Quadratization in Discrete Optimization and Quantum Mechanics

Nike Dattani∗

An open source Book on quadratizations for classical computing, quantum an-

nealing, and universal adiabatic quantum computing.

When optimizing discrete functions, it is often easier when the function is quadratic than if
it is of higher degree. But notice that the cubic and quadratic functions:
b1 b2 + b2 b3 + b3 b4 − 4b1 b2 b3 (cubic), (1)
b1 b2 + b2 b3 + b3 b4 + 4b1 − 4b1 b2 − 4b1 b3 (quadratic), (2)
arXiv:1901.04405v2 [quant-ph] 23 Sep 2019

where each bi can either be 0 or 1, both never go below the value of -2, and all minima occur
at (b1 , b2 , b3 , b4 ) = (1, 1, 1, 0). Therefore if we are interested in the ground state of a discrete
function of degree k, we may optimize either function and get exactly the same result. Part I
gives more than 40 different ways to do this, almost all of them published in the last 5 years.

The binary variables bi can be either of the eigenvalues of the matrix b below, which is
related to the Pauli z matrix by z = 2b − 11. The Pauli matrices x, y, z, 11 are listed below:
1 0 1 0 0 1 0 −i 1 0
b ≡ ( ), z ≡ ( ), x ≡ ( ), y ≡ ( ) , 11 ≡ ( ) . (3)
0 0 0 −1 1 0 i 0 0 1
Any Hermitian 2 × 2 matrix can be written as a linear combination of the Pauli matrices, so we
can therefore describe the Hamiltonian of any number of spin-1/2 particles by a function of
Pauli matrices acting on each particle, for instance:
x1 y2 z3 y4 + y1 x2 z3 y4 + x1 x2 y3 (cubic), (4)
x1 y4 + x2 y4 + x3 (quadratic), (5)
where the coefficients tell us about the strengths of couplings between these particles. The
Schrödinger equation tells us that the eigenvalues of the Hamiltonian are the allowed energy
levels and their eigenvectors (wavefunctions) are the corresponding physical states. More
generally these do not have to be spins but can be any type of qubits, and we can encode
the solution to any problem in the ground state of a Hamiltonian, then solve the problem
by finding the lowest energy state of the physical system (this is called adiabatic quantum
computing). Eqs. (5) and (4) have exactly the same energy spectra, so Eq. (5) is an example of
a type of quadratization.
Two-body physical interactions occur more naturally than many-body interactions so Parts
II-III give more than 30 different ways to quadratize general Hamiltonians (some of these
methods may use d × d matrices instead of only the 2 × 2 matrices in Eq. (3), meaning that we
can have types of qudits that are not qubits). All of these methods were published during the
last 15 years.
The optimization problems of Eqs. (1)-(2) are specific cases of the type in Eqs. (4)-(5), but
with only b matrices.

∗
[email protected]
 2

Part I
Diagonal Hamiltonians (pseudo-Boolean functions)

I. METHODS THAT INTRODUCE ZERO AUXILIARY VARIABLES

A. Deduction Reduction (Deduc-reduc; Tanburn, Okada, Dattani, 2015)

Summary
We look for deductions (e.g. b1 b2 = 0) that must hold true at the global minimum.
These can be found by a priori knowledge of the given problem, or by enumerating
solutions of a small subset of the variables. We can then substitute high-order terms
using the low-order terms of the deduction, and add on a penalty term to preserve the
ground states [1].
Cost
• 0 auxiliary variables needed.

• For a particular value of m, we have (m

n
) different m-variable subsets of the n
variable problem, and (m)2 evaluations of the objective function to find all
n m

possible m-variable deductions, whereas 2n evaluations is enough to solve the

entire problem. We therefore choose m ⋘ n.
Pros
• No auxiliary variables needed.
Cons
• When deductions cannot be determined naturally (as in the Ramsey number
determination problem, see Example XIII), deductions need to be found by ‘brute
force’, which scales exponentially with respect to m. For highly connected systems
(systems with a large number of non-zero coefficients), the value of m required to
find even one deduction can be prohibitively large.
Example
Consider the objective function:
H4−local = b1 b2 (4 + b3 + b3 b4 ) + b1 (b3 − 3) + b2 (1 − 2b3 − b4 ) + F (b3 , b4 , b5 , . . . , bN ) (6)

where F is any quadratic polynomial in bi for i ≥ 3. Since

H4−local (1, 1, b3 , b4 , ...) > H4−local (0, 0, b3 , b4 , ...) , H4−local (0, 1, b3 , b4 , ...) , H4−local (1, 0, b3 , b4 , ...) ,
(7)
it must be the case that b1 b2 = 0. Specifically, for the 4 assignments of (b3 , b4 ), we see
that b1 b2 = 0 at every minimum of H4−local − F .
Using deduc-reduc we have:
H2−local = 6b1 b2 + b1 (b3 − 3) + b2 (1 − 2b3 − b4 ) + F (b3 , b4 , b5 , . . . , bN ), (8)

which has the same global minima as H4−local but one fewer quartic and one fewer cubic
term. The coefficient of b1 b2 was chosen as 6 because 6 ≥ max (4 + b3 + b3 b4 ).
Bibliography
• Original paper, with more implementation details, and application to integer
factorization: [1].
 3

B. ELC Reduction (Ishikawa, 2014)

Summary
An Excludable Local Configuration (ELC) is a partial assignment of variables that
make it impossible to achieve the minimum. We can therefore add a term that corre-
sponds to the energy of this ELC without changing the solution to the minimization
problem. In practice we can eliminate all monomials with a variable in which a variable
is set to 0, and reduce any variable set to 1. Given a general objective function we can
try to find ELCs by enumerating solutions of a small subset of variables in the problem
[2].
Cost
• 0 auxiliary variables needed.

• For a particular value of m, we have (m

n
) different m-variable subsets of the n
variable problem, and (m)2m evaluations of the objective function to find all
n

possible m-variable deductions, whereas 2n evaluations is enough to solve the

entire problem. We therefore choose m ⋘ n.
• Approximate methods exist which have been shown to be much faster and give
good approximations to the global minimum [2].
Pros
• No auxiliary variables needed.
Cons
• No known way to find ELCs except by ‘brute force’, which scales exponentially
with respect to m.
• ELCs do not always exist.
Example
Consider the objective function:

H3−local = b1 b2 + b2 b3 + b3 b4 − 4b1 b2 b3 . (9)

If b1 b2 b3 = 0, no assignment of our variables will we be able to reach a lower energy than

if b1 b2 b3 = 1. Hence this gives us twelve ELCs, and one example is (b1 , b2 , b3 ) = (1, 0, 0)
which we can use to form the polynomial:

H2−local = H3−local + 4b1 (1 − b2 )(1 − b3 ) (10)

= b1 b2 + b2 b3 + b3 b4 + 4b1 − 4b1 b2 − 4b1 b3 . (11)

In both cases Eqs. (9) and (11), the only global minima occur when b1 b2 b3 = 1.
Bibliography
• Original paper and application to computerized image denoising: [2].
 4

C. Groebner Bases

Summary
Given a set of polynomials, a Groebner basis is another set of polynomials that have
exactly the same zeros. The advantage of a Groebner basis is it has nicer algebraic
properties than the original equations, in particular they tend to have smaller degree
polynomials. The algorithms for calculating Groebner bases are generalizations of
Euclid’s algorithm for the polynomial greatest common divisor.
Work has been done in the field of ’Boolean Groebner bases’, but while the variables
are Boolean the coefficients of the functions are in F2 rather than Q.
Cost
• 0 auxiliary variables needed.
• O (22 ) in general, O(dn ) if the zeros of the equations form a set of discrete points,
n 2

where d is the degree of the polynomial and n is the number of variables [3].
Pros
• No auxiliary variables needed.

• General method, which can be used for other rings, fields or types of variables.
Cons
• Best algorithms for finding Groebner bases scale double exponentially in n.

• Only works for objective functions whose minimization corresponds to solving

systems of discrete equations, as the method only preserves roots, not minima.
Example
Consider the following pair of equations:

b1 b2 b3 b4 + b1 b3 + b2 b4 − b3 = b1 + b1 b2 + b3 − 2 = 0. (12)
Feeding these to Mathematica’s GroebnerBasis function, along with the binarizing
b1 (b1 − 1) = . . . = b4 (b4 − 1) = 0 constraints, gives a Groebner basis:

{b4 b3 − b4 , b2 + b3 − 1, b1 − 1} . (13)
From this we can immediately read off the solutions b1 = 1, b2 = 1 − b3 and reduce
the problem to b3 b4 − b4 = 0. Solving this gives a final solution set of: (b1 , b2 , b3 , b4 ) =
(1, 0, 1, 0), (1, 0, 1, 1), (1, 1, 0, 0), which should be the same as the original 4-local prob-
lem.
Bibliography
• Reduction and embedding of factorizations of all bi-primes less than 200, 000: [4].
 5

D. Application of ELM (Dattani, 2018)

Summary
We use the formula from [5] for representing any function of three binary variables:
f (b1 , b2 , b3 ) = (f (1, 1, 1) + f (1, 0, 0) − f (1, 1, 0) − f (1, 0, 1) − f (0, 1, 1) − f (0, 0, 0) + (14)
f (0, 0, 1) + f (0, 1, 0)) b1 b2 b3 + (f (0, 1, 1) + f (0, 0, 0) − f (0, 0, 1) − f (0, 1, 0)) b2 b3 + (15)
(f (1, 0, 1) + f (0, 0, 0) − f (0, 0, 1) − f (1, 0, 0)) b1 b3 + (16)
(f (1, 1, 0) + f (0, 0, 0) − f (1, 0, 0) − f (0, 1, 0)) b1 b2 + (f (0, 1, 0) − f (0, 0, 0)) b2 (17)
+ (f (1, 0, 0) − f (0, 0, 0)) b1 + (f (0, 0, 1) − f (0, 0, 0)) b3 + f (0, 0, 0). (18)

If the cubic term is zero, then the function becomes quadratic. We can use ELM
(Energy Landscape Manipulation) to change the energy landscape without changing
the ground state [6]. In this case, we apply ELM in order to make the cubic term zero.
Cost
• No auxiliary variables.
• May require many evaluations of the cubic function, varying coefficients in order
to find the right ELM coefficients.
• For a particular value of m, we have (m
n
) different m-variable subsets of the n
variable problem, and (m)2 evaluations of the objective function to find all
n m

possible m-variable deductions, whereas 2n evaluations is enough to solve the

entire problem. We therefore choose m ⋘ n.
Pros
• Can be generalized to arbitrary k-local functions, but the ELM constraints may
become harder to achieve.
• Can quadratize an entire cubic function (or a cubic part of a more general function)
with no auxiliary qubits.
• Can reproduce the full spectrum.
Cons
• May not always be possible.
• May require a local search to find appropriate deductions.
Example
In order to reduce the number of constraints required for the cubic term to be zero, we
will assume that Deduc-Reduc told us that (1−b1 )(1−b2 )+(1−b2 )(1−b3 )+(1−b1 )(1−b3 ) = 0
when the overall function is minimized, which means the ground state only occurs
when at least two variables are 1. This does not allow us to assign the linear terms, but
assigns all quadratic terms to have bi bj = 1. This also suggests that b1 b2 b3 can also be
reduced to a linear term, but we do not know whether it is b1 , b2 , or b3 . So we have the
following constraint on the cubic term, after setting all f (b1 , b2 , b3 ) to zero if there is not
at least two 1’s:

f (1, 1, 1) − f (1, 1, 0) − f (1, 0, 1) − f (0, 1, 1) = 0. (19)

Bibliography
• The method was first presented in the first arXiv version of this book [7].
 6

E. Split Reduction (Okada, Tanburn, Dattani, 2015)

Summary
It is possible to reduce a lot of the problem by conditioning on the most connected
variables. We call each of these operations a split.
Cost
Usually slightly sub-exponential in the number of splits, as the number of problems
to solve at most doubles with every split, but often does not double (since entire cases
can get eliminated by some splits, as in the example below).
Pros
• This method can be applied to any problem and can be very effective on problems
with a few very connected variables.
Cons
• Multiple runs of the optimization procedure need to be made, and the number of
runs can often grow almost exponentially with respect to the number of splits.
Example
Consider the simple objective function

H = 1 + b1 b2 b5 + b1 b6 b7 b8 + b3 b4 b8 − b1 b3 b4 . (20)

In order to quadratize H, we first have to choose a variable over which to split. In this
case b1 is the obvious choice since it is present in the most terms and contributes to the
quartic term.
We then obtain two different problems:
H0 = 1 + b3 b4 b8 (21)
H1 = 1 + b2 b5 + b6 b7 b8 + b3 b4 b8 − b3 b4 . (22)
At this point, we could split H0 again and solve it entirely, or use a variable we saved in
the previous split to quadratize our only problem.
To solve H1 , we can split again on b8 , resulting in two quadratic problems:

H1,0 = 1 + b2 b5 − b3 b4 (23)
H1,1 = 1 + b2 b5 + b6 b7 . (24)
Now both of these problems are quadratic. Hence we have reduced our original,
hard problem into 3 easy problems, requiring only 2 extra (much easier) runs of our
minimization algorithm, and without needing any auxiliary variables.

Note that the number of quadratic problems to solve is 3, which is smaller than 22
which would be the "exponential" cost if the number of problems were (hypothetically)
to double with each split. This is a good example of the typical sub-exponential scaling
of split-reduc.
Bibliography
• Original paper and application to Ramsey number determination: [8].
 7

II. METHODS THAT INTRODUCE AUXILIARY VARIABLES TO QUADRATIZE A

SINGLE NEGATIVE TERM (NEGATIVE TERM REDUCTIONS, NTR)

A. NTR-KZFD (Kolmogorov & Zabih, 2004; Freedman& Drineas, 2005)

Summary
For a negative term −b1 b2 ...bk , introduce a single auxiliary variable ba and make the
substitution:
− b1 b2 . . . bk → (k − 1)ba − ∑ bi ba . (25)
i

Cost
• 1 auxiliary variable for each k-local term.
Pros
• All resulting quadratic terms are submodular (have negative coefficients).
• Can reduce arbitrary order terms with only 1 auxiliary.
• Reproduces the full spectrum.
Cons
• Only works for negative terms.
Example

H6−local = −2b1 b2 b3 b4 b5 b6 + b5 b6 , (26)

has a unique minimum energy of -1 when all bi = 1.

H2−local = 2 (5ba − b1 ba − b2 ba − b3 ba − b4 ba − b5 ba − b6 ba ) + b5 b6 (27)

has the same unique minimum energy, and it occurs at the same place (all bi = 1), with
ba = 1.
Alternate Forms

−b1 b2 . . . bk = min ((k − 1 − ∑ bi )ba ) (28)

ba i

→ ((k − 1 − ∑ bi )ba ) . (29)

i

Alternate Names
• "Standard quadratization" of negative monomials [9].
• sk (b, ba ) [9]
Bibliography
• 2004: Kolmogorov and Zabih presented this for cubic terms [10].
• 2005: Generalized to arbitrary order by Freedman and Drineas [11].
• Discussion: [12], [13].
 8

B. NTR-ABCG (Anthony, Boros, Crama, Gruber, 2014)

Summary
For a negative term −b1 b2 ...bk , introduce a single auxiliary variable ba and make the
substitution:
k−1 k−1 k
− b1 b2 . . . bk → ∑ bi − ∑ bi bk − ∑ bi ba + (k − 1)bk ba . (30)
i i i

Cost
• 1 auxiliary variable for each k-local term.
• 1 non-submodular term for each k-local termi (and it is quadratic).
Pros
• Can reduce arbitrary order terms with only 1 auxiliary.
• Reproduces the full spectrum.
Cons
• Only works for negative terms.
• Turns a symmetric term into a non-symmetric term (but only bk is asymmetric).
Example

H6−local = −2b1 b2 b3 b4 b5 b6 + b5 b6 , (31)

has a unique minimum energy of -1 when all bi = 1.

H2−local has the same unique minimum energy, and it occurs at the same place (all
bi = 1), with ba = 1.
Alternate Forms

−b1 b2 . . . bk → (k − 1)bk ba − ∑ bi (ba + bk − 1) (32)

i
k−1
= (k − 2)bk ba − ∑ bi (ba + bk − 1) (33)
i

Alternate Names
• "Extended standard quadratization" of negative monomials (see Eqs. 25-26 of [9]).
• sk (b, ba )+ [9].
Bibliography
• 2014, First presentation: [13, 14].
• Further discussion: [9].
 9

C. NTR-ABCG-2 (Anthony, Boros, Crama, Gruber, 2016)

Summary
For a negative term −b1 b2 ...bk , introduce a single auxiliary variable ba and make the
substitution:

−b1 b2 . . . bk → (2k − 1) ba − 2 ∑ bi ba (34)

i

Cost
• 1 auxiliary variable for each k-local term.

• 1 non-submodular term for each k-local term (and it is linear).

Pros
• Can reduce arbitrary order terms with only 1 auxiliary.

• Reproduces the full spectrum.

• The non-submodular term is linear as opposed to NTR-ABCG-1 whose non-
submodular term is quadratic.
• Symmetric with respect to all non-auxiliary variables.
Cons
• Only works for negative terms.
• Turns a symmetric term into a non-symmetric term (but only bk is asymmetric).

• Coefficients of quadratic terms are twice the size of their size in NTR-KZFD or
NTR-ABCG-1, and roughly twice the size for the linear term.
Example

H6−local = −2b1 b2 b3 b4 b5 b6 + b5 b6 , (35)

has a unique minimum energy of -1 when all bi = 1.

H2−local has the same unique minimum energy, and it occurs at the same place (all
bi = 1), with ba = 1.
Alternate Forms

1 k
−b1 b2 . . . bk = 2ba (k − − ∑ bi ) (36)
2 i=1

(34) can be generalized as follows:

−b1 b2 . . . bk → (Ck − 1) ba − C ∑ bi ba (37)

i

where C ≥ 1 is a constant. NTR-KZFD is a particular case of (37) where C = 1.

Bibliography
• Discussion: [13].
 10

D. NTR-GBP (“Asymmetric cubic reduction”, Gallagher, Batra, Parikh, 2011)

Summary

−b1 b2 b3 → ba (−b1 + b2 + b3 ) − b1 b2 − b1 b3 + b1 (38)

→ ba (−b2 + b1 + b3 ) − b1 b2 − b2 b3 + b2 (39)
→ ba (−b3 + b1 + b2 ) − b2 b3 − b1 b3 + b3 (40)

Cost
• 1 auxiliary variable per negative cubic term.
Pros
• Asymmetric which allows more flexibility in cancelling with other quadratics.
Cons
• Only works for negative cubic monomials.
Example
−b1 b2 b3 + b1 b3 − b2 = minba (ba − b1 ba − b3 ba + b2 ba + 2b1 b3 ) − b2 (41)

Alternate Forms

−b1 b2 b3 = min (ba − b1 + b2 + b3 − b1 b2 − b1 b3 + b1 ) (42)

ba

• By starting with (40), and flipping ba (i.e. setting ba → 1 − b̄a and relabelling ba → b̄a
since ba does not appear anywhere else in the function being quadratized), we
see that NTR-GBP can actually be derived from NTR-ABCG with k = 3).
Bibliography
• First introduced in: [15].
 11

E. NTR-RBL (Rocchetto, Benjamin, Li, 2016)

Summary
Using a ternary variable tq ∈ −1, 0, 1 we have:

− z1 z2 z3 → (1 + 4ta + z1 + z2 + z3 ) − 1.
2
(43)

Cost
• 1 auxiliary ternary variable
Pros
• One of the only methods designed specifically for z variables.
• Symmetric with respect to all variables.
Cons
• The auxiliary variable required is ternary (a qutrit).
• Requires all possile quadratic terms and they are all non-submodular.
• Only reproduces the ground state manifold.
Example

−b1 b2 b3 = minba ba − b1 + b2 + b3 − b1 b2 − b1 b3 + b1 (44)

Alternate Forms
4 4 4
−z1 z2 z3 z4 → 16 t2a + 4 ta ∑ zi + 2 ∑ ∑ zi zj + 4 (45)
i=1 i=1 j>i

Bibliography
• Original paper where they introduce this gadget for the end points in the LHZ
lattice: [16].
 12

F. NTR-LHZ (Lechner, Hauke, Zoller, 2015)

Summary
Extra binary or ternary variable is added to ensure that the energy of the even parity
sector is zero and the energy of the odd sector is higher:
Cost
• 1 auxiliary ternary variable (see appendix for transformation to a binary variable)
Pros
• One of the only methods designed specifically for z variables.
Cons
• Only reproduces the ground state manifold, not higher excited states.

• Requires all possile quadratic terms and they are all non-submodular.
• Only reproduces the ground state manifold.
Example
−z1 z2 z3 z4 = −16 b1 b2 b3 b4 + 8 (b1 b2 b3 + b1 b2 b4 + b1 b3 b4 + b2 b3 b4 ) −
4 (b1 b2 + b1 b3 + b1 b4 + b2 b3 + b2 b4 + b3 b4 ) + 2 (b1 + b2 + b3 + b4 ) − 1
4 4 4
→ 16 t2a + 8 ta ∑ bi + 8 ∑ ∑ bi bj + 16 (46)
i=1 i=1 j>i

Alternate Forms
4 4 4
−z1 z2 z3 z4 → 16 t2a + 4 ta ∑ zi + 2 ∑ ∑ zi zj + 4 (47)
i=1 i=1 j>i

Bibliography
• Original paper: [17] (Eq. 4).
• Also given in Eq. 10 of [16] for i + 2 < j.
 13

III. METHODS THAT INTRODUCE AUXILIARY VARIABLES TO QUADRATIZE A

SINGLE POSITIVE TERM (POSITIVE TERM REDUCTIONS, PTR)

A. PTR-BG (Boros and Gruber, 2014)

Summary
By considering the negated literals b̄i = 1 − bi , we recursively apply NTR-KZFD to
b1 b2 . . . bk = −b̄1 b2 . . . bk + b2 b3 . . . bk . The final identity is:
k−2 k
b1 b2 . . . bk → ( ∑ bai (k − i − 1 + bi − ∑ bj )) + bk−1 bk (48)
i=1 j=i+1

Cost
• k − 2 auxiliary variables for each k-local term.
Pros
• Works for positive monomials.
Cons
• k − 1 non-submodular quadratic terms.
Example
b1 b2 b3 b4 → ba1 (2 + b1 − b2 − b3 − b4 ) + ba2 (1 + b2 − b3 − b4 ) + b3 b4 (49)

Bibliography
• Summary: [18].
 14

B. PTR-Ishikawa (Ishikawa, 2011)

Summary
This method re-writes a positive monomial using symmetric polynomials, so all
possible quadratic terms are produced and they are all non-submodular:
nk k
b1 ...bk → (∑ bai (ci,d (− ∑ bj + 2i) − 1) + ∑ bi bj ) (50)
i=1 j=1 i<j

1, i = nd and k is odd,
where nk = ⌊ k−1
2
⌋ and ci,k = {
2, else.
Cost
• ⌊ k−1
2
⌋ auxiliary variables for each k-order term
• O(kt) for a k-local objective function with t terms.
Pros
• Works for positive monomials.
• About half as many auxiliary variables for each k-order term as the previous
method.
• Reproduces the full spectrum.
Cons
• O(k 2 ) quadratic terms are created, which may make chimerization more costly.
k(k−1)
• 2 non-submodular terms.
• Worse than the previous method for quartics, with respect to submodularity.
Example
b1 b2 b3 b4 → (3 − 2b1 − 2b2 − 2b3 − 2b4 )ba + b1 b2 + b1 b3 + b1 b4 + b2 b3 + b2 b4 + b3 b4 (51)

Alternate Forms
For even k, and equivalent expression is given in [19]:

b1 b2 . . . bk → ∑ bi + ∑ bi bj + ∑ ba2i (4i − 2 − ∑ bj ) (52)

i ij 2i j

→ ∑ bi + 2 ∑ ba2i (2i − 1) + ∑ bi bj − ∑ bj ba2i (53)

i 2i ij 2i,j

Alternate Names
• "Ishikawa’s Symmetric Reduction" [15].
• "Ishikawa Reduction"
• "Ishikawa"
Bibliography
• Original paper and application to image denoising: [12].
• Equivalent way of writing it for even k, shown in [19].
 15

C. PTR-BCR-1 (Boros, Crama, and Rodríguez-Heck, 2018)

Summary
This is very similar to the alternative form of Ishikawa Reduction, but works for odd
values of k, and is different from Ishikawa Reduction:

b1 b2 . . . bk → ∑ bi + ∑ (4i − 3) ba2i−1 + ∑ bi bj − ∑ bj ba2i−1 (54)

i 2i−1 ij 2i−1,j

Cost
• Same number of auxiliaries as Ishikawa Reduction.
Pros
• Same as for Ishikawa Reduction.
Cons
• Same as for Ishikawa Reduction.

• Only works for odd k, but for even k we have an analogous method which is
equivalent to Ishikawa Reduction.
Alternate Forms

b1 b2 . . . bk → ∑ bi + ∑ bi bj + ∑ ba2i−1 (4i − 3 − ∑ bj ) (55)

i ij 2i−1 j

Bibliography
• Original paper: [19].
 16

D. PTR-BCR-2 (Boros, Crama, and Rodríguez-Heck, 2018)

Summary
Let ⌈ k4 ⌉ ≤ m ≤ ⌈ k2 ⌉,

m−1
b1 b2 ⋯bk → αb ∑ bi + αba,1 ∑ bai + αba,2 bam + αbb ∑ bi bj + αbba,1 ∑ ∑ bi baj +
i i ij i j
m−1 m−1
(56)
α bba,2
∑ bi bam + α ba,1 ba,1
∑ bai baj + α ba,1 ba,2
∑ bai bam ,
i ij i

where:

⎛ α
b αbba,1 ⎞ ⎛ −1/2 −1 ⎞
⎜α ba,1 α bb a,2
⎟ ⎜ 1 −2 ⎟
⎜ ba,2 ba,1 ba,1 ⎟ = ⎜ 1 ⎟. (57)
⎜α α ⎟ ⎜ 2 (n − m + n2 − 2mn + m2 ) −(n − m)⎟
⎝ αbb αba,1 ba,2 ⎠ ⎝ 1/2 4(n − m) ⎠

Cost
⌈ k4 ⌉ auxiliary qubits per positive monomial.
Pros
• Smallest number of auxiliary coefficients that scales linearly with k.
• Smaller coefficients than the logarithmic reduction.
Cons
• Introduces many non-submodular terms.
Example
We quadratize a quartic term with only 1 auxiliary (half as many as in PTR-Ishikawa):
1
b1 b2 b3 b4 → (b1 + b2 + b3 + b4 − 2ba1 ) (b1 + b2 + b3 + b4 − 2ba1 − 1) (58)
2

Bibliography
• Original appears in: Theorem 7 of [19], and Theorem 10 of [20].
 17

E. PTR-BCR-3 (Boros, Crama, and Rodríguez-Heck, 2018)

Summary
Pick m such that k < 2m+1 ,

b1 b2 . . . bk → α + αb ∑ bi + αbai ∑ 2i−1 bai + αbb ∑ bi bj + αbba ∑ bi baj + αbai baj bai baj , (59)
i i ij ij

where,

⎛ α αbb ⎞ ⎛ (2m − k)
2
1 ⎞
⎜ α α ⎟ = ⎜ 2 (2 − k) 2 ⎟ .
b bb a m j−1 (60)
⎝αba αba ba ⎠ ⎝−2 (2m − k) 2i+j−2 ⎠

Cost
⌈log k⌉ auxiliary qubits per positive monomial.
Pros
• Logarithmic number of auxiliary variables.
Cons
• Introduces all terms non-submodular except for the term linear in auxiliaries.
Example

1
b1 b2 b3 b4 → (4 + b1 + b2 + b3 + b4 − ba1 − 2ba2 ) (3 + b1 + b2 + b3 + b4 − ba1 − 2ba2 ) (61)
2

Alternate Forms

2
b1 b2 . . . bk → (2 − k + ∑ bi − ∑ 2
m i−1
bai ) (62)
i i

Bibliography
• Original paper (Theorem 4, special case of Theorem 1): [19].
 18

F. PTR-BCR-4 (Boros, Crama, and Rodríguez-Heck, 2018)

Summary
Pick m such that k ≤ 2m+1 ,

m m
1 m+1
b1 . . . b k → (2 − k + ∑ bi − ∑ 2i bai ) (2m+1 − k + ∑ bi − ∑ 2i bai − 1) . (63)
2 i i i i

Cost
⌈log k⌉ − 1 auxiliary qubits per positive monomial.
Pros
• Logarithmic number of auxiliary variables.
Cons
• Introduces many non-submodular terms.
Example

1
b 1 b2 b3 b4 → (b1 + b2 + b3 + b4 − 2ba ) (b1 + b2 + b3 + b4 − 2ba − 1) (64)
2

Alternate Forms
Let X = ∑ bi and N = 2m+1 − k,

n n
1
b1 . . . bk = min (N + X − ∑ 2i b′i ) (N + X − ∑ 2i b′i − 1) . (65)
b1 ,...bn 2
′ ′
i=1 i=1

Bibliography
• Original paper (Theorem 5): [19], Also in (Theorem 9): [20].
 19

G. PTR-BCR-5 (Boros, Crama, and Rodríguez-Heck, 2018)

Summary
Pick m such that k ≤ 2m+1 ,

m m
1 m+1
b1 . . . b k → (2 − k + ∑ bi − ∑ 2i bai ) (2m+1 − k + ∑ bi − ∑ 2i bai − 1) . (66)
2 i i i i

Cost
⌈log n⌉ − 1 auxiliary qubits per positive monomial.
Pros
• Logarithmic number of auxiliary variables.

• "As mentioned in Section 1, Theorem 9 provides a significant improvement over

the best previously known quadratizations for the Positive monomial, and the
upper bound on the number of auxiliary variables precisely matches the lower
bound presented in Section 3."
Cons
• Introduces many non-submodular terms.
Example

1
b 1 b2 b3 b4 → (b1 + b2 + b3 + b4 − 2ba ) (b1 + b2 + b3 + b4 − 2ba − 1) (67)
2

Alternate Forms
Let X = ∑ bi and N = 2m+1 − k,

n n
1
b1 . . . bk = min (N + X − ∑ 2i b′i ) (N + X − ∑ 2i b′i − 1) . (68)
b1 ,...bn 2
′ ′
i=1 i=1

Bibliography
• Original paper (Theorem 9): [20].
 20

H. PTR-BCR-5 (Boros, Crama, and Rodríguez-Heck, 2018)

Summary
Pick m such that k ≤ 2m+1 ,

m m
1 m+1
b1 . . . b k → (2 − k + ∑ bi − ∑ 2i bai ) (2m+1 − k + ∑ bi − ∑ 2i bai − 1) . (69)
2 i i i i

Cost
⌈log n⌉ auxiliary qubits per positive monomial.
Pros
• Logarithmic number of auxiliary variables.

• "As mentioned in Section 1, Theorem 9 provides a significant improvement over

the best previously known quadratizations for the Positive monomial, and the
upper bound on the number of auxiliary variables precisely matches the lower
bound presented in Section 3."
Cons
• Introduces many non-submodular terms.
Example

1
b 1 b2 b3 b4 → (b1 + b2 + b3 + b4 − 2ba ) (b1 + b2 + b3 + b4 − 2ba − 1) (70)
2

Alternate Forms
Let X = ∑ bi and N = 2m+1 − k,

n n
1
b1 . . . bk = min (N + X − ∑ 2i b′i ) (N + X − ∑ 2i b′i − 1) . (71)
b1 ,...bn 2
′ ′
i=1 i=1

Bibliography
• Original paper (Remark 5): [20].
 21

I. CCG-based (Yip, Xu, Koenig and Kumar, 2019)

Summary
The CCG-based quadratization algorithm is an iterative algorithm. The CCG (Con-
straint composite graph) is a combinatorial structure associated with an optimization
problem posed as the weighted constraint satisfaction problem.
Cost
• Each positive monomial of degree i generates 2 auxiliary variables when it is
reduced to the sum of a quadratic polynomial and a monomial of degree i − 1,
which can then be combined with existing monomials of degree i − 1 if they are
composed of the same variables. This combination of monomials can take place in
each iteration, until the whole pseudo-Boolean function (PBF) becomes quadratic.
• Same as Ishikawa’s method, use 1 auxiliary variable for each negative monomial.
• For a k-local objective function, use a factor of k less of auxiliary variables asymp-
totically compared to Ishikawa’s method.
Pros
• Due to the recombinations of terms during its iterative reduction process, the
resulting number of auxiliary variables is less than Ishikawa’s method especially
for PBFs with many positive monomials and many terms (the difficult case).
• The higher the degree of the PBF is, the more advantageous the CCG-based
quadratization method is. It works particularly well for problems in real life such
as planning problem that requires a high-degree PBF formulation.
• Due to the nature of recombinations of terms during each iteration, the number
of quadratic terms in the finalized quadratic PBFs is less than Ishikawa’s method.
Cons
• Can lead to more auxiliary variables for sparse PBFs (the number of monomials
is much less than the maximum number the PBFs can have) and PBFs with low
degree.
Example
Each degree-d monomial in the PBF f (⃗
x) in one reduction step is substituted as:
d−1
ax1 . . . xd = min [axa + LxL + J ∑ (1 − xi )(1 − xa )
xa ,xL
i=1
(72)
+ J(1 − xd )(1 − xL ) + J(1 − xL )(1 − xa )]

−L(1 − xd ) − a + ax1 . . . xd−1 ,

d
−ax1 . . . xd = min [axa′ + J ∑(1 − xi )(1 − xa′ )] − a (73)
xa′
i=1

where J ≥ L > a > 0. xa and xL are the two auxiliary variables introduced for positive
monomial and xa′ is the auxiliary variable introduced for negative monomial.
Bibliography
• 2019, original paper: [21].
• 2019, (Theorem 1): [21], is inspired by 2008, (Theorem 3): [22].
 22

J. PTR-KZ (Kolmogorov & Zabih, 2004)

Summary
This method can be used to re-write positive or negative cubic terms in terms of 6
quadratic terms. The identity is given by:

b1 b2 b3 → 1 − (ba + b1 + b2 + b3 ) + ba (b1 + b2 + b3 ) + b1 b2 + b1 b3 + b2 b3 (74)

Cost
• 1 auxiliary variable per positive or negative cubic term.
Pros
• Works on positive or negative monomials.
• Reproduces the full spectrum.
Cons
• Introduces all 6 possible non-submodular quadratic terms.
Alternate Names
• "Reduction by Minimum Selection" [12, 15].
Bibliography
• Original paper: [10].
 23

K. PTR-KZ (in terms of z)

Summary
The formula is almost the same as in the version of PTR-KZ on the previous page
(which is written in terms of b), but with a factor of 2, a change of sign for the linear
terms, and a slight change in the constant term. This formula can be obtained directly
from the PTR-KZ quadratization formula in terms of b, by starting with 8b1 b2 b3 on the
left-side, making the substitution bi → (1 + zi )/2, and removing all terms that appear on
both sides of the equation. The result is:

±z1 z2 z3 → 3 ± (z1 + z2 + z3 + za ) + 2za (z1 + z2 + z3 ) + z1 z2 + z1 z3 + z2 z3 . (75)

Cost
• 1 auxiliary variable per positive or negative cubic term.
Pros
• Works on positive or negative monomials.

• Reproduces the full spectrum.

Cons
• Introduces all 6 possible non-submodular quadratic terms.
Bibliography
• 2004: published by Kolmogorov and Zabih in terms of b variables: [10].
• 31 March 2016: published by Chancellor, Zohren, and Warburton in terms of z,
and without the constant term: [23].
• 8 April 2016: published independently by Leib, Zoller, and Lechner in terms of z,
and without the constant term [24, 25].
 24

L. PTR-GBP (“Asymmetric reduction”, Gallagher, Batra, Parikh, 2011)

Summary
Similar to other methods of reducing one term, this method can reduce a positive cu-
bic monomial into quadratic terms using only one auxiliary variable, while introducing
fewer non-submodular terms than the symmetric version.
The identity is given by:

b1 b2 b3 → ba − b2 ba − b3 ba + b1 ba + b2 b3 (76)
→ ba − b1 ba − b3 ba + b2 ba + b1 b3 (77)
→ ba − b1 ba − b2 b a + b 3 ba + b1 b2 . (78)
(79)

Cost
1 auxiliary variable per positive cubic term.
Pros
• Works on positive monomials.
• Fewer non-submodular terms than Ishikawa Reduction.
Cons
• Only been shown to work for cubics.
Example
b1 b2 b3 + b1 b3 − b2 → (ba − b1 ba − b3 ba + b2 ba + 2b1 b3 ) − b2 (80)

Bibliography
• Original paper and application to computer vision: [15].
 25

M. PTR-RBL-(3→2) (Rocchetto, Benjamin, Li, 2016)

Summary
Using a ternary variable tq ∈ −1, 0, 1 we have:

z1 z2 z3 → (1 + 4ta + z1 + z2 + z3 )2 − 1. (81)

Cost
• 1 auxiliary ternary variable
Pros
• One of the only methods designed specifically for z variables.
• Symmetric with respect to all variables.
Cons
• The auxiliary variable required is ternary (a qutrit).
• Requires all possile quadratic terms and they are all non-submodular.
• Only reproduces the ground state manifold.
Example

−b1 b2 b3 = minba ba − b1 + b2 + b3 − b1 b2 − b1 b3 + b1 (82)

Alternate Forms
4 4 4
−z1 z2 z3 z4 → 16 t2a + 4 ta ∑ zi + 2 ∑ ∑ zi zj + 4 (83)
i=1 i=1 j>i

Bibliography
• Original paper where they introduce this gadget for the end points in the LHZ
lattice: [16].
 26

N. PTR-RBL-(4→2) (Rocchetto, Benjamin, Li, 2016)

Summary
Extra binary or ternary variable is added to ensure that the energy of the even parity
sector is zero and the energy of the odd sector is higher:
4 4 4
z1 z2 z3 z4 → 16 t2a + 4 ta ∑ zi + 2 ∑ ∑ zi zj + 4. (84)
i=1 i=1 j>i

Cost
• 1 auxiliary ternary variable
Pros
• One of the only methods designed specifically for z variables.
Cons
• Only reproduces the ground state manifold, not higher excited states.
• Requires all possile quadratic terms and they are all non-submodular.

• Only reproduces the ground state manifold.

Example
−z1 z2 z3 z4 = −16 b1 b2 b3 b4 + 8 (b1 b2 b3 + b1 b2 b4 + b1 b3 b4 + b2 b3 b4 ) −
4 (b1 b2 + b1 b3 + b1 b4 + b2 b3 + b2 b4 + b3 b4 ) + 2 (b1 + b2 + b3 + b4 ) − 1
4 4 4
→ 16 t2a + 8 ta ∑ bi + 8 ∑ ∑ bi bj + 16 (85)
i=1 i=1 j>i

Bibliography
• Original paper: [16].
 27

O. PTR-CZW (Chancellor, Zohren, Warburton, 2017)

Summary
Auxilliary qubits can be made to “count” the number of logical qubits in the 1
configuration. By applying single qubit terms to the auxilliary qubits, the spectrum of
any permutation symmetric objective function can be reproduced.
Cost
• For a k local coupler requires k auxilliary qubits.
Pros
• Natural flux qubit implementation [26].
• Single gadget can reproduce any permutation symmetric spectrum.
• High degree of symmetry means this method is natural for some kinds of quan-
tum simulations [27].
Cons
• Requires coupling between all logical bits and from all logical bits to all auxilliary
bits.
• Requires single body terms of increasing strength as k is increased.
Example
A 4 qubit gadget guarantees that the number of auxillary bits in the −1 state is equal
to the number of logical bits in the 1 state
4 i−1 4 4 4 4
H4−count = 4 ∑ ∑ bi bj + 4 ∑ ∑ bi baj − 15 ∑ bi − 8 ∑ bai + (5 ba1 + ba2 − 3 ba3 − 7 ba4 ) + 26 (86)
i=2 j=1 i=1 j=1 i=1 i=1

This gadget can be expressed more naturally in terms of z:

4 i−1
1 4 4 4
1
H4−count = ∑ ∑ zi zj − ∑ zi + ∑ ∑ zi zaj + (5za1 + za2 − 3 za3 − 7za4 ) . (87)
i=2 j=1 2 i=1 i=1 j=1 2

To replicate the spectrum of b1 b2 b3 b4 , we add

H2−local = −ba4 + λH4−count . (88)
where λ is a large number.
For the spectrum of
z1 z2 z3 z4 = 16 b1 b2 b3 b4 − 8 (b1 b2 b3 + b1 b2 b4 + b1 b3 b4 + b2 b3 b4 )+
4 (b1 b2 + b1 b3 + b1 b4 + b2 b3 + b2 b4 + b3 b4 ) − 2 (b1 + b2 + b3 + b4 ) + 1, (89)
we implement,
H2−local = 2 ba1 − 2 ba2 + 2 ba3 − 2 ba4 + λH4−count , (90)

Bibliography
• Paper on flux qubit implementation: [28]
• Paper on MAX-k-SAT mapping: [28] (published in a journal earlier than [26] but
put on arXiv 1 month later).
• Talk including use in quantum simulation: [27]
 28

P. Bit flipping (Ishikawa, 2011)

Summary
For any variable b, we can consider the negation b̄ = 1 − b. The process of exchanging
b for b̄ is called flipping. Using bit-flipping, an arbitrary function in n variables can be
represented using at most 2(n−2) (n − 3) + 1 variables, though this is a gross overestimate.
Can be used in many different ways:
1. Flipping positive terms and using II A, recursively;

2. For α < 0, we can reduce αb̄1 b̄2 ...b̄k very efficiently to submodular form using II A.
A generalized version exists for arbitrary combinations of flips in the monomial
which makes reduction entirely submodular [12];
3. When we have quadratized we can minimize the number of non-submodular
terms by flipping.
4. We can make use of both bi and b̄i in the same objective function by adding on a
sufficiently large penalty term: λ(bi + b̄i − 1)2 = λ(1 + 2bi b̄i − bi − b̄i ). This is similar
to the ideas in reduction by substitution or deduc-reduc. In this way, given a
quadratic in n variables we can make sure it only has at most n nonsubmodular
terms if we are willing to use the extra n negation variables as well (so we have
2n variables in total).

Cost
• None, as replacing bi with it’s negation b̄i costs nothing except a trivial symbolic
expansion.
Pros
• Cheap and effective way of improving submodularity.

• Can be used to combine terms in clever ways, making other methods more
efficient.
Cons
• Unless the form of the objective function is known, spotting these ’factorizations’
using negations is difficult.
• We need an auxiliary variable for each bi for which we also want to use b¯i in the
same objective function.
Example
By bit-flipping b2 and b4 , i.e. substituting b2 = 1 − b̄2 and b4 = 1 − b̄4 , we see that:
H = 3b1 b2 + b2 b3 + 2b1 b4 − 4b2 b4 (91)
= −3b1 b̄2 − b̄2 b3 − 2b1 b̄4 − b̄2 b̄4 + 5b1 + b3 + 4b̄2 + 4b̄4 − 4. (92)
The first expression is highly non-submodular while the second is entirely submodular.
Bibliography
• Original paper: [12].
 29

IV. METHODS THAT QUADRATIZE MULTIPLE TERMS WITH THE SAME

AUXILIARIES (CASE 1: SYMMETRIC FUNCTION REDUCTIONS, SFR)

A symmetric function is one where if we switch any of the variable names (for
example b1 → b5 → b8 → b1 ), the function’s output is unaffected.

A. SFR-ABCG-1 (Anthony, Boros, Crama, Gruber, 2014)

Summary
Any n-variable symmetric function f (b1 , b2 , . . . bn ) ≡ f (b) can be quadratized with
n − 2 auxiliaries:

1
f (b) → −α0 − α0 ∑ bi + a2 ∑ bi bj + 2 ∑ (αi − c) bai (2i − − ∑ bj ) (93)
i ij i 2 j
min (α2j ) , i ∈ even
c={ (94)
min (α2j−1 ) , i ∈ odd
a2 = Determined from Page 12 of [29] (95)

Cost
• n − 2 auxiliaries for any n-variable symmetric function.
• n2 non-submodular quadratic terms (of the non-auxiliary variables).
• n − 2 non-submodular linear terms (of the auxiliary variables).
Pros
• Quadratization is symmetric in all non-auxiliary variables (this is not always true,
for example some of the methods in the NTR section).
• Reproduces the full spectrum.
• When there’s a large number of terms, there’s fewer auxiliary variables than
quadratizing each positive monomial separately.
Cons
• Only works on a specific class of functions, although the quadratizations of
arbitrary functions can be related to the quadratizations of symmetric functions
on a larger number of variables.
• All quadratic terms of the non-auxiliary variables are non-sub-modular.

• All linear terms of the auxliaries are non-submodular.

• Not meant so much to be practical, but rather an easy proof of an upper bound
on the number of needed auxiliaries.
Bibliography
• 2014, original paper (Theorem 4.1, with αi from Corollary 2.3): [14].
 30

B. SFR-BCR-1 (Boros, Crama, Rodríguez-Heck, 2018)

Summary
Any n-variable symmetric function f (b1 , b2 , . . . bn ) ≡ f (b) that is non-zero only when
∑ i = c where n/2 ≤ c ≤ n, can be quadratized with m = ⌈log2 c⌉ + 1 auxiliary variables:
b

m−1 m−1 m−1 m−1

f (b) → α + αb ∑ bi + αb1a ∑ bai + αb2a bam + αbb ∑ bi bj + αbb bba ba ba
1 ∑∑ bi baj + α2 ∑ bi bam + α1
a b b
∑ bai baj + α2a a ∑ bai bam ,
i i ij i j i ij i
(96)

where:

⎛α αbb ⎞ ⎛ (c + 1)2 1 ⎞
⎜ αb α1 ⎟
bba
⎜ −2(c + 1) −2 i
⎟
⎜ ba ⎟ ⎜ m−1 ) ⎟
⎜α α2bba ⎟ = ⎜ (c + i (1 + ⎟.
⎜ 1 ⎟ ⎜ 1)2 2 2
⎟
(97)
⎜ α ba α1ba ba ⎟ ⎜(1 + m−1 ) (2m−1 − 2c − 1) ⎟
⎜ 2 ⎟ 2 2i+j−1

⎝ ⋅ α2ba ba ⎠ ⎝ ⋅ − (1 + 2 ) 2 ⎠
m−1 i

Cost
• m = ⌈log2 c⌉ + 1 auxiliary variables.

• n2 + m2 non-submodular quadratic terms (all possible quadratic terms involving

only non-auxiliary or only auxiliary variables).

• m non-submodular linear terms (all possible linear terms involving auxiliaries).

Pros
• Small number of auxiliary terms
Cons
• Only works for a special class of functions
• Introduces many linear and even more quadratic non-submodular terms.
Example
For n = 4 and c = 2, we have m = 2, and
(b1 b2 + b1 b3 + b1 b4 + b2 b3 + b2 b4 + b3 b4 ) − 3(b1 b2 b3 + b1 b2 b4 + b1 b3 b4 + b2 b3 b4 ) + 6b1 b2 b3 b4 (98)
→ (−3 + b1 + b2 + b3 + b4 − ba1 + 3ba2 )
2
(99)
using the alternate form below.
Alternate Forms
m−1 2
f (b1 , b2 , . . . , bn ) → (−(c + 1) + ∑ bi − ∑ 2 i−1
bai + (1 + 2m−1
) bam ) (100)
i i

Bibliography
• 2018, original paper (Theorem 1): [19].
 31

C. SFR-BCR-2 (Boros, Crama, Rodríguez-Heck, 2018)

Summary
Any n-variable symmetric function f (b1 , b2 , . . . bn ) ≡ f (b) that is non-zero only when
∑ i = c where 0 ≤ c ≤ n/2, can be quadratized with m = ⌈log2 (n − c)⌉ + 1 auxiliary
b
variables:

m−1 m−1 m−1 m−1

f (b) → α + αb ∑ bi + αb1a ∑ bai + αb2a bam + αbb ∑ bi bj + αbb bba ba ba
1 ∑∑ bi baj + α2 ∑ bi bam + α1
a b b
∑ bai baj + α2a a ∑ bai bam ,
i i ij i j i ij i
(101)

where:

⎛α αbb ⎞ ⎛ (c − 1)2 1 ⎞
⎜ αb bba
α1 ⎟ ⎜ ⎟ −2(c − 1) +2i ⎟
⎜ ba ⎜ m−1 ) ⎟
⎜α α2bba ⎟ = ⎜ (1 − i −2 (1 + ⎟.
⎜ 1 ⎟ ⎜ c)2 2
⎟
(102)
⎜αba ba ba ⎟ ⎜ (1 + 2 ) (2 + 2c − 1) ⎟
⎜ 2 α1 ⎟ m−1 m−1 2i+j−1

⎝ ⋅ ba ba ⎠
α2 ⎝ ⋅ − (1 + 2 ) 2 ⎠
m−1 i

Cost
• m = ⌈log2 (n − c)⌉ + 1 auxiliary variables.
• n2 + m2 non-submodular quadratic terms (all possible quadratic terms involving
only non-auxiliary or only auxiliary variables).
• m non-submodular linear terms (all possible linear terms involving auxiliaries).
Pros
• Small number of auxiliary terms
Cons
• Only works for a special class of functions
• Introduces many linear and even more quadratic non-submodular terms.
Example
For n = 4 and c = 2, we have m = 2, and
(b1 b2 + b1 b3 + b1 b4 + b2 b3 + b2 b4 + b3 b4 ) − 3(b1 b2 b3 + b1 b2 b4 + b1 b3 b4 + b2 b3 b4 ) + 6b1 b2 b3 b4(103)
→ (1 − b1 − b2 − b3 − b4 − ba1 + 3ba2 )
2
(104)
using the alternate form below.
Alternate Forms
m−1 2
f (b1 , b2 , . . . , bn ) → ((c − 1) − ∑ bi − ∑ 2 i−1
bai + (1 + 2
m−1
) bam ) (105)
i i

Bibliography
• 2018, original paper (Theorem 1): [19].
 32

D. SFR-BCR-3 (Boros, Crama, Rodríguez-Heck, 2018)

Summary
Any n-variable symmetric function f (b1 , b2 , . . . bn ) ≡ f (b) that is non-zero only when
∑ i = c where n/2 ≤ c ≤ n, can be quadratized with m = ⌈log2 c⌉ auxiliary variables f →:
b

m−1 m−1 m−1 m−1

f (b) → α + αb ∑ bi + αb1a ∑ bai + αb2a bam + αbb ∑ bi bj + αbb bba ba ba
1 ∑∑ bi baj + α2 ∑ bi bam + α1
a b b
∑ bai baj + α2a a ∑ bai bam ,
i i ij i j i ij i
(106)

where:

⎛α 2 (c + 3c + 2)
αbb ⎞ ⎛ 1 2 1
2 ⎞
⎜ αb α1 ⎟
bba
⎜ −c − 2
3
−2i ⎟
⎜ ba ⎟ ⎜ m) ⎟
⎜α α2bba ⎟ = ⎜ (3 + i−1 (1 + ⎟.
⎜ 1 ⎟ ⎜ c)2 2
⎟
(107)
⎜αba α1ba ba ⎟ ⎜(1 + m ) (2m−1 − c − 1) ⎟
⎜ 2 ⎟ 2 2i+j−1

⎝ ⋅ α2ba ba ⎠ ⎝ ⋅ − (1 + 2 ) 2 ⎠
m i

Cost
• m = ⌈log2 c⌉ auxiliary variables.

• n2 + m2 non-submodular quadratic terms (all possible quadratic terms involving

only non-auxiliary or only auxiliary variables).

• m non-submodular linear terms (all possible linear terms involving auxiliaries).

Pros
• Small number of auxiliary terms
Cons
• Only works for a special class of functions
• Introduces many linear and even more quadratic non-submodular terms.
Example
For n = 4 and c = 2, we have m = 1, and
(b1 b2 + b1 b3 + b1 b4 + b2 b3 + b2 b4 + b3 b4 ) − 3(b1 b2 b3 + b1 b2 b4 + b1 b3 b4 + b2 b3 b4 ) + 6b1 b2 b3 b4 (108)
−3 + b1 + b2 + b3 + b4 + 3ba1
→( ) (109)
2
using the alternate form below.
Alternate Forms
−(c + 1) + ∑i bi − ∑m−1 2i bai + (1 + 2m ) bam
f (b1 , b2 , . . . , bn ) → ( i
) (110)
2

Bibliography
• 2018, original paper (Theorem 2): [19].
 33

E. SFR-BCR-4 (Boros, Crama, Rodríguez-Heck, 2018)

Summary
Any n-variable symmetric function f (b1 , b2 , . . . bn ) ≡ f (b) that is non-zero only when
∑ i = c where 0 ≤ c ≤ n/2, can be quadratized with m = ⌈log2 (n − c)⌉ auxiliary variables
b
f →:

m−1 m−1 m−1 m−1

f (b) → α + αb ∑ bi + αb1a ∑ bai + αb2a bam + αbb ∑ bi bj + αbb bba ba ba
1 ∑∑ bi baj + α2 ∑ bi bam + α1
a b b
∑ bai baj + α2a a ∑ bai bam ,
i i ij i j i ij i
(111)

where:

⎛α 2 (c − 3c + 2)
αbb ⎞ ⎛ 1 2 1
2 ⎞
⎜ αb α1 ⎟
bba
⎜ −c + 3
+2 i
⎟
⎜ ba ⎟ ⎜ 2
m) ⎟
⎜α α2bba ⎟ = ⎜ (3 − i−1 − (1 + ⎟.
⎜ 1 ⎟ ⎜ c)2 2
⎟
(112)
⎜αba ba ba ⎟ ⎜(1 + 2 ) (2 + c − 1) ⎟
⎜ 2 α1 ⎟ m m−1 2i+j−1

⎝ ⋅ ba ba ⎠
α2 ⎝ ⋅ − (1 + 2 ) 2 ⎠
m i

Cost
• m = ⌈log2 (n − c)⌉ auxiliary variables.
• n2 + m2 non-submodular quadratic terms (all possible quadratic terms involving
only non-auxiliary or only auxiliary variables).
• m non-submodular linear terms (all possible linear terms involving auxiliaries).
Pros
• Small number of auxiliary terms
Cons
• Only works for a special class of functions
• Introduces many linear and even more quadratic non-submodular terms.
Example
For n = 4 and c = 2, we have m = 1, and
(b1 b2 + b1 b3 + b1 b4 + b2 b3 + b2 b4 + b3 b4 ) − 3(b1 b2 b3 + b1 b2 b4 + b1 b3 b4 + b2 b3 b4 ) + 6b1 b2 b3 b4 (113)
1 − b1 − b2 − b3 − b4 + 3ba1
→( ) (114)
2
using the alternate form below.
Alternate Forms
(c − 1) − ∑i bi − ∑m−1 2i bai + (1 + 2m ) bam
f (b1 , b2 , . . . , bn ) → ( i
) (115)
2

Bibliography
• 2018, original paper (Theorem 2): [19].
 34

F. SFR-BCR-5 (Boros, Crama, Rodríguez-Heck, 2018)

Summary
For an n-variable symmetric function that is a function of the sum √ of all variables
f (b1 , b2 , . . . bn ) = f (∑ bi ), for some large value of λ > max(f ), and ⌈ n + 1⌉:

2 2
m ⎛ m m
f (∑ bi ) → ∑ f ((i − 1) (m + 1) + (j − 1)) bai bac+j + λ (1 − ∑ bai ) + (1 − ∑ bac+i ) +
ij ⎝ i i
2 2
m m m m ⎞
(∑ bi − ((m + 1) ∑(i − 1)yai + ∑(i − 1)bac+i )) + (∑ bi − ((m + 1) ∑(i − 1)yai + ∑(i − 1)bac+i ))
i i i i i i ⎠
(116)

where:

⎛ α αbb ⎞ ⎛ (c + 1)2 1 ⎞
⎜ α α
b bb a,1
⎟ ⎜ −2(c + 1) −2i ⎟
⎜ ba,1 bba,2 ⎟ = ⎜ ⎟. (117)
⎜α α ⎟ ⎜ 2(c + 1) −2 (2m − 2m−1 + 1)⎟
⎝αba,2 αba ba ⎠ ⎝2(c + 1) (2c − 2m−1 + 1) 2i+j−2 ⎠

Cost √
• m = ⌈ n + 1⌉ auxiliary variables.
• n2 + m2 non-submodular quadratic terms (all possible quadratic terms involving
only non-auxiliary or only auxiliary variables).
• m non-submodular linear terms (all possible linear terms involving auxiliaries).
Pros
• Small number of auxiliary terms
Cons
• Only works for a special class of functions
• Introduces many linear and even more quadratic non-submodular terms.
Example
By bit-flipping b2 and b4 , i.e. substituting b2 = 1 − b̄2 and b4 = 1 − b̄4 , we see that:
H = 3b1 b2 + b2 b3 + 2b1 b4 − 4b2 b4 (118)
= −3b1 b̄2 − b̄2 b3 − 2b1 b̄4 − b̄2 b̄4 + 5b1 + b3 + 4b̄2 + 4b̄4 − 4. (119)
The first expression is highly non-submodular while the second is entirely submodular.
Alternate Forms
1
((c − 1) − ∑i bi − (2(n − c) − 2m−1 + 1) bam − ∑m−1 2i−1 bai )
f (b1 , b2 , . . . , bn ) → (
2 i
) (120)
2

Bibliography
• 2018, original paper (Theorem 9): [19] (contains a typo which was corrected in
Theorem 6 of [20].
 35

G. SFR-BCR-6 (Boros, Crama, Rodríguez-Heck, 2018)

Summary
For an n-variable symmetric function that is a function of a weighted sum of
all variables f (b1 , b2 , . . . bn ) = f (∑ wi bi ), for some large value of λ > max(f ), and
max (f (∑ wi bi )) < (m + 1)2 :

⎛
m m m 2
f (∑ wi bi ) → ∑ αij bai bam+i + λ 1 + (∑ wi bi − (m − 1) ∑ bai + ∑ bac+i )
ij ⎝ i i i
(121)
m−1 m−1
+ ∑ (1 − bai ) bai+1 + ∑ (1 − bai+m ) bai+m+1 )
i i

where:

α β
∑ ∑ αij = f (α(m + 1) + β) (122)
i j

Cost √
• 2m auxiliary variables, where m > max (f (∑ wi bi )) − 1.
• n2 + m2 non-submodular quadratic terms (all possible quadratic terms involving
only non-auxiliary or only auxiliary variables).
• m non-submodular linear terms (all possible linear terms involving auxiliaries).
Pros
• Small number of auxiliary terms
Cons
• Only works for a special class of functions

• Introduces many linear and even more quadratic non-submodular terms.

Example
By bit-flipping b2 and b4 , i.e. substituting b2 = 1 − b̄2 and b4 = 1 − b̄4 , we see that:

H = 3b1 b2 + b2 b3 + 2b1 b4 − 4b2 b4 (123)

= −3b1 b̄2 − b̄2 b3 − 2b1 b̄4 − b̄2 b̄4 + 5b1 + b3 + 4b̄2 + 4b̄4 − 4. (124)

The first expression is highly non-submodular while the second is entirely submodular.
Alternate Forms
1
((c − 1) − ∑i bi − (2(n − c) − 2m−1 + 1) bam − ∑m−1 2i−1 bai )
f (b1 , b2 , . . . , bn ) → ( 2 i
) (125)
2

Bibliography
• 2018, original paper (Theorem 10): [19].
 36

H. SFR-ABCG-2 (Anthony, Boros, Crama, Gruber, 2014)

Summary
For any n-variable, k-local function that is non-zero only if ∑ bi = 2m − 1, we call it
the "partity function" and it can be quadratized as follows:

n−1
f (b1 , b2 , . . . , bn ) → ∑ bi + 2 ∑ bi bj + 4 ∑ bai (2i − 1 − ∑ bj ) . (126)
i ij 2i−1 j

Cost
• m = 2⌊n/2⌋ auxiliary variables.
• ⌊1.5n⌋ non-submodular linear terms.
• n2 non-submodular quadratic terms.
Pros
• Smaller number of auxiliary variables than the most naive methods.
Cons
• Only works for a special class of functions
• Introduces many linear and even more quadratic non-submodular terms (every-
thing is non-submodular except for 0.5n2 quadratic terms involving the auxiliaries
with the non-auxiliaries).
• non-submodular terms can be rather large compared to the submodular terms
(about 4n times as big).
Example
By bit-flipping b2 and b4 , i.e. substituting b2 = 1 − b̄2 and b4 = 1 − b̄4 , we see that:
H = 3b1 b2 + b2 b3 + 2b1 b4 − 4b2 b4 (127)
= −3b1 b̄2 − b̄2 b3 − 2b1 b̄4 − b̄2 b̄4 + 5b1 + b3 + 4b̄2 + 4b̄4 − 4. (128)
The first expression is highly non-submodular while the second is entirely submodular.
Bibliography
• 2014, original paper (Theorem 4.6): [13, 14].
 37

I. SFR-ABCG-3 (Anthony, Boros, Crama, Gruber, 2014)

Summary
The complement of the parity function can be quadratized as follow:

n−1 n
f (b1 , b2 , . . . , bn ) → 1 + 2 ∑ bi bj − ∑ bi + 4 ∑ bai (i − ∑ bj ) (129)
ij i 2i j

Cost
2 ⌋ auxiliary variables.
• m = 2⌊ n−1
• ⌊0.5n⌋ non-submodular linear terms.
• n2 non-submodular quadratic terms.
Pros
• Smaller number of auxiliary variables than the most naive methods.
• Fewer non-submodular linear terms than in the analogous quadratization for its
complement (the parity function).
Cons
• Only works for a special class of functions
• Introduces many linear and even more quadratic non-submodular terms (every-
thing is non-submodular except for 0.5n2 quadratic terms involving the auxiliaries
with the non-auxiliaries, and all n linear terms involving only the non-auxiliaries).

• non-submodular terms can be rather large compared to the submodular terms

(about 4n times as big).
Example
By bit-flipping b2 and b4 , i.e. substituting b2 = 1 − b̄2 and b4 = 1 − b̄4 , we see that:
H = 3b1 b2 + b2 b3 + 2b1 b4 − 4b2 b4 (130)
= −3b1 b̄2 − b̄2 b3 − 2b1 b̄4 − b̄2 b̄4 + 5b1 + b3 + 4b̄2 + 4b̄4 − 4. (131)
The first expression is highly non-submodular while the second is entirely submodular.
Bibliography
• 2014, original paper (Theorem 4.6): [13, 14].
 38

J. SFR-BCR-7 (Boros, Crama, Rodríguez-Heck, 2018)

Summary √
For a symmetric function such that f (∣c∣) = 0 for c > n, then with with 2⌈ n + 1⌉
auxiliary variables, we have:

n−1 n
f (b1 , b2 , . . . , bn ) → 1 + 2 ∑ bi bj − ∑ bi + 4 ∑ bai (i − ∑ bj ) (132)
ij i 2i j

Cost √
• m = ⌈ n + 1⌉ auxiliary variables.
• ⌊0.5n⌋ non-submodular linear terms.

• n2 non-submodular quadratic terms.

Pros
• Smaller number of auxiliary variables than the most naive methods.
• Fewer non-submodular linear terms than in the analogous quadratization for its
complement (the parity function).
Cons
• Only works for a special class of functions

• Introduces many linear and even more quadratic non-submodular terms (every-
thing is non-submodular except for 0.5n2 quadratic terms involving the auxiliaries
with the non-auxiliaries, and all n linear terms involving only the non-auxiliaries).
• non-submodular terms can be rather large compared to the submodular terms
(about 4n times as big).
Example
By bit-flipping b2 and b4 , i.e. substituting b2 = 1 − b̄2 and b4 = 1 − b̄4 , we see that:
H = 3b1 b2 + b2 b3 + 2b1 b4 − 4b2 b4 (133)
= −3b1 b̄2 − b̄2 b3 − 2b1 b̄4 − b̄2 b̄4 + 5b1 + b3 + 4b̄2 + 4b̄4 − 4. (134)
The first expression is highly non-submodular while the second is entirely submodular.
Bibliography
• 2018, original paper (Theorem 6): [20].
 39

K. SFR-BCR-8 (Boros, Crama, Rodríguez-Heck, 2018)

Summary
For a symmetric function such that f (∣c∣) = 0 for c > n, then with with max (⌈log(c)⌉, ⌈log(n − c)⌉)
auxiliary variables, we have:

n−1 n
f (b1 , b2 , . . . , bn ) → 1 + 2 ∑ bi bj − ∑ bi + 4 ∑ bai (i − ∑ bj ) (135)
ij i 2i j

Cost
• max (⌈log(c)⌉, ⌈log(n − c)⌉) auxiliary variables.
• ⌊0.5n⌋ non-submodular linear terms.
• n2 non-submodular quadratic terms.
Pros
• Smaller number of auxiliary variables than the most naive methods.
• Fewer non-submodular linear terms than in the analogous quadratization for its
complement (the parity function).
Cons
• Only works for a special class of functions

• Introduces many linear and even more quadratic non-submodular terms (every-
thing is non-submodular except for 0.5n2 quadratic terms involving the auxiliaries
with the non-auxiliaries, and all n linear terms involving only the non-auxiliaries).

• non-submodular terms can be rather large compared to the submodular terms

(about 4n times as big).
Example
By bit-flipping b2 and b4 , i.e. substituting b2 = 1 − b̄2 and b4 = 1 − b̄4 , we see that:
H = 3b1 b2 + b2 b3 + 2b1 b4 − 4b2 b4 (136)
= −3b1 b̄2 − b̄2 b3 − 2b1 b̄4 − b̄2 b̄4 + 5b1 + b3 + 4b̄2 + 4b̄4 − 4. (137)
The first expression is highly non-submodular while the second is entirely submodular.
Bibliography
• 2018, original paper (Theorem 7): [20].
 40

L. SFR-BCR-9 (Boros, Crama, Rodríguez-Heck, 2018)

Summary
For a symmetric function such that f (∣c∣) = 0 for c > n, then with with max (⌈log(c)⌉, ⌈log(n − c)⌉)
auxiliary variables, we have:

n−1 n
f (b1 , b2 , . . . , bn ) → 1 + 2 ∑ bi bj − ∑ bi + 4 ∑ bai (i − ∑ bj ) (138)
ij i 2i j

Cost
• max (⌈log(c)⌉, ⌈log(n − c)⌉) auxiliary variables.
• ⌊0.5n⌋ non-submodular linear terms.
• n2 non-submodular quadratic terms.
Pros
• Smaller number of auxiliary variables than the most naive methods.
• Fewer non-submodular linear terms than in the analogous quadratization for its
complement (the parity function).
Cons
• Only works for a special class of functions

• Introduces many linear and even more quadratic non-submodular terms (every-
thing is non-submodular except for 0.5n2 quadratic terms involving the auxiliaries
with the non-auxiliaries, and all n linear terms involving only the non-auxiliaries).

• non-submodular terms can be rather large compared to the submodular terms

(about 4n times as big).
Example
By bit-flipping b2 and b4 , i.e. substituting b2 = 1 − b̄2 and b4 = 1 − b̄4 , we see that:
H = 3b1 b2 + b2 b3 + 2b1 b4 − 4b2 b4 (139)
= −3b1 b̄2 − b̄2 b3 − 2b1 b̄4 − b̄2 b̄4 + 5b1 + b3 + 4b̄2 + 4b̄4 − 4. (140)
The first expression is highly non-submodular while the second is entirely submodular.
Bibliography
• 2018, original paper (Theorem 8): [20].
 41

M. PFR-BCR-1 (Boros, Crama, Rodríguez-Heck, 2018)

Summary
The parity function for even n can be quadratized with:

n−1 n
f (b1 , b2 , . . . , bn ) → 1 + 2 ∑ bi bj − ∑ bi + 4 ∑ bai (i − ∑ bj ) (141)
ij i 2i j

Cost
• ⌈log(n)⌉ − 1 auxiliary variables.
• ⌊0.5n⌋ non-submodular linear terms.
• n2 non-submodular quadratic terms.
Pros
• Smaller number of auxiliary variables than the most naive methods.
• Fewer non-submodular linear terms than in the analogous quadratization for its
complement (the parity function).
Cons
• Only works for a special class of functions
• Introduces many linear and even more quadratic non-submodular terms (every-
thing is non-submodular except for 0.5n2 quadratic terms involving the auxiliaries
with the non-auxiliaries, and all n linear terms involving only the non-auxiliaries).

• non-submodular terms can be rather large compared to the submodular terms

(about 4n times as big).
Example
By bit-flipping b2 and b4 , i.e. substituting b2 = 1 − b̄2 and b4 = 1 − b̄4 , we see that:
H = 3b1 b2 + b2 b3 + 2b1 b4 − 4b2 b4 (142)
= −3b1 b̄2 − b̄2 b3 − 2b1 b̄4 − b̄2 b̄4 + 5b1 + b3 + 4b̄2 + 4b̄4 − 4. (143)
The first expression is highly non-submodular while the second is entirely submodular.
Bibliography
• 2018, original paper (Theorem 11): [20].
 42

N. PFR-BCR-2 (Boros, Crama, Rodríguez-Heck, 2018)

Summary
The parity function for odd n can be quadratized with:

n−1 n
f (b1 , b2 , . . . , bn ) → 1 + 2 ∑ bi bj − ∑ bi + 4 ∑ bai (i − ∑ bj ) (144)
ij i 2i j

Cost
• ⌈log(n)⌉ − 1 auxiliary variables.
• ⌊0.5n⌋ non-submodular linear terms.
• n2 non-submodular quadratic terms.
Pros
• Smaller number of auxiliary variables than the most naive methods.
• Fewer non-submodular linear terms than in the analogous quadratization for its
complement (the parity function).
Cons
• Only works for a special class of functions
• Introduces many linear and even more quadratic non-submodular terms (every-
thing is non-submodular except for 0.5n2 quadratic terms involving the auxiliaries
with the non-auxiliaries, and all n linear terms involving only the non-auxiliaries).

• non-submodular terms can be rather large compared to the submodular terms

(about 4n times as big).
Example
By bit-flipping b2 and b4 , i.e. substituting b2 = 1 − b̄2 and b4 = 1 − b̄4 , we see that:
H = 3b1 b2 + b2 b3 + 2b1 b4 − 4b2 b4 (145)
= −3b1 b̄2 − b̄2 b3 − 2b1 b̄4 − b̄2 b̄4 + 5b1 + b3 + 4b̄2 + 4b̄4 − 4. (146)
The first expression is highly non-submodular while the second is entirely submodular.
Bibliography
• 2018, original paper (Theorem 11): [20].
 43

O. Lower bounds for SFRs (Anthony, Boros, Crama, Gruber, 2014)

• There exist symmetric functaons on n variables for which no quadratization can

√
be done without at least Ω ( n) auxiliary variables (Theorem 5.3 from [13, 14].

• There exist symmetric functions on n variables for whcih no quadratization

linear in the auxiliaries can be done without at least Ω ( log2n(n) ) auxiliary variables
(Theorem 5.5 from [13, 14]).

• The parity function on n variables cannot be quadratized without

√ quadratic terms
√
involving the auxiliary variables, unless there is at least n/4 − 1 + 1 = Ω ( n)
auxiliary variables (Theorem 5.6 from [13, 14]).
• Theorem 5 of [20] gives an even tighter bound of ⌈log(n)⌉ − 1 for the minimum
number of auxiliary variables for the parity function.
• Corollary 5 of [20] gives m ≥ log (1/2 − µ) + log(n) − 1.

P. Lower bounds for positive monomials (Boros, Crama, Rodríguez-Heck, 2018)

• A positive monomial with n variables cannot be quadratized with fewer than

⌈log(n)⌉ − 1 auxiliary variables, unless there is some extra deduction we can make
about the optimization problem, as in for example deduc-reduc (Corollary 1 from
[20]).
• ALCN (at least c out of n) and ECN (exact c out of n) functions also cannot be
quadratized with fewer than ⌈log(n)⌉ − 1 auxiliary variables (Corollaries 2 and 3
from [20]).
• ECN (exact c out of n) functions also cannot be quadratized with fewer than
max (⌈log(c)⌉, ⌈log(n − c)⌉) − 1 auxiliary variables (Corollaries 2 and 3 from [20]).
 44

Q. Lower bounds for ZUCs (Boros, Crama, Rodríguez-Heck, 2018)

• There exist ZUC (zero until c) functions such that every quadratization must in-
volve at least Ω (2n/2 ) auxiliary variables, no matter what the value of c (Theorem
2 from [20]). This is true for almost all ZUC functions because the set of ZUCs
requiring fewer auxiliary variables has Lebesgue measure zero.
• For any c ≥ 0, the number of auxiliary variables is m ≥ ⌈log(c)⌉ − 1 (Theorem 3 of
[20]).

R. Lower bounds for d-sublinear functions (Boros, Crama, Rodríguez-Heck, 2018)

β(q1 )
• The number of auxiliary variables m is such that 2m+1 ≥ 2 − d + 1 (Theorem 12
from [20]).
• For any c ≥ 0, the number of auxiliary variables is m ≥ ⌈log(c)⌉ − 1 (Theorem 3 of
[20]).
 45

V. METHODS THAT QUADRATIZE MULTIPLE TERMS WITH THE SAME

AUXILIARIES (CASE 2: ARBITRARY FUNCTIONS)

A. Reduction by Substitution (Rosenberg 1975)

Summary
Pick a variable pair (bi , bj ) and substitute bi bj with a new auxiliary variable baij .
Enforce equality in the ground states by adding some scalar multiple of the penalty
P = bi bj − 2bi baij − 2bj baij + 3baij or similar. Since P > 0 if and only if baij ≠ bi bj , the
minimum of the new (k − 1)-local function will satisfy bai j = bi bj ), which means that at
the minimum, we have precisely the original function. Repeat (k − 2) times for each
k-local term and the resulting function will be 2-local. For an arbitrary cubic term we
have:

bi bj bk → ba bk + bi bj − 2bi ba − 2bj ba + 3ba . (147)

Cost
• 1 auxiliary variable per reduction.
• At most k t auxiliary variables for a k-local objective function of t terms, but
usually fewer.
Pros
• Variable can be used across the entire objective function, reducing many terms at
once.
• Very easy to implement.

• Reproduces not only the ground state, but the full spectrum.
Cons
• Inefficient for single terms as it introduces many auxiliary variables compared to
Ishikawa reduction, for example.
• Introduces quadratic terms with large positive coefficients, making them highly
non-submodular.
• Determining optimal substitutions can be expensive.
Example
We pick the pair (b1 , b2 ) and combine.

b1 b2 b3 + b1 b2 b4 ↦ b3 ba + b4 ba + b1 b2 − 2b1 ba − 2b1 ba + 3ba (148)

Bibliography
• Original paper: [30]
• Re-discovered in the context of diagonal quantum Hamiltonians: [31].
• Used in: [32, 33].
 46

B. FGBZ Reduction for Negative Terms (Fix-Gruber-Boros-Zabih, 2011)

Summary
We consider a set C of variables which can occur in multiple terms throughout the
objective function. Each application ‘rips out’ this common component from each term
[18, 34].

∑ αH ∏ bj → ∑ αH (1 − ∏ bj − ∏ bj ) ba (149)
H j∈H H j∈C j∈H∖C

Cost
• One auxiliary variable per application.

• In combination with II A, it can reduce t positive terms of degree k in n variables

using n + t(k − 1) auxiliary variables in the worst case.
Pros
• Can reduce the connectivity of an objective function, as it breaks interactions
between variables.
Cons
• Cannot reduce the degree of the original function if ∣C∣ ≤ 1, and cannot quadratize
anything for the other values of ∣C∣ (but it can reduce their degree).
Example
First let C = b1 and use the positive weight version:
b1 b2 b3 + b1 b2 b4 ↦ 2ba1 b1 + (1 − ba1 )b2 b3 + (1 − ba1 )b2 b4 (150)
= 2ba1 b1 + b2 b3 + b2 b4 − ba1 b2 b3 − ba1 b2 b4 (151)

now we can use II A:

−ba1 b2 b3 − ba1 b2 b4 ↦ 2ba2 − ba1 ba2 − ba2 b2 − ba2 b3 + 2ba2 − ba1 ba2 − ba2 b2 − ba2 b4 (152)
= 4ba2 − 2ba1 ba2 − 2ba2 b2 − ba2 b3 − ba2 b4 . (153)

Bibliography
• Original paper and application to image denoising: [34].
 47

C. FGBZ Reduction for Positive Terms (Fix-Gruber-Boros-Zabih, 2011)

Summary
We consider a set C of variables which can occur in multiple terms throughout the
objective function. Each application ‘rips out’ this common component from each term
[18, 34]:

∑ αH ∏ bj → ∑ αH ba ∏ bj + ∑ αH (1 − ba ) ∏ bj . (154)
H j∈H H j∈C H j∈H∖C

Cost
• One auxiliary variable per application.

• In combination with II A, it can reduce t positive terms of degree k in n variables

using n + t(k − 1) auxiliary variables in the worst case.
Pros
• It is a ‘perfect’ transformation, meaning that after minimizing over ba , the original
degree-k function is recovered.
• Can reduce the connectivity of an objective function, as it breaks interactions
between variables.
Cons
• If ∣C∣ = 1 the first sum will result in a quadratic but the second sum will have
degree k. If ∣C∣ = k − 1 the second sum will be quadratic but the first term will
have degree k. For any other value of ∣C∣, both sums will be super-quadratic, but
the part of the second sum involving ba will be negative and therefore can be
quadratized easily.
Example
With C = b1 we can get:
b1 b2 b3 + b1 b2 b4 → 2ba1 b1 + (1 − ba1 )b2 b3 + (1 − ba1 )b2 b4 (155)
= 2ba1 b1 + b2 b3 + b2 b4 − ba1 b2 b3 − ba1 b2 b4 (156)

now we can use II A to quadratize the two negative cubic terms:

−ba1 b2 b3 − ba1 b2 b4 ↦ 2ba2 − ba1 ba2 − ba2 b2 − ba2 b3 + 2ba2 − ba1 ba2 − ba2 b2 − ba2 b4 (157)
= 4ba2 − 2ba1 ba2 − 2ba2 b2 − ba2 b3 − ba2 b4 . (158)

Bibliography
• Original paper and application to image denoising: [34].
 48

D. Pairwise Covers (Anthony-Boros-Crama-Gruber, 2017)

Summary
Here we consider a set C of variables which occur in multiple monomials throughout
the objective function. Each application ’rips out’ this common component from each
term [34][18].
Let H be a set of monomials, where C ⊆ H for each H ∈ H and each monomial H
has a weight αH . The algorithm comes in 2 parts: when all αH > 0 and when all αH < 0.
Combining the 2 gives the final method:

1. αH > 0

∑ αH ∏ bj = min ( ∑ αH ) ba ∏ bj + ∑ αH (1 − ba ) ∏ bj (159)
H∈H j∈H ba H∈H j∈C H∈H j∈H∖C

2. αH < 0
∑ αH ∏ bj = min ∑ αH (1 − ∏ bj − ∏ bj ) ba (160)
H∈H j∈H ba H∈H j∈C j∈H∖C

Cost
• One auxiliary variable per application.

• In combination with II A, it can be used to make an algorithm which can reduce t

positive monomials of degree d in n variables using n + t(d − 1) auxiliary variables
in the worst case.
Pros
• Can reduce the connectivity of an objective function, as it breaks interactions
between variables.
Cons
• αH > 0 method converts positive terms into negative ones of same order rather
than reducing them, though these can then be reduced more easily.

• αH < 0 method only works for ∣C∣ > 1, and cannot quadratize cubic terms.
Example
First let C = b1 and use the positive weight version:

b1 b2 b3 + b1 b2 b4 ↦ 2ba1 b1 + (1 − ba1 )b2 b3 + (1 − ba1 )b2 b4 (161)

= 2ba1 b1 + b2 b3 + b2 b4 − ba1 b2 b3 − ba1 b2 b4 (162)

now we can use II A:

−ba1 b2 b3 − ba1 b2 b4 ↦ 2ba2 − ba1 ba2 − ba2 b2 − ba2 b3 + 2ba2 − ba1 ba2 − ba2 b2 − ba2 b4 (163)
= 4ba2 − 2ba1 ba2 − 2ba2 b2 − ba2 b3 − ba2 b4 . (164)

Bibliography
• Theorem 4 of: [9].
 49

E. Flag Based SAT Mapping

Summary
This method uses gadgets to produce separate 3-SAT clauses which allow variables
which ‘flag’ the state of pairs of other variables.
Cost
• 24 auxiliary variables to quadratize z1 z1 z3 .
Pros
• Very general and therefore conducive to proofs.
Cons
• Extremely inefficient in terms of number of auxiliary variables.
Example
To create a system which maps b1 b2 b3 , we use the following gadget (note that this is
given in terms of z in the orginal work and translated to b here):
3 3 3 3
23
H1 (b1 , b2 , b3 ) = 2 ∑ bi bai + 2 ∑ bai baj − 4 ∑ bai − 2 ∑ bi + . (165)
i=1 i<j i=1 i=1 2

Implementing αH1 , creates a situation where b3 is a ‘flag’ for b1 and b2 in other words
b3 is constrained to be 1 in the low energy manifold if b1 = 0 and b2 = 0. It follows from
the universality of 3 − SAT that these ‘flag’ clauses can be combined to map any spin
Hamiltonian. To do this, we also need anti-ferromagnetic couplings to express the
‘negated’ variable, to do this, we define,

H2 (b1 , b¬1 ) = 2 b1 b¬1 − bi − b¬1 + 1. (166)

As an explicit example, consider reproducing the spectrum of z1 z2 z3 = (2 b1 − 1)(2 b2 −

1)(2 b3 − 1). In this case we need to assign a higher energy to the (1, 1, 1), (0, 0, 1),
(0, 1, 0), and (1, 0, 0) states. A flag (ba4,1 ) which is forced into a higher energy state
if these conditions are satisfied can be constructed from two instances of H1 and an
auxilliary qubit, combining these leads to

3
⎛ 3 2
z1 z2 z3 →α ∑ H2 (bi , b¬i ) + ∑ H2 (bai , b¬ai ) + H2 (ba4,1 , ba4,2 ) + H1 (ba1,1 , ba1,2 , b¬a1 ) + H1 (ba1 , ba1,3 , ba4,1 ) + (167)
⎝i=1 i=1

H1 (b1 , b2 , b¬a2 ) + H1 (ba2 , b3 , ba4,2 + H1 (b1 , b2 , b¬a3 ) + H1 (ba3 , ba1,3 , ba4,1 ) + H1 (ba1,1 , ba1,2 , b¬a4 ) + (168)
H1 (ba4 , b3 , ba4,2 ) + H1 (b1 , ba1,2 , b¬a5 ) + H1 (ba5 , b3 , ba4,1 ) + H1 (ba1,1 , b2 , b¬a6 ) + H1 (ba6 , ba1,3 , ba4,2 ) + (169)
H1 (b1 , b2 , b¬a7 ) + H1 (ba7 , ba1,3 , ba4,1 ) + H1 (ba1,1 , ba1,2 , b¬a8 ) + H1 (ba8 , b3 , ba4,2 )) − 2 ba4,1 + 1. (170)

Each of the next four lines assigns a value to the flag variable ba4,1 for a state and
ba4,2 , for instance the leftmost two terms of the second line enforce that ba4,1 = 0 if
(b1 , b2 , b3 ) = (0, 0, 0), while the right two terms enforce that ba4,1 = 1 if (b1 , b2 , b3 ) = (1, 1, 1).
Because there are 23 = 8 possible bitstrings for (b1 , b2 , b3 ), and each term to enforce a flag
state requires two instances of H1 (and two auxilliary variables), a total of 16 instances
are required as well as 16 auxilliary variables.
Bibliography
• Paper showing the universality of the Ising spin models: [35].
 50

F. Lower bounds for arbitrary functions (Anthony, Boros, Crama, Gruber, 2015)

• There exist functions on n variables for which no quadratization can be done

n/2 √
without at least 2 8 = Ω ( n) auxiliary variables (Theorem 5.3 from [9].

• There exist symmetric functions on n variables for which no quadratization linear

in the auxiliaries can be done without at least Ω ( 2n ) auxiliary variables (Theorem
n

5.5 from [9]).

 51

VI. STRATEGIES FOR COMBINING METHODS

A. SCM-BCR (Boros, Crama, and Rodríguez-Heck, 2018)

Summary
Split a k-local monomial with odd k into a (k − 1)-local term (with even degree) and
a new odd k-local term which has negative coefficient:

k−1 k−1
b1 b2 ⋯bk → ∏ bi − ∏ bi (1 − bk ) (171)
i=1 i=1

We can use any of the PTR methods for even k on the first term, and we can use any of
the NTR methods on the second term. Can be generalized to split into different-degree
factors when seeking an "optimum" quadratization. Can be generalized into more
splits.
Cost
• Depends on the methods used for the PTR and NTR procedures.
Pros
• Very flexible.
Cons
• First turns one term into two terms, so might not be preferred when we wish to
minimize the number of terms.
Bibliography
• Original paper: [19].
 52

B. Decomposition into symmetric and anti-symmetric parts

Summary
Split a any function f into a symmetric part and anti-symmetric part:

f (b1 , b2 , . . . , bn ) = fsymmetric + fanti-symmetric , (172)

1
fsymmetric ≡ (f (b1 , b2 , . . . , bn ) + f (1 − b1 , 1 − b2 , . . . , 1 − bn )) (173)
2
1
fanti-symmetric ≡ (f (b1 , b2 , . . . , bn ) − f (1 − b1 , 1 − b2 , . . . , 1 − bn )) (174)
2
We can now use any of the methods described only for symmetric functions, on the
symmetric part, and use the (perhaps less powerful) general methods on the anti-
symmetric part.
Cost
• Depends on the methods used.
Pros
• Allows non-symmetric functions to benefit from techniques designed only for
symmetric functions.
Cons
• May result in more terms than simply quadratizing the non-symmetric function
directly.
Bibliography
• Discussed in: [36].
 53

Part II
Hamiltonians quadratic in z and linear in x (Transverse Field Ising Hamiltonians)

The Ising Hamiltonian with a transverse field in the x direction is possible to implement
in hardware:

(z) (x) (zz)

H = ∑ (αi zi + αi xi ) + ∑ (αij zi zj ) . (175)
i ij

C. ZZZ-TI-CBBK: Transvese Ising from ZZZ, by Cao, Babbush, Biamonte, and Kais (2015)

There is only one reduction in the literature for reducing a Hamiltonian term to the
transverse Ising Hamiltonian, and it works on 3-local zzz terms, by introducing an
auxiliary qubit with label a:

αzi zj zk → αI + αiz zi + αjz zj + αkz zk + αaz za + αax xa + αia

zz
zi za + αja
zz
zj za + αka
zz
zk za (176)
2/5
αI = 1
2 (∆+ ( α6 ) ∆3/5 )
3/5 1/5
αiz = − 21 (( 7α ( α ) ∆2/5 ) − ( α∆
6 + 6 6 ) )
4

(z)
αjz = αi
(z)
αkz = αi
2/5
αaz = 21 (∆− ( α6 ) ∆3/5 )
1/5
αax = ( α∆
6 )
4

3/5 1/5
= − 21 (( 7α ( α ) ∆2/5 ) + ( α∆
6 + 6 6 ) )
4
zz
αia
zz (zz)
αja = αia
zz (zz)
αka = αja

Including all coefficients and factorizing, we get:

1/5
1 − za α∆4 /5
1
α∆4
αzi zj zk → (∆ + (zi + zj + zk )) ( )+ xa (177)
6 2 6
α /5 3 α /5 2 1 + za
2 3
7α
+ (( ) ∆ /5 − ( + ( ) ∆ /5 ) (zi + zj + zk )) ( ) (178)
6 6 6 2

The low-lying spectrum (eigenvalues and eigenvectors) of the right side of Eq. (176)
will match those of the left side to within a spectral error of  as long as ∆ = O (−5 ).
Cost
• 1 auxiliary qubit
• 8 auxiliary terms not proportional to 11.
 54

Part III
General Quantum Hamiltonians

VII. NON-PERTURBATIVE GADGETS

A. NP-OY (Ocko & Yoshida, 2011)

Summary
For the 8-body Hamiltonian:

H8-body = −J ∑ (xij3 xij+1,2 xij4 xij+1,4 xi+1j1 xi+1j+1,1 xi+1j3 xij+1,2 +

ij (179)
zij1 zij2 zij3 zij4 + zi−1j4 zij1 + zij2 zij−1,3 + zij4 zi+1j1 + zij3 zij+1,2 ) ,

we define auxiliary qubits labeled by aijk , two auxiliaries for each pair ij: labeled aij1
and aij2 . Then the 8-body Hamiltonian has the same low-lying eigenspace as the 4-body
Hamiltonian:

H4-body = − ∑ α (zij1 zij2 zij3 zij4 + zi,j,−1,4 zij1 + zij2 zi,j−1,3 + zij4 zi+1,j,1 + zij3 zi,j+1,2
ij

(1 − zaij1 + zaij2 + zaij1 zaij2 ) (zai,j+1,1 + zai,j+1,2 + zai,j+1,1 zai,j+1,2 − 1) +

(1 + zaij1 − zaij2 + zaij1 zaij2 ) (1 − zai+1,j1 − zai+1,j2 − zai+1,j1 zai+1,j2 )) +
U (180)
(zaij1 + zaij2 + zaij1 zaij2 − 1) +
2
t
((xaij2 + zaij1 xaij2 ) xij3 xij4 + (xaij1 xaij2 + yaij1 yaij2 ) xi,j+1,2 xi,j+1,4 +
2
(xaij2 − zaij1 xaij2 ) xi+1,j+1,1 xi+1,j+1,2 + (xaij1 xaij2 − yaij1 yaij2 ) xi+1,j,1 xi+1,j,3 )) .

Now by defining the following ququits (spin-3/2 particles, or 4-level systems):

ijki′ j ′ k′ = zijk zi′ j ′ k′

szz (181)
aij 1 = (1 − za1ij + za2ij + za1ij za2ij )
szz (182)
aij 2 = (za1ij + za2ij + za1ij za2ij − 1)
szz (183)
aij 3 = (1 + za1ij − za2ij + za1ij za2ij )
szz (184)
aij 1 = (xa2ij + za1ij xa2ij )
sxz (185)
aij 2 = (xa2ij − za1ij xa2ij )
sxz (186)
ijki′ j ′ k′ = xijk xi′ j ′ k′
sxx (187)
sxy
aij 1 = (xa1ij xa2ij + ya1ij ya2ij ) (188)
xy
saij 2 = (xa1ij xa2ij − ya1ij ya2ij ) (189)
 55

NP-OY (Ocko & Yoshida, 2011) [Continued]

We can write the 4-body Hamiltonian on qubits as a 2-body Hamiltonian on ququits:

H2-body = − ∑ (α (szz zz zz zz zz zz zz zz zz zz
ij1ij2 sij3ij4 + sij−1,4ij1 + sij2ij−1,3 + sij4i+1j1 + sij3ij+1,2 + saij 1 saij+1 2 − saij 3 sai+1j 3 )
ij
(190)
U t xz xx
+ szz
a ,1 + (saij 1 sij3ij4 + sxy xx xz xx xy xx
aij 1 sij+1,2ij+1,4 + saij 2 si+1,j+1,1i+1,j+1,2 + saij 2 si+1j1,i+1,j3 )) .
2 ij 2

The low-lying eigenspace of H2−body is exactly the same as for H4−local .

Cost
• 2 auxiliary ququits for each pair ij.

• 6 more total terms (6 terms in the 8-body version becomes 12 terms:

11 of them 2-body and 1 of them 1-body).
Pros
• Non-perturbative. No prohibitive control precision requirement.
• Only two auxiliaries required for each pair ij.

• 8-body to 2-body transformation can be accomplished in 1 step, rather than a 1B1

gadget which would take 6 steps or an SD + (3 → 2) gadget combination which
would take 4 steps.
Cons
• Increase in dimention from working with only 2-level systems (spin-1/2 particles
or 2 × 2 matrices) to working with 4-level systems (spin-3/2 particles).
• Until now, only derived for a very specific Hamiltonian form.
• This appraoch may become more demanding for Hamiltonians that are more
than 8-local.
Bibliography
• Original paper: [37].
 56

B. NP-SJ (Subasi & Jarzynski, 2016)

Summary
Determine the k-local term, Hk−local , whose degree we wish to reduce, and factor it
into two commuting factors: Hk′ −local H(k−k′ )−local , where k ′ can be as low as 0. Separate
all terms that are at most (k −1)-local into ones that commmute with one of these factors
(it does not matter which one, but without loss of generality we assume it to be the
(k − k ′ )-local one) and ones that anti-commute with it:

commuting
H<k−local + H<k−local
anti−commuting
+ αHk′ −local H(k−k′ )−local (191)
Introduce one auxiliary qubit labeled by a and the Hamiltonian:

αxa Hk′ −local + H<k−local

commuting
+ za H<k−local
anti−commuting
(192)
anti−commuting
no longer contains Hk−local but H<k−local is now one degree higher.
Cost
• 1 auxiliary qubit to reduce k-local term to (k ′ + 1)-local where k ′ can even be
0-local, meaning the k-local term is reduced to a 1-local one.
anti−commuting
• Raises the k-locality of H<k−local by 1 during each application. It can become
(> k)-local!
Pros
• Non-perturbative
• Can linearize a term of arbitrary degree in one step.
• Requires very few auxiliary qubits.
Cons
• Can introduce many new non-local terms as an expense for reducing only one
k-local term.
• If the portion of the Hamiltonian that does not commute with the (k − k ′ )-local
term has termms of degree k − 1 (which can happen if k ′ = 0) they will all become
k-local, so there is no guarantee that this method reduces k-locality.
• If any terms were more than 1-local, this method will not fully quadratize the
Hamiltonian (it must be combined with other methods).
• It only works when the Hamiltonian’s terms of degree at most k − 1 all either
commute or anti-commmute with the k-local term to be eliminated.
Example

4z5 − 3x1 + 2z1 y2 x5 + 9x1 x2 x3 x4 − x1 y2 z3 x5 → 9xa1 + 4za2 z5 − 3za3 x1 − za3 xa2 + 2xa3 x5
(193)

Bibliography
• Original paper, and description of the choices of terms and factors used for the
given example [38].
 57

C. NP-Nagaj-1 (Nagaj, 2010)

Summary The Feynman Hamiltonian can be written as [39]:

1 1
(x1 x2 − iy1 x2 + ix1 y2 + y1 y2 ) U2−local + (x1 x2 + iy1 x2 − ix1 y2 + y1 y2 ) U2−local
†
, (194)
4 4
where U2−local is an arbitrary 2-local unitary matrix that acts on qubits different from
the ones labeled by "1" and "2". This Hamiltonian that is 4-local on qubits can be
transformed into one that is 2-local in qubits and qutrits. Here we show the 2-local
Hamiltonian for the case where U2−local = CNOT ≡ 12 (11 + z3 + x4 − z3 x4 ) . We start with
the specific 4-local Hamiltonian:
1
H4−local = (x1 x2 + y1 y2 + x1 x2 z3 + x1 x2 x4 + y1 y2 z3 + y1 y2 x4 − x1 x2 z3 x4 − y1 y2 z3 x4 ) ,
4
(195)
and after adding 4 auxiliary qubits labeled by a1 to a4 and 6 auxiliary qutrits labeled by
a5 to a10 and acted on by the Gell-Mann matrices λ1 to λ9 , we get the following 2-local
Hamiltonian:
H2−local = 1/2 (2λ6,a8 + x1 λ1,a5 + y1 λ2,a5 + λ6,a5 − z3 λ6,a5 + xa1 λ4,a5 + ya1 λ5,a5 + xa1 λ1,a6 +
(196)
ya1 λ2,a6 + 2x4 λ6,a6 + xa2 λ4,a6 + ya2 λ5,a6 + xa2 λ1,a7 + ya2 λ2,a7 + λ6,a7 − za1 λ6,a7 + (197)
x2 λ4,a7 + y2 λ5,a7 + x1 λ1,a8 + y1 λ2,a8 + z3 λ6,a8 + xa5 λ4,a9 + ya5 λ5,a9 + λ6,a9 + (198)
xa6 λ4,a9 + ya6 λ5,a9 + xa6 λ1,a10 + ya6 λ2,a10 + λ6,a10 + z3 λ6,a10 + x2 λ4,a10 + y2 λ5,a10 ) ,
(199)
whose low-lying spectrum is equivalent to the spectrum of H2−local .
Cost
• 6 auxiliary qutrits and 4 auxiliary qubits
• 2 quartic, 4 cubic, and 2 quadratic terms becomes 27 quadratic terms and 5 linear
terms in the Pauli-GellMann basis.
Pros
• Exact (non-perturbative). No special control precision demands.
• All coefficients are equal to each other, with a value of 1/2, except one which is
equal to 1.
• With more auxiliary qubits, can be further reduced to only containing qubits.
Cons
• Involves qutrits in all 32 terms.
• Only derived (so far) for the Feynman Hamiltonian.
• High overhead in terms of number of auxiliary qubits and number of terms.
Example
The transformation presented above was for the case of U2−local = CNOT ≡ 12 (11 + z3 + x4 − z3 x4 ),
but similar transformations can be derived for any arbitrary unitary matrix U2−local .
Bibliography
• Original paper: [40].
 58

D. NP-Nagaj-2 (Nagaj, 2012)

Summary
Similar to NP-Nagaj-1 but instead of using qutrits, we use two qubits for each qutrit,
according to:
1 1
∣0⟩ → ∣00⟩, ∣1⟩ → √ (∣01⟩ + ∣10⟩) , ∣2⟩ → √ (∣01⟩ − ∣10⟩) . (200)
2 2
which leads to the following transformations:
1
∣01⟩⟨10∣ij + h.c. → √ (∣01⟩⟨10∣ij1 + ∣01⟩⟨10∣ij2 ) + h.c. (201)
2
1
∣02⟩⟨10∣ij + h.c. → √ (∣01⟩⟨10∣ij1 − ∣01⟩⟨10∣ij2 ) + h.c. (202)
2
∣1⟩⟨2∣j + h.c. → zj1 − zj2 , (203)

and the following 2-local Hamiltonian involving only qubits:

H2−local = 1/2 (za5 − za6 + za9 − za10 − za3 za5 + za3 za6 − za3 za9 + za3 za10 +
za11 − za12 + za15 − za16 + za3 za11 − za3 za12 + za3 za15 − za3 za16 ) + x4 za7 − x4 za8 + za13 − za14 +
1/√2 (x λ
1 1,a5 + y1 λ2,a5 + x1 λ1,a6 + y1 λ2,a6 + xa1 λ1,a7 + ya1 λ2,a7 + xa1 λ1,a8 + y1 λ2,a8 +
xa2 λ1,a7 + ya2 λ2,a7 + xa2 λ1,a8 + ya2 λ2,a8 + xa2 λ1,a9 + ya2 λ2,a9 + xa2 λ1,a10 +
(204)
ya2 λ2,a10 + x1 λ1,a11 + y1 λ2,a11 + xa4 λ1,a15 + ya4 λ2,a15 + xa1 λ4,a5 + ya1 λ5,a5 +
xa2 λ4,a7 + ya2 λ5,a7 + x2 λ4,a9 + y2 λ5,a9 + xa3 λ4,a11 + ya3 λ5,a11 + xa4 λ4,a13 −
ya4 λ5,a13 + xa2 λ4,a15 + ya2 λ5,a15 − xa1 λ4,a6 − ya1 λ5,a6 − xa2 λ4,a8 − ya2 λ5,a8 − x2 λ4,a10 −
−y2 λ5,a10 − xa3 λ4,a12 − ya3 λ5,a12 − xa4 λ4,a14 − ya4 λ5,a14 − x2 λ4,a16 − y2 λ5,a16 ) ,

whose low-lying spectrum is equivalent to the spectrum of H2−local .

Cost
• 16 auxiliary qubits.
Pros
• Exact (non-perturbative). No special control precision demands.
• Only involves qubits (as opposed to NP-Nagaj-1 which contains qutrits and
NP-OY which contains ququits.
Cons
• Only derived (so far) for the Feynman Hamiltonian.
• High overhead in terms of number of auxiliary qubits and number of terms.
Example
The transformation presented above was for the case of U2−local = CNOT ≡ 12 (11 + z3 + x4 − z3 x4 ),
but similar transformations can be derived for any arbitrary unitary matrix U2−local .
Bibliography
• Original paper: [41].
 59

VIII. PERTURBATIVE (3 → 2) GADGETS

The first gadgets for arbitrary Hamiltonians acting on some number of qubits, were
designed to reproduce the spectrum of a 3-local Hamiltonian in the low-lying spectrum
of a 2-local Hamiltonian.

A. P(3 → 2)-DC (Duan, Chen, 2011)

Summary
For any group of 3-local terms that can be factored into a product of three 1-local
factors, we can define three auxiliary qubits (regardless of the number of qubits we
have in total) labeled by ai and make the transformation:

3 2

∏ ∑ αij si → α + αi ∑ (∑ αij sij )

ss
+ αisx ∑ ∑ αij sij xai + αzz ∑ zai zaj (205)
i j i j i j ij

1
α= (206)
8∆
1
α =
ss
(207)
6∆1/3
1
αsx = − 2/3 (208)
6∆
1
α =−
zz
(209)
24∆
The result will be a 2-local Hamiltonian whose low-lying spectrum is equivalent to the
spectrum of H3−local to within  as long as ∆ = Θ (−3 ).
Cost
• 1 auxiliary qubit for each group of 3-local terms that can be factored into three
1-local factors.
• ∆ = Θ (−3 )
Pros
• Very few auxiliary qubits needed
Cons
• Will not work for Hamiltonians that do not factorize appropriately.
Bibliography
• Original paper: [42]
 60

B. P(3 → 2)-DC2 (Duan, Chen, 2011)

Summary
For any 3-local term (product of Pauli matrices si ) in the Hamiltonian, we can define
one auxiliary qubit labeled by a and make the transformation:

3
a ∏ si → α + αs s3 + αz za + αss (s1 + s2 ) + αsz s3 za + +αsx (s1 xa + s2 xa )
2
(210)
i

α αs αz − 1 a ( 4∆12/3 − 1) a ( 4∆12/3 − 1)
( ss sz sx ) = ( 2∆
1 a 1 ) (211)
α α α ∆1/3 4∆2/3 ∆2/3

The result will be a 2-local Hamiltonian whose low-lying spectrum is equivalent to the
spectrum of H3−local to within  as long as ∆ = Θ (−3 ).
Cost
• 1 auxiliary qubit for each 3-local term.
• ∆ = Θ (−3 )
Example

x1 z2 y3 − 3x1 x2 y4 + z1 x2 → α + αz (za1 + za2 ) + αy (y3 + y4 ) + α12

zx
z1 x2 + αzx z2 xa1 + α11
xx
x1 x2
(212)
+ αxx (x1 xa1 + x1 xa2 + x2 xa2 ) + αyz (y3 za1 + y4 za2 ) (213)

Bibliography
• Original paper: [42]
 61

C. P(3 → 2)-KKR (Kempe, Kitaev, Regev, 2004)

SummaryFor any 3-local term (product of commuting matrices si ) in the Hamilto-

nian, we can define three auxiliary qubits labeled by ai and make the transformation:

3
∏ si → α + αiss ∑ s2i + αisx ∑ si xai + αzz ∑ zai zaj (214)
i i i ij

1
α=− (215)
8∆
1
αss = − 1/3 (216)
6∆
1
αsx = (217)
6∆2/3
1
αzz = (218)
24∆
The result will be a 2-local Hamiltonian whose low-lying spectrum is equivalent to the
spectrum of H3−local to within  as long as ∆ = Θ (−3 ).
Cost
• 3 auxiliary qubits for each 3-local term.
• ∆ = Ω (−3 )
Example
zx xx xx xx xx yz yx
x1 z2 y3 − 3x1 x2 y4 + z1 x2 → α + α2a z x + α12 x1 x2 + α1a x xa11 + α1a x xa21 + α2a x xa22 + α3a y3 xa13 + α4a y4 xa23
12 2 a1 2 11 1 21 1 22 2 13 23
(219)

Bibliography
• Original paper on arXiv: [43]

• Journal publication two years later: [44]

 62

D. P(3 → 2)-OT (Oliveira-Terhal, 2005)

Summary
For any 3-local term which is a product of 1-local matrices si , we can define one
auxiliary qubit labeled by a and make the transformation:

3
a ∏ si → α + α1s s21 + α2s s22 + α3s s3 + αaz za + α12
ss
s1 s2 + α13 s1 s3 + α23
ss 2
s2 s3 + α3a
ss 2 sz
s3 za + α1a
sx
s1 xa + α2a
sx
s2 xa
i
(220)

⎛ 2/32 1/3 −∆
∆ 1/3
ss ⎞
⎛ αs α12 ⎞ ⎜− α ∆
a 1 ⎟
⎟ ⎜ ⎟
ss
⎜ α1 α13 ⎜
2 2
⎜ αs ss ⎟ a2 ⎟
2/3 ∆1/3
⎟ ⎜ ⎟
α 1
⎜ 2 α23
⎜ s sz ⎟ =⎜ ⎟
2
⎜ α3 α3a ⎟ ⎜− 2
α1/3 ∆2/3 α1/3 ∆2/3 ⎟ . (221)
⎜ z ⎟ ⎜⎜ ⎟
∆2/3 ⎟
2
⎜ αa sx ⎟
α2a ⎜ − ∆2 − α1/3
√ ⎟
⎝N/A sx ⎠ ⎜ 2 ⎟
α2a ⎝ N/A
1/3
α √∆ 2/3
⎠
2

A k-local Hamiltonian with a 3-local term replaced by this 2-local Hamiltonian will
have an equivalent low-lying spectrum to within  as long as ∆ = Ω (−3 ).
Cost
• 1 auxiliary qubit for each 3-local term.
• ∆ = Ω (−3 )
Example
Bibliography
• Original paper where α = 1: [45]. For arbitrary α see the 2005 v1 from arXiv, or
[46]. Connection to improved version: [47].
 63

IX. PERTURBATIVE 1-BY-1 GADGETS

A 1B1 gadget allows k-local terms to be quadratized one step at a time, where at
each step the term’s order is reduced by at most one. In each step, a k-local term is
reduced to (k − 1)-local, contrary to SD (sub-division) gadgets which can reduce k-local
terms to (1/2)-local in one step.

A. P1B1-OT (Oliveira & Terhal, 2008)

Summary
We wish to reduce the k-local term:

k
Hk−local = α ∏ sj . (222)
j

Define one auxiliary qubit labeled by a and make the transformation:

α /3 1 − za α /3 ∆r
1 1

Hk−local → − ( ) ∆2(1−r) sk ( ) + ( ) √ (sk−1 − sk−2 ) xa (223)

2 2 2 2
1 α /3
2
√ −1/4 k−2 2
α
+ ( ) (∆ sk−1 + sgn(α) 2∆ ∏ sj ) + (1 + 2sgn2 α∆ /2−2r ) sk .
r−1 3

2 2 j 4
(224)

The result will be a (k − 1)-local Hamiltonian with the same low-lying spectrum as
Hk−local to within  as long as ∆ = Ω (−3 ).
Cost
• Only 1 auxiliary qubit.
• ∆ = Ω (−3 )
Example
Bibliography
• Described in: [47], based on: [45].
 64

B. P1B1-CBBK (Cao, Babbush, Biamonte, Kais, 2015)

Summary
Define one auxiliary qubit labeled by a and make the transformation:

α /2 1 1 − za
3

Hk−local → (∆ + ( ) ∆ /2 sk ) ( ) (225)
2 2
α2/3 √ 1 + za
− (1 + sgn2 α) ((2α) /3 sgn2 α + α1/3 sk − 2∆1/2 ) ( )
2 3
(226)
2 2
α /3 3/4 k−2
1
√ k−1
+ ( ) ∆ (∏ sj − sgn(α)sk−1 ) xa + sgn(α) 2α /3 (∆ /2 + ∆ /2 ) ∏ sj . (227)
3 2 1 3

2 j j

The result is (k − 1)-local and its low-lying spectrum is the same as that of Hk−local when
∆ is large enough.
Cost
• Only 1 auxiliary qubit.
• ∆ = Ω (−3 )
Example
Bibliography
• Described in: [47], based on: [45].
 65

X. PERTURBATIVE SUBDIVISION GADGETS

Instead of recursively reducing k-local to (k − 1)-local one reduction at a time, we

can reduce k-local terms to (k/2)-local terms directly for even k, or to (k + 1)/2-local
terms directly for odd k. Since when k is odd we can add an identity operator to the
k-local term to make it even, we will assume in the following that k is even, in order to
avoid having to write floor and ceiling functions.

A. PSD-OT (Oliveira & Terhal, 2008)

Summary
We factor a k-local term into a product of three factors: operators H1 , H2 acting on
non-overlapping spaces, and scalar α. Then introduce an auxiliary qubit labelled by a
and make the transformation:

√
1 − za α 2 α 2 α∆
Hk−local → ∆ + H1 + H2 + (−H1 + H2 ) xa . (228)
2 2 2 2
The resulting Hamiltonian has a degree of 1 larger than the degree of whichever
factor H1 or H2 has a larger degree, and the low-lying spectrum is equivalent to the
original one to within O (α) for sufficiently large ∆.
Cost
• 1 auxiliary qubit for each k-local term that can be factored into two non-
overlapping subspaces, is enough to reduce the degree down to k/2 + 1.
√ 6
α(∣∣H(else) +Ω( 2)max(∣∣H1 ∣∣,∣∣H2 ∣∣)∣∣)
• ∆=  2 = Ω (α−2 ).
Pros
• Potentially very few auxiliary qubits needed.
Cons
• Requires the ability to factor k-local terms into non-overlapping subspaces that
are at most (k − 2)-local in order to reduce k-locality. This is not possible for
z1 x2 x3 + z2 z3 x4 , for example.
• ∆ needs to be rather large.

• Cannot reduce 3-local to 2-local unless we generalize to a factor of 3 non-

overlapping subspaces instead of 2. Needs to be combined with 3− → 2 gadgets,
for example.
• A lot of work may be needed to find the optimal reduction, since each k-local
term can be factored in many ways, and some of these ways may affect the ability
to reduce other k-local terms.
Example
Bibliography
• Original paper where α = 1: [45]. For arbitrary α see the 2005 v1 from arXiv, or
[46]. Connection to improved version: [47].
 66

B. PSD-CBBK (Cao, Babbush, Biamonte, Kais 2015)

Summary
For any k-local term, we can subdivide it into a product of two (k/2) −local terms:

Hk−local = αH1,(k/2)−local H2,(k/2)−local + H(k−1)−local . (229)

Define one qubit a and make the following Hamiltonian is (k/2)-local:

1 − za 1 + za √
∆ + ∣α∣ + ∣α∣∆/2 (sgn(α)H1,(k/2−local) − H2,(k/2−local) ) xa (230)
2 2
The result is a (k/2)-local Hamiltonian with the same low-lying spectrum as Hk−local for
large enough ∆. The disadvantage is that ∆ has to be larger.
Cost
• ∆ ≥ ( 2∣α∣
 + 1) (∣α∣ +  + 22∣∣H(k−1)−local )

Pros
• only one qubit to reduce k to ⌈k/2⌉ + 1
Cons
• Only beneficial for k ≥ 5.
Example
Bibliography
• Original paper: [47].
 67

C. PSD-CN (Cao & Nagaj, 2014)

Summary
For a sum of terms that are k-local, with each term j written as a product H1j H2j ,
introduce Ncore ‘core’ auxiliary qubits labeled by ai and Ndirect ‘direct’ auxiliary qubits
labeled by aij for each term j. Make all core auxiliary qubits couple to all others, and
make the direct auxiliary qubits couple each H1j and H2j to the core auxiliary qubits.

∑ aj H1j H2j → α ∑ij (1 − zaij zai + αj xaij (H1j − H2j )) + α ∑i (1 − zai + ∑j (1 − zai zaj ))
j
(231)
If we would like the spectrum of the RHS to be close to that of the LHS, with a
difference of O(), then for any d ∈ (0, 1) we can choose
⎧
⎪
1
⎫
1 ⎪⎞
⎛ ⎪ −2 ∥Helse ∥2 d
−2 ⎪
Ndirect ∈ Ω max ⎨ d , ( ) , (M  ) ⎬ ,
3 1−d
(232)
⎝ ⎪
⎪ 2M maxj ∣aj ∣
4 ⎪
⎪ ⎠
⎩ ⎭
Ncore ∈ Ω (M 3 Ndirect
d
−1 ) , (233)
αj , α ∈ O() (234)

.
Pros
• For spectral error , uses only O() coupling between the qubits (See Equation
234).
Cons
• Uses poly(−1 ) ancilla qubits (See Equations 232 and 233).
• The construction only describes the asymptotic scaling of the parameters rather
than concrete assignments of them. More work is needed for finding tight non-
asymptotic error bounds in perturbative expansion.
Example
Bibliography
• Original paper: [48].
 68

XI. PERTURBATIVE DIRECT GADGETS

Here we do not reduce k by one order at a time (1B1 reduction) or by k/2 at a time
(SD reduction), but we directly reduce k-local terms to 2-local terms.

A. PD-JF (Jordan & Farhi, 2008)

Summary
Express a sum of k-local terms as a sum of products of Pauli matrices sij , and define
k auxiliary qubits laelled by aij for each term i, and make the transformmation:

k
−k(−)k 1⎛ 2 k ⎞ k
∑ αi ∏ sij → ∑ k − ∑ zaij zail +  (αi si1 xi1 + ∑ sij xij ) − f ()Π, (235)
i j (k − 1)! i 2 ⎝ jl ⎠ j

for some polynomial f (λ). The result is a 2-local Hamiltonian with the same low-lying
spectrum to within k+1 for sufficiently smmall .
Cost
• Number of auxiliary qubits is tk for t terms.
• Unknown requirement for .
Pros
• All done in one step, so easier to implement than 1B1 and SD gadgets.
Cons
• Requires 2 more auxiliary qubits per term than 1B1-KKR.
• Unknown polynomial f (λ)
Example
Bibliography
• Original paper [49].
 69

B. PD-BFBD (Brell, Flammia, Bartlett, Doherty, 2011)

Summary
The 4-body Hamiltonian:

H4-local = − ∑ (z4i+1,j z4i+2,j z4i+3,j z4i+4,j + x4i+3,j x4i+4,j x4i+6,j x4i+4,j+1 ) (236)
ij

is transformed into the 2-body Hamiltonian:

H2-local = − ∑ (x8i+4,j x8i+6,j + x8i+3,j+1 x8i+5,j+1 + z8i+4,j z8i+3,j+1 + z8i+6,j z8i+5,j+1 + (237)
ij

= −λ (x8i+1,j x8i+3,j + x8i+2,j x8i+4,j + x8i+5,j x8i+6 + x8i+7 x8i+8 (238)

= z8i+1,j z8i+3,j + z8i+2,j z8i+4,j + z8i+5,j z8i+6 + z8i+7 z8i+8 )) (239)

For λ = O (−5 ), the 2-local Hamiltonian has the same low-lying spectrum as the 4-local
Hamiltonian, to within an error of .
Cost
• In total, uses four times the number of qubits of the original Hamiltonian.
• Unknown requirement for λ.
Pros
• All done in one step, so easier to implement than implementing two 1B1 gadgets.

• Very symmetric
Cons
• Ordinary 1B1 or SD followed by 3→2 gadgets would require half as many total
qubits.
• Many 2-local terms.

• Perturbative, as opposed to NR-OY which is similar but does not involve any λ
parameter.

• Required value of λ for it to work, is presently unknown.

Bibliography
• Original paper: [50].
 70

Part IV
Appendix

XII. TRANSFORMATIONS FROM TERNARY TO BINARY VARIABLES

Hternary = −λ(z1 z2 + z1 − z2 ) (240)

In this implementation the variable 12 (z1 + z2 ) plays the role of t, assuming λ is large
and positive. For instance, when coupled to a binary variable t z3 → 12 (z1 + z2 ) z3 .
 71

XIII. FURTHER EXAMPLES

Example Here we show how deductions can arise naturally from the Ramsey number
problem. Consider R(4, 3) with N = 4 nodes. Consider a Hamiltonian:

H = (1 − z12 )(1 − z13 )(1 − z23 ) + . . . + (1 − z23 )(1 − z24 )(1 − z34 ) + z12 z13 z14 z23 z24 z34 . (241)

See [8] for full details of how we arrive at this Hamiltonian.

Since we are assuming we have no 3-independent sets, we know that (1 − z12 )(1 −
z13 )(1 − z23 ) = 0, so z12 z13 z23 = z12 z13 + z12 z23 + z13 z23 − z12 − z13 − z23 + 1. This will be our
deduction.
Using deduc-reduc we can substitute this into our 6-local term to get:

H = 2(1 − z12 )(1 − z13 )(1 − z23 ) + . . . + (1 − z23 )(1 − z24 )(1 − z34 ) + (242)
z14 z24 z34 (z12 z13 + z12 z23 + z13 z23 − z12 − z13 − z23 + 1). (243)
We could repeat this process to remove all 5- and 4-local terms without adding any
auxiliary qubits. Note in this case the error terms added by deduc-reduc already appear
in our Hamiltonian.
 72

XIV. 2 → 2 GADGETS

This review has only focused on k−local to 2−local transformations where k > 2.
There is also a large number of 2−local to 2−local transformations in the literature,
which are used for various pruposes. Some of these are listed here:

• Gadgetization of any 2−local Hamiltonian into {11, z, x, zz, xx} or {11, z, x, zx} [51].
Used for the proof that xx + zz or xz is universal is enough for universal quantum
computation. In other words, any computation can be transformed into a problem
of finding the ground state of a 2−local Hamiltonian containing terms from
{11, z, x, zz, xx} or from {11, z, x, zx} with real coefficients, and the ground state
can be found by adiabatic quantum computing with only polynomial time and
space overhead over the best alternative algorithm for the problem.

• Transformation of any 2−local Hamiltonian into {11, z, x, zz, xx + yy}, without any
perturbative gadgets, and only requiring the qubits to be connected in an almost
2D lattice [52].
• "Cross gadget", "fork gadget", and "triangle gadget" described in [45].

• Gadgetizeation of a 2−local Hamiltonian with very strong couplings, into a 2−local

Hamiltonian with strengths in O (1/poly(−1 ,n)), and poly (−1 , n) auxiliary qubits
and poly (−1 , n) new quadratic terms. [48].
• yy creation gadget: Simulation of yy terms using {11, z, x, zz, xx}, with coupling
strength restriction defined according to ∆ = Θ (−4 ) [47].

A. Minor-embedding quadratic functions for different graphs

• Minor-embedding general problems for the Chimera [53] graph [22, 54].
• Minor-embedding quadratization gadgets for the Pegasus [55] graph [56].
 73

XV. FURTHER REFERENCES

• Gadgets for pseudo-Boolean optimization problems, with reduced precision

requirements: [57].

• Formalization of pseudo-Boolean gadgets in quantum language. [31].

• By adding more couplers and more auxiliary qubits, we can bring the error down
arbitrarily low: [48].
• More toric code gadgets: [58].

• Parity adiabatic quantum computing (LHZ lattice): [17].

• Extensions of the LHZ scheme: [16].
• Minimizing k-local discrete functions with the help of continuous variable calcu-
lus [59].
• ORI graph which attempts to give optimal quadratizations [15].

• Survey on pseudo-Boolean optimization [60].

• Linearization of equations before they are squared [61], and its application to
factoring numbers [62].
• Mentioned in [5] as an early application of quadratization: [63].
• Characterizaton of NTRs for cubics: [64].

• Relation between cones of nonnegative quadratic pseudo-Boolean functions and

the Boolean quadric polytope [65].

• Relation between cones of nonnegative quadratic pseudo-Boolean functions and

the Boolean quadric polytope [65].

• Effective non-Hermitian Hamiltonian with 3-body interactions which helps to

calcualte electronic structure energies closer to the complete basis set limit [66].
 74

XVI. CIRCUITS THAT EFFECTIVELY IMPLEMENT DEGREE-k TERMS FOR

SUPERCONDUCTING QUBITS

• Presentation by Northrop Grumman about a zzz coupler [67] and associated

patent [68, 69].

• Presentaiton that included discussion about engineering multi-qubit interactions

[70], presentation by the same lab about the design and experimental demonstra-
tion of a zzzz coupling [71], and associated patent [72–74].
• Design of an effective zzzz coupling without any auxiliary logical qubits [75].

• Design of a tunable zzz coupling in which all zz couplings are cancelled, and its
experimental demonstration [76].
 75

XVII. CONTRIBUTORS

Richard Tanburn

• Richard was the original creator and maintaner of the Git repository.

• Richard created the Tex commands used throughout the document, and con-
tributed majorly to the overall layout.

• Richard wrote the original versions of the following sections: (1) Deduc-Reduc,
(2) ELC Reduction, (3) Groebner Bases, (4) Split Reduction, (5) NTR-KZFD, (6)
NTR-GBP, (7) PTR, (8) PTR-Ishikawa, (9) PTR-KZ, (10) PTR-GBP, (11) Bit flipping,
(12) RBS, and (13) FGBZ.

• Richard also wrote the "Further Example" of Deduc-Reduc in the Appendix.

Nicholas Chancellor

• Nick made contributions to the following sections: (1) RMS (in terms of z), (2)
PTR-RBL-(3→2), (3) PTR-RBL-(4→2), (4) SBM, (5) Flag based SAT Mapping, and
to the qutrit → qubit transformation (6).

Szilard Szalay

• Szilard re-derived Nike’s transformations for the sections: (1) SFR-BCR-1, (2)
SFR-BCR-2, (3) SFR-BCR-3, and (4) SFR-BCR-4 from the notation of the original
paper, into the format consistent with the rest of the book. In doing so he corrected
errors in Nike’s work and also fixed them in the main document.

Ka Wa Yip

• Ka Wa Yip added the page about his own method co-authored with Xu, Koenig
and Kumar.

Yudong Cao

• Yudong wrote the first version of the following section: (1) PSD-CN.

Daniel Nagaj

• Daniel provided a .tex document to Nike in May 2018 which helped Nike to write
the following sections: (1) NP-Nagaj-1, (2) NP-Nagaj-2. The document that Daniel
provided made it easier for Nike to write these sections than the original papers.
 76

Aritanan Gruber

• Aritanan informed us in August 2015 of what we ended up making the following

sections: (1) PTR-Ishikawa.

Charles Herrmann

• Charles informed us in May 2018 of the papers which contained results which
became the following sections: (1) PTR-BCR-1, (2) PTR-BCR-2, (3) PTR-BCR-3,
(4) PTR-BCR-4, (5) SFR-BCR-1, (6) SFR-BCR-2, (7) SFR-BCR-3, (8) SFR-BCR-4, (9)
SFR-BCR-5, (10) SFR-BCR-6.

Elisabeth Rodriguez-Heck

• Elisabeth provided us with a 2-page PDF document with valuable comments on

the entire Book.

• Elisabeth also pointed us to what became the following section: (1) ABCG Reduc-
tion.

Hou Tin Chau

• Tin made the examples for the following sections: SFR-BCR-1,2,3,4.

• Tin fixed a typo in the alternative forms of the following sections: SFR-BCR-3,4.

Andreas Soteriou

• Andreas found typos on the opening page in the arXiv version which surprisingly
no one else found (or pointed out), and he diligently fixed them.

• Andreas created the example involving x,y, and z presented on the opening page
in the September 2019 version (I plan to have this example further improved at a
later time).

Jacob Biamonte

• Jacob made valuable edits during a proof-reading of the book.

 77

ACKNOWLEDGMENTS

• It is with immense pleasure that we thank Emile Okada of Cambridge University,

who during his first year of udnergraduate study, worked with Nike Dattani and
Richard Tanburn on quadratization of pseudo-boolean functions for quantum
annealing, and in the first half of 2015 played a role in the development of the
Deduc-Reduc and Split-Reduc and Groebner bases methods presented in this
review.

• We thank Gernot Schaller of University of Berlin, who in December 2014 provided

Nike Dattani with insights into his quadratization methods mentioned in this
review paper, as well as for sharnig his Mathematica code which could be used to
generate such quadratization formulas and others.

• We thank Mohammad Amin of D-Wave for pointing Nike Dattani to the paper
[33] on determining Ramsey numbers on the D-Wave device, which contained
what we call in this review "Reduction by substitution", later found through
Ishikawa’s paper to be from the much older 1970s paper by Rozenberg.

• We thank Catherine McGeoch of Amherst University and D-Wave, for helpful

discussions with Nike Dattani in December 2014 about how to map quadratic
pseudo-Boolean optimization problems onto the chimera graph of the D-Wave
hardware and for pointing us to the important references of Vicky Choi. While
chimerization is very different from quadratization, understanding that roughly
n2 variables would be needed to map a quadratic function of n variables, helped
Nike Dattani and Richard Tanburn to appreciate how impotrant it is to be able to
quadratize with as few variables as possible, and having this in mind throughout
our studies helped inspire us in our goals towards "optimal quadratization".

• We thank Aritanan Gruber and Endre Boros of Rutgers University, who in August
2015 shared with Nike Dattani and Richard Tanburn some of their wisdom about
sub-modularity, and Aritanan Gruber for pointing us to the then very recent paper
of Hiroshi Ishikawa on what we call in this review "ELC reductions", which was
also a valuable paper due to the references in it. We also thank him for helping us
in our quest to determine whether or not "deduc-reduc" was a re-discovery by
Richard, Emile, and Nike, or perhaps a novel quadratization scheme.

• We thank Toby Cathcart-Burn of Oxford University, who during the third year
of undergraduate study, worked with Nike Dattani and Richard Tanburn and in
Autumn 2015 and Winter 2016 helped us gain insights about the application of
deduc-reduc and bit flipping to the problem of determining Ramsey numbers
via discrete optimization, and for insights into the trade-offs between Ishikawa’s
symmetric reduction and reduction by substitution.

• We thank Hiroshi Ishikawa of Waseda University, who Nike Dattani had the
memorable opportunity to visit in November 2015, and through discussions
about the computer vision problem and neural network problem (two examples
of real-world discrete optimization problems that benefit from quadratization),
provided insights about the role of quadratization for calculations on classical
computers. In particular, solving the computer vision problem in which he had
experience solving on classical computers, was very different from the integer
factoring problem and Ramsey number problem which we had been attempting
 78

to quadratize for D-Wave and NMR devices. He taught us that far more total
(original plus auxiliary) variables can be tolerated on classical computers than on
D-Wave machines or NMR systems, and approximate solutions to the optimiza-
tion problems are acceptable (unlike for the factorization and Ramsey number
problems in which we were interested). This gave us more insight into what
trade-offs one might wish to prioritize when quadratizing optially. We also thank
him for helping us in our quest to determine whether or not "deduc-reduc" was
a re-discovery by Richard, Emile, and Nike, or perhaps a novel quadratization
scheme.

• Last but indubitably not least, we thank Jacob Biamonte of Skolkovo Institute of
Technology, who Nike Dattani enjoyed visiting in Hangzhou in January 2017 and
meeting at Harvard University in April 2018. Jacob provided us plenty of insights
about perturbative gadgets, made valuable comments on early versions of our
manuscript, and has been a major supporter of this review paper since the idea
was presented to him in December 2016. At many points during the preparation
of this review, we had prioritized other commitments and put preparation of this
review aside. Jake’s frequent encouragement was often what got us working on
this review again. Without him, we are not certain this paper would have been
completed by this time (or ever!).
 79

[1] Richard Tanburn, Emile Okada, and Nike Dattani, “Reducing multi-qubit interactions in
adiabatic quantum computation without adding auxiliary qubits. Part 1: The "deduc-reduc"
method and its application to quantum factorization of numbers,” (2015), arXiv:1508.04816.

[2] Hiroshi Ishikawa, “Higher-Order Clique Reduction without Auxiliary Variables,” in 2014
IEEE Conference on Computer Vision and Pattern Recognition (IEEE, 2014) pp. 1362–1369.

[3] Magali Bardet, “On the Complexity of a Gröbner Basis Algorithm,” Algorithms Seminar ,
85–92 (2002).

[4] Raouf Dridi and Hedayat Alghassi, “Prime factorization using quantum annealing and
computational algebraic geometry,” (2016), arXiv:1604.05796.

[5] Asem M. Ali, Aly A. Farag, and Georgy L. Gimel’farb, “Optimizing Binary MRFs with
Higher Order Cliques,” (Springer, Berlin, Heidelberg, 2008) pp. 98–111.

[6] Richard Tanburn, Oliver Lunt, and Nikesh S. Dattani, “Crushing runtimes in adiabatic
quantum computation with Energy Landscape Manipulation (ELM): Application to Quan-
tum Factoring,” Physical Review A (submitted). (2015), arXiv:1510.07420.

[7] Nike Dattani, Quadratization in discrete optimization and quantum mechanics (2019)
arXiv:1901.04405.

[8] Emile Okada, Richard Tanburn, and Nikesh S. Dattani, “Reducing multi-qubit interactions
in adiabatic quantum computation without adding auxiliary qubits. Part 2: The "split-
reduc" method and its application to quantum determination of Ramsey numbers,” , 5
(2015), arXiv:1508.07190.

[9] Martin Anthony, Endre Boros, Yves Crama, and Aritanan Gruber, “Quadratic reformu-
lations of nonlinear binary optimization problems,” Mathematical Programming 162,
115–144 (2017).

[10] V. Kolmogorov and R. Zabih, “What energy functions can be minimized via graph cuts?”
IEEE Transactions on Pattern Analysis and Machine Intelligence 26, 147–159 (2004).

[11] D. Freedman and P. Drineas, “Energy Minimization via Graph Cuts: Settling What is Pos-
sible,” in 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition
(CVPR’05), Vol. 2 (IEEE, 2005) pp. 939–946.

[12] H Ishikawa, “Transformation of General Binary MRF Minimization to the First-Order

Case,” IEEE Transactions on Pattern Analysis and Machine Intelligence 33, 1234–1249
(2011).

[13] Martin Anthony, Endre Boros, Yves Crama, and Aritanan Gruber, “Quadratization of
symmetric pseudo-Boolean functions,” Discrete Applied Mathematics 203, 1–12 (2016).

[14] Martin Anthony, Endre Boros, Yves Crama, and Aritanan Gruber, “Quadratization of
symmetric pseudo-Boolean functions,” (2014), arXiv:1404.6535.

[15] Andrew C. Gallagher, Dhruv Batra, and Devi Parikh, “Inference for order reduction in
Markov random fields,” in CVPR 2011 (IEEE, 2011) pp. 1857–1864.
 80

[16] Andrea Rocchetto, Simon C Benjamin, and Ying Li, “Stabilizers as a design tool for new
forms of the Lechner-Hauke-Zoller annealer,” Science Advances 2 (2016), 10.1126/sci-
adv.1601246.

[17] Wolfgang Lechner, Philipp Hauke, and Peter Zoller, “A quantum annealing architecture
with all-to-all connectivity from local interactions,” Science Advances 1 (2015), 10.1126/sci-
adv.1500838.

[18] Endre Boros and Aritanan Gruber, “On Quadratization of Pseudo-Boolean Functions,”
(2014), arXiv:1404.6538.

[19] Endre Boros, Yves Crama, and Elisabeth Rodríguez-Heck, “Quadratizations of symmetric
pseudo-boolean functions: sub-linear bounds on the number of auxiliary variables,” in
ISAIM (2018).

[20] Endre Boros, Yves Crama, and Elisabeth Rodríguez-Heck, “Compact quadratizations for
pseudo-boolean functions,” in unpublished (2018).

[21] Ka Wa Yip, Hong Xu, T. K. Satish Kumar, and Sven Koenig, “Quadratic reformulation of
nonlinear pseudo-boolean functions via the constraint composite graph,” in the Interna-
tional Conference on Integration of Artificial Intelligence and Operations Research Techniques in
Constraint Programming (2019) pp. 643–660.

[22] Vicky Choi, “Minor-embedding in adiabatic quantum computation: I. the parameter

setting problem,” Quantum Information Processing 7, 193–209 (2008).

[23] Nicholas Chancellor, Stefan Zohren, and Paul A. Warburton, “Circuit design for multi-
body interactions in superconducting quantum annealing system with applications to a
scalable architecture,” (2016), 10.1038/s41534-017-0022-6, arXiv:1603.09521.

[24] Martin Leib, Peter Zoller, and Wolfgang Lechner, “A Transmon Quantum Annealer: De-
composing Many-Body Ising Constraints Into Pair Interactions,” (2016), arXiv:1604.02359.

[25] Martin Leib, Peter Zoller, and Wolfgang Lechner, “A transmon quantum annealer: de-
composing many-body Ising constraints into pair interactions,” Quantum Science and
Technology 1, 15008 (2016).

[26] N Chancellor, S Zohren, and P A Warburton, “Circuit design for multi-body interactions in
superconducting quantum annealing systems with applications to a scalable architecture,”
npj Quantum Information 3, 21 (2017).

[27] Nicholas Chancellor, “Max-k-SAT, Multi-Body Frustration, & Multi-Body Sampling on a

Two Local Ising System,” AQC 2016 https://www.youtube.com/watch?v=aC-6hg_h3EA
(2016).

[28] N. Chancellor, S. Zohren, P.A. Warburton, S.C. Benjamin, and S. Roberts, “A Direct
Mapping of Max k-SAT and High Order Parity Checks to a Chimera Graph,” Scientific
Reports 6 (2016), 10.1038/srep37107.

[29] Martin Anthony, Endre Boros, Yves Crama, and Aritanan Gruber, “Quadratic reformula-
tions of nonlinear binary optimization problems,” .

[30] I. G. Rosenberg, “Reduction of Bivalent Maximization to the Quadratic Case,” Cahiers du

Centre d’Etudes de Recherche Operationnelle 17, 71–74 (1975).
 81

[31] J D Biamonte, “Nonperturbative k-body to two-body commuting conversion Hamiltonians

and embedding problem instances into Ising spins,” Phys. Rev. A 77, 52331 (2008).

[32] Alejandro Perdomo, Colin Truncik, Ivan Tubert-Brohman, Geordie Rose, and Alán Aspuru-
Guzik, “Construction of model Hamiltonians for adiabatic quantum computation and its
application to finding low-energy conformations of lattice protein models,” Phys. Rev. A
78, 12320 (2008).

[33] Zhengbing Bian, Fabian Chudak, William G. Macready, Lane Clark, and Frank Gaitan,
“Experimental Determination of Ramsey Numbers,” Physical Review Letters 111, 130505
(2013).

[34] Alexander Fix, Aritanan Gruber, Endre Boros, and Ramin Zabih, “A graph cut algorithm
for higher-order Markov Random Fields,” in 2011 International Conference on Computer
Vision (IEEE, 2011) pp. 1020–1027.

[35] Gemma De las Cuevas and Toby S. Cubitt, “Simple universal models capture all classical
spin physics,” Science 351, 1180—-1183 (2016).

[36] Kahl and Strandmark, “Generalized Roof Duality for Pseudo-Boolean Optimization,”
International Conference on Computer Vision (2011).

[37] Samuel A. Ocko and Beni Yoshida, “Nonperturbative Gadget for Topological Quantum
Codes,” Physical Review Letters 107, 250502 (2011).

[38] Yigit Subasi and Christopher Jarzynski, “Nonperturbative embedding for highly nonlocal
Hamiltonians,” Physical Review A 94, 012342 (2016).

[39] Richard P. Feynman, “Quantum Mechanical Computers,” Optics News 11, 11 (1985).

[40] Daniel Nagaj, “Fast universal quantum computation with railroad-switch local Hamiltoni-
ans,” Journal of Mathematical Physics 51, 062201 (2010).

[41] Daniel Nagaj, “Universal two-body-Hamiltonian quantum computing,” Physical Review

A 85, 032330 (2012).

[42] Qian-Heng Duan and Ping-Xing Chen, “Realization of universal adiabatic quantum com-
putation with fewer physical resources,” Physical Review A 84, 042332 (2011).

[43] Julia Kempe, Alexei Kitaev, and Oded Regev, “The Complexity of the Local Hamiltonian
Problem,” (2004), arXiv:0406180 [quant-ph].

[44] Julia Kempe, Alexei Kitaev, and Oded Regev, “The Complexity of the Local Hamiltonian
Problem,” SIAM Journal on Computing 35, 1070–1097 (2006).

[45] Roberto Oliveira and Barbara M. Terhal, “The complexity of quantum spin systems on a
two-dimensional square lattice,” Quantum Info. Comput. 8, 900–924 (2008).

[46] Sergey Bravyi, David P. DiVincenzo, Daniel Loss, and Barbara M. Terhal, “Quantum Sim-
ulation of Many-Body Hamiltonians Using Perturbation Theory with Bounded-Strength
Interactions,” Physical Review Letters 101, 070503 (2008).

[47] Yudong Cao, Ryan Babbush, Jacob Biamonte, and Sabre Kais, “Hamiltonian gadgets with
reduced resource requirements,” Physical Review A 91, 012315 (2015).
 82

[48] Yudong Cao and Daniel Nagaj, Quantum Information & Computation, Vol. 15 (Rinton Press,
2015) pp. 1197–1222.

[49] Stephen P. Jordan and Edward Farhi, “Perturbative gadgets at arbitrary orders,” Physical
Review A 77, 062329 (2008).

[50] Courtney G Brell, Steven T Flammia, Stephen D Bartlett, and Andrew C Doherty, “Toric
codes and quantum doubles from two-body Hamiltonians,” New Journal of Physics 13,
053039 (2011).

[51] Jacob D. Biamonte and Peter J. Love, “Realizable Hamiltonians for universal adiabatic
quantum computers,” Physical Review A 78, 012352 (2008).

[52] Seth Lloyd and Barbara M Terhal, “Adiabatic and Hamiltonian computing on a 2D lattice
with simple two-qubit interactions,” New Journal of Physics 18, 023042 (2016).

[53] Hartmut Neven, Vasil S Denchev, Marshall Drew-Brook, Jiayong Zhang, William G
Macready, and Geordie Rose, NIPS 2009 Demonstration: Binary Classification using Hardware
Implementation of Quantum Annealing, Tech. Rep. (2009).

[54] Vicky Choi, “Minor-embedding in adiabatic quantum computation: II. Minor-universal

graph design,” Quantum Information Processing 10, 343–353 (2011).

[55] Nikesh S. Dattani, Szilard Szalay, and Nicholas Chancellor, “Pegasus: The second connec-
tivity graph for large-scale quantum annealing hardware,” Pegasus: The second connectiv-
ity graph for large-scale quantum annealing hardware (2019), arXiv:1901.07636.

[56] Nike Dattani and Nick Chancellor, “Embedding quadratization gadgets on Chimera and
Pegasus graphs,” (2019), arXiv:1901.07676.

[57] Ryan Babbush, Bryan O’Gorman, and Alán Aspuru-Guzik, “Resource efficient gadgets for
compiling adiabatic quantum optimization problems,” Annalen der Physik 525, 877–888
(2013).

[58] Courtney G Brell, Stephen D Bartlett, and Andrew C Doherty, “Perturbative 2-body parent
Hamiltonians for projected entangled pair states,” New Journal of Physics 16, 123056
(2014).

[59] Jianbing Shen, Jianteng Peng, Xingping Dong, Ling Shao, and Fatih Porikli, “Higher Order
Energies for Image Segmentation,” IEEE Transactions on Image Processing 26, 4911–4922
(2017).

[60] Endre Boros and Peter L. Hammer, “Pseudo-Boolean optimization,” Discrete Applied
Mathematics 123, 155–225 (2002).

[61] Gernot Schaller and Ralf Schützhold, “The role of symmetries in adiabatic quantum
algorithms,” Quantum Information & Computation 10, 109–140 (2010).

[62] Nanyang Xu, Jing Zhu, Dawei Lu, Xianyi Zhou, Xinhua Peng, and Jiangfeng Du, “Quan-
tum Factorization of 143 on a Dipolar-Coupling Nuclear Magnetic Resonance System,”
Physical Review Letters 108, 130501 (2012).

[63] William H. Cunningham, “Minimum cuts, modular functions, and matroid polyhedra,”
Networks 15, 205–215 (1985).
 83

[64] Yves Crama and Elisabeth Rodriguez Heck, Short Prime Quadratizations of Cubic Negative
Monomials (HEC ULg, 2014).

[65] Endre Boros and Isabella Lari, “Cones of nonnegative quadratic pseudo-boolean func-
tions,” in International Symposium on Artificial Intelligence and Mathematics, ISAIM 2014, Fort
Lauderdale, FL, USA, January 6-8, 2014 (2014).

[66] Aron J. Cohen, Hongjun Luo, Kai Guther, Werner Dobrautz, David P. Tew, and Ali Alavi,
“Similarity transformation of the electronic Schrödinger equation via Jastrow factorization,”
The Journal of Chemical Physics 151, 061101 (2019).

[67] Joel Strand, Anthony Przybysz, David Ferguson, and Ken Zick, “ZZZ coupler for native
embedding of MAX-3SAT problem instances in quantum annealing hardware,” Bulletin of
the American Physical Society Volume 62, Number 4 (2017).

[68] David Ferguson, Anthony Przybysz, and Joel Strand, “ZZZ coupler for superconducting
qubits,” US Patent: 20180261752A1 (2017).

[69] David Ferguson, Anthony Przybysz, and Joel Strand, “ZZZ coupler for superconducting
qubits,” US Patent: 20180342663A1 (2018).

[70] Andrew Kerman, “Design and simulation of complex superconducting circuits for ad-
vanced quantum annealing hardware,” Bulletin of the American Physical Society Volume
63 (2018).

[71] Tim Menke, Cyrus Hirjibehedin, Steven Weber, Gabriel Samach, Simon Gustavsson, Alan
Aspuru-Guzik, William Oliver, and Andrew Kerman, “A many-body coupler for coherent
4-local interaction of superconducting flux qubits,” APS March Meeting 2019, abstract
id.A42.011 (2019).

[72] Andrew Kerman, “Four Spin Couplers for Quantum Information Processing,” US Patent:
20180309452A1 (2018).

[73] Andrew Kerman, “Four Spin Couplers for Quantum Information Processing,” US Patent:
10187065B2 (2019).

[74] Andrew Kerman, “Four Spin Couplers for Quantum Information Processing,” US Patent:
20190158098A1 (2019).

[75] M. Schöndorf and F. K. Wilhelm, “Non-pairwise interactions induced by virtual transitions,”

(2018), arXiv:1811.07683.

[76] Denis Melanson, Antonio J. Martinez, Salil Bedkihal, and Adrian Lupascu, “Tunable
three-body coupler for superconducting flux qubits,” (2019), arXiv:1909.02091.