Centralised multi link measurement compression with side

information
Sayantan Chakraborty‡ Arun Padakandla∗ * Pranab Sen‡ †

Abstract
arXiv:2203.16157v1 [quant-ph] 30 Mar 2022

We prove new one shot achievability results for measurement compression of quantum instruments with side
information at the receiver. Unlike previous one shot results for this problem, our one shot bounds are nearly
optimal and do not need catalytic randomness. In fact, we state a more general problem called centralised multi
link measurement compression with quantum side information and provide one shot achievability results for it.
As a simple corollary, we obtain one shot measurement compression results for quantum instruments with side
information that we mentioned earlier. All our one shot results lead to the standard results for this problem in the
asymptotic iid setting. We prove our achievability bounds by first proving a novel sequential classical quantum
multipartite covering lemma, which should be of independent interest.

1 Introduction
In order to obtain important statistical information about quantum systems, an experimenter must perform measure-
ments on the system. This information can be used in subsequent physical operations on the system. However, the
measurement process is inherently noisy. This noise could arise due to extrinsic noise in the measurement procedure
itself (which is uncorrelated to the state) and/or intrinsic noise arising from the uncertainty introduced by the quantum
state being measured. An important information theoretic problem is to filter out the extrinsic noise so that we can
compress the number of bits required to describe the measurement outcome.
We model the measurement procedure using the positive operator valued measure (POVM) formalism. Consider
the following setting: The experimenter Alice possesses a quantum state ρ and a POVM Λ. The action of the POVM
on the quantum state is to produce a classical register (in Alice’s possession) which contains an index corresponding
to the measurement outcome, as well as a post measurement quantum state (to which Alice does not have access). The
task is for Alice to send the contents of her classical register to Bob, such that the correlations between this register
and the post measurement state are conserved, using as few bits as possible. This problem is formally known as
measurement compression.
In his seminal paper, Winter [21] observed that the above task can be achieved if one can construct an approximate
decomposition of the POVM of the following form :
ε X
Λ≈ PCOIN (c)Λ(c)
c

where each Λ(c) is a POVM with fewer possible classical outcomes as compared to Λ. The variable c corresponds to
the outcome of a coin toss which is distributed according to PCOIN . As long as Alice and Bob both have access to the
coin c, Alice only needs to send enough bits to be able to describe the set of possible outcomes of Λ(c). Note that this
distribution is independent of the state ρ, and thus is a source of extrinsic noise. The measurement procedure can then
be summarised as follows :
1. Toss a coin ∼ PCOIN . Let the outcome be c.
*∗Department of Electrical Engineering and Computer Science, University of Tennessee at Knoxville, U.S.A. Email: arunpr@utk.edu.
†‡
School of Technology and Computer Science, Tata Institute of Fundamental Research, Mumbai, India. Email:
{kingsbandz,pranab.sen.73}@gmail.com. Sayantan is partially supported by a Google India late Ph.D. fellowship. Sayantan and Pranab
are supported by the Department of Atomic Energy, Government of India, under project no. RTI4001.

1
 2. Measure ρ with Λ(c)
Let the number of bits needed to describe the state of the coin be C and the number of bits of communication
from Alice to Bob be R. Winter characterised the region of all (R, C) pairs for which the above decomposition exists.
A cleaner exposition of Winter’s results can be found in a later paper by Wilde et al. [20], who also extend Winter’s
original contribution by considering the case when Bob has access to some part of the quantum state being measured.
This protocol is known as Measurement Compression with Quantum Side Information or MC-QSI in short. In this
scenario Alice and Bob share the quantum state ρAB , where Alice possesses the system A and bob possesses the
system B. Alice then measures her share of the state and sends her classical register to Bob. Wilde et al. showed that
in this setting, Bob can use his share of the quantum state to reduce the number of bits Alice needs to send to him even
further.
All of the above works consider the measurement compression problem in the asymptotic iid regime i.e. in the
limit of a large number of repeated identical measurements on several copies of the original quantum state. A more
general variant of the problem would be if only one copy of the state was available for measurement. This is the
so called one-shot regime. In a recent paper, Anshu, Jain and Warsi [3] provide an achievable region for MC-QSI
in this regime. In contrast to all previous works, that paper shows a reduction of the MC-QSI problem to the well
studied problem of quantum message compression with side information [2], which allows them to leverage known
tools for the latter problem. However, even though Anshu et al.’s results imply the best known asymptotic iid bounds
for measurement compression, their one-shot bounds are unsatisfactory, due to the following issues:
1. The protocol uses shared randomness in a catalytic manner i.e. it needs some a large amount of shared random-
ness to begin with but regenerates part of this randomness at the end of the protocol. In the asymptotic iid limit
with n copies, this protocol is applied iteratively n/m times to blocks of states containing m copies each, where
m < n. By choosing m appropriately, the authors could show that that the amount of extra initial randomness
required is o(n). Thus, the amount of extra initial shared randomness per iteration goes to zero in the asymptotic
iid limit. Nevertheless in the one shot scenario the extra initial shared randomness can potentially be very large,
almost as large as the alphabet size;

2. The second and perhaps more serious issue is that the bounds achieved by Anshu et al. are optimal only in the
case when the probability distribution induced by the original POVM is uniform on the alphabet X. In all other
cases, there is a trade-off between the amount of shared randomness required to run the protocol and the extent
to which Alice can compress her message to Bob.
In this paper we define an even more general problem called centralised multi-link measurement compression with
side information and prove a one shot achievability result for it. The problem that we discussed above will be called
the point to point message compression problem. As corollaries of our result for centralised multi link measurement
compression, we obtain
1. New one shot achievability result for point to point measurement compression with side information in both
feedback as well as non-feedback cases;

2. Our protocol does not require any extra initial shared randomness to begin with;

3. It recovers the optimal bounds of Wilde et al. in the asymptotic iid limit;

4. It achieves the natural one-shot bounds that one expects for this problem, for any probability distribution on the
set of outcomes, without making any compromises between the amount of shared randomness required and the
rate of compression.
To motivate this problem we consider a natural generalisation of the measurement compression, to the case of
quantum instruments. Instruments are the most general model for quantum measurements, which include both a
classical output and a post measurement quantum state [6, 7]. A quantum instrument is a CPTP map NINSTR of the
following form: X
NINSTR (ρ) := |xi hx|X ⊗ Nx (ρ)
x

2
where each Nx is a completely positive trace non-increasing map with Kraus decompositions as follows
X
†
Nx (ρ) = Nx,y ρ Nx,y
y

and
X
†
Nx,y ρ Nx,y ≤ I
y

One can then ask the question whether we can prove a compression theorem for quantum instruments as well. This
theorem was claimed without proof in [9] and proven rigorously in [20]. Wilde et al. solve this problem by suitably
reducing it to an instance of the original measurement compression problem. Their approach at simulating the action
of NINSTR is to identify its operation as a tracing out of the Y −register of the state
X
|xi hx|X ⊗ |yi hy|Y ⊗ Nx,y ρ Nx,y †
.
x,y

†
Naturally, they design a POVM simulation protocol wherein Alice simulates the POVM {Nx,y Nx,y }x,y . Recognizing
that Bob only needs to recover the X−register they propose Alice discards her simulation of the Y −register and
thereby achieve rates that one would naturally expect.
We now define our new problem called the centralised multilink measurement compression problem with side
information. Taking a cue from the discussion above we consider a POVM ΛXY = {Λx,y } which outputs two
classical symbols x and y according to some joint distribution which depends on the state ρ. There exist two separate
noiseless channels, called X and Y channels, from Alice to Bob, and two independent public coin registers, called
X and Y public coins, between them. We use the word link to refer to a noiseless channel. During the protocol at
most one of the links may be turned OFF by an adversary without Alice or Bob’s knowledge. We ask whether we can
design a single simulation protocol that enables Bob to recover either X or Y or both X, Y depending on which link
is ON. See Fig 1.
To be precise, we seek one simulation protocol that enables Bob to recover (i) the X register when only the X-link
of rate R1 is active and they share C1 bits of randomness, (ii) the Y −register when only the Y -link of rate R2 is active
and they share C2 bits of randomness, and (iii) recover both X and Y registers when both links are active and they
share a total of C1 + C2 random bits. We require that Alice’s and Bob’s strategies should be agnostic to which links
are operational i.e. their encoding and decoding strategies should continue to work even if one link fails.

Figure 1: Fig.1

A careful consideration of this problem reveals the following challenge. The requirement that each link enable
recovery of the corresponding classical register precludes the use of designing conditional codebooks for the POVM
simulation protocol. Otherwise, if the codebook for Y were to be conditionally dependant on the codebook for X,

3
Bob would be unable to recover Y when the link of rate R1 is inactive, and vice-versa. Therefore, this strategy, which
works well for the purposes of Wilde et al. , fails in our case.
On the other hand, if the codebooks have to be designed independently, then it is unclear how Bob would be
able to recover both X, Y when both links are active. We overcome this issue by proving a novel quantum covering
lemma, referred to herein as measure transformed sequential covering lemma. This lemma is one of the main technical
contributions of our work. The issue in proving a covering lemma of this sort is that most arguments invariably run
into the issue of simultaneous smoothing, which is an outstanding open problem in quantum information theory [10].
Recently however, Chakraborty, Nema and Sen [5] showed how one can get around the smoothing problem in certain
situations. We adapt their construction and generalise it for our purposes. In fact, the proof of the covering lemma
presented in this work is much easier and far more general than that in [5], and it may be useful in other applications
as well.
Before stating our one shot achievability results for centralised measurement compression, we need to recall the
concept of rate splitting in the quantum setting [4]. The joint probability distribution on the measurement outcomes
(x, y) of the given POVM Λ is defined by p(x, y) := Tr[((Λx,y ⊗ IBR )(ρ))ABR ], which in turn defines p(x) and
p(y|x) in the natural fashion. Now, the splitting function creates two new random variables U and V supported on
alphabet X , defines x := max{u, v}, and splits the probability distribution p(x) into a joint distribution pθ (u, v)
parametrised by a real number θ ∈ [0, 1]. This splitting function satisfies the following properties for any θ ∈ [0, 1]:

1. U θ , V θ are independent random variables;

2. The probability distribution of the random variable max{U θ , V θ } is the same as that of the random variable X;
3. When θ = 0, U 0 = X and V 0 is constant. When θ = 1, V 1 = X and U 1 is constant.

We prove the following theorem for one shot achievability results for centralised measurement compression. The
ε (·), I ε (· : ·), I ε (· : ·), the definitions
theorem is stated in terms of smooth one shot entropic quantities like Hmax max H
of which can be found in the full version of this paper. They are the natural one shot analogues of Shannon entropic
quantities like entropy and mutual information.
Theorem 1.1. (Centralised multi link measurement compression) Consider the ε-error centralised multi-link
measurement compression problem with side information at Bob in the feedback case. One achievable rate region is
obtained as the union over a parameter θ ∈ [0, 1] of the regions Sθ defined by:

RX = RU + RV
ε ε
RU > Imax (U : RB) − IH (U : B)
+ O(log ε−1 )
ε ε
RY > Imax (Y : RBU ) − IH (Y : B)
+ O(log ε−1 )
Sθ : ε ε
RV > Imax (V : RBU Y ) − IH (V : B)
+ O(log ε−1 )
CX = CU + CV
ε ε
CU +RU > Hmax (U ) − IH (U : B)
ε ε
CY +RY > Hmax (Y ) − IH (Y : B)
ε ε
CV +RV > Hmax (V ) − IH (V : B),
where the entropic quantities are calculated for the control state
X
pθ (u)pθ (v)p(y|u, v) |u, v, yi hu, v, y|U V Y
(u,v,y)∈X ×X ×Y

((Λmax{u,v},y ⊗ IBR )(ρ))ABR

⊗ .
Tr[((Λmax{u,v},y ⊗ IBR )(ρ))ABR ]

4
 above state is obtained by splitting random variable X into independent random variables U , V in the state
TheP
|x, yi hx, y|XY ⊗ ((Λx,y ⊗ IBR )(ρ))ABR according to the parameter θ. Another achievability region is
(x,y)∈X ×Y
obtained by rate splitting Y instead of X. The total achievable region is the union of the two regions. The encoding
and decoding strategies are agnostic to which links are actually functioning.
Our general theorem above lends itself to two important corollaries which are stated below.
Corollary 1.2. In the asymptotic iid setting of the centralised measurement compression problem, the following rate
region per channel use is achievable:

RX > I(X : BR) − I(X : B)

RY > I(Y : BR) − I(Y : B)
RX + RY > I(XY : BR) + I(X : Y )
−I(X : B) − I(Y : B)
RX + CX > H(X) − I(X : B)
RY + CY > H(Y ) − I(Y : B),

|x, yi hx, y|XY ⊗ ((Λx,y ⊗ IBR )(ρ))ABR .

P
where all entropic quantities are computed with respect to
(x,y)∈X ×Y

Corollary 1.3. For the point to point measurement compression problem with side information for quantum instru-
ments , an achievable region can be obtained by assuming that the Y links is not functional. In the one shot setting we
get
ε
RX > Imax ε
(X : RB) − IH (X : B) + O(log ε−1 )

and
ε ε
RX + CX > Hmax (X) − IH (X : B).

In the asymptotic iid setting this reduces to

RX > I(X : RB) − I(X : B) + O(log ε−1 )

and

RX + CX > H(X) − I(X : B).

|x, yi hx, y|XY ⊗ ((Λx,y ⊗ IBR )(ρ))ABR .

P
Above, all entropic quantities are computed with respect to
(x,y)∈X ×Y

1.1 Measurement Compression with Quantum Side Information

In this section we provide a brief exposition of the ideas which allow us to prove Corollary 1.3. The techniques
developed in this section allow us to design a measurement compression protocol in which the receiver Bob can use
quantum side information to reduce the umber of bits that Alice needs to send him. The setup is as follows :
Alice and Bob share a joint quantum state ρAB , where Alice has access to system A and Bob has access to system
B. Alice is given a POVM ΛA = ΛA

x , where each POVM element Λx acts on the system A and corresponds to a
classical outcome x ∈ X . We define the true post measurement state

σ RBX X̂ := TrA IRB ⊗ ΛA ϕRAB



X
= |x, xi hx, x|X X̂ ⊗
x∈X
IRB ⊗ ΛA

RAB 
TrA x ϕ

5
We allow Alice and Bob access to a noiseless forward classical channel and also to public coins. The system X̂ is
used by Bob to store his guess for the outcome of the measurement. In the case of the true post measurement state,
one can imagine that Alice sent Bob the full description of the outcome through the classical channel, which Bob then
stored in X̂. Now suppose that Alice and Bob use a protocol P which allows them to reduce the number of bits which
need to be communicated to Bob. Let the post measurement state created by P be σ̃ RBX X̂ . We require that for any
such protocol,
kσ − σ̃k1 ≤ ε
The first idea is that Alice designs a class of POVMs {θk }, where each POVM θk is indexed by the setting of the
public coin k ∈ [2C ]. Each θk contains fewer outcomes than the original Λ which allows Alice to save on the number
of random bits that she needs to describe the outcome. In particular, each θk acts on the system A and produces as
output an index in the set [2R+B ] i.e.
θk := θk (`) | ` ∈ [2R+B ]


. Winter’s original construction of these POVMs relied on the Operator Chernoff [1] bound to show the existence of
these POVMs. Unfortunately, the Operator Chernoff bound is not easily adapted to the one shot setting. Instead we
use a new one shot covering lemma (Section 3.1) along with a one shot operator inequality (Section 3.2) to show the
existence of the compressed POVM. These two techniques are new and are powerful enough to allow us to emulate the
Operator Chernoff bound even in the one shot setting. We believe that these two new tools maybe useful elsewhere.
We note that recently Padakandla [13] provided another alternate way of designing the compressed POVM in the one
shot setting.
The second idea is that after Alice applies a POVM θk and observes some outcome `, she hashes the outcome into
a bit string by using a hash function f : [2R+B ] → [2R ], which is chosen uniformly at random from a 2-universal hash
family. This hash value is what Alice sends to Bob. Suppose that the hash value sent by Alice is m.
Upon receiving Alice’s bit string, Bob chooses a POVM γ(m, k), based on Alice’s message and the setting of the
public coin. Bob will use this POVM to measure his system B to find the index which is consistent with the outcome
`. The idea is that as long as the set f −1 (m) is not too large, Bob should be able to find the correct measurement
outcome with a high probability of success.
The above protocol is essentially a classical data compression with quantum side information protocol (CDC-
QSI), which has been studied in the asymptotic iid setting by Devetak and Winter [8] and Renes [14] and in the
one shot setting by Renes and Renner [15]. However, it is not immediately obvious how one can adapt Renes and
Rener’s protocol to our setting. Instead, we design a new one shot protocol for CDC-QSI, which is more suitable for
our setting. The one shot rates our new protocol achieves are different from those of Renes and Renner, but recover
to the same asymptotic iid bounds achieved by them. The upshot of our protocol is that it is composable with the
measurement compression protocol, which is in general not true. The formal statement is as follows :

Lemma 1.4. Classical Message Compression with Quantum Side Information

Given a classical quantum state of the form
X
σ XB := PX (x) |xi hx|X ⊗ σ B
x

where Alice possesses the system X and Bob possess the system B, there exists a CDC-QSI protocol with probability
of error at most ε and the rate of communication RA from Alice to Bob is at least
 
ε ε 1
RA ≥ Hmax (X) − IH (X : B) + O log
ε

where all entropic quantities are computed in terms of the state σ XB .

1.2 Centralised Measurement Compression

In this section we briefly explain our strategy to design a protocol for centralised multi link measurement compression.
Due to the composability of our CDC-QSI lemma with any measurement compression scheme, we combine these two

6
protocols to achieve the bounds claimed in Theorem 1.1. The technique that we use to design the compressed POVMs
θk first requires the creation of a random codebook C := {X(k, `)} sampled iid from the distribution PX (x) :=
Tr[ρΛx ]. We further require that, on expectation over choices of codebook, the following sample average state is
close to X 1 1 √ √ ε
ρΛx(k,`) ρ ≈ ρ
L PX (x(k, l))
`∈[L]

In the centralised case however, since we are forced to construct the two codebooks for the X and Y links indepen-
dently from the distributions PX and PY , the sample average matrices that we must consider are of the form
1 X 1 √ √
ρλx(k1 ,`1 ) ρ
L1 · L2 PX (x(k1 , `1 ))PY (y(k2 , `2 )) y(k2 ,`2 )
`1 ,`2

Even though the above matrix may not even be a quantum state, we still require that in expectation, it should be close
to ρ. We show this using our measure transformed sequential quantum covering lemma :
Lemma 1.5. Measure Transformed Sequential Covering Lemma
Suppose we are given a joint distribution PXY on classical alphabets X ⊗ Y, with marginals PX and PY . Suppose
we are also given the following quantum state:
X X Y
ρXY E := PXY (x, y) |xi hx| ⊗ |yi hy| ⊗ ρx,y

Let {x(1), x(2), . . . , x(K)} and {y(1), y(2), . . . , y(L)} be iid samples from the distribution PX ⊗ PY . Then


1 X PXY (x(k), y(l)) E

E ρ x(k),y(l) − σ ≤ε
x(1),x(2),...,x(K) K · L PX (x(k)) · PY (y(l))

k,l
y(1),y(2),...,y(L) 1

whenever
ε ε
log K > Imax (X : E) and log L > Imax (Y : XE)

where X
σ E := PXY (x, y)ρE
x,y
x,y

1.3 Organisation
The paper is organised as follows: In Section 2 we define the one shot quantities that we will need for the proof,
and state some useful facts. In Section 3 we present the main technical tools that we will use in the proof of our
measurement compression theorem, including the measure transformed sequential covering lemma Section 3.1, a
one shot operator inequality that arises from this covering lemma Section 3.2 and the one shot classical message
compression protocol with quantum side information Section 3.3. Then in Section 4 we prove a version of our
centralised multi link measurement compression theorem in the case when there is no side information available to the
receiver Bob. In Section 4.2, we show how one can compose a measurement compression theorem with the protocol
for classical message compression with quantum side information. Finally, we prove our full centralised multi link
measurement compression theorem with side information in Section 4.3. In Section 5 we show how our one-shot
bounds can be extended to the desired rate region in the asymptotic iid setting.

2 One Shot Entropic Quantities

In this section we present the important one shot entropic quantities that we will use throughout the paper, and some
useful facts regarding these quantities.

7
2.1 Smooth Max Entropy
Definition 2.1. Smooth Max Entropy
Given a distribution PX on the alphabet X and some error parameter ε > 0, the ε-smooth max entropy or just smooth
ε (X) is defined as
max entropy of PX , denoted by Hmax
X
ε
Hmax (X) := log λ∗ (x)
x

where vector λ∗ (indexed by x) is the optimiser for the following linear program :
X
min λ(x)
x
X
PX (x)·λ(x) ≥ 1 − ε
x
(1)
λ(x) ≤ 1
λ(x) ≥ 0

Fact 2.2. Given the distribution PX , consider an ordering x0 , x1 , . . . x|X | on the alphabet X such that
xi < xj =⇒ PX (xi ) ≤ PX (xj )
Define the set S ⊂ X such that ( )
X
S := xi | PX (xi ) ≤ ε ∧ xi ≤ xi+1
xi
Then, the linear program in Eq. (1) is optimised by a vector λ∗ such that
λ∗ (x) = 0 x∈S
∈ (0, 1) x = x|S|+1
=1 otherwise
Definition 2.3. Optimising Sub-distribution
Given the setup of Fact 2.2, we define the sub-distribution PX0 as
PX0 (x) = PX (x) x ∈ Sc
=0 otherwise
Fact 2.4. Given the setup of Fact 2.2 and the sub-distribution P 0 ,
0
P − P ≤ ε
1
and for all x in the support of P 0 (x),
1
PX (x) ≥ ε · (1 − ε)
2Hmax (X)

2.2 Hypothesis Testing Relative Entropy

Definition 2.5. Smooth Hypothesis Testing Relative Entropy Given quantum state ρA and σ A , the smooth hypoth-
esis testing relative entropy between ρ and σ is defines as
ε
DH (ρA || σ A ) := − log OPT
where OPT is the optimum attained for the following semi-definite program
max Tr[Πσ]
Tr[Πρ] ≥ 1 − ε
Π≥0
Π≤I

8
Definition 2.6. Hypothesis Testing Mutual Information Given the quantum state ρAB , the hypothesis testing mutual
information IHε (A : B) is defined as

ε ε
IH (A : B)ρ := DH (ρAB || ρA ⊗ ρB )

Fact 2.7. Given a classical quantum state

X
PX (x) |xi hx|X ⊗ ρB
x
x

ε (X : B) is of the form
the operator ΠOPT which is the optimiser for IH ρ
X
ΠOPT = |xi hx|X ⊗ ΠBx
x

where for each x,

0 ≤ Πx ≤ I

2.3 Smooth Max Relative Entropy

Definition 2.8. Max Relative Entropy Given a quantum states ρA and σ A , the max relative entropy Dmax (ρA || σ A )
is defined as n o
Dmax (ρA ||σ A ) := inf λ | ρA ≤ 2λ σ A

Definition 2.9. Purified Distance Given two quantum states ρA and σ A , the purified distance P (ρ, σ) between the
two states is defined as p
P (ρ, σ) := 1 − F 2 (ρ, σ)
where F (· | ·) is the fidelity.

Definition 2.10. ε-Ball Given a quantum state ρA , the ε− ball around this state is defined as

B ε (ρA ) := ρ0A | P (ρ, ρ0 ) ≤ ε



Definition 2.11. Smooth Max Relative Entropy Given two quantum states ρA and σ A , the smooth max relative
ε
entropy Dmax (ρA || σ A ) is defined as
ε
Dmax (ρA || σ A ) := inf Dmax (ρ0A || σ A )
ρ0A ∈B ε (ρA )

Definition 2.12. Smooth Max Information Given a state ρAB the smooth max information Imax
ε (A : B) is given by
ρ

ε ε
Imax (A : B)ρ := Dmax (ρAB || ρA ⊗ ρB )

We will need a slightly perturbed version of this quantity to state our theorems. We call this the tilde smooth max
information.

Definition 2.13. Tilde Smooth Max Information Given a state ρAB the tilde smooth max information between A
and B is defined as
I˜max
ε
(A : B)ρ := inf Dmax (ρ0AB || ρA ⊗ ρ0B )
ρ0AB :P (ρAB ,ρ0AB )≤ε

9
2.4 Quantum Asymptotic Equipartition Property
We will find the following facts useful while extending our results to the asymptotic iid setting. We collectively
refer to them as the Quantum Asymptotic Equipartition Property (QAEP). The proofs of these facts can be found in
[17, 18, 12, 16].

Fact 2.14. Given a classical probability distribution PX and some integer n ∈ N, the following holds
1 ε
lim lim Hmax (X) = H(X)
ε→0 n→∞ n

Fact 2.15. Asymptotic Equipartition Property Given the quantum states ρA and σ A , and some integer n ∈ N, the
following hold:
1 ε ⊗n
lim lim DH (ρ || σ ⊗n ) = D(ρ || σ)
ε→0 n→∞ n
1 ε
lim lim Dmax (ρ⊗n || σ ⊗n ) = D(ρ || σ)
ε→0 n→∞ n

where D(ρ || σ) is the quantum von Neumann relative entropy.

2.5 Other Useful Facts

Fact 2.16. [19] For any state ρAB , ε > 0 and 0 < γ < ε, the following is true
3
I˜max
ε ε−γ
(A : B)ρ ≤ Imax (A : B)ρ + log 2
γ
Fact 2.17. [2, 19]Smoothed Convex Split Lemma
Let ρAB be any quantum state and let τ A1 ...AK B be the so called convex split state, defined as
1 X Ak B O A[K]\k
τ A1 ...AK B := ρ ρ
K
k∈[K]

Then for all

√ 1
log K ≥ I˜max
ε−η
(A : B)ρ + 2 log
η
the following holds true
A1 ...AK B
√
− ρB ρA[K] ≤ 2 · 2 ε − η
O

τ

1
√
where ε ∈ (0, 1) and η ∈ (0, ε).

3 Technical Tools
In this section we will prove the two main technical tools we will need to prove our centralised multi link measurement
compression theorem:

1. the measure transformed sequential covering lemma,

2. a new operator inequality implied by the covering lemma, and

3. the protocol for classical message compression with quantum side information.

10
3.1 A Measure Transformed Sequential Covering Lemma
In this section we will prove Lemma 1.5. To do this we will first need to prove the following sequential convex split
lemma:

Theorem 3.1. Successive Cancellation Convex Split Suppose we are given the tripartite state ρABR . Define the
convex split state as follows
1 X
τ A1 ...AK B1 ...BL R := ρAk B` R ρA[K]\k ρB[L]\`
O O
K ·L
k,`∈[K],[L]

Then whenever
√
ε
1
log K ≥ Imax
4
(A : BR)ρ + O(log )
√
ε
ε
1
log L ≥ Imax
4
(B : R)ρ + O(log )
ε
the following holds

A1 ...AK B1 ...BL R R
O
Ai
O
Bj √
τ − ρ ρ ρ ≤6 ε
i∈[K] j∈[L]

1

Proof. Consider the expression

A1 ...AK B1 ...BL R
− ρR ρAi ρBj
O O
τ (1)

i∈[K] j∈[L]

1


1 X
B` R
O
B[L]\`
O
Ai
1 X
B` R
O
B[L]\`
O
Ai R
O
Ai
O
Bj

≤ τ −
ρ ρ ρ + ρ ρ ρ −ρ ρ ω (2)
L i∈[K] L `∈[L] i∈[K] i∈[K] j∈[L]
`∈[L]
1

1

Let us first consider the first term in the RHS. Note that we can write this term as follows


τ − 1
X
B` R B[L]\` Ai
O O
ρ ρ ρ
L j6=` i∈[K]
`∈[L]
1


1 X 1 X
Ak B ` R
O
A[K]\k
O
B[L]\` B` R
O
B[L]\`
O
Ai

≤ ρ ρ ρ −ρ ρ ρ
L K

i∈[K]
`∈[L] k∈[K]
1

Taking the systems B1 . . . B`−1 B`+1 . . . BL systems from each of the corresponding norm expressions outside the
norm, we see that RHS


1 X
1
X
Ak B ` R
O
A[K]\k B` R
O
Ai

= ρ ρ − ρ ρ
L K

i∈[K]

`∈[L] k∈[K]
1

By the convex split lemma, for

√ 1
log K ≥ I˜max
ε−η
(A : BR)ρ + 2 log
η
we see that, for all ` ∈ [L], the terms

√

1 X A B RO A B` R
O
Ai


K ρ k `
ρ [K]\k − ρ ρ ≤ 2 · (2 ε − η)
k∈[K] i∈[K]
1

11
 √
where 0 < η < ε. We will now consider the second term in the RHS of Eq. (1). Notice that,

X
1 B` R
O
Bj
O
Ak R
O
Ai
O
Bj


L ρ ρ ρ − ρ ρ ρ
`∈[L] j6=` i∈[K] i∈[K] j∈[L]
1
X
1 B` R
O
Bj R
O
Bj

= L ρ ρ − ρ ρ
`∈[L] j6=` j∈[L]
1

Then by another application of the convex split lemma, for all

√ 1
log L ≥ I˜max
ε−η
(B : R)ρ + 2 log
η
the following holds
√
X
1 B` R
O
Bj R
O
Bj


L ρ ρ − ρ ρ ≤ 2 · (2 ε − η)
`∈[L] j6=` j∈[L]
1
Collating all these arguments we see that


A1 ...AK B1 ...BL R
√
− ρR ρAi ρBj ≤ 4 · 2 ε − η)
O O
τ

i∈[K] j∈[L]

1

Finally, invoking Fact 2.16 we see that the above bound holds true if we set
√ 1 3
( ε−(η+γ))
log K ≥ Imax (A : BR)ρ + 2 log + log 2
η γ
√ 1 3
( ε−(η+γ))
log L ≥ Imax (B : R)ρ + 2 log + log 2
η γ
√
whenever γ ∈ (0, ε − η). To simplify the bounds we set
√
ε
η=
√2
ε
γ=
4
which gives us the bounds
√
ε
1
log K ≥ Imax
4
(A : BR)ρ + O(log )
√
ε
ε
1
log L ≥ Imax
4
(B : R)ρ + O(log )
ε
√
with a corresponding error of 6 ε in the 1-norm. This concludes the proof.

We will now use Theorem 3.1 to prove Lemma 1.5 as a corollary. We restate the lemma below for convenience.

Corollary 3.2. Measure Transformed Sequential Covering Lemma

Suppose we are given a joint distribution PXY on classical alphabets X ⊗ Y, with marginals PX and PY . Suppose
we are also given the following quantum state:
X
ρXY E := PXY (x, y) |xi hx|X ⊗ |yi hy|Y ⊗ ρx,y

12
Let {x(1), x(2), . . . , x(K)} and {y(1), y(2), . . . , y(L)} be iid samples from the distribution PX ⊗ PY . Then


1 X PXY (x(k), y(l)) E

E ρx(k),y(l) − σ ≤ε
x(1),x(2),...,x(K) K · L PX (x(k) · PY (y(l)))

k,l
y(1),y(2),...,y(L) 1

whenever
ε
log K > Imax (X : E)
ε
log L > Imax (Y : XE)

where X
σ E := PXY (x, y)ρE
x,y
x,y

Proof. We apply Theorem 3.1 after instantiating terms appropriately. For our case, the classical systems X and Y
play the roles of the systems A and B in Theorem 3.1. Then, recall that
1 X
τ X1 ...XK Y1 ...YL R := ρXk Y` R ρX[K]\k ρY[L]\`
O O
K ·L
k,`∈[K],[L]

where
X
ρX := PX (x) |xi hx|X
x
X
Y
ρ := PY (y) |yi hy|Y
y

Note that the states ρX and ρY are marginals of the control state ρXY E . Let us consider one term in the expansion of
τ , for some fixed k ∈ [K] and ` ∈ [L] O O
ρXk Y` R ρX[K]\k ρY[L]\` (1)
To describe this fixed term we use the following convention : we write xi to denote a sample from the system Xi , and
similarly for yj . Let us fix a certain setting of these samples
Y Y
|xi i hxi |Xi ⊗ |yj i hyj |Yj
i∈[K] j∈[L]

Since the Xk and the Y` systems are entangled with the R system for the state in Eq. (1), the term that appears in the
expansion of the state in Eq. (1) is
Y
PXY (xk , y` ) |xk i hxk |Xk ⊗ |y` i hy` |Y` ⊗ PX (xi )PY (yj ) ⊗ |xi i hxi |Xi |yj i hyj |Yj ⊗ ρR
xk ,y` (2)
i∈[K], j∈[L]
i6=k, j6=`

The full expansion of the state in Eq. (1) is a sum over all strings x1 , x2 , . . . , xK and y1 , y2 , . . . , yL , of terms like the
one in Eq. (2). Now, when we sum over all fixed states as in Eq. (1) to get τ , we can group together those matrices
of the kind in Eq. (2) which have the same string x1 , x2 , . . . , xK , y1 , y2 , . . . , yL . Note that this choice of fixed string
fixes the pure states on the systems X1 X2 . . . XK Y1 Y2 . . . YL . However, the corresponding state on the system R will
be of the following form
1 X Y
PXY (xk , y` ) PX (xi ) · PY (yj )ρR xk ,y`
K ·L
k,` i6=k
j6=`

13
the above argument, along with Theorem 3.1, essentially implies the following :

O O
X X ...X Y Y ...Y R
τ 1 2 K 1 2 L − σ R Xi Yj

ρ ρ
i∈[K] j∈[L]
1


X
1 X Y
PX (xi ) · PY (yj )ρR

= x 1 , x2 , . . . xk ⊗ y 1 , y2 , . . . y L ⊗ PXY (x ,
k ` y ) xk ,y`

x1 ,x2 ,...xK K · L
y1 ,y2 ,...yL k,` i6=k
j6=`


X Y
R
− x1 , x2 , . . . xk ⊗ y1 , y2 , . . . yL ⊗ PX (xi ) · PY (yj )σ
x1 ,x2 ,...xK

i∈[K], j∈[L]
y ,y ,...y

1 2 L 1
where we have omitted the ket bra notation on the classical string for brevity. Then, using the block diagonal nature
of these matrices, we see that the above expression is


X Y 1 X PXY (xk , y` ) R R

= PX (xi ) · PY (yj )
ρxk ,y` − σ
x1 ,x2 ,...xK i∈[K], j∈[L] K · L k,` PX (xk ) · PY (y` )
y1 ,y2 ,...yL 1

We now change notation from xi and yj to x(i) and y(j) to emphasise the fact that they are the i-th and j-th samples
drawn from the distributions PX and PY respectively. Then, invoking Theorem 3.1, we see that the above expression
is


1 X PXY (x(k), y(`)) R R

= E ρx(k),y(`) − σ
x(1),x(2),...x(K) K · L PX (x(k)) · PY (y(`))

k,`
y(1),y(2),...y(L) 1

≤ε
for appropriately chosen values of K and L.

This concludes the proof of the claim.

3.2 An Operator Inequality

In this section we prove an operator inequality that we will find useful in proving our measurement compression
theorem.
Lemma 3.3. Operator Inequality for the Covering Lemma Suppose we are given some states ρA A A
1 , ρ2 , . . . , ρL and
a distribution P (i) on the set of indices [L] such that

X
P (i)ρA − ρA
≤ε

i

i 1
for a given average state ρ. Then there exist a good subset GOOD ⊂ [L] such that
Pr[GOOD] ≥ 1 − O(ε1/4 )
P

and quantum states {ρ0i | i ∈ GOOD} such that

0
A
ρi − ρi ≤ 2ε1/4

1
and
0
X
P (i)ρiA ≤ (1 + ε1/4 )ρA
i∈GOOD

14
Proof. Let the state |ρiAB be a purification of ρA . Similarly, let |ρi iAB denote purifications of the states ρA
i . Consider
the purification : Xp
|ψiABC := P (i) |ρi iAB |iiC
i

of the sample average state X

P (i)ρA
i
i

Then by Uhlmann’s theorem and the closeness in 1-norm of the sample average state to ρA , we see that there exists a
pure state |ϕiABC such that √
kϕ − ψk1 ≤ 2 ε
We will expand |ϕiABC by fixing the computational basis on C. This expansion can be written in the following form
: X
|ϕiABC = |vi iAB |iiC
i

where |vi i’s are some vectors on which we make no assumptions. We claim that each vector |vi iAB has length at most
1. This is not hard to see, since taking the inner product of |ϕiABC with itself, we see that
X
hvi |vi i = 1
i

The above observation can be used to define a distribution Q(i) on [L] in the following way :

Q(i) := hvi |vi i

We also define vectors |ρ0i iAB by normalising the |vi i’s :

AB 1
|ρ0i i := p |vi iAB
Q(i)

This allows us to express |ϕiABC as follows :

AB
Xp
|ϕiABC = Q(i) |ρ0i i |iiC
i

Note that by definition each |ρ0i iAB is a vector of length 1, and hence by the Schmidt decomposition and tracing out
0
the system B, can be seen as a purification of some quantum state ρiA .

Claim 3.4. There exists a subset of [L] of probability (under P ) of at least 1 − O(ε1/4 ) on which

P (i) ≤ (1 + ε1/4 )Q(i)

√
Proof. Note that by the 2 ε closeness of ϕ and ψ and the monotonicity of 1-norm :

√
 
X  P (i) 
Q(i) 1 −
 ≤2 ε
Q(i) 
i

Define the set    

1 − P (i)  ≤ ε1/4
 
INDEX := i |  Q(i) 
From Markov’s inequality, we see that
Pr [INDEXc ] ≤ 2ε1/4
Q

15
 √
Since kQ − P k1 ≤ 2 ε, this implies that
Pr [INDEXc ] ≤ O(ε1/4 )
P

Therefore for all i ∈ INDEX, we have that

P (i) ≤ (1 + ε1/4 )Q(i)
This proves the claim.

Claim 3.5. There exists a subset of [L] of probability under P at least 1 − O(ε1/4 ) such that
0
ρi − ρi ≤ 2ε1/4

1

Proof. Consider the expression √

kϕ − ψk1 ≤ ε
We will measure the matrices inside the 1-norm along the computational basis on the system C. This produces states
that are block diagonal. Appealing to the monotonicity of the 1-norm we see that this implies
√

X 0 P (i)
Q(i) ρi −
ρi ≤2 ε
Q(i)
1
i

Define the set  

0 P (i) 1/4
CLOSE := i | ρi −
ρi ≤ ε

Q(i) 1
By arguments similar to those used in Claim 3.4, we see that
Pr [CLOSEc ] ≤ O(ε1/4 )
P

Then, recalling the definition of the set INDEX from Claim 3.4, we see that
h \ i
Pr INDEX CLOSE ≥ 1 − O(ε1/4 )
P

Define \
GOOD := INDEX CLOSE

Then, for all i ∈ GOOD we see that

ρi − ρi ≤ ρ0i − P (i) ρi + P (i) ρi − ρi
0
1 Q(i) Q(i)
1
≤ ε1/4 + ε1/4
= 2ε1/4

We now define |ϕ̃i by throwing away those indices from the expansion of |ϕi which are not in GOOD.
AB
X p
|ϕ̃iABC := Q(i) |ρ0i i |ii
i∈GOOD

Then note that

0 0
X X
TrBC [ϕ̃] = Q(i)ρiA ≤ Q(i)ρiA = ρ
i∈GOOD i∈[L]

Finally, by using the properties of GOOD, we observe that

0 0
X X
P (i)ρiA ≤ (1 + ε1/4 ) Q(i)ρiA
i∈GOOD i∈GOOD
1/4
≤ (1 + ε )ρ
This concludes the proof.

16
Corollary 3.6. Suppose we are given some positive semi-definite matrices σ1A , σ2A , . . . , σLA and a distribution P (i) on
the set of indices [L] such that
X
A A
P (i)σi − ρ ≤ ε



i 1
for a given average state ρ. Define
P (i) · Tr[σi ]
P 0 (i) := P
P (i) · Tr[σi ]
i

Then there exist a good subset GOOD ⊂ [L] such that

Pr0 [GOOD] ≥ 1 − O(ε1/4 )

P

and quantum states {ρ0i | i ∈ GOOD} such that

0A
ρi − σi ≤ 2ε1/4
Tr[σi ] 1

and
0
X
P (i) · Tr[σi ] · ρiA ≤ (1 + O(ε1/4 ))ρA
i∈GOOD

Proof. Firstly note that from the hypothesis given in the statement, using monotonicity of the 1-norm

X
P (i) · Tr[σi ] − 1 ≤ ε



i 1

which implies that X

1−ε≤ P (i) · Tr[σi ] ≤ 1 + ε
i

Let us define the distribution

P (i) · Tr[σi ]
P 0 (i) := P
P (i) · Tr[σi ]
i

We also define
σi
ρi :=
Tr[σi ]
We will need the following fact:
Fact 3.7. Given a state ρ and a positive semi-definite matrix σ, the following holds:

ρ − σ ≤ 2 kρ − σk

1
Tr[σ] 1

Using this fact we can now rewrite the condition in the theorem statement as:

X
P 0 (i)ρi − ρ ≤ 2ε



i 1

Then, invoking Lemma 3.3, we can infer the existence of a set of indices GOODsuch that

Pr0 [GOOD] ≥ 1 − O(ε1/4 )

P

17
and states ρ0i such that 0
ρi − ρi ≤ O(ε1/4 )

1
and X
P 0 (i)ρ0i ≤ (1 + O(ε1/4 ))ρ
i∈GOOD

Recall that
P (i) · Tr[σi ] 1
P 0 (i) = P ≥ P (i) · Tr[σi ]
P (i) · Tr[σi ] 1+ε
i
Therefore, we conclude that
0
X
P (i) · Tr[σi ] · ρiA ≤ (1 + O(ε1/4 ))ρA
i∈GOOD

3.3 Classical Message Compression with Quantum Side Information

In this section we will prove Lemma 1.4. To define the above task, abbreviated as CQSI, we consider two parties Alice
and Bob and a classical quantum control state of the form
X
ρXB := PX (x) |xi hx|X ⊗ ρBx
x

where the classical system X belongs to Alice and the quantum system B belongs to Bob. Alice and Bob also share
a forward noiseless classical channel from Alice to Bob. Alice wishes to send the contents of the classical register to
Bob using as little classical communication as possible.
To be precise, Alice and Bob wish to create the state σ X X̂B via classical communication, where the system X
belongs to Alice and X̂B belongs to Bob such that

X X̂B
ρ − σ X X̂B ≤ ε

1

where we define X
ρX X̂B := PX (x) |xi hx|X ⊗ |xi hx|X̂ ⊗ ρB
x
x
We also allow Alice and Bob to share public random coins and they are allowed to use private coin algorithms
for encoding and decoding. We wish to minimise the amount of classical communication from Alice to Bob. For
convenience we restate Lemma 1.4 below.

Theorem 3.8. Given the control state

X
ρXB = PX (x) |xi hx|X ⊗ ρB
x
x

there exists a classical message compression protocol with quantum side information whenever the rate of communi-
cation R satisfies
ε ε 1
R ≥ Hmax (X) − IH (X : B)ρ + O(log )
ε
Proof. The rough idea is as follows :

1. Alice hashes the set X into a smaller set L using a 2-universal hash family F, where each f ∈ F maps X → L.

2. Each index ` ∈ L then corresponds to a subset of symbols, usually called a ’bucket’, which is essentially the
pre-image of ` with respect to some randomly chosen hash function f ∈ F.

3. Alice and Bob are both provided the description of this randomly chosen has function before the protocol starts.

18
 4. One can imagine that Alice never receives a symbol which has low probability. Thus, we can essentially discard
those x’s of low probability, which have at most ε-mass under the distribution PX . This implies that we can
imagine that Alice receives her symbols from the sub-distribution P 0 (x) instead of P (x).

5. Upon receiving the symbol x, Alice sends the hash f (x) to Bob via the classical noiseless channel.

6. Bob then knows the subset f −1 (x) ⊂ X . To decode the correct x sent by Alice, Bob does a measurement on
his quantum system B.

7. To be able to distinguish among the members x0 ∈ f −1 (x), the quantum states ρx0 need to be sufficiently far
apart.

8. We show that this condition holds as long as

ε
|f −1 (x)| ≤ 2IH (X:B)ρ

ε
9. Recall that the support of P 0 is of size at most 2Hmax (X) . Thus, the hash function essentially divides this set into
ε ε
2Hmax (X)−IH (X:B)ρ buckets. This concludes the protocol.

The reason why the above protocol works is the ’random codebooks’ created by the hash function. Without this
randomness, we could partition the support of P 0 arbitrarily. But then we would not be able to guarantee that the
symbols in the pre-image of any hash is distinguishable on average over the choice of the hash function.

Bob’s Decoding
We wish to analyse the probability of a decoding error. However, we will not work with the distribution PX but with
the sub-distribution PX0 . To see that this only incurs an extra ε error, note that
X
Pr[decoding error] = PX (x) · Pr[decoding error | x]
x
X X
= PX (x) · Pr[decoding error | x] + PX (X)(x) · Pr[decoding error | x]
x∈supp(P 0 ) x∈supp(P
/ 0)
X
≤ PX0 (x) · Pr[decoding error | x] + ε
x

We first define the bucket corresponding to the the hash function f and the hash ` as

A(f, `) := x | f (x) = `, x ∈ supp(P 0 )



n o
Let us denote the elements in A(f, `) as a`1 , a`2 , . . . , a`|A| . Now consider the operator ΠOPT from the definition of
ε (X : B) and the associated operators Π . Upon receiving the has `, Bob sequentially measures his system B with
IH ρ x
the operators Πa` .
i
We wish to analyse the probability of a decoding error. However, we will not work with the distribution PX but
with the sub-distribution PX0 . To see that this only incurs an extra ε error, note that
X
Pr[decoding error] = PX (x) · Pr[decoding error | x]
x
X X
= PX (x) · Pr[decoding error | x] + PX (X)(x) · Pr[decoding error | x]
x∈supp(P 0 ) x∈supp(P
/ 0)
X
≤ PX0 (x) · Pr[decoding error | x] + ε
x

19
Suppose that when the symbol x is sent, the corresponding index of this symbol in the set A(f, `) is a`m . Then,
conditioned on Alice having received x, the probability of incorrect decoding is given by
h i
1 − Tr Πa`m (I − Πa` ) . . . (I − Πa` ) · ρa`m
m−1 1
h i h i
= Tr ρa`m − Tr Πa`m (I − Πa` ) . . . (I − Πa` ) · ρa`m
m−1 1

Using Sen’s non-commutative union bound, the above expression can by bounded by
v
u h
u i m−1X h i
≤ Tr (I − Πa`m )ρa`m +
t Tr Πa` ρa`m
i
i=1

Now notice that, the sets A(f, 0

S `) form a disjoint cover of the set SUPP (PX ) over the indices 0`. Thus, taking an average
over the elements of the set A(f, `) is the same as taking an average over the set SUPP (PX ). Using this observation
`
along with the concavity of the square root, we see that the average error probability over choices of x is at most
v
i m−1
uX X !
h X h i
PX0 (a`m ) Tr (I − Πa`m )ρa`m +
u
t Tr Πa` ρa`m
i
`∈L a`m ∈A(f,l) i=1
v
uX X X m−1
X h i
PX0 (x) Tr [(I − Πx )ρx ] + PX0 (a`m )
u
=t Tr Πa` ρa`m
i
x `∈L a`m ∈A(f,l) i=1

The first term inside the square root is at most ε, by the property of ΠOPT that

" #
X X
Tr ΠOPT PX (x) |xi hx| ⊗ ρx ≥ 1 − ε
x

which implies that

" #
X
Tr PX (x) |xi hx|X ⊗ Πx ρx ≥ 1 − ε
x

It is easy to see from the above that X

PX0 (x) Tr [(I − Πx )ρx ]
x
X
≤ PX (x) Tr [(I − Πx )ρx ]
x
≤ε
To analyse the second term inside the square root, consider the following :

X X m−1
X h i
PX0 (a`m ) Tr Πa` ρa`m
i
`∈L a`m ∈A(f,l) i=1
X X X h i
≤ PX0 (a`m ) Tr Πa` ρa`m
i
`∈L a`m ∈A(f,l) i6=m
X X
= PX0 (x) I{f (x0 )=f (x)} Tr [Πx0 ρx ]
x x0 6=x
x0 ∈SUPP(P 0 )

20
where Ix0 6=x is the indicator for when f (x0 ) = f (x). We will now take an expectation over the choice of the hash
function f . Note that the above term is inside a square root, so to do this we use the concavity if square root :
 
X X
E PX0 (x) I{x0 6=x} Tr [Πx0 ρx ]
f
x x0 6=x
X X
PX0 (x)
 
= E I{f (x0 )=f (x)} Tr [Πx0 ρx ]
f
x x0 6=x
x0 ∈SUPP(P 0 )
X X
= PX0 (x) Pr[f (x0 ) = f (x)] Tr [Πx0 ρx ]
x x0 6=x
x0 ∈SUPP(P 0 )
X X
≤2−R PX0 (x) Tr [Πx0 ρx ]
x x0 6=x
x0 ∈SUPP(P 0 )
X X
≤2−R PX0 (x) Tr [Πx0 ρx ]
x x0 ∈SUPP(P 0 )

To bound this term, we multiply and divide by PX0 (x0 ) inside the second summation. This shows us that the above
expression is equal to
X PX0 (x0 ) X
=2−R PX0 (x) Tr [Πx0 ρx ]
P 0 (x0 )
x x0 ∈SUPP(P 0 ) X
 
ε
X X
≤2−R · 2Hmax (X) PX0 (x) Tr  PX0 (x0 )Πx0 ρx 


x x0 ∈SUPP(P 0 )

ε
where we have used Fact 2.4 to upper bound each P 0 1(x0 ) term by 2Hmax (X) . We will now switch back to the distri-
X
bution PX by adding the terms corresponding to the x’s which not in the support of PX0 . This implies that the above
expression can be upper bounded by
" #
ε
X X
−R Hmax (X) 0


≤2 · 2 PX (x) Tr PX (x )Πx0 ρx
x x0
" ! !#
ε
X X
−R+Hmax
=2 (X)
Tr PX (x0 )Πx0 PX (x)ρx
x0 x
" ! !#
ε 0 X
X X
−R+Hmax (X) 0 0
=2 Tr PX (x ) |x i hx | ⊗ ΠB
x0 I ⊗ X
PX (x)ρB
x
x0 x
" ! ! !#
ε 0 X 00 X
X X X
−R+Hmax (X) 0 00 00
=2 Tr |x i hx | ⊗ ΠB
x0 PX (x ) |x i hx | ⊗ PX (x)ρB
x
x0 x00 x
ε
−R+Hmax (X)
Tr ΠOPT ρX ⊗ ρB
 
=2
ε ε
≤2−R+Hmax (X)−IH (X:B)ρ

Thus, this shows that as long as

ε ε
R ≥ Hmax (X) − IH (X : B)ρ − log ε
√
the average decoding error over choices of x and the hash function f is at most 2ε + ε.
To finish the proof, consider the left polar decomposition of the operator
q
Πa`m (I − Πa` ) . . . (I − Πa` ) = Ua`m Θa`m
m−1 1

21
where Θa`m is some positive operator. Then, when Bob recovers the correct symbol x = a`m , the post measurement
state is given by
1 q q  †
h i Ua`m Θa`m ρa`m Θa`m Ua`m
Tr Θa`m ρa`m

Bob then applies the unitary Ua`m to this state to get the state

1 q q
h i Θa`m ρa`m Θa`m
Tr Θa`m ρa`m

Suppose that a`m = x ∈ SUPP(PX0 ). Then, by the Gentle Measurement Lemma, we have that


ρx − 1 p p p
Θx ρx Θx ≤ 2 Tr [(I − Θx )ρx ]
Tr [Θx ρx ]
1

It is now easy to see that the following bounds hold

X 1 p p
ρx −
PX (x)
Tr [Θx ρx ]
Θx ρx Θx

x 1

X 1
PX0 (x)
p p
≤ ρx − Tr [Θx ρx ] Θx ρx Θx + 2ε

x 1
X
PX0 (x) Tr [(I − Θx )ρx ] + 2ε
p
≤2
x
s
X
≤2 PX0 (x) Tr [(I − Θx )ρx ] + 2ε
x

which implies, by our previous computations that, the above expression can be upper bounded by
√
q
≤2 2ε + ε + 2ε =: ε0

The protocol then is that Bob, after decoding with the above POVM elements, places the classical symbol in the
system X̂. Let σ X X̂B be the state after the protocol ends. We set the marginal σ X̂B to be some junk if the decoding
failed. Note that σ is always classical on the system X. Then,

X X̂B
ρ − σ X X̂B

1
X
= PX (x) |xi hx|X̂ ⊗ ρB − σ X̂B

x
1
x
h i
X̂ B X̂B
= E |xi hx| ⊗ ρx − σ

PX 1
h i
X̂ B X̂B
= E |xi hx| ⊗ ρx − σ | correct decoding · Pr[correct decoding]

PX 1
h i
+ E |xi hx|X̂ ⊗ ρB − σ X̂B
| incorrect decoding · Pr[incorrect decoding]

x
PX 1

X 1 p p
0
≤ PX (x) ρx −
Θx ρx Θx +ε
x
Tr [Θ x ρ x ] 1

≤2ε0

This concludes the proof of the lemma.

22
4 Centralised Multi Link Measurement Compression
In this section we prove our centralised multi-link measurement compression theorem, in the presence of quantum
side information. We precisely define the problem below: Suppose that Alice possesses register A of a pure state
|ρiABR . Let ΛA→AXY := {Λx,y A→A }
x,y be a POVM where for each classical outcome (x, y), Λx,y is a genuine POVM
element i.e. a Hermitian operator on A with eigenvalues between zero and one. There exist two separate noiseless
channels, called X and Y channels, from Alice to Bob, and two independent public coin registers, called X and Y
public coins, between them. A noiseless channel together with its corresponding public coin is called a link. During
the protocol at most one of the links may be turned OFF by an adversary without Alice or Bob’s knowledge. Suppose
Alice were to measure register A of state |ρiABR with the POVM ΛA→AXY obtaining classical outcome (x, y). In
the centralised measurement compression protocol, Alice compresses the outcome pair and conveys the messages
through the corresponding noiseless channels with the help of the corresponding public coins. Let RX , RY denote the
rates of the noiseless channels for X and Y links, and CX , CY the rates for the public coins for X and Y links. We
require Bob to be able to decode with the help of the public coins and produce a state on ABRXY with the following
properties:

PX link is
1. If the ON and Y link is OFF, the state at the end ofP
the protocol should be ε-close in Schatten `1 -norm
to |xi hx|X ⊗ ((Λx ⊗ IBR )(ρ))ABR , where ΛA→Ax := y Λx,y ;
x∈X

P Y link is
2. If the ON and X link is OFF, the state at the end of
Pthe protocol should be ε-close in Schatten `1 -norm
to |yi hy|Y ⊗ ((Λy ⊗ IBR )(ρ))ABR , where ΛA→Ay := x Λx,y ;
y∈Y

3. If both
P links are ON, XY the state at the end of the protocol should be ε-close in Schatten `1 -norm to the state
|x, yi hx, y| ⊗ ((Λx,y ⊗ IBR )(ρ))ABR .
(x,y)∈X ×Y

Moreover, Alice’s and Bob’s strategies should be agnostic to which links are operational i.e. their encoding and
decoding strategies should continue to work even if one link fails.
We will prove Theorem 1.1 and Corollary 1.2. We do this in several steps:

1. We will first prove the centralised multi link measurement compression theorem in the one shot setting when
Bob does does not posses any qunatum side information. To that end we will first show that a corner point
in the rate region claimed in Theorem 1.1 (in the case when there is no side information) is achievable. See
Proposition 4.1.

2. Using the techniques of Proposition 4.1 we show that a rate region of the kind claimed in Theorem 1.1 (in the
case when there is no side information) is achievable by using the rate splitting technique shown in [5]. See
Proposition 4.6.

3. We then derive the rate region when quantum side information is present at the decoder by composing the
measurement compression theorem with our protocol to do classical message compression with quantum side
information. This is a highly non-trivial task and we do this in Section 4.2.

4. Finally, we observe that the rate region claimed in Corollary 1.2 is easily obtained by using the Quantum
Asymptotic Equipartition Property (Section 2.4) for all the one shot quantities used in the proof of the one shot
theorem.

23
4.1 One Shot Centralised Multi Link Measurement Without Side Information
Proposition 4.1. Given the state ρA and the POVM ΛXY := ΛA

xy , the following rate point is achievable for
centralised multi link measurement compression:
ε
RX + CX > Hmax (X)
ε
RY + CY > Hmax (Y )
ε
RX > Imax (X : R)
ε
RY > Imax (Y : RX)

where all entropic quantities are computed with respect to the state
X √ √
|xi hx|X ⊗ |yi hy|Y ⊗ ρΛxy ρR
x,y

Proof. POVM Construction

We are given the POVM

ΛXY := ΛA

xy

and the state ρA . Consider the true post measurement state

X √ √
|xi hx|X ⊗ |yi hy|Y ⊗ ρΛxy ρR
x,y

We define
PXY (x, y) := Tr[Λxy ρ]
Let log K1 and log K2 be the number of public coins available to Alice and Bob with respect to the X− and Y −
channels respectively. Similarly, let log L1 and log L2 be the number of bits that Alice needs to send Bob along the
two channels respectively. We will show that the following region is achievable for centralised multi link measurement
compression.
ε
log K1 + log L1 > Hmax (X)
ε
log K2 + log L2 > Hmax (Y )
ε
log L1 > Imax (X : R)
ε
log L2 > Imax (Y : RX)

Construct CX := {x(k1 , `1 )} and CY := {y(k2 , `2 )} randomly and independently from the distributions PX and PY ,
where k1 ∈ [K1 ], `1 ∈ [L1 ] and k2 ∈ [K2 ], `2 ∈ [L2 ]. Define
√ √
ρΛx(k1 ,`1 ),y(k2 ,`2 ) ρ
ρ(k1 , k2 , `1 , `2 ) :=
PXY (x(k1 , `1 ), y(k2 , `2 ))
From the choices of L1 and L2 and the measure transformed successive cancellation covering lemma we have that


X 1 1 X PXY (x(k1 , `1 ), y(k2 , `2 ))
E ρ(k 1 , k2 , `1 , `2 ) − ρ ≤ε (close)
CX ×CY K1 · K2 L1 · L2 PX (x(k1 , `1 )) · PY (y(k2 , `2 ))
k1 ,k2 `1 ,`2
1

Definition 4.2. We call a block (k1 , k2 ) ‘nice’ if the following condition holds for that block, with respect to some
fixed codebook CX × CY :

√

1 X PXY (x(k1 , `1 ), y(k2 , `2 ))

L1 · L2 ρ(k1 , k2 , `1 , `2 ) − ρ ≤ ε
PX (x(k1 , `1 )) · PY (y(k2 , `2 ))
`1 ,`2
1

24
Let us fix a nice block (k1 , k2 ) with respect to some fixed codebook. To ease the notation define

PXY (x(k1 , `1 ), y(k2 , `2 ))

t(k1 , k2 , `1 , `2 ) :=
PX (x(k1 , `1 ))PY (y(k2 , `2 )
We are now in a position to apply Corollary 3.6. To do this we define Q to be the uniform distribution on the set
[L1 ] × [L2 ]. Then let
Q(`1 , `2 ) · t(k1 , k2 , `1 , `2 )
Q0 (k1 , k2 , `1 , `2 ) := P
Q(`1 , `2 ) · t(k1 , k2 , `1 , `2 )
`1 ,`2

Then, Corollary 3.6 implies that there exists a subset GOOD ⊂ [L1 ] × [L2 ] such that

Pr0 [GOOD] > 1 − O(ε1/4 )

Q

and for all (`1 , `2 ) ∈ GOOD there exist states ρ0 (k1 , k2 , `1 , `2 ) such that
0
ρ (k1 , k2 , `1 , `2 ) − ρ(k1 , k2 , `1 , `2 ) ≤ O(ε1/4 )

1

and
X 1
· t(k1 , k2 , `1 , `2 )ρ0 (k1 , k2 , `1 , `2 ) ≤ (1 + O(ε1/4 ))ρ
L1 · L2
(`1 ,`2 )∈GOOD

For the fixed block indices k1 and k2 we define the POVM elements for (`1 , `2 ) ∈ GOOD:

1 t(k1 , k2 , `1 , `2 ) −1/2 0
Γ`1 ,`2 (k1 , k2 ) := 1/4
ρ ρ (k1 , k2 , `1 , `2 )ρ−1/2
1 + O(ε ) L1 · L2

Some Important Observations :

 
X t(k1 , k2 , `1 , `2 )  X X
= Q(`1 , `2 ) · t(k1 , k2 , `1 , `2 ) · Q0 (k1 , k2 , `1 , `2 )
L1 · L2
(`1 ,`2 )∈
/ GOOD `1 ,`2 (`1 ,`2 )∈
/ GOOD
1/4 1/4
≤(1 + O(ε )) · O(ε )
=O(ε1/4 )

Then, we define the POVM

[
Γ(k1 , k2 ) := {Γ`1 ,`2 (k1 , k2 ) | (`1 , `2 ) ∈ GOOD} {Γ0 (k1 , k2 )}

where X
Γ0 (k1 , k2 ) := Isupp(ρ) − Γ`1 ,`2 (k1 , k2 )
(`1 ,`2 )∈GOOD

Claim 4.3.
Tr[Γ0 (k1 , k2 )ρ] ≤ O(ε1/4 )

Proof.
X 1 t(k1 , k2 , `1 , `2 ) 0
Tr[Γ0 (k1 , k2 )ρ] = Tr[ρ − 1/4
ρ (k1 , k2 , `1 , `2 )]
1 + O(ε ) L1 · L2
(`1 ,`2 )∈GOOD
1 X t(k1 , k2 , `1 , `2 )
=1 −
1 + O(ε1/4 ) L1 · L2
(`1 ,`2 )∈GOOD

25
Recall that
X t(k1 , k2 , `1 , `2 )
≥1 − ε
L1 · L2
(`1 ,`2 )
X t(k1 , k2 , `1 , `2 ) X t(k1 , k2 , `1 , `2 )
=⇒ ≥1 − ε −
L1 · L2 L1 · L2
(`1 ,`2 )∈GOOD (`1 ,`2 )∈
/ GOOD
X t(k1 , k2 , `1 , `2 )
=⇒ ≥1 − ε − O(ε1/4 )
L1 · L2
(`1 ,`2 )∈GOOD

Therefore,

1 − O(ε1/4 )
Tr[Γ0 (k1 , k2 )ρ] ≤ 1 − ≤ O(ε1/4 )
1 + O(ε1/4 )

Closeness of the Post Measurement States

Alice will use the compressed POVMs only for those blocks which are nice. For all other blocks she aborts the
protocol. Notice that the compressed POVMs that we designed output a classical index (`1 , `2 ). There exist functions
f and g such that
f (k1 , `1 ) :=x(k1 , `1 )
g(k2 , `2 ) :=y(k2 , `2 )
The error expression is then given by

X √ √
|xi hx|X ⊗ |yi hy|Y ⊗ ρΛx,y ρR

E
CX ×CY
x,y



X 1 √ √
− |x(k1 , `1 )i hx(k1 , `1 )| ⊗ |y(k2 , `2 )i hy(k2 , `2 )| ⊗ ρΓ`1 ,`2 (k1 , k2 ) ρ
K1 · K2
(k1 ,k2 ) nice

(`1 ,`2 )∈GOOD ∪{0} 1
(post-measurement)

Definition 4.4. We define random variable:

X √ √
T (CX , CY ) := |xi hx|X ⊗ |yi hy|Y ⊗ ρΛx,y ρR


x,y



X 1 √ √
− |x(k1 , `1 )i hx(k1 , `1 )| ⊗ |y(k2 , `2 )i hy(k2 , `2 )| ⊗ ρΓ`1 ,`2 (k1 , k2 ) ρ
K1 · K2
(k1 ,k2 ) nice

(`1 ,`2 )∈GOOD ∪{0} 1

and the event: n o

E := # of nice blocks ≥ (1 − ε1/4 ) · K1 · K2

Consider the following claim which is easy to prove:

Claim 4.5.
Pr[E] ≥ 1 − ε1/4

26
 The error expression can then be analysed as:
E[T ] = E[T · 1E ] + E[T · 1E c ]
≤ E[T · 1E ] + 2 · Pr[E c ]
≤ E[T · 1E ] + 2 · ε1/4
The first inequality used the fact that T is at most 2, since it is the 1-norm between two states. We will thus bound
T (CX , CY ) for only those codebooks for which the event E holds.
We will massage the second term inside the 1-norm expression above. First, we discard the 0-th outcome which
adds at most O(ε1/4 ). Next, notice that
√ √ 1 t(k1 , k2 , `1 , `2 ) 0
ρΓ`1 ,`2 (k1 , k2 ) ρ = 1/4
ρ (k1 , k2 , `1 , `2 )
1 + O(ε ) L1 · L2
We will apply the triangle inequality twice. In the first we replace all the ρ0 (k1 , k2 , `1 , `2 ) states with ρ(k1 , k2 , `1 , `2 ).
This gives rise to the term
1 X 1 t(k1 , k2 , `1 , `2 )
ρ0 (k1 , k2 , `1 , `2 ) − ρ(k1 , k2 , `1 , `2 )

1 + O(ε ) 1/4 K1 · K2 L1 · L2 1
(k1 ,k2 ) is nice
(`1 ,`2 )∈GOOD

Due to the conditions in Corollary 3.6, this can be bounded above by

1 X 1 t(k1 , k2 , `1 , `2 )
O(ε1/4 )
1 + O(ε1/4 ) K1 · K2 L1 · L2
(k1 ,k2 ) is nice
(`1 ,`2 )∈GOOD

We already know from Eq. (close) that for a nice block index (k1 , k2 )
X t(k1 , k2 , `1 , `2 )
≤1+ε
L1 · L2
`1 ,`2

Therefore,
O(ε1/4 ) X 1 t(k1 , k2 , `1 , `2 ) O(ε1/4 )
≤ (1 + ε)
1 + O(ε1/4 ) K1 · K2 L1 · L2 1 + O(ε1/4 )
(k1 ,k2 ) is nice
(`1 ,`2 )∈GOOD

≤O(ε1/4 )
These steps have allowed us to massage the second term inside the norm in Eq. (post-measurement) into
1 X 1 t(k1 , k2 , `1 , `2 )
1/4
|x(k1 , `1 )i hx(k1 , `1 )|
1 + O(ε ) K 1 · K2 L1 · L2
(k1 ,k2 ) is nice
(`1 ,`2 )∈GOOD

⊗ |y(k2 , `2 )i hy(k2 , `2 )| ⊗ ρ(k1 , k2 , `1 , `2 )

We further massage this term by adding those terms in the sum above which correspond to those (`1 , `2 ) which are
not in the set GOOD. This adds an extra
1 X 1 X t(k1 , k2 , `1 , `2 )
1/4
≤ O(ε1/4 )
1 + O(ε ) K1 · K2 L1 · L2
(k1 ,k2 )is nice (`1 ,`2 )∈
/ GOOD

Collating the arguments above, we see that the true post measurement state in Eq. (post-measurement) can be replaced
with the state
1 X 1 t(k1 , k2 , `1 , `2 )
TGOOD := 1/4
|x(k1 , `1 )i hx(k1 , `1 )|
1 + O(ε ) K1 · K2 L1 · L2
(k1 ,k2 ) is nice
(`1 ,`2 )

⊗ |y(k2 , `2 )i hy(k2 , `2 )| ⊗ ρ(k1 , k2 , `1 , `2 )

27
Finally, we will add the terms corresponding to the blocks which are not nice, i.e,

1 X 1 t(k1 , k2 , `1 , `2 )
TBAD := |x(k1 , `1 )i hx(k1 , `1 )|
1 + O(ε1/4 ) K1 · K2 L1 · L2
(k1 ,k2 ) not nice
(`1 ,`2 )

⊗ |y(k2 , `2 )i hy(k2 , `2 )| ⊗ ρ(k1 , k2 , `1 , `2 )

Recall that since we only consider those codebooks for which the event E holds, there are at most ε1/4 · K1 · K2 blocks
which are not nice. Consider the random variable
1 X 1 t(k1 , k2 , `1 , `2 )
kTBAD k1 = (bad blocks)
1 + O(ε1/4 ) K1 · K2 L1 · L2
(k1 ,k2 ) not nice
(`1 ,`2 )

Bounding kTBAD k1
From the classical measure transformed covering lemma we know that

X √ √
|xi hx|X ⊗ |yi hy|Y ⊗ ρΛx,y ρR

E
CX ×CY
x,y



X t(k1 , k2 , `1 , `2 )
− |x(k1 , `1 )i hx(k1 , `1 )| ⊗ |y(k2 , `2 )i hy(k2 , `2 )| ⊗ ρ(k1 , k2 , `1 , `2 )
K1 L1 · K2 L2
(k1 ,k2 )

(`1 ,`2 ) 1
≤ε

This implies that

X √ √
|xi hx|X ⊗ |yi hy|Y ⊗ ρΛx,y ρR

E
CX ×CY
x,y


X t(k1 , k2 , `1 , `2 )

− |x(k1 , `1 )i hx(k1 , `1 )| ⊗ |y(k2 , `2 )i hy(k2 , `2 )| ⊗ ρ(k1 , k2 , `1 , `2 )
· 1E
K1 L1 · K2 L2
(k1 ,k2 )

(`1 ,`2 ) 1
≤ε

Thus a random codebook satisfies the event E and the condition

X √ √
|xi hx|X ⊗ |yi hy|Y ⊗ ρΛx,y ρR



x,y



X t(k1 , k2 , `1 , `2 )
− |x(k1 , `1 )i hx(k1 , `1 )| ⊗ |y(k2 , `2 )i hy(k2 , `2 )| ⊗ ρ(k1 , k2 , `1 , `2 )
K1 L1 · K2 L2
(k1 ,k2 )

(`1 ,`2 ) 1
√
≤ ε

28
 √
with probability at least 1 − ε. Fix such a codebook. Then for this good codebook it is easy to see that

X √ √
|xi hx|X ⊗ |yi hy|Y ⊗ ρΛx,y ρR



x,y



1 X t(k1 , k2 , `1 , `2 )
− 1/4
|x(k1 , `1 )i hx(k1 , `1 )| ⊗ |y(k2 , `2 )i hy(k2 , `2 )| ⊗ ρ(k1 , k2 , `1 , `2 )
1 + O(ε ) K1 L1 · K2 L2
(k1 ,k2 )

(`1 ,`2 ) 1
1/4
≤ O(ε )

We can write the second term in the 1-norm above as a sum TGOOD and TBAD . Then using the monotonicity of the
1-norm under trace, we see that

|1 − Tr[TGOOD ] − Tr[TBAD ]|1 ≤ O(ε1/4 )

From our previous analysis, we know that

  
 
 
 
Tr[TGOOD ] − Tr 
X 1 √ √ 
|x(k1 , `1 )i hx(k1 , `1 )| ⊗ |y(k2 , `2 )i hy(k2 , `2 )| ⊗ ρΓ`1 ,`2 (k1 , k2 ) ρ
  K1 · K2 

 (k1 ,k 2 ) nice 

(`1 ,`2 )∈GOOD ∪{0} 1
1/4
≤O(ε )

Recall that for our choice of codebook, there are at most ε1/4 fraction of bad blocks. This implies that the second term
in the expression above is at least 1 − ε1/4 . Therefore, we can conclude that

Tr[TGOOD ] ≥ 1 − O(ε1/4 )

Using this lower bound on the trace of TGOOD , it is easy to see that

Tr[TBAD ] ≤ O(ε1/4 )

Collating all the arguments above, we see that we can replace the true post measurement state in Eq. (post-measurement)
to TGOOD +TBAD with an additive error of at most O(ε1/4 ). Recall that we have already fixed a codebook which bounds
√
the 1-norm between the ideal post measurement state and TGOOD + TBAD by ε. Thus, this implies that there exists
√
with probability at least 1 − ε a codebook which ensures that
1/4
E [T (CX , CY )] ≤ O(ε )
CX ×CY

This concludes the proof.

We will now show how one can achieve a larger rate region in the one shot setting for centralised multi link
measurement compression by using the technique of quantum rate splitting, as shown in [5].

Proposition 4.6. Suppose Alice is given the state ρA and the POVM ΛX Y as in Proposition 4.1. Suppose that PX Y
is the distribution induced by the POVM. Let (PUθ , PVθ , max) be a split of the marginal PX , as defined in [5, 11] , for
some parameter θ ∈ [0, 1]. Then one achievable rate region is obtained as the union over a parameter θ ∈ [0, 1] of

29
the regions Sθ defined by:
RX = RU + RV
ε
RU > Imax (U : R)
+ O(log ε−1 )
ε
RY > Imax (Y : RU )
+ O(log ε−1 )
Sθ : ε
RV > Imax (V : RU Y )
+ O(log ε−1 )
CX = CU + CV
ε
CU +RU > Hmax (U )
ε
CY +RY > Hmax (Y )
ε
CV +RV > Hmax (V ),
where the entropic quantities are calculated for the control state
X
pθU (u)pθV (v)p(y|u, v) |u, v, yi hu, v, y|U V Y
(u,v,y)∈X ×X ×Y

((Λmax{u,v},y ⊗ IR )(ρ))AR
⊗ .
Tr[((Λmax{u,v},y ⊗ IR )(ρ))AR ]
above state is obtained by splitting random variable X into independent random variables U , V in the state
TheP
|x, yi hx, y|XY ⊗((Λx,y ⊗IR )(ρ))AR according to the parameter θ. Another achievability region is obtained
(x,y)∈X ×Y
by rate splitting Y instead of X. The total achievable region is the union of the two regions. The encoding and
decoding strategies are agnostic to which links are actually functioning.
Proof. First consider the post measurement state
X
|x, yi hx, y|XY ⊗ ((Λx,y ⊗ IR )(ρ))AR
(x,y)∈X ×Y

This can be rewritten as:

X ((Λx,y ⊗ IR )(ρ))AR
PX (x)PY |X (y|x) |x, yi hx, y|XY ⊗
Tr [((Λx,y ⊗ IR )(ρ))AR ]
(x,y)∈X ×Y

where
PXY (x, y) := Tr ((Λx,y ⊗ IR )(ρ))AR
 

One can then split the distribution PX into the distributions pθU (u) and pθV (v) such that
max (U, V ) ∼ PX
where the pθU (u) and pθV (v) are both defined on the classical alphabet X . Refer to [5] for details about this splitting
operation. This leads to the control state:
X
pθU (u)pθV (v)p(y|u, v) |u, v, yi hu, v, y|U V Y
(u,v,y)∈X ×X ×Y

((Λmax{u,v},y ⊗ IR )(ρ))AR
⊗ .
Tr[((Λmax{u,v},y ⊗ IR )(ρ))AR ]
where
p(y|u, v) := PY |X (y| max(u, v))
Next, we construct the POVM elements as follows:

30
 CU ], ` ∈ [2RU ] iid from pθ , Y (k , ` ) | k ∈ [2CY ], ` ∈ [2RY ]
 
1. Alice picks samples U (k 1 , `1 ) | k1 ∈ [2 1 U 2 2 1 1
iid from PY and V (k3 , `3 ) | k3 ∈ [2CV ], `1 ∈ [2RV ] iid from pθV .


2. Define
Λu,v,y := Λmax(u,v),y

The rest of the proof is similar to the proof of Proposition 4.1, where in this case we have to use the sequential covering
lemma for 3 parties. Using the techniques in that proof, it is not hard to see that the rate region Sθ is achievable. This
concludes the proof. This concludes the proof.

4.2 Measurement Compression with Side Information

In this section we will show how to compose the centralised measurement compression theorem with our protocol for
classical data compression with quantum side information. We will first demonstrate our technique for the simpler
case when the POVM only outputs one classical symbol x i.e. the point to point case.

Proposition 4.7. Given the shares quantum state ρAB , where the receiver Bob possesses the B system, the following
rates are achievable for one shot measurement compression with quantum side information:
ε /2
ε
RX > Imax (X : RB) − IH0 (X : B) + O(log ε−1 ) + 1

and
ε ε /2
RX + CX > Hmax (X) − IH0 (X : B) + O(log 1/ε) + 1.

where ε0 := ε1/10 . Above, all entropic quantities are computed with respect to |xi hx|X ⊗ ((Λx ⊗ IBR )(ρ))ABR .
P
x∈X

Proof. We will first construct a point to point measurement compression protocol which ignores that Bob has any
side information. This follows as a corollary of Proposition 4.1. Recall that this construction consists of a random
codebook CX which is divided into K blocks, where each block contains L elements. Note that
ε
log K + log L > Hmax (X) − O(log ε)
ε
log L > Imax (X : RB) − O(log ε)

where the entropic quantities are computed with respect to the control state
X
|xi hx|X ⊗ ((Λx ⊗ IBR )(ρ))ABR .
x∈X

Suppose that CX obeys the condition that the fraction of nice blocks is at least 1 − O(ε1/4 ), where a ‘nice’ block refers
to a block which satisfies Definition 4.2. Suppose that for every nice block index k, let GOOD(k) be the set of indices
` which obey the conditions of Lemma 3.3. Then, our protocol ensures that with respect to this codebook, the global
post measurement state is
1 X 1 0 RB
ρ̃KLRB := |ki hk|K ⊗ |`i h`|L ⊗ ρk,`
Tr[ρ̃] L·K
k is a nice block
`∈GOOD(k)

with the promise that for all nice indices k,

X 1 0 RB
ρ ≤ (1 + ε1/4 ) TrA ((Λ ⊗ IBR )(ρ))ABR
L k,`
`∈GOOD(k)

Recall that conditioned on the codebook, there exists a deterministic map f : [K] × [L] → X such that

31
 XBR
− ρ̃XBR ≤ O(ε1/4 )

ρ

Let P̃X be the marginal of ρ̃XBR on the system X. Then it is not hard to see that there exists a subset INDEX of
x’s with probability at least 1 − O(ε1/8 ) under the distribution PX such that, for all x ∈ INDEX

P̃X (x) ≤ (1 + O(ε1/8 ))PX (x)

We now require that Alice and Bob work with a post-measurement state which has P̃X as the marginal distribution
on the system X. To do this, Alice simply discards those (k, `) pairs which map to those x’s under f which are not in
INDEX . We will call this post measurement state σ KLB :

1 X 1 0 RB
σ KLRB := |ki hk|K ⊗ |`i h`|L ⊗ σk,`
Tr[σ] K ·L
k is nice
`∈GOOD(k)
f (k,`)∈INDEX

where
0
RB RB 0
σk,` := ρk,`
and for all nice k,
X 1 0 RB
σ ≤ (1 + ε1/4 ) TrA ((Λ ⊗ IBR )(ρ))ABR (op-ineq)
L k,`
`∈GOOD(k)
f (k,`)∈INDEX

Since Pr[INDEX] ≥ 1 − O(ε1/8 ), it is not hard to see that

KLRB
− σ KLRB 1 ≤ O(ε1/8 )

ρ̃

Note that the definition of σ implies that, under the map f :

1
σX ≤ ρX
1 − O(ε1/8 )

since the trace of σ is at least 1 − O(ε1/8 ).

For the tools we develop in this section, we will not require the system R. Thus we will only with the control state
σ KLB . Note that the operator inequality still holds since trace out preserves operator inequalities. Eq. (op-ineq) can
now be written as: X 1 0B
σ ≤ (1 + ε1/4 )ρB
L k,`
`∈GOOD(k)
f (k,`)∈INDEX

Now, consider the following claim:

Claim 4.8. Control state :
0 1 X 1 0 0B
σ KK LB := |ki hk|K ⊗ |ki hk|K ⊗ |`i h`|L ⊗ σk,`
Tr[σ] L·K
k is a nice
`∈GOOD(k)
f (k,`)∈INDEX

with the property that for every nice index k,

X 1 0B
σ ≤ (1 + ε1/4 )ρB
L k,`
`∈GOOD(k)
f (k,`)∈INDEX

and under the map f

1
σX ≤ ρX
1 − O(ε1/8 )

32
Then :
ε0
IH (KL : BK 0 )σ ≥ log K + IH
ε0
(X : B)ρXB − 1
where
ε0 := ε1/10

Proof. Let ΠOPT be the optimising operator in the definition of

ε0 KLB
DH (σ || σ KL ⊗ ρB )

Without loss of generality we can assume that

X 0
ΠOPT = |ki hk|K ⊗ |ki hk|K ⊗ |`i h`|L ⊗ ΠB
k,`
k is a nice
`∈GOOD(k)
f (k,`)∈INDEX

For the purposed of brevity we will refer to the (k, `) which obey the conditions in the definition of σ as ‘fine’. Then
  !
0
X 1 X 1 K 0 0
ΠOPT σ KL ⊗ σ K B = c0 · ΠOPT  |ki hk|K ⊗ |`i h`|L  ⊗ |k 0 i hk 0 | ⊗ σkB0
L·K K
(k,`) fine k fine

where
0
X 1 0
σkB := σB
L k,`
` fine

and
1 1 1
c0 := 2 ≤
2 ≤ 1/8 )
(Tr[σ]) 1/8
1 − O(ε ) 1 − O(ε

Then RHS is equal to

X 1 0
= c0 · |ki hk|K ⊗ |ki hk|K ⊗ |`i h`|L ⊗ Πk,` σkB (1)
K2 ·L
(k,`) fine
X 1 0
≤ c0 · (1 + ε1/4 ) |ki hk|K ⊗ |ki hk|K ⊗ |`i h`|L ⊗ Πk,` ρB
K2 · L
(k,`) fine

Then taking trace on both sides and using the definition of ΠOPT we get
 
ε0 0
X 1 0
2−IH (KL:K B) ≤c0 · (1 + ε1/4 ) Tr  2
|ki hk|K ⊗ |ki hk|K ⊗ |`i h`|L ⊗ Πk,` ρB 
K ·L
(k,`) fine
ε0 KLB || σ KL ⊗ρB )
=c0 · (1 + ε1/4 ) · 2− log K · 2−DH (σ
ε0 XB || σ X ⊗ρB )
≤c0 · (1 + ε1/4 ) · 2− log K · 2−DH (σ
where the last inequality is via the data processing inequality.
Next, suppose that ΠXB is the optimising operator for
ε0 XB
DH (σ || σ X ⊗ ρB )

Also observe that

Tr[ΠXB σ XB ] ≥ Tr[ΠXB ρXB ] − σ XB − ρXB 1

≥1 − ε0 − O(ε1/8 )
ε0
≥1 −
2

33
where the last line is due to the fact that ε0 = ε1/10 . This implies that
ε0 0 ε0 /2 XB
|| σ X ⊗ρB )
2−IH (KL:K B) ≤c0 · (1 + ε1/4 ) · 2− log K · 2−DH (ρ

Finally, since we know that

1
σX ≤ ρX
1 − O(ε1/8 )
one can see that
ε0 0 1 ε0 /2 XB X B
2−IH (KL:K B) ≤c0 · (1 + ε1/4 ) · 1/8
· 2− log K · 2−DH (ρ || ρ ⊗ρ )
1 − O(ε )
1 − log K
ε /2
−IH0 (X:B)ρXB
=c0 · (1 + ε1/4 ) · · 2 · 2
1 − O(ε1/8 )
For small enough ε, we can sue the following bound
1
c0 · (1 + ε1/4 ) · ≤2
1 − O(ε1/8 )
Therefore, it is now clear that
ε /2
ε0
IH (KL : K 0 B)σ ≥ log K + IH0 (X : B)ρXB − 1
This concludes the proof.
Composing the Two Protocols :
We will now use the claims proved above to show that the measurement compression theorem and the one-shot
protocol for classical message compression with side information can be composed in the case when Bob has some
side information. In that case, the actual post measurement state that we will work with is
0 1 X 1 0 0 RB
σ KLK RB := |ki hk|K ⊗ |ki hk|K ⊗ |`i h`|L ⊗ σk,`
Tr[σ] K ·L
where the summation is over the appropriate set of (k, `) as defined before. We will treat the systems K 0 and B as
side information available to Bob. One can now see that the arguments of Claim 4.8 and Theorem 3.8 together imply
that Alice has to send at most the following number of bits:
ε0
Hmax ε0
(KL)σ − IH (KL : K 0 R)σ
ε0 ε /2
≤Hmax (KL) − log K − IH0 (X : B)ρXB + 1
ε /2
≤ log K + log L − log K − IH0 (X : B)ρXB + 1
ε /2
= log L − IH0 (X : B)ρXB + 1
From the measurement compression theorem, we know that
ε
log L > Imax (X : RB) − O(log ε)
Thus, this implies that Alice needs to send at most
ε ε /2
Imax (X : RB) − IH0 (X : B)ρXB − O(log ε) + 1
bits. This concludes the proof.
Remark 4.9. It is important to note that Claim 4.8 holds even when the control state is a weighted sum of the following
form:
0 1 X w(k, `) 0 0B
σ KK LB = |ki hk|K ⊗ |ki hk|K ⊗ |`i h`|L ⊗ σk,`
Tr[σ] L·K
This is precisely the case that we will face when we use Claim 4.8 to prove Theorem 1.1, due to the presence of the
terms t(k1 , k2 , `1 , `2 ) (refer to the proof of Proposition 4.1). However, it is easy to verify that even in this case the
proof of Claim 4.8 remains unchanged.
We are finally ready to prove Theorem 1.1.

34
4.3 Proof of Theorem 1.1
Proof. The proof is not hard given the discussion in the previous sections. We will show that the region Sθ , when
θ = 0 is achievable. This argument can then be extended to other values of θ using the similar reasoning to what
we used in proving Proposition 4.6. To show that S0 is achievable, we first invoke Proposition 4.1 to show that there
exists a measurement compression scheme which achieves the rates given by S0 in the case when there is no side
information. To be precise, S0 (in the case when the system B is trivial) is given by:
ε
RX > Imax (X : RB)
ε
RY > Imax (Y : XRB)
ε
CX + RX > Hmax (X)
ε
CY + RY > Hmax (Y )

where we have ignored the additive O(log 1ε ) factors. We will call the above region

S0unassisted

Now when Bob possesses some quantum side information in the B system, we can repeat the proof of Proposition 4.7
independently for the X− and Y − channels. This is possible because of two reasons:

1. The encoding maps for the X− and Y − channels are independent by design.

2. Even when both channels are ON, Bob can first decode the X− channel and then decode the Y − channel
sequentially, since the first decoding does not perturb the state in the B system too much. To be precise, this
will lead to at most an additive error of some g(ε), where g(ε) is ε to some rational power. This is due to the
Gentle Measurement Lemma and the proof of Theorem 3.8.

An important point to note that is that we cannot do any better in the sum rate, even though we could potentially get
even more savings by repeating the proof of Proposition 4.7 in the case when both channels are ON. This is because
we require the protocol to be agnostic to the actions of the adversary. For example, suppose Alice were to send a
bit string along the X− channel which was shorter than the one prescribed by our protocol. She would then have to
compensate by sending more bits along the Y − channel. However, the adversary could then simply turn OFF the Y −
channel, leading to a situation where it is not certain whether Bob can decode.
This is why, the rate region will look
ε0 ε0
S0unassisted − (IH (X : B), IH (Y : B))

The operation S − (x0 , y0 ) is defined as

S − (x0 , y0 ) := {(a − x0 , b − y0 ) | (a, b) ∈ S}

For other values of θ, we can now split the distribution PX , as in the proof of Proposition 4.6. Bob can then repeat his
sequential decoding algorithm pretending as if there are three channels namely the U − channel, the Y − channel and
the V − channel. This shows that the rate region claimed in the statement of Theorem 1.1 is indeed achievable. This
concludes the proof.

5 Asymptotic IID Analysis

In this section we will prove Corollary 1.2:

Proof. As in the proof of Theorem 1.1, we first show that all points in the region
ε0 ε0
S0assisted = S0unassisted − (IH (X : B), IH (Y : B))

35
is achievable when only 1 copy of the state ρAB is available. Now, we use the quantum asymptotic equipartition
property for each of the one shot quantities in this expression to show that, when the state available for measurement
is ρAB⊗n , with n → ∞. Thus, after normalising by n and taking the limit, the region S0assisted is equivalent to:

RX > I(X : RB) − I(X : B)

RY > I(Y : XRB) − I(Y : B)
CX + RX > H(X) − I(X : B)
CY + RY > H(Y ) − I(Y : B)

One can similarly prove that the following region, which we call S1assisted , corresponding to θ = 1 is also achievable:

RX > I(X : Y RB) − I(X : B)

RY > I(Y : RB) − I(Y : B)
CX + RX > H(X) − I(X : B)
CY + RY > H(Y ) − I(Y : B)

One can then use a time sharing argument to show that the full achievable region is as follows:

RX > I(X : RB) − I(X : B)

RY > I(Y : RB) − I(Y : B)
RX + RY > I(XY : RB) + I(X : Y )
− I(X : B) − I(Y : B)
CX + RX > H(X) − I(X : B)
CY + RY > H(Y ) − I(Y : B)

This concludes the proof.

References
[1] R. Ahlswede and A. Winter. Strong converse for identification via quantum channels. IEEE Transactions on
Information Theory, 48(3):569–579, 2002.

[2] Anurag Anshu, Vamsi Krishna Devabathini, and Rahul Jain. Quantum communication using coherent rejection
sampling. Phys. Rev. Lett., 119:120506, Sep 2017.

[3] Anurag Anshu, Rahul Jain, and Naqueeb Ahmad Warsi. Convex-split and hypothesis testing approach to one-
shot quantum measurement compression and randomness extraction. IEEE Transactions on Information Theory,
65(9):5905–5924, 2019.

[4] Sayantan Chakraborty, Aditya Nema, and Pranab Sen. Novel one-shot inner bounds for unassisted fully quantum
channels via rate splitting. In 2021 IEEE International Symposium on Information Theory (ISIT), pages 628–633,
2021.

[5] Sayantan Chakraborty, Aditya Nema, and Pranab Sen. One-shot inner bounds for sending private classical
information over a quantum mac. In 2021 IEEE Information Theory Workshop (ITW), pages 1–6, 2021.

[6] E. B. Davies and J. T. Lewis. An operational approach to quantum probability. Communications in Mathematical
Physics, 17(3):239–260, Sep 1970.

[7] Edward B. Davies. Quantum Theory of Open Systems. Academic Press, 1976.

[8] I. Devetak and A. Winter. Classical data compression with quantum side information. Phys. Rev. A, 68:042301,
Oct 2003.

36
 [9] Igor Devetak, Aram W. Harrow, and Andreas J. Winter. A resource framework for quantum shannon theory.
IEEE Transactions on Information Theory, 54(10):4587–4618, 2008.

[10] Lukas Drescher and Omar Fawzi. On simultaneous min-entropy smoothing. IEEE International Symposium on
Information Theory - Proceedings, 12 2013.

[11] A.J. Grant, B. Rimoldi, R.L. Urbanke, and P.A. Whiting. Rate-splitting multiple access for discrete memoryless
channels. IEEE Transactions on Information Theory, 47(3):873–890, 2001.

[12] Ke Li. Second-order asymptotics for quantum hypothesis testing. The Annals of Statistics, 42(1):171 – 189,
2014.

[13] Arun Padakandla. Optimal simulation of quantum measurements via the likelihood povms, 2021.

[14] Joseph M. Renes and Jean-Christian Boileau. Physical underpinnings of privacy. Phys. Rev. A, 78:032335, Sep
2008.

[15] Joseph M. Renes and Renato Renner. One-shot classical data compression with quantum side information and
the distillation of common randomness or secret keys. IEEE Transactions on Information Theory, 58(3):1985–
1991, 2012.

[16] Marco Tomamichel, Roger Colbeck, and Renato Renner. A fully quantum asymptotic equipartition property.
IEEE Transactions on Information Theory, 55(12):5840–5847, 2009.

[17] Marco Tomamichel and Masahito Hayashi. A hierarchy of information quantities for finite block length analysis
of quantum tasks. IEEE Transactions on Information Theory, 59(11):7693–7710, 2013.

[18] Marco Tomamichel and Vincent Y. F. Tan. Second-order asymptotics for the classical capacity of image-additive
quantum channels. Communications in Mathematical Physics, 338(1):103–137, Aug 2015.

[19] Mark M. Wilde. Position-based coding and convex splitting for private communication over quantum channels.
Quantum Information Processing, 16(10):264, Sep 2017.

[20] Mark M Wilde, Patrick Hayden, Francesco Buscemi, and Min-Hsiu Hsieh. The information-theoretic costs of
simulating quantum measurements. Journal of Physics A: Mathematical and Theoretical, 45(45):453001, oct
2012.

[21] Andreas Winter. “extrinsic” and “intrinsic” data in quantum measurements: Asymptotic convex decomposition
of positive operator valued measures. Communications in Mathematical Physics, 244(1):157–185, Jan 2004.

37