Optimization of Quantum Discord to Connection with Non-commutivity

Chitradeep Gupta

Abstract
In a pair of correlated quantum systems, a measurement in one corresponds to a change in the state of the
other. In the process, information is lost. Measurement along which set of projectors would accompany minimum
loss in information content is the optimization problem of quantum discord. This optimization problem needs to
be addressed in any computation of discord and is also an important aspect of our understanding of any quantum
to classical transition. The zero discord condition is a useful tool to explore this, we exploit the currently known
standard to move to a stronger measure that addresses the correlated observables for an m x n density matrix.
In such a context one could show that discord minimizes at the diagonal basis of the reduced density matrices
and present an analytical expression for quantum discord.
’ ...instead it is your equations that should speak for you’
- Prof. Ravishankar , 1st presentation , June’17

1 Introduction

A seminal paper in 2001 by Ollivier and Zurech[1] established that quantum correlations could exist beyond en-
tanglement through what we today know as quantum discord. While shedding light on such new correlations, the
paper also establishes a new frontier in exploring classical to quantum transitions in terms of information content. A
quantum state upon measurement undergoes decoherence and loses information content that isn’t seen in classical
systems and this loss is being addressed as quantum discord[1]. The states achieved upon minimizing this loss are
called the most classical states or Pointer States[2]. Ever since its discovery quantum discord is being used for
quantum speed up over classical algorithms[3,4,5], entanglement distillation[6,7], quantum teleportation[5], super-
dense coding[5] and in many other areas of quantum computation. In all such applications however, researchers
had to compute discord through a fairly difficult[8] optimization process over all possible sets of measurement by
which it is defined. This computation is known to be NP complete[8] and the running time of any algorithm grows
exponentially with the size of the input. Analytical results[9,10,11] for simple low dimensional states are known
but the standard method of optimization with state parameters runs into increasing number of transcedental equa-
tions that are not easy to solve. Calculational difficulties aside, the optimization problem of quantum discord is
also deeply intertwined with our understanding of any form of departure from classicality. The problem asks us
to find the sets of measurement that shall accompany minimum information loss. A classical system should there-
fore correspond to zero discord. The states achieved upon minimization would tell us about the most classical states.

In this letter the author wishes to address the important open problem of minimizing discord in the product space
of two correlated systems with more emphasis given towards the resulting correlations in their observables. So we
are treating the problem not as one unknown but two where discord itself gets through a refinement in its mean-
ing and formalism. With such an intention, we improve the underlying equations of quantum discord to
move to a stronger measure, perform the required optimization for an in general m × n matrix with the help of
standard information inequalities and produce an analytical expression of discord that surprisingly turns
out to be symmetric at the point of minima even though for a general set of projections it isn’t. We look
into this emergent symmetry and its connection to classicality. This is followed by a comparative analysis of both
the measures and we demonstrate that their character is exactly similar for symmetric states. The new zero discord
condition helps us to establish a strong equivalence between discord and non-commutivity and shows how that this
a faithful measure of quantumness. The solution to the optimization problem also provides a clear identification
scheme for finding pointer states.

2 Background and Motivation

The formal development of quantum discord in the literature opens with the insight that mutual information I(S : A)
between two systems S and A that is symmetric in a classical setting fails to be so in a quantum context with
J(S : A){ΠA } where the set of projectors along which the measurement is being made needs to be specified.
j

I(S : A) = H(A) + H(S) − H(S, A)

J(S : A){ΠA } = H(S) − H(S|A){ΠA

j }
j

The difference between the two gives us discord δ(S : A){ΠA

j }
for a specific set of projectors that in turn speaks of

1
the information loss as the difference in the information content of the pre and post-measurement states. Discord
between S and A is then defined through an optimization over the entire projector space of A.
δ(S : A){ΠA
j }
= I(S : A) − J(S : A){ΠA }
j

= H(A) + H(S|A){ΠA
j }
− H(S, A)
δ(S : A) = min(δ(S : A){ΠA j }
)
P
The zero discord condition[1] is ρS,A = ρD = ΠA ρS,A ΠAj ; the equivalence of the pre-measurement and post-
PS,A j j Aj
measurement states. A ρS,A of the form j Pj ρS × Πj would satisfy such a condition and with a spectral decompo-
sition one could arrive at the following relation.
X
ρS,A = ΠSj,i ΠA A S
j ρS,A Πj Πj,i
i,j

Zero discord is equivalent to finding the set of measurements {ΠA

j , Πj,i } for which the density matrix is not disturbed
S

and no information is lost for any other observer unaware of the outcomes. However, the set of measurements {ΠSj,i }
in S for a given j is dependent on the outcome of the measurement in A. In other words, what outcome one gets
for A determines which observable one needs to measure in S. A more classical measure for correlations
should answer the question of how much the distribution of a certain particular observable in S changes based on
the outcome in A which leads us to define a mutual information of the form J(S : A){ΠA ,ΠS } , where
j i

J(S : A){ΠA ,ΠS } = H(S) − H(S|A){ΠA S

j ,Πi }
j i

D
H(S|A){ΠA S
j ,Πi }
is defined as H(ρ S|{ΠA
){ΠSi } . While there is gain in information about S upon conditioning with A by
j }

H(ρS )−H(ρS|{ΠA
j }
), there is loss in information about S upon projection along {ΠSi } by H(ρS|{ΠA
j }
)−H(ρD ) S .
S|{ΠA } {Πi }
j
J(S : A){ΠA ,ΠS } is then the sum of both, expressing the amount of information gained about a certain observable
j i
in S from a measurement of a certain other observable in A. The extra information that is lost upon projection
is exclusively a quantum phenomena which would go to zero in the classical case. This establishes the otherwise
classical equivalence of the two mutual informations J(S : A){ΠA ,ΠS } and I(S : A) which was the original starting
j i
point of Ollivier and Zurech in 2001. In such a setting, one could define a discord α(S : A) as follows,
α(S : A){ΠA S
j ,Πi }
= I(S : A) − J(S : A){ΠA ,ΠS }
j i

= H(A) + H(S|A){ΠA S
j ,Πi }
− H(S, A)
α(S : A) = min(α(S : A){ΠA S )
j ,Πi }

α(S : A){ΠA S
j ,Πi }
is the information loss when a series of successive measurements is made of two particular observables
in the respective Hilbert Spaces of A, and then in S. α(S : A) is defined by the optimization of that over the two
projector spaces. Notice that one of the intentions behind doing this is to move to a larger measure than the standard
version so that the minimization that follows stays relevant. From a strictly problem solving perspective, this is just
standard protocol since we do acknowledge that optimization of δ(S : A){ΠA j }
is difficult[8].

3 Minimization in Projector Space

Minimization of α depends on the minimization of H(S|A){ΠA S .

j ,Πi }
At this point we would like to introduce a
lemma, trying to understand what this term means explicitly. The Hilbert spaces of S and A are spanned by the
states {|ni >} and {|kj >}. The notation P (x, y) stands for the probability of some state |x > to collapse at |y >.

Lemma 1 H(S|A){ΠA ,ΠS } = H(ρD A − H(ρA ){ΠA } = D(P (SA, ni kj )||P (A, kj ))
D
S,A ){ΠS
i ×Πj } j
j i

Proof:
X
H(ρD ) S =
S|{ΠA } {Πi }
P (A, kj )H(ρD ) S
S|ΠA {Πi }
j j
j
X
=− P (A, kj )P (S, ni |A, kj )log(P (S, ni |A, kj ))
j,i
X
=− P (SA, ni kj )log(P (SA, ni kj )/P (A, kj ))
j,i

= H(ρD S A
D
S,A ){Πi ×Πj } − H(ρA ){Πj }
A

2
The next lemma helps us establish that any form of measurement or projection in the Hilbert space is always ac-
companied with a loss in information content. It is zero only when the projection is along the diagonal basis.

{Πj } ) ≥ H(ρ)
Lemma 2 H(ρD

Proof: {λi } are the eigenvalues of ρ with eigenstates {k̄i } while projection is taken along {kj }. We apply Jensen’s
inequality[14] to complete the proof.
X
H(ρD{Πj } ) = H( Πj λi Π̄i Πj )
j,i
X
= H( P (kj , k̄i )λi Πj )
j,i
XX X
=− ( P (kj , k̄i )λi )log( P (kj , k̄i )λi )
j i i
X
≥− P (kj , k̄i )λi logλi
j,i
X
≥− λi logλi = H(ρ)
i

Theorem 1 Minimization - α(S : A){ΠA S

j ,Πi }
for a separable density matrix ρS,A minimizes along the projectors of
the diagonal basis of its reduced density matrices.

Proof:PThe density matrix we choose to work with is a mixture of otherwise classical density matrices of the
form j Pj ΠSj × ΠA
j . The resultant matrix is of the form -
X
ρS,A = Pi Pji ΠSi,j × ΠA
i,j
i,j

In this case, the minimization of α by lemma 1 is just that of D(P (SA, ni0 kj 0 ||P (A, kj 0 )), as in D(Pi0 j 0 ||Pj 0 )). The
reduced density matrices have eigenbasis {Π̄A S
j 00 } and {Π̄i00 }.
X
Pi 0 j 0 = Pi Pji P (nij , ni0 )P (kij , kj 0 ) (1)
i,j
X X
ρA = Pi Pji ΠA
i,j = P̄j 00 Π̄A
j 00
i,j j 00
X
P̄j 00 = Pi Pji P (kij , k̄j 00 ) (2)
i,j
X
Pj 0 = Pi Pji P (kij , kj 0 ) (3)
i,j
X
Pj 0 = P̄j 00 P (k̄j 00 , kj 0 )
j 00
X
= Pi Pji P (kij , k̄j 00 )P (k̄j 00 , kj 0 ) (4)
i,j,j 00

We arrived at (4) using (2). Comparing (4) and (3) gives us the following equations that we shall use in differentiating
the states of the diagonal basis from the rest.
X
P (kij , kj 0 ) = P (kij , k̄j 00 )P (k̄j 00 , kj 0 ) (5)
j 00
X
P (nij , ni0 ) = P (nij , n̄i00 )P (n̄i00 , ni0 ) (6)
i00

It is important to notice here that the chain rule multiplication of probabilities in (5) and (6) followed by a summation
over the exclusive alternatives wouldn’t work for any other basis other than {|n̄i00 >} or {|k̄j 00 >} because of the
interference terms[15]. Equations (5) and (6) ask us to recognize that the states of S and A are essentially a classical
mixture only in their diagonal basis and no other. Now that we have all the required probabilities in order, let us
minimize.

3
We use equations (1) and (3) to find H(S|A){ΠA0 ,ΠS0 } , introduce (5) at the next step and employ log-sum inequality[16]
j i
with respect to the index {j 00 }. The j 0 -terms within the logarithm cancel out and a summation over {j 0 } gives us
the minimization at {Π̄A
j 00 }.

X Pi 0 j 0
H(S|A){ΠA0 ,ΠS0 } = − Pi0 j 0 log( )
j i Pj 0
i0 ,j 0

Pi Pji P (nij , ni0 )P (kij , kj 0 )

P
i,j
XX
=− ( Pi Pji P (nij , ni0 )P (kij , kj 0 ))log( P i
)
i,j Pi Pj P (kij , kj )
0
i0 ,j 0 i,j
i
P
i,j,j 00 Pi Pj P (nij , ni )P (kij , k̄j )P (k̄j , kj )
X X 0 00 00 0
i
=− ( Pi Pj P (nij , ni0 )P (kij , k̄j 00 )P (k̄j 00 , kj 0 ))log( P i
)
i0 ,j 0 i,j,j 00 i,j,j 00 Pi Pj P (kij , k̄j )P (k̄j , kj )
00 00 0

i
P
i,j Pi Pj P (nij , ni )P (kij , k̄j )P (k̄j , kj )
X X 0 00 00 0
≥− ( Pi Pji P (nij , ni0 )P (kij , k̄j 00 )P (k̄j 00 , kj 0 ))log( P i
)
i0 ,j 0 ,j 00 i,j i,j Pi Pj P (kij , k̄j 00 )P (k̄j 00 , kj 0 )
i
P
i,j Pi Pj P (nij , ni )P (kij , k̄j )
X X 0 00
i
≥− ( Pi Pj P (nij , ni0 )P (kij , k̄j 00 ))log( P i
)
i0 ,j 00 i,j i,j Pi Pj P (kij , k̄j 00 )

≥ H(S|A){Π̄A00 ,ΠS0 }
j i

We use equations (1) and (3) to find H(S|A){Π̄A00 ,ΠS0 } , introduce (6) at the next step and employ Jensen’s inequality
j i
for concave functions[14] with respect to the distribution {P (n̄i00 , ni0 )} with the varying index as {i00 }. A summation
over {i0 } then gives us the minimization at {Π̄Si00 } and completes the proof.
i
P
i,j Pi Pj P (nij , ni )P (kij , k̄j )
X X 0 00
i
H(S|A){Π̄A00 ,ΠS0 } = − ( Pi Pj P (nij , ni0 )P (kij , k̄j 00 ))log( P i
)
i,j Pi Pj P (kij , k̄j )
j i 00
i0 ,j 00 i,j
i
P
i,j,i00 Pi Pj P (nij , n̄i )P (n̄i , ni )P (kij , k̄j )
X X 00 00 0 00
i
=− ( Pi Pj P (nij , n̄i00 )P (n̄i00 , ni0 )P (kij , k̄j 00 ))log( P i
)
i,j Pi Pj P (kij , k̄j )
00
i0 ,j 00 i,j,i00
i
P
i,j Pi Pj P (nij , n̄i )P (kij , k̄j )
X X 00 00
i
≥− P (n̄i00 , ni0 )( Pi Pj P (nij , n̄i00 )P (kij , k̄j 00 ))log( P i
)
i,j Pi Pj P (kij , k̄j )
00
i0 ,j 00 ,i00 i,j
i
P
i,j Pi Pj P (nij , n̄i )P (kij , k̄j )
X X 00 00
i
≥− ( Pi Pj P (nij , n̄i00 )P (kij , k̄j 00 ))log( P i
)
i,j Pi Pj P (kij , k̄j )
00
i00 ,j 00 i,j

≥ H(S|A){Π̄A00 ,Π̄S00 }
j i

The most general m×n separable density matrix i Pi ρiS ×ρiA upon decomposition would give i,j,l Pi Pji Pli ΠSi,j ×ΠA
P P
i,l .
The extra index l is being ignored for simplification. The reader is still encouraged to check for herself that even in
the most general case the proof gets through.

Theorem 2 Symmetrization - α(S : A) is symmetric in S and A.

Proof: α(S : A) = α(S : A){Π̄A S

j ,Π̄i }
= H(A) + H(S|A){Π̄A S
j ,Π̄i }
− H(S, A)
D
= H(A) + H(ρS,A ){Π̄Si ×Π̄A j }
− H(ρA ){Π̄A
D
j }
− H(S, A)
= H(ρD ) S A
S,A {Π̄i ×Π̄j } − H(S, A)

The symmetry that broke down at J(S : A){ΠA S

j ,Πi }
is again restored in α(S : A) while giving us an analytical
expression of discord in equation (7).

α(S : A) = H(ρD
S,A ){Π̄S A − H(S, A)
i ×Π̄j }
(7)

This is the expression we had strived for, notice the projectors in equation (7) is a function of the given density
matrix. The symmetry that arises at the point of minima with the most classical states shows how the sequence of
measurements in the classical realm is inconsequential. Equation (7) along with lemma 2 helps us attain quite a few
key insights that we would like to discuss at length in the next section.

4
4 Implications

(i) The zero discord condition- α(S : A) = 0 only when the projectors {Π̄Si × Π̄A
j } becomes the diagonal basis of
the given density matrix, there by restricting it to the form i,j Pij ΠSi × ΠA
P
j with product eigenstates. In other
words, unless [ρS,A , N × K] = 0 for some N and K in the respective Hilbert spaces of S and A, information loss
through discord is inevitable. This shows that non-commutivity is the underlying cause of quantum discord.

(ii) The strongness condition- α(S : A) ≥ δ(S : A). This is a direct consequence of lemma 2 since α only dif-
fers from δ by an extra projection. The zero discord condition of α(S : A) in (i) assigns both the discords δ(S : A)
and δ(A : S) to be zero together. This shows that the new measure we defined is a stronger measure of
quantum discord.

(iii) Identification of Pointer States- The minimization of α shows that the diagonal basis of the reduced
density matrices are a strong candidate for Pointer States. This is in alignment with lemma 2 where we saw
that measurement along the diagonal basis doesn’t lead to any information loss and hence, are most classical. Equa-
tion (7) in that sense, talks about the closest one could get to such a scenario in the product space of two correlated
systems when the diagonal basis of the joint density matrix is essentially entangled with a nonzero discord.

5 Instances and Discussions

A comparative study of both the discords is being done as we see it play out for real systems.

α for the Werner states ρw = 41 (I + xσ~1 .σ~2 )

1+x 1 + x 1 − 3x 1 − 3x 1 − x 1−x
α(x) = log + log − log
4 4 4 4 2 4
The curve for α(x) that we have obtained behaves just
as δ(x) does in Werner states as demonstrated in [17].

α for the example in Zurech’01 ρz = 21 (|00 >< 00| + |11 >< 11|) + z2 (|00 >< 11| + |11 >< 00|)

1+z 1+z 1−z 1−z

α(z) = log + log + log2
2 2 2 2
The curve for α that we have obtained corresponds to the spine of the plot given in Zurech’01[1]. The variation
over the projector space by θ and how it minimizes at that spine follows from theorem 1. So it seems α and δ
agree fairly well. For asymmetric states in S and A, they of course shouldn’t, contributions from both δ(S : A)

5
and δ(A : S) play into α because of the symmetrization. While δ(S : A) was originally designed by Ollivier and
Zurek[1] to address the measurement problem from the perspective of information for a state S being measured by
an apparatus A, α(S : A) is designed to address the correlations between two states S and A.

6 In Closing

A stronger measure for discord is formalized to address the quantumness in correlations between two systems. The
optimization problem is then solved for an in general m × n matrix to produce an analytical expression of discord
that turns out to be symmetric. This helps us to identify the most classical states (pointer states) as the diagonal
basis of the reduced density matrices that when measured would disturb the systems the least with minimum loss in
information content. Notice that this proposition is in alignment with the symmetry that emerges upon optimization.
This being the synopsis of our main results, one can’t help but wonder about the exact scope of theorem 1. We
chose to work with a separable density matrix because it is known now that entanglement isn’t necessary for a
nonzero discord[1]. Having said that the proof heavily relies on the separability of ρS,A . If such a theorem cannot
be established for entangled states at all then one could devise a new separability criterion from it which from
where we stand, seems highly unlikely. At the end, the confidence in theorem 1 is reinforced via equation (7) which
establishes a strong equivalence between discord and non-commutivity. Breakdown of commutation in terms of
geometry becomes loss of information through quantum discord.

7 Acknowledgement

The author thanks Prof. V. Ravishankar for his helpful comments and for giving the student the time.

8 References

[1] Ollivier, H. and Zurek, W. (2001) ”Quantum Dis- cord for two-qubit X states” Phys. Rev. A, 81(4):042105
cord: A Measure of the Quantumness of Correlations”
Phys. Rev. Lett. 88, 017901 [10] Girolami, D. and Adesso, G. ”Quantum discord
for general two-qubit states: Analytical progress” Phys.
[2] Zurek, W. (2003). ”Decoherence, einselection and Rev. A 83, 052108
the quantum origins of the classical”. Reviews of Mod-
ern Physics. 75 (3): 715–775 [11] Vinjanampathy, S. and Rau, A. (2012) ”Quantum
discord for qubit–qudit systems” J. of Phys. A: Mathe-
[3] Datta, A. Shaji, A and Caves,C. (2008) ”Quantum matical and Theoretical
discord and the power of one qubit” Phys. Rev. Lett.,
100(5):050502 [12] Lu, X. Ma, J. Xi, Z. and Wang, X. (2011) ”Op-
timal measurements to access classical correlations of
[4] Henderson, L. and Vedral, V. (2001) ”Classical quan- two-qubit states” Phys. Rev. A, 83(1):012327
tum and total correlations” J. Phys. A: Math. Gen.,
34(35):6899 [13] Maziero, J. Celeri, L. and Serra, R. (2010) ”Suit-
ability of symmetric and asymmetric versions of the
[5] Madhok, V. and Datta, A. (2013) ”Quantum dis- quantum discord” arXiv:1004.2082[quant-ph]
cord as a resource in quantum communication” Int. J.
Mod. Phys. B, 27(1–3):1345041 [14] Cover, T. and Thomas, J. (1991) ”Elements of In-
formation Theory” Chapter-16, Theorem-16.1.2
[6] Streltsov, A. Kampermann, H. and Bruss, D. ( 2011)
”Linking quantum discord to entanglement in a measure- [15] Feynman, R. (1948) ”Space-Time Approach to Non-
ment” Phys. Rev. Lett., 106(16):160401 Relativistic Quantum Mechanics” Rev. Mod. Phys. 20,
367
[7] Devetak, I. and Winter, A. (2004) ”Distilling com-
mon randomness from bipartite quantum states” IEEE [16] Cover, T. and Thomas, J. (1991) ”Elements of In-
Trans. Inf. Theory, 50(12):3183–3196 formation Theory” Chapter-16, Theorem-16.1.1

[8] Huang, Y. (2014) ”Computing quantum discord is [17] Dı́az-Solórzano, S. and Castro, E. (2018) ”Exact ana-
NP-complete” New Journal of Physics 16, 033027 lytical solution of Entanglement of Formation and Quan-
tum Discord for Werner state and Generalized Werner-
[9] Ali, M. Rau, A. and Alber, G. (2010) ”Quantum dis- Like states” arXiv:1802.05877 [quant-ph], Figure-2