction (a) =⇒ (b) is obvious. For the direction (b) =⇒ (a), consider all sets of integers {`r : r ∈ Θx} satisfying (8). There are two possibilities.

13 (i) All such sets satisfy (9). Then δ = 0 is a common solution to (7). (ii) Some set does not satisfy (9). Then ( ) X X `r : `rθr ≡ 0 (mod 2π)

r∈Θx r∈Θx forms a non-trivial additive subgroup of Z. Thus, it is cyclic and has a generator g. Pick {`r} satisfying (7) and X `r = g.

r∈Θx

We claim that P 2π `rσr δ = r∈Θx , mg 0 is a common solution to (7). To see this, pick any set of integers {`r} satisfying (8). Deﬁne a new set of integers by P `0 η := s∈Θx s ` − `0 . r g r r Then X X ηrθr ≡ 0 (mod 2π), ηr = 0.

r∈Θx r∈Θx Thus by (b), X ηrσr ≡ 0 (mod m).

r∈Θx Therefore, P X 2π X 2π `sσs X 2π X δ `0 − `0 σ ≡ s:s∈Θx `0 − `0 σ r m r r mg r m r r r∈Θx r∈Θx r∈Θx r∈Θx 2π X ≡ ( η σ ) m r r r∈Θx ≡ 0 (mod m)

Note that the condition in Theorem 5.2 is symmetric about x and y. 5.3 Corollary. Let x and y be m-strongly cospectral states. If there is pretty good state transfer from x to y, then there is pretty good state transfer from y to x.

14 6 Quantum walk with respect to (X,W )

We apply Theorem 5.2 to the quantum walk deﬁned in Section 2. Let X be a graph and W a normalized complex adjacency matrix. Recall that

∗ U = R(2Nt Nt − I), where R is the arc-reversal permutation matrix, and Nt is the normalized tail-arc incidence with (Nt)a,(a,b) = wab. It was shown in Section 3 that the spectral decomposition of U is largely determined by the following Hermitian adjacency matrix of X: H = W ◦ W ∗. ∗ ∗ By Lemma 3.4, strongly cospectrality of states Nt ea and Nt eb, which is a property of the eigenprojections of U, translates into a property of the eigenprojections of H. For convenience, we will extend a few deﬁnitions for states to vertices. Let the spectral decomposition of H be X H = λEλ. λ

The eigenvalue support of a vertex a, denoted Λa, is the set

Λa := {λ : Eλea 6= 0}.

We say two vertices a and b are strongly cospectral if for each eigenvalue λ ∈ Λa, there is a unimodular αλ ∈ C such that

Eλea = αλEλeb; if, in addition, each αr is an m-th root of unity and at least one of them is primitive, then a and b are m-strongly cospectral. The following result is a direct consequence of Lemma 3.4. 6.1 Lemma. Two vertices a and b are (m-)strongly cospectral if and only if ∗ ∗ the states Nt ea and Nt eb are (m-)strongly cospectral. We devote the rest of the paper to perfect and pretty good state transfer between m-strongly cospectral vertices. Let

k 2πik/m Λab := {λ ∈ Λa : Eλea = e Eλeb}.

15 It is clear that k Λa = Λb = tkΛab for m-strongly cospectral vertices a and b. Moreover, there are at least two classes in this partition, as otherwise we would have

X 2πik/m X 2πik/m ea = Eλea = e Eλeb = e eb.

λ∈Λa λ∈Λa Our next characterizations is based on this partition of eigenvalue support. 6.2 Theorem. Let a and b be two vertices that are m-strongly cospectral. Let β±1 = 1 if ±1 ∈ Λa, and β±1 = 0 otherwise. There is pretty good state ∗ ∗ transfer between Nt ea and Nt eb if and only if for any set of integers

0 {`λ : λ ∈ Λa\{1, −1}} ∪ {`λ : λ ∈ Λa\{1, −1}} ∪ {`1, `−1}, the two conditions

X 0 (`λ − `λ) arccos(λ) + `−1β−1π ≡ 0 (mod 2π) (10) λ∈Λ\{±1} and X (`λ + `λ0 ) + `1β1 + `−1β−1 = 0 (11) λ∈Λ\{±1} imply   X X 0 k  (`λ + `λ) + `1β1σ1 + `−1β−1σ−1 ≡ 0 (mod m). (12) k∈ m k Z λ∈Λab Proof. By Theorem 3.3, each eigenvalue λ of H contributes to one or two eigenvalues e±iθ of U, where θ = arccos(λ). The statement now follows from Theorem 5.2. Theorem 6.2 imposes a strong condition on the structure of the eigenvalue support. 6.3 Theorem. Let a and b be two vertices that are m-strongly cospectral. ∗ ∗ If pretty good state transfer occurs between Nt ea and Nt eb, then m is even, and there is k ∈ Zm such that

k m/2+k Λa = Λab t Λab .

16 ∗ k k Proof. Pick k ∈ Zk for which Λab is non-empty, and let µ ∈ Λab. Suppose there exists n ∈ Zm\{k, m/2 + k} (13) n n for which Λab is also non-empty. Let η ∈ Λab. We deﬁne a set of integers

0 {`λ : λ ∈ Λa\{1, −1}} ∪ {`λ : λ ∈ Λa\{1, −1}} ∪ {`1, `−1} based on η. If η = 1 or η = −1, we choose

0 0 `µ = `µ = 1, `η = −2, and `λ = `λ = 0, ∀λ 6= µ, η. and if η 6= ±1, we choose

0 0 0 `µ = `µ = 1, `η = `η = −1, and `λ = `λ = 0, ∀λ 6= µ, η. Then Equations (10) and (11) hold. Now, the left hand side of Equation (12) becomes 2(k − n), which is divisible by m only when m = 2 or k − n = m/2, both contradicting the assumption (13). We now give a characterization for perfect state transfer, which is a special case of pretty good state transfer, for m-strongly cospectral vertices.

6.4 Theorem. Let m be an even positive integer, and let a and b be two vertices that are m-strongly cospectral. Perfect state transfer occurs at time ∗ ∗ t between Nt ea and Nt eb if and only if the following hold.

(i) There is k ∈ Zm such that

k m/2+k Λa = Λab t Λab .

(ii) For each eigenvalue λ ∈ (−1, 1) of H,

t arccos(λ)

is a multiple of π.

k m/2+k (iii) For each pair of eigenvalues λ, µ ∈ Λab or λ, µ ∈ Λab , t(arccos(λ) − arccos(µ))

is an even multiple of π.

17 k m/2+k (iv) For each pair of eigenvalues λ ∈ Λab and µ ∈ Λab , t(arccos(λ) − arccos(µ))

is an odd multiple of π.

Proof. Condition (i) in Theorem 4.4 is equivalent to the existence of k ∈ Zm such that k m/2+k Λa = Λab t Λab . For Condition (ii) in Theorem 4.4, note that each eigenvalue λ of H con- ±iθ tributes to one of two eigenvalues e where cos(θ) = λ. Hence, for λr, λs ∈ Λab, the condition is equivalent to

it arccos(λr) it arccos(λs) k m/2+k e = e , λr, λs ∈ Λab or λr, λs ∈ Λab

it arccos(λr) it arccos(λs) k m/2+k e = −e , λr ∈ Λab, λs ∈ Λab it arccos(λr) −it arccos(λr) e = e , λr ∈ (−1, 1).

7 Constructions

We have reduced a problem on the quantum walk to a problem on certain Hermitian adjacency matrix H of the underlying graph. However, to recover a walk from H, we need to ﬁnd W so that H = W ◦ W ∗ and W ◦ W is row-stochastic.

7.1 Lemma. Let H = (hab)ab be a Hermitian adjacency matrix of a graph such that for every vertex a, X |hab| = 1. b∼a For each edge {a, b}, pick a direction (a, b), a unimodular complex number δab, and let p hab wab = |hab|δab, wba = . wab ∗ Let W = (wab)ab. Then H = W ◦ W and W ◦ W is row-stochastic.

18 Proof. One can verity that wabwba = hab and X X wabwab = |hab| = 1. b∼a b∼a Lemma 7.1 finds more than one choice of W for the same H. From a physical point of view, this means we can engineer pretty good state transfer of different states on the same graph, by assigning different phases δab. A real number θ is called a pure geodetic angle if any one of its six squared trigonometric functions is either rational or infinite. From Theorem 6.2 Equation (10), we see that pretty good state transfer is related to linear dependence of certain angles over the rationals. Such linear relations have been well-studied for geodetic angles; we cite two useful results.

7.2 Theorem. [2, Ch 11] If q ∈ Q and qπ is a pure geodetic angle, then √ 1 tan(qπ) ∈ 0, ± 3, ±√ , ±1 . 3 7.3 Theorem. [3] If the value of a rational linear combination of pure geodetic angles is a rational multiple of π, then so is the value of its restriction to those angles whose tangents are rational multiple of any given square root.

We construct several inﬁnite families of quantum walks that admit pretty good state transfer.

7.1 Normalized adjacency matrix The ﬁrst two families are with respect to (X,W ), where X is distance regular and W ◦ W ∗ is the normalized adjacency matrix of X. By Lemma 7.1, more than one choices of W will work; the simples example is W = (wab)ab where p wab = 1/ deg(a) deg(b).

7.4 Theorem. Let X be the cocktail party graph nK2, and let H be the normalized adjacency matrix of X. Pick any W such that H = W ◦ W ∗ and W ◦ W is row-stochastic. The quantum walk with respect to (X,W ) has pretty good state transfer between the antipodal vertices.

19 Proof. The cases where n = 2 and n = 3 were proved in [24]; in fact, 2K2 and 3K2 have perfect state transfer. We show that pretty good state transfer occurs for larger n. Let H be the normalized adjacency matrix of X. Since X is regular with valency 2n − 2, 1 H = A(X), 2n − 2 and so strong cospectrality relative to H is equivalent to strong cospectrality relative to A(X). It is known (see, for example, [4]) that the antipodal vertices a and b of X are 2-strongly cospectral, and, relative to A(X),

0 1 Λab = {2n − 2, −2}, Λab = {0}. Now suppose 1 π (` − `0 ) arccos − + (` − `0 ) ≡ 0 (mod 2π). −2 −2 n − 1 0 0 2 By Theorem 7.2, arccos(−1/(n − 1)) is linearly independent with π over the 0 0 rationals. Hence `−2 = `−2, and so `0 − `0 is even. It follows from Theorem 6.2 and 0 `0 + `0 ≡ 0 (mod 2) that pretty good state transfer occurs on X. To construct the second family, we derive some tools to treat rational eigenvalue support.

7.5 Lemma. Let d, λ1, λ2, . . . , λk be positive integers. Suppose

(i) λr ∈ (0, d/2) ∪ (d/2, d) for each r, and (ii) the square free parts of

2 2 2 2 2 2 d − λ1, d − λ2, ··· , d − λk are pairwise distinct. Then the angles

π, arccos(λ1/d), arccos(λ2/d), ··· , arccos(λk/d) are linearly independent over Q.

20 p 2 2 Proof. Let θr = arccos(λr/d). Then tan(θr) is a rational multiple of d − λr. Suppose for some qr ∈ Q,

q1θ1 + q1θ2 + ··· + qkθk ∈ Qπ.

By Theorem 7.3, each θr√is a rational√ multiple of π. It follows from Theorem 7.2 that tan(θr) ∈ {0, ± 3, ±1/ 3, ±1}. However, this is impossible as λr is an integer and λr 6= d/2.

7.6 Lemma. Let p be an odd prime. For any two distinct integers j, k ∈ {1, 2, ··· , (p − 1)/2}, the square free parts of j(p − j), k(p − k) are distinct. Proof. First note that two integers m and n have the same square free part if and only if mn is a perfect square. Now write p = a + b = c + d for some positive integers a, b, c, d. Suppose ac and bd have the same square free part. Then abcd is a perfect square. It follows that ac and bd have the same square free part, say ∆1, and ad and bc have the same square free part, say ∆2. As p2 = (a + b)(c + d) = (ac + bd) + (ad + bc), 2 gcd(∆1, ∆2) divides p . If p were a factor of ∆1, it would divide a or c, which contradicts our choice of a, b, c, d. Hence, ∆1 and ∆2 are coprime, and so gcd(ac, ad) is a perfect square. On the other hand, c and d are also coprime, which implies gcd(ac, ad) = a gcd(c, d) = a. Therefore, a is a perfect square. A similar argument shows that b, c and d are all perfect squares. Since every prime can be expressed in at most one way as a sum of two squares, either a = c or a = d. We now show that pretty good state transfer occurs on every hypercube Qp with prime p.

7.7 Theorem. Let p be a prime. Let X be the hypercube Qp, and let H be the normalized adjacency matrix of X. Pick any W such that H = W ◦ W ∗, and W ◦ W is row-stochastic. The quantum walk with respect to (X,W ) has pretty good state transfer between the antipodal vertices.

21 Proof. We prove this for odd prime p, as Q2 = 2K2. It is known [4] that the adjacency eigenvalues of Qp are

λr = p − 2r, r = 0, 1, ··· , p, and, relative to A(X), the antipodal vertices a and b are 2-strongly cospectral with 1 0 Λab = {p − 2r : r odd}, Λab = {p − 2r : r even} p Let θr = arccos(λr/p). Then tan(θr) is a rational multiple of (p − r)r. By Lemma 7.5, Lemma 7.6 and symmetry of the eigenvalues,

π, θ1, θ3, ··· , θp are linearly independent over Q. Now

X 0 (`r − `r)θr + `−pπ ≡ 0 (mod 2π). 0

We may rewrite the left hand side as ! X 0 0 X 0 (`r − `r − `p−r + `p−r)θr + `−p + (`p−r − `p−r) π. 0

It follows that 0 0 `p−r − `p−r = `r − `r for every odd r, and

X 0 `−p + (`r + `r) ≡ 0 (mod 2), 0

Thus, Equation (12) in Theorem 6.2 holds.

It remains√ open to determine n for which the quantum walk relative to (Qn,A(Qn)/ n) has pretty good state transfer.

7.2 Cyclic covers A graph Y is called a cover of X if there is a surjective homomorphism from Y to X that is locally bijective. For instance, the hypercube Q3 covers the

22 complete graph K4, where the covering map sends each pair of antipodal vertices in Q3 to a distinct vertex in K4. In general, covers can be constructed through symmetric arc functions, that is, a function f from the arcs of X to Sym(r), such that

f(a, b) = f(b, a)−1 for any two adjacent vertices a and b. Let Y f be the graph with vertex set V (X) × {1, 2, ··· , r} and edge set

{((a, i), (b, f(a, b)f(i))) : a ∼ b}

Then Y f is a cover of X, called an r-fold cover. An r-fold cover is cyclic if the image of f generates a cyclic group. If H is a Hermitian adjacency matrix of X whose entries are r-th roots of unity, then H naturally deﬁnes a symmetric arc function and hence an r-fold cover of X. Moreover, the spectrum of H is a subset of the adjacency spectrum of the cover. An important topic in algebraic graph theory is to construct covers of graphs whose adjacency spectrum diﬀers very little from that of the underlying graph; we refer interested readers to [8] and [10] for more details. Given any graph X, the Seidel adjacency matrix of X is J − I − 2A, which can been seen as a signed adjacency matrix, and hence a double cover, of the complete graph. While complete graphs on more than two vertices do not have strongly cospectral vertices relative to the adjacency matrix, they do relative to some signed adjacency matrices. We use these matrices to construct walks that admit pretty good state transfer on complete graphs. First, we cite an old result due to Erdos. 7.8 Theorem. [7] A product of consecutive positive integers is never a perfect square. 7.9 Theorem. Let n be a positive integer that is not twice a perfect square. Let X = nK2 and let H be the Seidel adjacency matrix of X. Let W be any matrix such that H = (2n − 1)(W ◦ W ∗) and W ◦ W is a row-stochastic matrix. The quantum walk with respect to (K2n,W ) has pretty good state transfer between every pair of adjacent vertices in X. Proof. The Seidel adjacency matrix of X is

H = In ⊗ I2 + (Jn − In) ⊗ J2 − In ⊗ J2,

23 and it has spectral decomposition 1 1 1 1 H = (2n−3)· (J ⊗J )+1· I ⊗ I − J −3· I − J ⊗ J . 2n n 2 n 2 2 n n n 2 2 Let 0 and 1 be two adjacent vertices in X. We see that 0 and 1 are 2-stongly cospectral, with 0 1 Λ01 = {2n − 3, −3}, Λ01 = {1}. Since each row of H has 2n − 1 non-zero entries, we need to apply Theorem 6.2 to 2n − 3 3 1 θ = arccos , θ = arccos − , θ = arccos . 1 2n − 1 2 2n − 1 3 2n − 1 Note that the tangents of these angles are, respectively, rational multiples of square roots of

a1 = 2(n − 1), a2 = (n + 1)(n − 2), a3 = n(n − 1).

We claim that a1 and a3 do not have the same square free factor, and neither do a2 and a3. Indeed, by our choice of n, the product a1a3 is not a perfect square, and by Theorem 7.8, a2a3 is not a perfect square. Therefore, due to Theorem 7.3,

0 0 0 (`1 − `1)θ1 + (`2 − `2)θ2 + (`3 − `3)θ3 ≡ 0 (mod 2π)

0 implies that (`3 − ` )θ3 is a rational multiple of π. Moreover, tan(θ3) = p 3 2 n(n − 1), which shows that θ3 is not a rational multiple of π. Hence, 0 `3 = `3, and so Equation (12) in Theorem 6.2 holds. We now consider a special class of cyclic 4-fold covers, that is, Hermitian adjacency matrices with entries in {±1, ±i}, associated to oriented graphs. Given an orientation of X, the skew-adjacency matrix, denoted S(X), is the matrix whose ab-entry is 1 if (a, b) is an arc in the orientation, −1 if (b, a) is an arc in the orientation, and 0 otherwise. As H = iS(X) is Hermitian, the eigenvalues of S(X) are purely imaginary and symmetric about 0. Moreover, if Eλ is the projection onto the λ-eigenspace of H, then Eλ is the projection onto the (−λ)-eigenspace of H. The following lemma concerns an orientation of the Cartesian product K2X; its skew-adjacency spectrum has been studied in [6]. Here, we prove this construction preserves 2-strong cospectrality.

24 7.10 Lemma. Let S(X) be a skew-adjacency matrix of an oriented graph X. Let H(X) = iS(Y ). Let Y = K2X and deﬁne an orientation on Y by H(X) iI H(Y ) = . −iI −H(X) √ Then the eigenvalues of H(Y ) are ± λ2 + 1 for each eigenvalue λ of H(X). Moreover, any pair of 2-strongly cospectral vertices in X are also 2-strongly cospectral in Y .

Proof. First note that for any eigenvalue λ and eigenprojection Eλ,

Eλea = αEλeb if and only if each eigenvector z for λ satisﬁes hea, zi = αheb, zi. We now consider the eigenspaces of H(Y ). Suppose for some µ ∈ R and vectors x, y, x x H(Y ) = µ . y y Then (µI − H(X))x = iy, (µI + H(X))y = −ix, from which it follows that

H(X)2x = (µ2 − 1)x, H(X)2y = (µ2 − 1)y.

This proves the first part of the lemma. To see the second part, suppose a and b are 2-strongly cospectral relative to H(X). Then for each eigenvalue 0 λ of H(X) in the eigenvalue support of a, either both ±λ lie in Λab, or both x ±λ lie in Λ1 . Therefore, every eigenvector for H(Y ) with eigenvalue ab y p 2 0 µ satisfies hea, zi = αheb, zi, where α = 1 if µ − 1 ∈ Λab, and α = −1 if p 2 1 µ − 1 ∈ Λab. We show how cyclic covers may be used to generate pretty good state transfer. Let X be the line graph of the cube Q3. Since there are no strongly cospectral vertices relative to A(X), pretty good state transfer does not occur on the quantum walk with respect to (X,A(X)/2). However, the following orientation gives a Hermitian matrix relative to which 1 and 10 are strongly cospectral, and together with Lemma 7.10, it provides an infinite family of quantum walks with pretty good state transfer.

25 Figure 2: An orientation of the line graph of Q3

7.11 Theorem. Let X be the line graph of Q3 and let S(X) be the skew- adjacency matrix of the orientation shown in Figure 2. Deﬁne a sequence of Hermitian matrices as follows: Hd−1 iI H4 = iS(X),Hd = . −iI Hd−1

Let Xd be the underlying graph of Hd. Pick any Hermitian matrix Wd such ∗ that Hd = d(Wd ◦ Wd ) and Wd ◦ Wd is row-stochastic. Then for each odd prime p, pretty good state transfer occurs on the quantum walk with respect to (Xp,Wp). Proof. A direct computation shows that vertices 1 and 10 are 2-strongly cospectral in X, and

0 1 Λ1,10 = {±2}, Λ1,10 = {0, ±12}.

By Lemma 7.10, the same vertices are 2-strongly cospectral in Xd, and, relative to Hd, √ √ √ 0 1 Λ1,10 = {± d}, Λ1,10 = {± d − 4, ± d + 8}.

Since each row of Hd has exactly d non-zero entries, to check pretty good state transfer we consider the following linear relation:

3 X 0 0 ((`r − `r)θr + (`−r − `−r)(π − θr)) ≡ 0 (mod 2π), r=1

26 where √ ! √ √ d d − 4 d + 8 θ = arccos , θ = arccos , θ = arccos . 1 d 2 d 3 d

Hence the tangents of θ1, θ2, θ3 are rational multiples of square roots of the following, respectively,

2 2 2 a1 = (d − d)d, a2 = (d − (d − 4))(d − 4), a3 = (d − (d + 8))(d + 8).

Now suppose d = p for some odd prime p. By an argument similar to Lemma 7.6, if a1a2 or a1a3 is a perfect square, then d is a perfect square, and if a2a3 is a perfect square, then both d − 4 and d + 8 are perfect squares. Neither of these can happen when d = p is an odd prime. Thus,

0 0 `r − `r = `−r − `−r, r = 1, 2, 3 and by Theorem 6.2, pretty good state transfer occurs between vertices 1 and 10. We expect more examples of pretty good state transfer arising from cyclic covers of graphs; in particular, the Hermitian adjacency matrices deﬁned in [17], [13] and [18] are worth investigating.

7.3 Real-weighted graphs Compared to pretty good state transfer, perfect state transfer is rarer - the eigenvalue condition in Theorem 4.4 is particularly hard to satisfy when H is the normalized adjacency matrix. In this section, we use other real weights to obtain quantum walks that admit perfect state transfer.

7.12 Theorem. Let s be a positive integer and set n = 2s−1 −1. There exist two sets of odd integers

{p1, p2, . . . , pn}, {q1, q2, . . . , qn} such that gcd(pj, qj) = 1 for each j, and n X pj cos < 1. q j=1 j

27 Deﬁne p π p π p π p π D = diag 1, −1, cos 1 , − cos 1 , cos 2 , − cos 2 , ··· . q1 q1 q2 q2 Let P be the Hadamard matrix

1 1 ⊗s P = 1 −1 and H = PDP T .

Then H is a real-weighted adjacency matrix of Kn+1,n+1. Moreover, for each W such that H = (2n + 2)(W ◦ W ∗) and W ◦ W is row-stochastic, there is perfect state transfer between vertices 0 and 1 on the quantum walk with respect to (Kn+1,n+1,W ). Proof. Starting from s = 1, one can construct the set

{p1, p2, . . . , pn}, {q1, q2, . . . , qn} inductively such that pj is coprime to qj and the sum n X pj cos < 1. (14) q j=1 j

Now, it is easy to tell from P that all vertices are 2-strongly cospectral relative to H. In particular for 0 and 1, we have 0 pjπ Λ01 = {1} ∪ cos : j = 1, 2, ··· , n , qj 1 pjπ Λ01 = {−1} ∪ − cos : j = 1, 2, ··· , n . qj

Thus, t = q1q2 ··· qn satisﬁes the condition in Theorem 4.4. A bit calculation shows that H has constant row sum 2n + 2 and, due to (14), all non-zero entries are positive. Therefore, there exists W such that H = (2n + 2)(W ◦ W ∗) and W ◦ W is row-stochastic.

28 7.4 m-strongly cospectral vertices with m > 2 In all the examples we have seen so far, pretty good state transfer occurs between 2-strongly cospectral vertices. This raises the question whether m can take any value. Theorem 6.3 already rules out all odd values; we now show that pretty good state transfer can happen for any even integers m. 7.13 Theorem. Let m be any positive even integer and let α = e2πi/m. Let p > 3 be a prime. Deﬁne a Hermitian adjacency matrix H of K4 by  0 2α p p  2α 0 pα pα H =   .  p pα 0 2  p pα 2 0

For every W such that H = 2(p + 1)W ◦ W ∗ and W ◦ W is row-stochastic, there is pretty good state transfer between vertices 0 and 1 on the quantum walk with respect to (K4,W ). Proof. The spectral decomposition of H is

H = PDP ∗, where √   1 1 √2 0 1 α α − 2α 0  P =  √  2 1 −1 0 √2  1 −1 0 − 2 and D = diag(2 + 2p, 2 − 2p, −2, −2). Hence vertices 0 and 1 are m-strongly cospectral, and

1 m/2+1 Λ01 = {2 + 2p, 2 − 2p}, Λ01 = {−2}. Let 1 − p 2 θ = 0, θ = arccos − , θ = arccos − . 1 2 1 + p 3 1 + p According to Theorem 6.2, we need to show that

0 0 (`2 − `2)θ2 + (`3 − `3)θ3 ≡ 0 (mod 2π)

29 and 0 0 `2 + `2 + `3 + `3 + `1 = 0 imply 0 0 (`2 + `2) + (m/2 + 1)(`3 + `3) + `1 ≡ 0 (mod m). (15) √ Since tan(θ1) is rational multiple of p, while tan(θ3) is a rational multiple p 2 of p + 2p + 3, the pure geodetic angles θ1 and θ2 are linearly independent with π over the rationals. Thus,

0 0 `2 = `2, `3 = `3, and so (15) holds.

8 Future work

We have developed the theory for pretty good state transfer in discrete-time quantum walks. For the quantum walk with respect to (X,W ), the coin matrix of the quantum walk translates into a Hermitian adjacency matrix of X, which in turn determines the transport properties of the walk. While we have constructed some inﬁnite families of walks that admit pretty good state transfer, this is merely a start to exploring the rich connection between the spectrum of Hermitian adjacency matrices and the behavior of quantum walks. Below we list a few questions for future research in this area.

1. Strong cospectrality relative to the adjacency matrix A(X) has been studied in [11]. Since A(X) is real symmetric, any two vertices that are m-strongly cospectral relative to A(X) must be 2-strongly cospectral. It is worth extending the theory in [11] to m-strong cospectrality for larger m, in particular relative to (i) the signed adjacency matrices, (ii) the Hermitian adjacency matrix with entries in {1, ±i, 0} [17, 13], or (iii) the Hermitian adjacency matrix with entries in {1, e±2πi/3, 0} [18].

2. The basis for the rational vector space generated by geodetic angles was studied in [3]. It will be interesting to apply this result to determine pretty good state transfer between vertices whose eigenvalue supports consist of integers or quadratic integers.

30 3. There are strongly cospectral vertices that are not m-strongly cospectral. For example, relative to the following Hermitian adjacency matrix [13]:  0 i −i i   −i 0 i i  H =   ,  i −i 0 1  −i −i 1 0 vertices 2 and 3 are strongly cospectral, and there is an eigenvalue λ with −i arctan(1/2) Eλea = e Eλeb. Can pretty good state transfer occur on such type of vertices? If so, how do we test it eﬃciently?

References

[1] Leonardo Banchi, Gabriel Coutinho, Chris Godsil, and Simone Severini, Pretty good state transfer in qubit chains—The Heisenberg Hamiltonian, Journal of Mathematical Physics 58 (2017), no. 3.

[2] Jeﬀery Bergen, A Concrete Approach to Abstract Algebra, Academic Press, 2009.

[3] J. H. Conway, C. Radin, and L. Sadun, On angles whose squared trigonometric functions are rational, Discrete and Computational Geometry 22 (1999), no. 3, 321–332.

[4] Gabriel Coutinho, Chris Godsil, Krystal Guo, and Fr´ed´eric Van- hove, Perfect state transfer on distance-regular graphs and association schemes, Linear Algebra and Its Applications 478 (2015), 108–130.

[5] Gabriel Coutinho, Krystal Guo, and Christopher M. Van Bommel, Pretty good state transfer between internal nodes of paths, Quantum Information & Computation 17 (2017), no. 9-10.

[6] Denglan Cui and Yaoping Hou, On the skew spectra of Cartesian products of graphs, Electronic Journal of Combinatorics 20 (2013), no. 2, P19–P19.

31 [7] P. Erdos, Note on products of consecutive integers, Journal of the London Mathematical Society s1-14 (1939), no. 3, 194–198.

[8] C Godsil and A Hensel, Distance regular covers of the complete graph, Journal of Combinatorial Theory, Series B 56 (1992), no. 2, 205–238.

[9] Chris Godsil, Stephen Kirkland, Simone Severini, and Jamie Smith, Number-theoretic nature of communication in quantum spin systems, Physical Review Letters 109 (2012), no. 5.

[10] Chris Godsil, Maxwell Levit, and Olha Silina, Graph covers with two new eigenvalues, European Journal of Combinatorics 93 (2021), 103280.

[11] Chris Godsil and Jamie Smith, Strongly cospectral vertices, arXiv:1709.07975 (2017).

[12] Lov Grover, A fast quantum mechanical algorithm for estimating the me- dian, Proceedings of the 28th Annual ACM Symposium on the Theory of Computing (1996), 212–219.

[13] Krystal Guo and Bojan Mohar, Hermitian adjacency matrix of digraphs and mixed graphs, Journal of Graph Theory (2016).

[14] Yusuke Higuchi, Norio Konno, Iwao Sato, and Etsuo Segawa, Spectral and asymptotic properties of Grover walks on crystal lattices, Journal of Functional Analysis 267 (2014), no. 11, 4197–4235.

[15] Viv Kendon, Quantum walks on general graphs, International Journal of Quantum Information (2006), 791–805.

[16] Sho Kubota, Etsuo Segawa, and Tetsuji Taniguchi, Quantum walks de- ﬁned by digraphs and generalized Hermitian adjacency matrices, Quan- tum Information Processing 20 (2021), no. 3, 1–30.

[17] Xueliang Li and Jianxi Liu, Hermitian-adjacency matrices and Hermi- tian energies of mixed graphs, Linear Algebra and Its Applications 466 (2015), 182–207.

[18] Bojan Mohar, A new kind of Hermitian matrices for digraphs, Linear Algebra and Its Applications 584 (2020), 343–352.

32 [19] Etsuo Segawa and Yusuke Yoshie, Quantum search of matching on signed graphs, arXiv (2020).

[20] Mario Szegedy, Quantum speed-up of Markov chain based algorithms, 45th Annual IEEE Symposium on Foundations of Computer Science (2004), 32–41.

[21] Christopher M. Van Bommel, A complete characterization of pretty good state transfer on paths — Quantum Information & Computation, Quan- tum Information & Computation 19 (2019), no. 7-8.

[22] Luc Vinet and Alexei Zhedanov, Almost perfect state transfer in quantum spin chains, Physical Review A - Atomic, Molecular, and Optical Physics 86 (2012), no. 5, 052319.

[23] Hanmeng Zhan, Discrete Quantum Walks on Graphs and Digraphs, Ph.D. thesis, 2018.

[24] , An inﬁnite family of circulant graphs with perfect state transfer in discrete quantum walks, Quantum Information Processing 18 (2019), no. 12, 1–26.

[25] , Quantum walks on embeddings, Journal of Algebraic Combina- torics (2020), 1–27.