Calculating optimal bounds on the multi-parameter

estimation of Gaussian quantum states

Dominik Šafránek1, ∗
1
SCIPP and Department of Physics, University of California, Santa Cruz, CA 95064, USA
(Dated: January 3, 2018)
We derive several expressions for the quantum Fisher information matrix (QFIM) for multi-mode
Gaussian quantum states. This quantity determines the ultimate precision with what parameters
encoded into these states can be estimated. The derived expressions involve expression for a multi-
mode Gaussian state for the case when the Williamson’s decomposition of the covariance matrix
is known, and an expression for a multi-mode state in terms of infinite series, together with the
arXiv:1801.00299v1 [quant-ph] 31 Dec 2017

estimate of the remainder. We also discuss problematic behavior when some of the modes are pure,
and present a method that allows to use expressions that have been previously defined only for
mixed Gaussian states, to compute QFIM for states with any number of pure modes.

With growing need for precise and cost-effective detec-
tors, quantum metrology paves the way in which the new-
era quantum sensors should be designed. This theory
gives the ultimate bound on the precision with what pa-
rameters of a physical system can be measured by a quan-
tum probe. This bound, called the quantum Cramér-Rao
bound [1], might not be generally achievable, either be-
cause of practical limitations, or not even in principle [2].
But it still gives an estimate whether a certain design of
a new quantum detector is feasible, and whether it can
yield a better result than the current technology.
To determine this bound, it is necessary to compute
the quantum Fisher information matrix (QFIM), denoted
H(ǫ). Introduced by Holevo [3], Helstrom [4, 5], by Bu-
res [6], and later popularized by Braunstein and Caves [7],
QFIM describe limits in distinguishability of infinites-
imally close quantum states ρ̂ǫ and ρ̂ǫ+dǫ that differ
only by a small variation in parameters that parametrize
them. Generally, larger elements of QFIM predict bet-
ter distinguishability, which therefore leads to a better
precision in estimating vector of parameters ǫ.
There has been a wide range of applicability of this
quantity, such as optical interferometry used in the de-
tection of gravitational waves [8] and litography [9], ap-
plications in thermometry [10–12], estimating space-time
parameters [13–17], magnetic fields [18–20], squeezing pa-
rameters [21–23], time [24], and frequency [25, 26].
QFIM has been also used in the description of crit-
icality and quantum phase transitions under the name
of ‘fidelity susceptibility’ where they help to describe a
sudden change of a quantum state when an external pa-
rameter such as temperature is varied [27–30]. Moreover,
QFIM measures speed limits on the evolution of quantum
states [31, 32], speed limits of quantum computation [33],
and speed limits in charging of batteries [34].
Many recent experimental setups use Gaussian quan-
tum states. This is because these states are easily cre-
ated, and easily manipulated. They have been used for
∗
dsafrane@ucsc.edu
example in the aforementioned detection of Gravitational
waves [35], and in Bose-Einstein condensates [36–38].
Naturally, one often wants to calculate QFIM for Gaus-
sian states. That can be a daunting task, because
they appear very complicated in the density matrix for-
malilsm. Consequently, there has been a lot of effort
to make these calculations easier by utilizing the phase-
space formalism, which allows to elegantly describe any
Gaussian state just by its first and the second moment.
Numerous expressions have been derived. For a single
parameter estimation, it is QFI for a pure state [39], for
a nearly pure state [40], for a single-mode state [41], for
a two-mode state [42], for a multi-mode state [43], for
a multi-mode state in terms of Williamson’s decomposi-
tion of the covariance matrix [42], and for a multi-mode
state in terms of infinite series (limit formula) [42]; for
a multi-parameter estimation, it is QFIM for a single-
mode state [41], for a multi-mode state in terms of ten-
sors [44], for a multi-mode state in terms of inverses of
super-operators [45], and QFIM for a multi-mode state
in terms of Williamson’s decomposition of the covariance
matrix [46] that differs in form from the single-parameter
result of [42].
In this paper, we complete the story by deriving sev-
eral missing expressions for the QFIM for the multi-
parameter estimation of multi-mode Gaussian states, and
show how they connect to each other. We also dis-
cuss and resolve problematic behavior when some of the
modes are pure. We devise a regularization procedure,
which allows us to use expressions for QFIM that were
originally valid only for Gaussian states with all modes
mixed to calculate the QFIM for Gaussian states with
any number of pure modes.
This paper is structured as follows. We first write an
elegant matrix form of the QFIM for multi-mode mixed
states. We use this expression to derive QFIM in terms
of the Williamson’s decomposition of the covariance ma-
trix, generalizing the result of [42] to a multi-parameter
estimation. We discuss and resolve the problematic be-
havior of expressions when some of the modes are pure,
and explain the aforementioned regularization procedure.
Then we derive the QFIM for a multi-mode state in terms

of infinite series (limit formula), which can be used for
efficient numerical calculations. Finally, we show how all
of these expressions connect to each other, and how they
connect to the well-known result for pure states.
Notation
We consider a Bosonic system with a set annihilation
and creation operators {ân , â†n }. We collect them into a
vector of operators  ∶= (â1 , . . . , âN , â†1 , . . . , â†N )T , where
N denotes the number of modes. Now we can write
commutation relations in an elegant form, [Â , Â ] =
m n†
K mn id, where id denotes the identity operator and
K =[ ]
I 0
(1)
0 −I
is a constant matrix called the symplectic form, and I
denotes the identity matrix.
In a Bosonic system one can define a special class of
continuous variable states called Gaussian states [47],
ρ̂, which are fully characterized by its first moments
dm = tr[ρ̂Â ] (the displacement vector ) and the second
m
moments σ mn = tr[ρ̂ {(Â − d)m , (Â − d)n† }] (the covari-
ance matrix ), with {⋅, ⋅} denoting the anti-commutator.
These definitions are known as the complex form, and
we will use this convention throughout this paper. This
description is equivalent to the real form used by some
authors [47–49]. We show how to switch between these
descriptions in Appendix A. For more information on the
real and the complex form see for example [23, 42, 50, 51].
In this paper lower indices will denote different matri-
ces, while upper indices will denote elements of a matrix.
Bar as in σ will denote the complex conjugate.
Mixed states
In tasks where we know the covariance matrix and dis-
placement vector of a Gaussian state, it is possible to
use result derived in [44] to calculate the QFIM. This
result can be expressed in an elegant matrix form: for a
Gaussian state (d, σ), the QFIM can be calculated as
H ij (ǫ) = vec[∂i σ] M−1 vec[∂j σ] + 2∂i d† σ −1 ∂j d, (2a)
1 †
2 N
M = σ ⊗ σ − K ⊗ K, (2b)
where ⊗ denotes the Kronecker product, vec[⋅] is a vec-
torization of a matrix, and ∂i ≡ ∂ǫi .
The above formula requires differentiating the dis-
placement vector and the covariance matrix, and invert-
ing two matrices. However, the above formula cannot be
used when at least one of the modes is in a pure state (at
least not without any modification), because M is not
invertible in that case. We will discuss this issue later,
and show how this can be resolved (see Eq. (11)).
Proof of how the result of [44] transforms into this ele-
gant matrix form can be found in Appendix A. There we
also show the analogous expression in the real form, and
matrix expressions for the symmetric logarithmic deriva-
tives, which can be used for finding the optimal measure-
ment strategy.
2
When the Williamson’s decomposition of the
covariance matrix is known
There are situations we know the Williamson’s decom-
position of the covariance matrix. This is for example
when we have control over preparation of the initial state,
and we are planning to estimate parameters encoded into
this state by a Gaussian unitary channel. This is a case
when it is useful to use an expression for calculating the
QFIM that uses this knowledge of decomposition, with-
out the actual need of calculating this decomposition.
According to the Williamson’s theorem [52–54] any
positive-definite matrix can be diagonalized by symplec-
tic matrices, σ = SDS † . D = diag(λ1 , . . . , λN , λ1 , . . . , λN )
is a diagonal matrix consisting of symplectic eigenvalues,
which are defined as the positive eigenvalues of the ma-
trix Kσ. It follows from Heisenberg uncertainty relations
that for all k, λk ≥ 1. S is a symplectic matrix satisfy-
ing the defining property of the complex form of the real
symplectic group Sp(2N, R),
S =[ ] , SKS † = K.
α β
(3)
β α
We define matrices Pi ∶= S −1 ∂i S, which are elements
of the Lie algebra associated with the symplectic group,
satisfying the defining properties of this group,
Pi = [ ] , Pi K + KPi† = 0.
Ri Qi
(4)
Qi R i
Common symplectic matrices, representing for example
a squeezing operation, phase-change, or a beam-splitter,
are derived in Refs. [23, 51].
Rewriting Eq. (2) in terms of the Williamson’s de-
composition of the covariance matrix, switching to
element-wise notation, and simplifying using identi-
ties (3) and (4), we derive we derive an exact expression
for the quantum Fisher information matrix of Gaussian
states in terms of the Williamson’s decomposition of the
covariance matrix,
(λk −λl )2 (λk +λl )2
H ij (ǫ)= ∑ R[Ri Rjkl ] + j ]
kl kl
R[Qi Qkl
k,l=1 λ λ
k l −1 λ λ
k l +1
N
+∑
∂i λk ∂j λk
2
+ 2∂i d† σ −1 ∂j d. (5)
k=1 λk − 1
R denotes the real part, Ri = α† ∂i α − β † ∂i β is a skew-
Hermitian and Qi = α† ∂i β−β † ∂i α a (complex) symmetric
matrix. This formula represents a multi-parameter gen-
eralization of the result for a single-parameter estimation
published in Ref. [42]. The full derivation can be found
in Appendix B.
The above formula can be used even for states with
some pure modes, defined by λk = 1 for some mode k.
We will show how in the next section.
When none of the modes are pure, i.e., all sym-
plectic eigenvalues are larger than one, we can rewrite

Eq. (5) in a very elegant way. Defining Hermitian matrix
̃kl ∶= √λk −λl Rkl , symmetric matrix Q
R ̃kl ∶= √λk +λl Qkl ,
i λk λl −1 i i λk λl +1 i
and diagonal matrix L ∶= diag(λ1 , . . . , λN ), QFIM can be
written as
1 ̃ ̃† ̃ ̃† ̃ ̃ † ̃ ̃ †
H ij (ǫ) = tr[R i Rj + Rj Ri + Qi Qj + Qj Qi ]
2 (6)
+ tr[(L2 − I)−1 ∂i L∂j L] + 2∂i d† σ −1 ∂j d.
Looking at the derivatives in each term, we can conclude
that QFIM consists of three qualitatively different terms:
the first term is connected to the change of orientation
and squeezing of the Gaussian state with small variations
in ǫ, the second to the change of purity, and the third to
the change of displacement.
When some modes are pure
Both Eq. (2) and Eq. (5) are not well defined for states
that have at least one mode in a pure state, i.e., λk = 1 for
some mode k, which also results in matrix M not being
invertible. It has been shown [55], and we explain it in
detail in Appendix C, that there are two unique ways of
defining QFIM at these problematic points, depending
on the quantity we want to obtain.
First, we will illustrate this on Eq. (5). To obtain the
QFIM, which is defined [1] though the symmetric loga-
rithmic derivatives Li as H ij (ǫ) ≡ 21 tr[{Li , Lj }ρ̂], we
define problematic terms as
(λk −λl )2
∂i λk ∂j λk
= 0, = 0, (7)
λk λl −1 λ2k − 1
for ǫ such that λk (ǫ) = λl (ǫ) = 1.
The continuous quantum Fisher information ma-
trix (cQFIM) is defined [55] as four-times the Bu-
res metric as Hcij ≡ 4g ij , where ∑i,j g ij (ǫ)dǫi dǫj ≡
√ ν→1
2(1 − F (ρ̂ǫ , ρ̂ǫ+dǫ )), and F denotes the Uhlmann’s fi-
delity [56]. To obtain the cQFIM, we define
(λk −λl )2 ∂i λk ∂j λk
= 0, = ∂i ∂j λk , (8)
λk λl −1 λ2k − 1
for ǫ such that λk (ǫ) = λl (ǫ) = 1.
QFIM and cQFIM are identical everywhere, apart from
those problematic points [55]. At those points, QFIM can
be discontinuous, while cQFIM is in some sense continu-
ous [55]. Defining Hessian matrix Hkij ∶= ∂i ∂j λk , we can
use the above equations to write relation
Hc (ǫ) = H(ǫ) + ∑ Hk (ǫ). (9)
k∶λk (ǫ)=1
By writing k ∶ λk (ǫ) = 1 we mean that the sum goes only
over values of k for which λk (ǫ) = 1. For any k such that
λk (ǫ) = 1, Hk (ǫ) is positive semi-definite, and we can
therefore write Hc ≥ H. Hc = H if and only if for all k
such that λk = 1, Hk = 0.
Regularization procedure
Now we will show how to treat cases when some modes
are pure in general, not limiting ourselves to already re-
solved case of Eq. (5). We can devise a regularization
3
procedure that will allow us to use expressions that work
only for states where all the modes are mixed, such as
Eq. (2), to compute the QFIM for any state. Similar
method has been already used for regularizing QFIM for
non-Gaussian states [55].
It goes as follows. First we multiply the covariance
matrix by regularization parameter ν > 1, and use some
expression, such as Eq. (2), to calculate the QFIM for
state (d, νσ). Then we perform limit ν → 1. The re-
sulting value will represent the correct QFIM for state
(d, σ).
To prove that, however, we have to check that this
limit leads to the proper definition of the problematic
points, as given by Eq. (C2). We take Eq. (5) as a study
case, but because this formula is general, the result will
be valid for any other expression for QFIM. When co-
variance matrix σ has symplectic eigenvalues λk , covari-
ance matrix νσ has symplectic eigenvalues νλk . Sym-
pletic matrices from the decompositions of σ and νσ are
identical. Parameter ν therefore appears only as a mod-
ification of symplectic eigenvalues, which we will take
advantage of. Assuming λk (ǫ) = λl (ǫ) = 1 and per-
forming the limit, both problematic terms are set to zero
2
by taking the limit, limν→1 (νλ k −νλl )
= limν→1 ν 20−1 = 0,
νλk νλl −1
2
(ν∂i λk )
limν→1 (νλ 2
k ) −1
= limν→1 ν 20−1 = 0, which is exactly the
definition, Eq. (C2), that we wanted. ∂i λk (ǫ) = 0, be-
cause λk (ǫ) achieves a local minimum at point ǫ when
λk (ǫ) = 1.
This method therefore leads to the correct value of the
QFIM, and we can write expression for the QFIM for any
Gaussian quantum state as
H(ǫ) ≡ H(d(ǫ), σ(ǫ)) = lim H(d(ǫ), νσ(ǫ)). (10)
Applying this result to Eq. (2), QFIM for any state
can be computed as
ν2
H ij (ǫ) = lim vec[∂i σ] (ν 2 σ ⊗ σ − K ⊗ K)−1 vec[∂j σ]
†
ν→1 2
+ 2∂i d† σ −1 ∂j d.
(11)
Expression for the cQFIM can obtained by combining
Eq. (10) with Eq. (9).
Limit formula
We presented exact analytical expressions for QFIM,
however, in some cases a numerical value that approxi-
mates the exact value to any desired precision is enough.
Defining matrix A ∶= Kσ, and generalizing procedure de-
rived in Ref. [42] for a single-parameter estimation, we
derive the limit expression for the QFIM,
1 ∞
H ij (ǫ) = ∑ tr[A ∂i AA ∂j A] + 2∂i d σ ∂j d. (12)
−n −n † −1
2 n=1
This expression is proved in the next section, by showing
its relation to Eq. (2).
 4

Note that although the above infinite series converges which gives Eq. (2).
even when some symplectic eigenvalue are equal to one, Pure states
for such cases it is not absolutely convergent and it does
Combining Eq. (12), the regularization procedure (10),
and A2 (ǫ) = I (which holds for pure states because for
not give the correct expression for the QFIM. This can
them, λk (ǫ) = 1 for all k), we obtain the well-known
be shown by careful analysis of the elements in the se-
ries given by Eq. (14), and it was explained in detail in
Ref. [42]. The correct expression for cases when some result for pure states [39],
of the modes are pure can be obtained by combining
H ij (ǫ) = tr[σ −1 ∂i σσ −1 ∂j σ] + 2∂i d† σ −1 ∂j d.
the above equation with the regularization procedure, 1
Eq. (10). (16)
4
In applications, we would like to take just a few el-
ements of the series, and believe that their sum well It is important to note that this expression is discontin-
approximates the QFIM. To estimate the error when uous in ǫ, if the state becomes mixed when ǫ is slightly
varied, i.e., when ∂i ∂j λk (ǫ) ≠ 0 for some k (see Ref. [55]).
ij
doing this, we define remainder of the series as RM ∶=
2 ∑n=M+1
1 ∞ −n −n
tr[A ∂i AA ∂j A]. As shown in Appendix D,
To obtain cQFIM for states that are pure at point ǫ,
i.e., A2 (ǫ) = I, we can use expression
this remainder is bounded,
√ √
tr[(A∂i A)2 ] tr[(A∂j A)2 ]
∣RM ∣ ≤
ij
, (13)
2λmin (λ2min − 1) Hcij (ǫ) = (2tr[σ −1 ∂i ∂j σ] − tr[σ −1 ∂i σσ −1 ∂j σ])
2(M+1) 1
4 (17)
where λmin ∶= mink {λk } is the smallest symplectic eigen- + 2∂i d† σ −1 ∂j d,
value of the covariance matrix σ. The right hand side
therefore represents the maximal error when calculating To prove this expression, one needs to utilize the
the QFIM by using the first M elements of the series, Williamson’s decomposition of the covariance matrix,
Eq. (12). find tr[σ −1 ∂i ∂j σ] and tr[σ −1 ∂i σσ −1 ∂j σ] in terms of ma-
Relations between different expressions for trices K, Pi , Pij ∶= S −1 ∂i ∂j S (which will give expres-
QFIM sions similar to Eq. (14)), and use Eqs. (3), (4), and
Here we show how the limit expression, Eq. (12), re-
Pij K + Pi KPj† + Pj KPi† + KPji† = 0. When applied to
lates to Eqs. (5) and (2). Relation between Eqs. (5)
and (2) is shown in Appendix B. both Eq. (16) and Eq. (17), we find that Eq. (17) gives
Using identity the same expression as Eq. (16), plus an additional fac-
tor given by the second part of Eq. (9). This proves that
tr[A−n ∂i AA−n ∂j A] = 2tr[D−n+1K −n+1Pi D−n+1K −n+1Pj ] Eq. (17) represents the cQFIM for pure states.

− tr[D−n+2K nPi D−nK nPj )] − tr[D−n+2K nPj D−nK nPi )] Note that Eqs. (16) and (17) can be further simplified
by using σ −1 = KσK, which follows from A2 (ǫ) = I.
+ tr[D−n ∂i DD−n ∂j D]
Conclusion
(14)
In this paper we derived several new formulae for calcu-
and changing to element-wise notation, the infinite lating the quantum Fisher information matrix for multi-
sum (12) turns out to be geometric series in powers of mode Gaussian quantum states. Our main results are:
lk expression for the QFIM when the Williamson’s decom-
λk ’s, which can be evaluated. Then, using Rikl = −Ri ,
Qkl lk position of the covariance matrix is known, Eq. (5), which
i = Qi which follows from Eq. (4), we prove that
Eq. (12) simplifies to Eq. (5). can be used for example for finding optimal Gaussian
To obtain Eq. (2), we use σ † = σ, properties of vec- probe states for Gaussian unitary channels; the limit
† formula together with the estimate of the remainder,
torization, tr[A† B] = vec[A] vec[B], and properties of
Kronecker product, (AB) ⊗ (A′ B) = (A ⊗ A′ )(B ⊗ B ′ ),
Eqs. (12) and (13), which can be used for efficient nu-
(C T ⊗ A)vec[B] = vec[ABC], to transform the infinite
merical calculations, to any given precision; and regular-
ization procedure, Eq. (10), that allows to use expres-
sum in Eq. (12) into a Neumann series that can be eval-
sions valid only for mixed states, to calculate the QFIM
uated,
for any state. In addition, we discussed and resolved the
∞ problematic behavior of QFIM at the points of purity.
∑ tr[A−n ∂i AA−n ∂j A] Finally, we have shown how different formulas for QFIM
n=1 relate to each other, creating a coherent story among this
∞
= vec[∂i σ] ( ∑ (A ⊗ A)−n )(σ ⊗ σ) vec[∂j σ]
† −1 vast number of expressions, and providing a useful set of
(15) tools for calculating this quantity.
n=0

= vec[∂i σ] (I − A ⊗ A−1 )−1 (σ ⊗ σ) vec[∂j σ]

† −1 −1 Acknowledgement I would like to thank Tanja Fab-
sits and Karishma Hathlia, for reading the first version
= vec[∂i σ] (σ ⊗ σ − K ⊗ K)−1 vec[∂j σ],
†
of the manuscript, and for useful feedback.
 5

Appendix A: Derivation of formula for mixed states, Some authors also use the real form, which is defined
the real form, and the symmetric logarithmic with respect to the collection of quadrature operators
R = tr[ρ̂Q̂ ], σR
Q̂ = {x̂1 , x̂2 , . . . , p̂1 , p̂2 , . . .}, as dm
derivative. m mn
=
tr[ρ̂{∆Q̂ , ∆Q̂ }], ∆Q̂ ∶= Q̂ − dR , [Q̂ , Q̂ ] = i ΩR id.
m n m n mn
Here we use the general result of [44] to derive Eq. (2). The real form is connected to the complex form through
According to [44] while using the Einstein’s summation a unitary matrix
convention, the quantum Fisher information for N -mode
U=√ [ ]
Gaussian state can be calculated as 1 I +iI
(A8)
2 I −iI
Hi,j (ǫ) = M−1
1
∂j Σαβ ∂i Σµν + Σ−1 µ ν
µν ∂j λ ∂i λ . (A1)
2 αβ,µν as d = U dR , σ = U σR U † , and K = U iΩR U † . Using these
transformation relations, Eqs. (A6) and (A7), we derive
The displacement vector and the covariance matrix are
defined as λm = tr[ρ̂ÂG ] and Σmn = tr[ρ̂ {(ÂG −
m

H ij (ǫ) = vec[∂i σR ] M−1R vec[∂j σR ] + 2∂i dR σR ∂j dR ,

1 T T −1
λ)m , (ÂG − λ)n }], ÂG = (â1 , â†1 , . . . , âN , â†N )T , and the 2
symplectic form is given by [ÂG , ÂG ] =∶ Ωmn id. The
m n (A9)
αβ,µν where MR = σR ⊗ σR − ΩR ⊗ ΩR .
inverse of the tensor MG = Σαµ Σβν + 14 Ωαµ Ωβν is de- Using a similar approach, we can rewrite expressions
fined via for the symmetric logarithmic derivatives originally pub-
(M−1
G )µ̃ν̃,αβ MG
αβ,µν lished in Ref. [44] in an elegant matrix form,
= δµ̃µ δν̃ν , (A2)
Li = ∆ Ai ∆ − tr[σAi ] + 2∆A† σ −1 ∂i d,
† 1
(A10)
where δµ̃µ
denotes the Kronecker delta. Considering the 2
above definition, we can find a matrix form to Eq. (A1),
where ∆ ∶=  − d, vec[Ai ] ∶= M−1 vec[∂i σ]. The
quantum Fisher information matrix is then defined as
H ij (ǫ) = vec[∂i Σ] M−1
1
H ij (ǫ) = 12 tr[{Li , Lj }ρ̂].
T T −1
G vec[∂j Σ] + ∂i λ Σ ∂j λ, (A3)
2

where MG = Σ ⊗ Σ + 14 Ω ⊗ Ω, ⊗ denotes the Kronecker

product, and vec[⋅] is a vectorization of a matrix. Appendix B: Derivation of the formula for the case
when the Williamson decomposition is known
To obtain the result in our notation we need to consider
transformation relations
Here we use Eq. (2) to derive Eq. (5). Using the
σ = 2P ΣXP T , Williamson decomposition σ = SDS † , identities (3) and
(A6) we derive
K = P ΩXP T , (A4)
d = P λ, M−1 = ((S −1)T ⊗(KSK))(D⊗D−K ⊗K)−1((KSK)T ⊗S −1),
∂i σ = ∂i SDKS −1 K + S∂i DKS −1 K − SDKS −1 ∂i SS −1 K.
where X = ⊕N 2
σx (X is real and X = I) and P is a
i=1 (B1)
certain permutation matrix (P is real and P P T = I).
Using properties The first part of Eq. (2) then reads

XΣX = Σ, 1 †
(A5) vec[∂i σ] M−1 vec[∂j σ] =
XΩX = −Ω, 2
†
vec[Pi D−DKPi K +∂i D] M−1 diag vec[Pj D−DKPj K +∂j D],
the fact that Ω is real, and identities (B2)
(ABC) ⊗ (A′ B ′ C ′ ) = (A ⊗ A′ )(B ⊗ B ′ )(C ⊗ C ′ ), where Mdiag = D ⊗D −K ⊗K. Solving Eq. (A2) for the
(A6)
(C ⊗ A)vec[B] = vec[ABC],
T diagonal matrix Mdiag we find

δµ̃µ δν̃ν
(M−1
diag )µ̃ν̃,µν
we derive
= . (B3)
Dµµ Dνν − K µµ K νν
H ij (ǫ) = vec[∂i σ] M−1 vec[∂j σ] + 2∂i d† σ −1 ∂j d, (A7)
1 †
2 Changing to element-wise notation and using Einstein’s
summation convention (µ, ν ∈ {1, 2N }, k, l ∈ {1, N }) we
where M = σ ⊗ σ − K ⊗ K. expand Eq. (B2),
 6
1 †
vec[∂i σ] M−1 vec[∂j σ]
2
δµ̃µ δν̃ν
= (Pi D − DKPi K + ∂j D)µ̃ν̃ µµ νν (Pj D − DKPj K + ∂j D)µν
1
2 D D − K µµ K νν
1 (Pi D − D K Pi K + ∂i D δ )(Pj D − D K Pj K + ∂j D δ )
µν νν µµ µµ µν νν µµ µν µν νν µµ µµ µν νν µµ µν
=
2 Dµµ Dνν − K µµ K νν
= ( (Ri λl − λk Ri + ∂i λk δ kl )(Rjkl λl − λk Rjkl + ∂j λk δ kl )
1 1 kl kl
2 λk λl − 1

(Qi λl + λk Qi )(Qkl j λl + λk Qj )
1 kl kl kl
+
λk λl + 1
(B4)
(Ri kl λl − λk Ri kl + ∂i λk δ kl )(Rj λl − λk Rj + ∂j λk δ kl )
1 kl kl
+
λk λl − 1

i λl + λk Qi )(Qj λl + λk Qj ))
(Qkl
1 kl kl kl
+
λk λl + 1

= ( ((λl − λk )Ri + ∂i λk δ kl )((λl − λk )Rjkl + ∂j λk δ kl ) + (λl + λk )2 Qi Qkl

1 1 kl 1 kl
j
2 λk λl − 1 λk λl + 1

((λl − λk )Ri kl + ∂i λk δ kl )((λl − λk )Rj + ∂j λk δ kl ) + (λl + λk )2 Qkl

i Qj )
1 kl 1 kl
+
λk λl − 1 λk λl + 1
(λl − λk )2 (λl + λk )2
R[Ri Rjkl ] + j ]+
kl kl ∂i λk ∂j λk
= R[Qi Qkl ,
λk λl − 1 λk λl + 1 λ2k − 1

which in combination with term 2∂i d† σ −1 ∂j d gives We assume that defining problematic terms this way
Eq. (5). is valid also when only some symplectic eigenvalues are
equal to 1, but not all of them as in the case of pure
Gaussian states. Of course, this could also be precisely
Appendix C: Problematic behavior at the points of derived by carefully studying this intermediate case, for
purity example by using methods introduced in Ref. [43] for a
single-parameter estimation: solving the Stein equation
It follows from the Heisenberg uncertainty relations for the symmetric logarithmic derivatives, and utilizing
that λk ≥ 1, and λk = 1 if and only if the mode is in the Moore-Penrose pseudoinverse for writing the solution
a pure state. The multi-mode Gaussian state is pure for the QFIM.
when all symplectic eigenvalues are equal to one. Here Now, according to Ref. [55], the cQFIM can be calcu-
we show why terms in Eq. (5) that are problematic for lated from QFIM by performing a special limit:
pure modes, are defined by Eq. (C2) for QFIM, and by
Eq. (8) for cQFIM respectively. Hcij (ǫ) = lim H ij (ǫ + dǫ ei ) = lim H ij (ǫ + dǫ ej ), (C3)
dǫ→0 dǫ→0
Let us first take a look at the QFIM. We can ask how
to define problematic terms in Eq. (5), which has been where ei = (0, . . . , 0, 1, 0 . . . , 0) denotes a unit vector with
derived for λk > 1, so that it can also be valid for λk = 1. number 1 at the i’th position.
Rewriting formula for QFIM for pure Gaussian states, Let ǫ be a point such λk (ǫ) = 1 for some k, and σ(ǫ) ∈
Eq. (16), in terms of the Williamson’s decomposition of C (2) . The function λk must achieve the local minimum
the covariance matrix yields at point ǫ, because λ(ǫ) ≥ 1. The Taylor expansion must
N be of the form
H ij (ǫ) = 2 ∑ R[Qi Qkl
j ] + 2∂i d σ ∂j d.
kl † −1
(C1)
λk (ǫ + dǫ) = 1 + dǫT Hk dǫ + ⋯,
1
k,l=1 (C4)
2
Since for pure Gaussian states λk (ǫ) = 1 for all k, we can
easily see that Eq. (5) becomes identical with Eq. (C1), where Hkij = ∂i ∂j λk is the positive semi-definite matrix
when we define problematic terms as called Hessian. The Taylor expansion of the derivative
at the same point is
(λk −λl )2
(ǫ) = 0, (ǫ) = 0,
∂i λk ∂j λk
(C2)
∂i λk (ǫ+dǫ) = (eTi Hk dǫ+dǫT Hk ei )+⋯ = ∑ Hkij dǫj +⋯.
1
λk λl −1 λ2k − 1
2 j
for ǫ such that λk (ǫ) = λl (ǫ) = 1. (C5)
 7

Specifically, for dǫ ≡ dǫei we have where Bi = S † ∂i A(S † )−1 K −n−1 . We can derive inequali-
ties
λk (ǫ + dǫei ) = 1 + Hkii dǫ2 + ⋯, RRR RR RR RR
1
R R R
∣an ∣ = RRRR∑ n n Bikl Bjlk RRRR ≤ RRRR∑ n n ∣Bikl ∣ ∣Bjlk ∣RRRR
(C6a) 1 1
RRR k,l λ λ RRR RRR k,l λ λ RRR
2
∂i λk (ǫ + dǫei ) = Hkii dǫ + ⋯, (C6b) R R R R
R RR √ √
k l k l

1 RRRR R
∂j λk (ǫ + dǫei ) = Hkij dǫ + ⋯. ≤ 2n RR∑ ∣Bikl ∣ ∣Bjlk ∣RRRR ≤ 2n ∑ ∣Bikl ∣ ∑ ∣Bjkl ∣
(C6c) 1 2 2
R
λmin RR k,l R
RRR λmin k,l
√R
√
k,l

tr[Bi† Bi ] tr[Bj† Bj ],
According to Eq. (C3), and assuming ǫ is such that 1
λk (ǫ) = λl (ǫ) = 1, to obtain the cQFIM Hc , the problem- = 2n
λmin
atic terms in Eq. (5) must be defined as (D2)

(λk −λl )2 (λk (ǫ + dǫ ei )−λl (ǫ + dǫ ei ))2

(ǫ)∶= lim
where the last inequality is the Cauchy-Schwarz inequal-
λk λl −1 dǫ→0 λk (ǫ + dǫ ei )λl (ǫ + dǫ ei )−1 ity between Bikl and Bjlk considered as vectors with
(Hkii − Hlii )2 dǫ4 + ⋯
2N × 2N entries where N is number of modes, λmin ∶=
mini {λi } is the smallest symplectic eigenvalue. Defining
1
4
(Hkii + Hlii )dǫ2 + ⋯
= lim 1 = 0, (C7a)
dǫ→0
2 Hermitian matrix Ci ∶= S † ∂i AKS we have
∂i λk (ǫ + dǫ ei )∂j λk (ǫ + dǫ ei )
(ǫ)∶= lim
∂i λk ∂j λk
tr[(A∂i A)2 ] = tr[Ci† DCi D] = ∑ ∣Cikl ∣ λk λl
2
2
λk − 1 dǫ→0 λ2k (ǫ + dǫ ei ) − 1
k,l (D3)
Hii Hij dǫ2 + ⋯ λ2min tr[Ci† Ci ] = λ2min tr[Bi† Bi ].
= lim k ii k 2 = Hkij . ≥
dǫ→0 H dǫ + ⋯
(C7b)
k
Combining (D2) and (D3) gives
√ √
∣an ∣ ≤ tr[(A∂i A)2 ] tr[(A∂j A)2 ]λ−2n−2
min . (D4)
Appendix D: Estimation of the remainder in the
limit formula For λmin > 1 we can estimate the remainder,
√ √
tr[(A∂i A)2 ] tr[(A∂j A)2 ] ∞ −2n−2
Here we prove the bound on the remainder of the limit ∣RM ∣ ≤ ∑ λmin
formula, Eq. (13). We consider the Williamson decom- 2
√ √
n=M+1
position σ = SDS † . An element of the infinite series, (D5)
Eq. (12), can be written as tr[(A∂i A)2 ] tr[(A∂j A)2 ]
min (λmin − 1)
= .
2λ2M+2 2

an = tr[A−n ∂i AA−n ∂j A] = tr[D−n Bi D−n Bj ], (D1)

[1] M. G. A. Paris, International Journal of Quantum Infor-
mation 07, 125 (2009).
[2] M. Szczykulska, T. Baumgratz, and A. Datta, Adv.
Phys. X 1, 621 (2016).
[3] A. Holevo, Probabilistic and Statistical Aspects of Quan-
tum Theory (Springer Berlin Heidelberg, 2011).
[4] C. Helstrom, Physics letters A 25, 101 (1967).
[5] C. W. Helstrom, Quantum detection and estimation the-
ory, Vol. 123 (Academic press, 1976).
[6] D. Bures, Trans. Amer. Math. Soc. 135, 199 (1969).
[7] S. L. Braunstein and C. M. Caves, Phys. Rev. Lett. 72,
3439 (1994).
[8] B. P. e. a. Abbott (LIGO Scientific Collaboration and
Virgo Collaboration), Phys. Rev. Lett. 116, 061102
(2016).
[9] A. N. Boto, P. Kok, D. S. Abrams, S. L. Braunstein, C. P.
Williams, and J. P. Dowling, Phys. Rev. Lett. 85, 2733
(2000).
[10] A. Monras and F. Illuminati, Phys. Rev. A 83, 012315
(2011).
[11] L. A. Correa, M. Mehboudi, G. Adesso, and A. Sanpera,
Phys. Rev. Lett. 114, 220405 (2015).
[12] P. P. Hofer, J. B. Brask, M. Perarnau-Llobet, and
N. Brunner, Phys. Rev. Lett. 119, 090603 (2017).
[13] P. Nation, M. Blencowe, A. Rimberg, and E. Buks, Phys.
Rev. Lett. 103, 087004 (2009).
[14] S. Weinfurtner, E. W. Tedford, M. C. J. Penrice, W. G.
Unruh, and G. A. Lawrence, Phys. Rev. Lett. 106,
021302 (2011).
[15] M. Aspachs, G. Adesso, and I. Fuentes, Phys. Rev. Lett.
105, 151301 (2010).
[16] S. P. Kish and T. C. Ralph, Phys. Rev. A 96, 041801
(2017).
[17] M. Fink, A. Rodriguez-Aramendia, J. Handsteiner,
A. Ziarkash, F. Steinlechner, T. Scheidl, I. Fuentes,
J. Pienaar, T. C. Ralph, and R. Ursin, Nat. Commun.
8 (2017).
[18] W. Wasilewski, K. Jensen, H. Krauter, J. J. Renema,
M. Balabas, and E. S. Polzik, Phys. Rev. Lett. 104,
133601 (2010).
[19] J. Cai and M. B. Plenio, Phys. Rev. Lett. 111, 230503
(2013).
[20] R. Nair and M. Tsang, Phys. Rev. Lett. 117, 190801
(2016).
[21] M. Aspachs, J. Calsamiglia, R. Muñoz Tapia, and
E. Bagan, Phys. Rev. A 79, 033834 (2009).
[22] R. Gaiba and M. G. A. Paris, Phys. Lett. A 373, 934
(2009).

[23] D. Šafránek and I. Fuentes, arXiv:1603.05545 [quant-ph]
(2016).
[24] P. Komar, E. M. Kessler, M. Bishof, L. Jiang, A. S.
Sørensen, J. Ye, and M. D. Lukin, Nat. Phys. 10, 582
(2014).
[25] F. Fröwis, M. Skotiniotis, B. Kraus, and W. Dür, New
J. Phys. 16, 083010 (2014).
[26] J. Boss, K. Cujia, J. Zopes, and C. Degen, Science 356,
837 (2017).
[27] G.-S. Paraoanu and H. Scutaru, Phys. Rev. A 58, 869
(1998).
[28] P. Zanardi, L. CamposVenuti, and P. Giorda, Phys. Rev.
A 76, 062318 (2007).
[29] L. CamposVenuti and P. Zanardi, Phys. Rev. Lett. 99,
095701 (2007).
[30] S.-J. Gu, Int. J. Mod. Phys. B 24, 4371 (2010).
[31] N. Margolus and L. B. Levitin, Physica D: Nonlinear Phe-
nomena 120, 188 (1998).
[32] D. P. Pires, M. Cianciaruso, L. C. Céleri, G. Adesso, and
D. O. Soares-Pinto, Phys. Rev. X 6, 021031 (2016).
[33] S. Lloyd, Nature 406, 1047 (2000).
[34] F. C. Binder, S. Vinjanampathy, K. Modi, and J. Goold,
New J. Phys. 17, 075015 (2015).
[35] J. Aasi, J. Abadie, B. Abbott, R. Abbott, T. Abbott,
M. Abernathy, C. Adams, T. Adams, P. Addesso, R. Ad-
hikari, et al., Nat. Photonics 7, 613 (2013).
[36] J. Dunningham, M. Collett, and D. Walls, Phys. Lett.
A 245, 49 (1998).
[37] C. Gross, H. Strobel, E. Nicklas, T. Zibold, N. Bar-Gill,
G. Kurizki, and M. Oberthaler, Nature 480, 219 (2011).
[38] A. C. Wade, J. F. Sherson, and K. Mølmer, Phys. Rev.
A 93, 023610 (2016).
8
[39] O. Pinel, J. Fade, D. Braun, P. Jian, N. Treps, and
C. Fabre, Phys. Rev. A 85, 010101 (2012).
[40] N. Friis, M. Skotiniotis, I. Fuentes, and W. Dür, Phys.
Rev. A 92, 022106 (2015).
[41] O. Pinel, P. Jian, N. Treps, C. Fabre, and D. Braun,
Phys. Rev. A 88, 040102 (2013).
[42] D. Šafránek, A. R. Lee, and I. Fuentes, New J. Phys.
17, 073016 (2015).
[43] A. Monras, arXiv: 1008.2849 [quant-ph] (2013).
[44] Y. Gao and H. Lee, Eur. Phys. J. D 68, 347 (2014).
[45] L. Banchi, S. L. Braunstein, and S. Pirandola, Phys.
Rev. Lett. 115, 260501 (2015).
[46] R. Nichols, P. Liuzzo-Scorpo, P. A. Knott, and
G. Adesso, arXiv:1711.09132 [quant-ph] (2017).
[47] C. Weedbrook, S. Pirandola, R. Garcı́a-Patrón, N. J.
Cerf, T. C. Ralph, J. H. Shapiro, and S. Lloyd, Rev.
Mod. Phys. 84, 621 (2012).
[48] S. L. Braunstein and P. Van Loock, Rev. Mod. Phys. 77,
513 (2005).
[49] G. Adesso, A. R. Lee, and S. Ragy, Open Syst. Inf. Dyn.
21, 1440001 (2014).
[50] Arvind, B. Dutta, N. Mukunda, and R. Simon, Pramana
45, 471 (1995).
[51] D. Šafránek, arXiv:1610.03538 [quant-ph] (2016).
[52] J. Williamson, Am. J. of Math. 58, 141 (1936).
[53] M. A. de Gosson, Symplectic geometry and quantum me-
chanics, Vol. 166 (Springer Science & Business Media,
2006).
[54] R. Simon, S. Chaturvedi, and V. Srinivasan, J. Math.
Phys. 40, 3632 (1999).
[55] D. Šafránek, Phys. Rev. A 95, 052320 (2017).
[56] A. Uhlmann, Rep. Math. Phys. 9, 273 (1976).