Heisenbergs Uncertainty Principle

P. Busch

Perimeter Institute for Theoretical Physics, Waterloo, Canada,

and Department of Mathematics, University of York, York, UK
T. Heinonen

and P.J. Lahti

Department of Physics, University of Turku, Turku, Finland

(Dated: 5 December 2006)
Heisenbergs uncertainty principle is usually taken to express a limitation of operational possibil-
ities imposed by quantum mechanics. Here we demonstrate that the full content of this principle
also includes its positive role as a condition ensuring that mutually exclusive experimental options
can be reconciled if an appropriate trade-o is accepted. The uncertainty principle is shown to
appear in three manifestations, in the form of uncertainty relations: for the widths of the position
and momentum distributions in any quantum state; for the inaccuracies of any joint measurement
of these quantities; and for the inaccuracy of a measurement of one of the quantities and the
ensuing disturbance in the distribution of the other quantity. Whilst conceptually distinct, these
three kinds of uncertainty relations are shown to be closely related formally. Finally, we survey
models and experimental implementations of joint measurements of position and momentum and
comment briey on the status of experimental tests of the uncertainty principle.
PACS numbers: 03.65
Contents
I. Introduction 2
II. From no joint sharp values to approximate joint localizations 3
III. Joint and sequential measurements 5
A. Joint measurements 6
B. Sequential measurements 6
C. On measures of inaccuracy 7
1. Standard measures of error and disturbance 7
2. Geometric measure of approximation and disturbance 8
3. Error bars 8
IV. From no joint measurements to approximate joint measurements 9
A. Commuting functions of position and momentum 9
B. Uncertainty relations for covariant approximations of position and momentum 9
C. Uncertainty relations for general approximate joint measurements 11
V. From no measurement without disturbance to sequential joint measurements 12
VI. Illustration: the Arthurs-Kelly model 12
VII. On experimental implementations and tests of the uncertainty principle 14
A. Tests of preparation uncertainty relations 14
B. On implementations of joint and sequential measurements 15
C. On some alleged violations of the uncertainty principle 15
VIII. Conclusion 17
References 18

Electronic address: pb516@york.ac.uk

Electronic address: teiko.heinonen@utu.fi

Electronic address: pekka.lahti@utu.fi

2
It is the theory which decides what can be observed.
(Albert Einstein according to Werner Heisenberg (1977))
I. INTRODUCTION
It seems to be no exaggeration to say that Heisenbergs uncertainty principle, symbolized by the famous inequality
for position and momentum,
q p , (1)
epitomizes quantum physics, even in the eyes of the scientically informed public. Nevertheless, still now, 80 years
after its inception, there is no general consensus over the scope and validity of this principle. The aim of this article is
to demonstrate that based on recent work, it has nally become possible to elucidate the full content of the uncertainty
principle in precise terms. This will be done for the prime example of the position-momentum pair.
The uncertainty principle is usually described, rather vaguely, as comprising one or more of the following no-go
statements, each of which will be made precise below:
(A) It is impossible to prepare states in which position and momentum are simultaneously arbitrarily well localized.
(B) It is impossible to measure simultaneously position and momentum.
(C) It is impossible to measure position without disturbing momentum, and vice versa.
The negative characterization of the uncertainty principle as a limitations of quantum preparations and mea-
surements has led to the widespread view that this principle is nothing but a formal expression of the principle of
complementarity. This limited perspective has led some authors to question the fundamental status of the relation
(1) (Scully et al., 1991).
Here we will show that the uncertainty principle does have an independent content and role that in our view has
not yet been duly recognized. In fact, instead of resigning to accepting the negative verdicts (A), (B), (C), it is
possible to adopt a positive perspective on the underlying questions of joint preparation and measurement: according
to the uncertainty principle, the qualitative relationship of a strict mutual exclusiveness of sharp preparations or
measurements of position and momentum is complemented with a quantitative statement on a trade-o between
competing degrees of the concentration of the distributions of these observables in state preparations or between the
accuracies in joint measurements. Similarly, it turns out that the extent of the disturbance of (say) momentum in a
measurement of position can be controlled if a limitation in the accuracy of that measurement is accepted.
Only if taken together, the statements (A), (B), (C) and their positive counterparts can be said to exhaust the
content of the uncertainty principle for position and momentum. It also follows that the uncertainty principle comprises
three conceptually distinct types of uncertainty relations.
We will give a systematic exposition of these three faces of the uncertainty principle, with an emphasis on elucidating
its positive role. After a brief discussion of the well known uncertainty relation for preparations, we focus on the
less well established measurement uncertainty relations, the formulation of which requires a careful discussion of joint
measurements, measurement accuracy and disturbance.
We present a fundamental result, proved only very recently, which constitutes the rst rigorous demonstration of the
fact that the uncertainty relation for measurement inaccuracies is not only a sucient but also a necessary condition
for the possibility of approximate joint measurements of position and momentum.
Finally, we discuss some models and proposed realizations of joint measurements of position and momentum and
address the question of possible experimental tests of the uncertainty principle.
The idea of the uncertainty principle ensuring the positive possibility of joint albeit imprecise measurements, which
is rather latent in Heisenbergs works
1
has been made fully explicit and brought to our attention by his former student
Peter Mittelstaedt, our teacher and mentor, to whom we dedicate this treatise.
1
A judicious reading of Heisenbergs seminal paper of 1927 (Heisenberg, 1927) shows that both the double role and the three variants
of the uncertainty principle discussed here are already manifest, if only expressed rather vaguely. In fact, in the abstract, Heisenberg
immediately refers to limitations of joint measurements; later in the paper, he links this with a statement of the uncertainty relation
for the widths of a Gaussian wave function and its Fourier transform; nally he gives illustrations by means of thought experiments in
which the idea of mutual disturbance is prominent.
3
II. FROM NO JOINT SHARP VALUES TO APPROXIMATE JOINT LOCALIZATIONS
Throughout the paper, we will only consider the case of a spin zero quantum system in one spatial dimension,
represented by the Hilbert space H = L
2
(R). The states of the system are described by positive trace one operators
on H, the pure states being given as the one-dimensional projections. We occasionally write [ ) [ for the pure
state in question, and we call unit vectors H vector states. We denote the set of vector states by H
1
.
The position and momentum of the system are represented as the Schrodinger pair of operators

Q,

P, where

Q(x) = x(x),

P(x) = i

(x). We denote their spectral measures by the letters Q and P, respectively, and
recall that they are Fourier-Plancherel connected. The probability of obtaining the value of position in a (Borel)
subset X of R on measurement in a vector state is then given by the formula p
Q

(X) = [Q(X)) =
_
X
[(x)[
2
dx.
Similarly, the probability of obtaining the value of momentum in a (Borel) set Y on measurement in a vector state
is given by p
P

(Y ) :=
_
Y
[

(p)[
2
dp, where

is the Fourier-Plancherel transform of .
In the formalization of all three types of uncertainty relations, we will use two dierent measures of the width, or
degree of concentration, of a probability distribution. These are the standard deviation and the overall width. The
standard deviations of position and momentum in a state are
(Q, ) :=
_
[

Q
2
) [

Q)
2
_
1/2
, (P, ) :=
_
[

P
2
) [

P)
2
_
1/2
. (2)
The overall width of a probability measure p on R is dened, for given (0, 1), as the smallest interval length
required to have probability greater or equal to 1 ; thus:
W

(p) := inf
X
[X[ [ p(X) 1 , (3)
where X run through all intervals in R. For the overall widths of the position and momentum distribution in a vector
state we will use the notation
W

1
(Q, ) := W

1
(p
Q

), W

2
(P, ) := W

2
(p
P

). (4)
The no-go statement (A) says, broadly speaking, that the distributions p
Q

and p
P

of position and momentum

cannot simultaneously (i.e., in the same state ) be arbitrarily sharply concentrated. To appreciate this fact, we
note rst that position and momentum, being continuous quantities, cannot be assigned absolutely sharp values since
they have no eigenvalues. But both quantities can separately have arbitrarily sharply concentrated distributions. We
discuss two dierent ways of formalizing this idea, in terms of standard deviations and overall widths. Each of these
formalizations gives rise to a precise form of (A). We then proceed to complement the no-go statement (A) with
descriptions of the positive possibilities of simultaneous approximate localizations of position and momentum.
The rst formalization of arbitrarily sharp localizations makes use of the standard deviation of a distribution (stated
here for position):
for any q
0
R and any > 0, there is a vector state such that [

Q) = q
0
and (Q, ) < . (5)
Thus there is no obstacle to concentrating the distributions of position or momentum arbitrarily sharply at any points
q
0
, p
0
R, if these observables are considered separately, on dierent sets of states.
But the corresponding state preparation procedures are mutually exclusive; this is the operational content of the
negative statement (A). What is positively possible if one considers both observables together in the same state will
be described by an appropriate uncertainty relation. Property (5) gives rise to the following formalization of (A):
Theorem 1 For all states and for any > 0,
if (Q, ) < , then (P, ) > /2; and vice versa. (6)
This is a statement about the spreads of the position and momentum probability distributions in a given state: the
sharper one is peaked, the wider the other must be. This limitation follows directly from the uncertainty relation for
standard deviations, valid for all vector states :
(Q, ) (P, )

2
. (7)
The vector states
a,b
(x) = (2a/)
1/4
e
(a+ib)x
2
, a, b R, a > 0, give (Q,
a,b
)
2
= 1/4a and (P,
a,b
)
2
=

2
(a
2
+b
2
)/a, so that the following positive statement complementing the no-go Theorem 1 is obtained:
4
Theorem 2 For all positive numbers q, p for which q p /2, there is a state such that (Q, ) = q and
(P, ) = p.
The vector state
a,0
is a minimal uncertainty state in the sense that it gives (Q,
a,0
) (P,
a,0
) = /2. Every
minimum uncertainty state is of the form e
icx

a,0
(xd) for some c, d R. Minimal uncertainty states have a number
of distinctive properties (see, e.g., (Dym and McKean, 1972)). For instance, if is a vector state which satises
[[
2
[
a,0
[
2
and [

[
2
[
a,0
[
2
, for some a, then is a minimal uncertainty state.
We now turn to the second way of saying that position and momentum can separately be localized arbitrarily well
(expressed again only for position):
for any bounded interval X (however small), there exists a vector state such that p
Q

(X) = 1. (8)
The corresponding formalization of (A) then is given by the following theorem.
Theorem 3 For all vector states and all bounded intervals X, Y , p
Q

(X) = 1 implies 0 ,= p
P

(Y ) ,= 1, and vice
versa.
This means that whenever the position is localized in a bounded interval then the momentum cannot be conned to
any bounded interval (nor to its complement), and vice versa.
For any two bounded intervals X and Y and for any vector state , Theorem 3 implies that p
Q

(X) + p
P

(Y ) < 2.
However, for any such intervals, one can construct a vector state
0
for which the sum of the probabilities p
Q

0
(X) and
p
P

0
(Y ) attains its maximum value. The precise statement is given in the following theorem, which can be regarded
as a positive complement to Theorem 3.
Theorem 4 For any vector state and for any bounded intervals X and Y ,
p
Q

(X) + p
P

(Y ) 1 +

a
0
< 2, (9)
where a
0
is the largest eigenvalue of the positive trace class operator Q(X)P(Y )Q(X). There exists an optimizing
vector state
0
such that
p
Q

0
(X) + p
P

0
(Y ) = 1 +

a
0
. (10)
This result follows from the work of Landau and Pollak (1961) and Lenard (1972) (for details, see (Lahti, 1986)).
We will say that position Q is approximately localized in an interval X for a given state whenever p
Q

(X) 1
for some (preferably small) , 0 < < 1, and similarly for momentum. Then Eq. (10) describes the maximum degree
of approximate localization that can be achieved in any phase space cell of given size [X[ [Y [.
The largest eigenvalue a
0
is invariant under a scale transformation applied simultaneously to

Q and

P; it is therefore
a function of the product [X[ [Y [ of the interval lengths [X[ and [Y [. A simple calculation gives tr [Q(X)P(Y )Q(X)] =
[X[ [Y [/(2), so that we obtain
[X[ [Y [ 2 a
0
. (11)
If both position and momentum are approximately localized within X and Y , respectively, so that p
Q

(X) 1
1
and p
P

(Y ) 1
2
, then due to inequality (9), one must have 1
1

2


a
0
, and then (11) implies:
[X[ [Y [ 2 (1
1

2
)
2
if p
Q

(X) 1
1
, p
P

(Y ) 1
2
, and
1
+
2
< 1. (12)
It is convenient to express this uncertainty relation for approximate localization widths in terms of the overall
widths:
W

1
(Q, ) W

2
(P, ) 2 (1
1

2
)
2
if
1
+
2
< 1. (13)
If
1
+
2
1, then the product of widths has no positive lower bound (Landau and Pollak, 1961). In the case

1
+
2
< 1, the inequality is tight in the sense that even for fairly small values of
1
,
2
, the product of overall widths
can be in the order of 2; we quote a numerical example given in (Landau and Pollak, 1961): if
1
=
2
= .01, then
[X[ [Y [ can still be as small as 6.25 (2).
We will make repeated use of this uncertainty relation; since we are only interested in small values of
1
,
2
, we
will assume these to be less than 1/2; then the condition
1
+
2
< 1 is fullled and need not be stated explicitly.
5
Several other measures of uncertainty have been introduced to analyze the degree of (approximate) localizability
of position and momentum distributions p
Q

and p
P

, ranging from extensive studies on the support properties of [[

2
and [

[
2
to various information theoretic (entropic) uncertainty relations. It is beyond the scope of this paper to
review the vast body of literature on this topic. The interested reader may consult e.g. (Cowling and Price, 1984),
(Folland and Sitaram, 1997) or (Busch et al., 1997, Sect. V.4) for reviews and references.
To summarize: instead of leaving it at the negative statement that position and momentum cannot be arbitrarily
sharply localized in the same state, the uncertainty relation for state preparations oers precise specications of the
extent to which these two observables can simultaneously be approximately localized.
III. JOINT AND SEQUENTIAL MEASUREMENTS
In order to go beyond the no-go theorems of (B) and (C) and establish their positive complements, one needs to
use the full-edged apparatus of quantum mechanics. The general representation of observables as positive operator
measures will be required to introduce viable notions of joint and sequential measurements and an appropriate
quantication of measurement inaccuracy. Furthermore, some tools of measurement theory will be needed to describe
and quantify the disturbance of one observable due to the measurement of another.
In the present context of the discussion of position and momentum and their joint measurements, observables will
be described as positive operator measures X E(X) on the (Borel) subsets of R or R
2
. This means that the map
X [E(X)) =: p
E

(X) is a normalized probability measure for every vector state . An observable E will be
called sharp if it is a spectral measure, that is, if all of the eects E(X) are projections.
For an observable E on R, we will make use of the notation E[1], E[2] for the rst and the second moment operators,
dened (weakly) as E[k] :=
_
x
k
E(dx) (k = 1, 2).
2
We let (E, ) denote the standard deviation of p
E

,
(E, )
2
:=
_

_
x
_

p
E

(dx

)
_
2
p
E

(dx) = [E[2]) [E[1])

2
. (14)
It is a remarkable feature of an observable E, dened as a positive operator measure, that it need not be commu-
tative; that is, it is not always the case that E(X
1
)E(X
2
) = E(X
2
)E(X
1
) for all sets X
1
, X
2
. This opens up the
possibility of dening a notion of joint measurability for not necessarily commuting families of observables.
Indeed, it will become evident below that in the set of noncommuting pairs of observables, the jointly measurable
ones are necessarily unsharp. It is to be expected intuitively that the degree of mutual noncommutativity determines
the necessary degree of unsharpness required to allow a joint measurement. Here we present two ways of indicating
the inherent unsharpness of an observable E on R.
We dene the intrinsic noise operator of E as
N
i
(E) := E[2] E[1]
2
. (15)
This is a positive operator. If E[1] is selfadjoint, then the intrinsic noise N
i
(E) is zero exactly when E is a sharp
observable (Kiukas et al., 2006b, Theorem 5). A measure
i
(E; ) of intrinsic noise is then given by the expectation
value of the intrinsic noise operator (for all vector states for which this expression is well dened):

i
(E; ) := [N
i
(E)). (16)
The overall intrinsic noise is dened as

i
(E) := sup
H
1

i
(E; ). (17)
The next denition applies to observables E on R whose support is R.
3
The resolution width of E (at condence
level 1 ) is (Carmeli et al., 2006)

(E) := infd > 0 [ for all x Rthere exists H

1
with p
E

(J
x;d
) 1 . (18)
2
It is important to bear in mind that the domain of the symmetric operator E[k] is not necessarily dense; it consists of all vectors H
for which the function x x
k
is integrable w.r.t. the complex measure X |E(X) for all H. Here and in subsequent formulas
it is understood that the expectations of unbounded operators are only well dened for appropriate subsets of states.
3
This means that for every interval J there is a vector state such that p
E

(J) = 0.
6
Here J
x;d
denotes the interval [x
d
2
, x +
d
2
].
We note that positive resolution width is a certain indicator that the observable E is unsharp. This measure
describes the possibilities of concentrating the probability distributions to a xed condence level across all intervals.
However, the requirement of vanishing resolution width does not single out sharp observables (Carmeli et al., 2006).
A. Joint measurements
Two observables E
1
and E
2
on R are called jointly measurable if there is an observable M on R
2
such that
E
1
(X) = M(X R), E
2
(Y ) = M(R Y ) (19)
for all (Borel) sets X, Y . Then E
1
and E
2
are the marginal observables M
1
and M
2
of the joint observable M. If either
E
1
or E
2
is a sharp observable, then they are jointly measurable exactly when they commute mutually. In that case,
the unique joint observable M is determined by M(X Y ) = E
1
(X)E
2
(Y ). In general, the mutual commutativity of
E
1
and E
2
is not a necessary (although still a sucient) condition for their joint measurability.
The above notion of joint measurability is fully supported by the quantum theory of measurement, which ensures
that E
1
and E
2
are jointly measurable exactly when there is a measurement scheme which measures both E
1
and E
2
(Busch et al., 1996).
Considering that the (sharp) position and momentum observables Q and P do not commute with each other, we
recover immediately the well-known fact that these observables have no joint observable, that is, they are not jointly
measurable. This is a precise formulation of the no-go statement (B).
In preparation of developing a positive complementation to (B), we give an outline of the notion that sharp position
and momentum may be jointly measurable in an approximate sense. While there is no observable on phase space
whose marginals coincide with Q and P, one can explore the idea that there may be observables M on R
2
whose
marginals M
1
and M
2
are approximations (in some suitably dened sense) of Q and P, respectively. Such an M will
be called an approximate joint observable for Q and P. An appropriate quantication of the dierences between M
1
and Q and between M
2
and P may serve as a measure of the (in)accuracy of the joint approximate measurement
represented by M.
B. Sequential measurements
In order to analyze measurements of two observables E
1
and E
2
performed in immediate succession, it is necessary
to take into account the inuence of the rst measurement on the object system. The tool to describe the state
changes due to a measurement is provided by the notion of an instrument; see, for instance, (Davies, 1976, Chapter
4).
Let 1
1
be the instrument associated with a measurement of E
1
, that is, 1
1
determines the probability tr [E
1
(X)] =
tr [1
1
(X)()] for every state and set X. The number tr [1
1
(X)()E
2
(Y )] is the sequential joint probability that the
measurement of E
1
, performed on the system in state , gives a result in X and a subsequent measurement of E
2
leads to a result in Y . Using the dual instrument 1

1
(i.e., the Heisenberg picture), this probability can be written as
tr [1

1
(X)(E
2
(Y ))]. The map (X, Y ) tr [1

1
(X)(E
2
(Y ))] is a probability bimeasure and therefore extends uniquely
to a joint probability for each , dening thus a unique joint observable M on R
2
via
M(X Y ) = 1

1
(X)(E
2
(Y )). (20)
Its marginal observables are
M
1
(X) = 1

1
(X)(E
2
(R)) = E
1
(X), (21)
M
2
(Y ) = 1

1
(R)(E
2
(Y )) =: E

2
(Y ). (22)
Thus the rst marginal is the rst-measured observable E
1
and the second marginal is a distorted version E

2
of E
2
.
This general consideration shows that one must expect that a measurement of an observable E
1
will disturb (the
distribution of) another observable E
2
. In fact, it is a fundamental theorem of the quantum theory of measurement
that there is no nontrivial measurement without some state changes. In other words, if a measurement leaves all
states unchanged, then its statistics will be the same for all states; in this sense there is no information gain without
some disturbance.
If the rst-measured observable E
1
is sharp, then the distorted eects 1

1
(R)(E
2
(Y )) must commute with E
1
(X) for
all X, Y , whatever the second observable E
2
is. Thus, if we consider a sequential measurement of the (sharp) position
7
and momentum observables Q and P as an attempted joint measurement, we see that such an attempt is bound to fail.
If (say) one rst measures position Q, with an instrument 1
Q
, then all distorted momentum eects P

(Y ) := 1

Q
(Y )
are functions of the position operator

Q. In this sense, a measurement of (sharp) position completely destroys any
information about the momentum distribution in the input state. This result formalizes the no-go statement (C).
The formulation of a positive complement to (C) is based on the idea that one may be able to control and limit the
disturbance due to a measurement of Q, by measuring an observable Q

which is an approximation (in some sense)

to Q. One can then hope to achieve that the distorted momentum P

is an approximation (in some sense) to P.

We note that this amounts to dening a sequential joint observable M with marginals M
1
= Q

and M
2
= P

. Any
appropriate quantication of the dierence between M
1
and Q is a measure of the inaccuracy of the rst (approximate)
position measurement; similarly any appropriate quantication of the dierence between M
2
and P is a measure of
the disturbance of the momentum due to the position measurement.
In this way the problem of dening measures of the disturbance of (say) momentum due to a measurement of
position has been reduced to dening the inaccuracy of the second marginal of a sequential joint measurement of rst
position and then momentum.
C. On measures of inaccuracy
The above discussion shows that it is the noncommutativity of observables such as position and momentum which
forces one to allow inaccuracies if one attempts to make an approximate joint measurement of these observables. This
shows clearly that the required inaccuracies are of quantum-mechanical origin, which will also become manifest in the
models of approximate position measurements and phase space measurements presented below. With this observation
as proviso, we believe that it is acceptable to use the classical terms of measurement inaccuracy and error, particularly
because their operational denitions are essentially the same as in a classical measurement context.
In fact, every measurement, whether classical or quantum, is subject to noise, which results in a deviation of the
actually measured observable E
1
from that intended to be measured, E. We will refer to this deviation and any
measure of it as error or inaccuracy. In general there can be systematic errors, or bias, leading to a shift of the mean
values, and random errors, resulting in a broadening of the distributions. Any measure of measurement noise should
be operationally signicant in the sense that it should be determined by the probability distributions p
E

and p
E
1

.
In the following we will discuss three dierent approaches to quantifying measurement inaccuracy.
1. Standard measures of error and disturbance
Classical statistical analysis suggests the use of moments of probability distributions for the quantication of error
and disturbance in measurements. Thus, the standard approach found in the literature of dening a measure of error
is in terms of the average deviation of the value of a readout observable of the measuring apparatus from the value
of the observable to be measured approximately. If these observables are represented as selfadjoint operators Z and
A (acting on the apparatus and the object Hilbert spaces), respectively, this standard error measure is given as the
root mean square
(Z, A, ) := U( )[(Z A)
2
U( ))
1/2
, (23)
where U is the unitary map modelling the measuring interaction and is the initial state of the apparatus. This
measure of error has been studied in recent years in the foundational context, for example, by Appleby (1998a,b),
Hall (2004) and Ozawa (2004).
If we denote by E the observable actually measured by the given scheme, we dene the relative noise operator,
N
r
(E, A) := E[1] A; (24)
the standard error can then be rewritten as (Ozawa, 2004)
(E, A; )
2
= [(E[1] A)
2
) +[(E[2] E[1]
2
)) = [N
r
(E, A)
2
) +[N
i
(E)) (25)
(for any vector state for which the expressions are well-dened). We note that (E, A; ) = 0 for all exactly when
E is sharp and E[1] = A. The relative noise term cannot, in general, be determined from the statistics of measurements
of E and A alone, so that the standard error measure (E, A; ) does not always satisfy the requirement of operational
signicance (Busch et al., 2004). However, we will encounter important cases where this quantity does turn out to be
operationally well dened.
8
The standard error is a state-dependent quantity. This stands in contrast to the fact that estimates of errors
obtained in a calibration process are meant to be applicable to a range of states since in a typical measurement the
state is unknown to begin with. In order to obtain state independet measures, we dene the global standard error of
an observable E relative to A as
(E, A) := sup
H
1
(E, A; ). (26)
We will say that E is a standard approximation to A if E has nite global standard error relative to A. This
denition provides a possible criterion for selecting joint or sequential measurements schemes as approximate joint
measurements of Q and P; but it is not always possible to verify this criterion if the standard error fails to be
operationally signicant.
2. Geometric measure of approximation and disturbance
Following the work of Werner (2004), we dene a distance d(E
1
, E
2
) on the set of observables on R.
We rst recall that for any bounded measurable function h : R R, the integral
_
R
hdE denes (in the weak sense) a
bounded selfadjoint operator, which we denote by E[h]. Thus, for any vector state the number [E[h])) =
_
R
hdp
E

is well-dened.
Denoting by the set of bounded measurable functions h : R R for which [h(x) h(y)[ [x y[, the distance
between the observables E
1
and E
2
is dened as
d(E
1
, E
2
) := sup
H
1
sup
h
[[(E
1
[h] E
2
[h]) )[ . (27)
This measure is operationally signicant, using only properties of the distributions to be compared. Furthermore,
it is a global measure in that it takes into account the largest possible deviations of the expectations [E
1
[h]) and
[E
2
[h]). It gives a geometrically appealing quantication of how well a given observable can be approximated by
other observables.
We will say that an observable E
1
is a geometric approximation to E
2
if d(E
1
, E
2
) < . We shall apply this
condition of nite distance as a criterion for a joint or sequential measurement scheme to dene an approximate joint
measurement of Q and P. It is not clear whether this criterion is practical since the distance is not related in any
obvious way to concepts of measurement inaccuracy commonly applied in an experimental context.
3. Error bars
We now present a denition of measurement inaccuracy in terms of likely error intervals that follows most closely
the usual practice of calibrating measuring instruments. In the process of calibration of a measurement scheme, one
seeks to obtain estimates of the likely error and perhaps also the degree of disturbance that the scheme contains.
To estimate the error, one tests the device by applying it to a suciently large family of input states in which the
observable one wishes to measure with this setup has fairly sharp values. The error is then characterized as an overall
measure of the bias and the width of the output distribution across a range of input values. Error bars give the
minimal average interval lengths that one has to allow to contain all output values with a given condence level.
For each (0, 1), we say that an observable E
1
is an -approximation to a sharp observable E if for all > 0 there
is a positive number w < such that for all x R, H
1
, the condition p
E

(J
x;
) = 1 implies that p
E
1

(J
x,w
) 1.
The inmum of all such w will be called the inaccuracy of E
1
with respect to E and will be denoted J
,
(E
1
, E).
Thus,
J
,
(E
1
, E) := infw[ for all x R : p
E
1

(J
x,w
) 1 for all H
1
with p
E

(J
x;
) = 1. (28)
The inaccuracy describes the range within which the input values can be inferred from the output distributions, with
condence level 1 , given initial localizations within . We note that the inaccuracy is a increasing function of ,
so that we can dene the error bar width
4
of E
1
relative to E:
J

(E
1
, E) := inf

J
,
(E
1
, E) = lim
0
J
,
(E
1
, E). (29)
4
This denition and all subsequent results based on it can be found in (Busch and Pearson, 2006).
9
If J

(E
1
, E) is nite for all (0,
1
2
), we will say that E
1
approximates E in the sense of nite error bars. We
note that the niteness of either (E
1
, E) or d(E
1
, E) implies the niteness of J

(E
1
, E). Therefore, among the three
measures of inaccuracy, the condition of nite error bars gives the most general criterion for selecting approximations
of Q and P.
IV. FROM NO JOINT MEASUREMENTS TO APPROXIMATE JOINT MEASUREMENTS
Position Q and momentum P have no joint observable, they cannot be measured together. However, one may ask
for an approximate joint measurement, that is, for an observable M on R
2
such that the marginals M
1
and M
2
are
appropriate approximations of Q and P. In this section we study two important cases and then consider the general
situation.
A. Commuting functions of position and momentum
The rst approach is related with the fact that although Q and P are noncommutative, they do have commuting
spectral projections. Indeed, let Q
g
be a function of Q, that is, Q
g
(X) = Q(g
1
(X)) for all X B(R), with g : R R
being a (Borel) function. Similarly, let P
h
be a function of momentum. The associated operators are g(

Q) and h(

P).
The following result, proved by Busch et al. (1987, Theorem 1) and in a more general setting by Ylinen (1989),
characterizes the functions g and h for which Q
g
(X)P
h
(Y ) = P
h
(Y )Q
g
(X) for all X and Y .
Theorem 5 Let g and h be essentially bounded Borel functions such that neither g(

Q) nor h(

P) is a constant operator.
The functions Q
g
of position and P
h
of momentum commute if and only if g and h are both periodic with minimal
positive periods a, b satisfying
2
ab
N.
If Q
g
and P
h
are commuting observables, then they have the joint observable M, with M(X Y ) = Q
g
(X)P
h
(Y ),
meaning that Q
g
and P
h
can be measured jointly. The price for this restricted form of joint measurability of position
and momentum as given by Theorem 5 is that they are to be coarse-grained by periodic functions g and h with
appropriately related minimal periods a, b.
The functions g and h can be chosen as characteristic functions of appropriate periodic sets. This allows one to
model a situation known in solid state physics, where an electron in a crystal can be conned arbitrarily closely to
the atoms while at the same time its momentum is localized arbitrarily closely to the reciprocal lattice points.
Simultaneous localization of position and momentum in periodic sets thus constitutes a sharp joint measurement
of functions of these observables. However, bounded functions Q
f
of Q provide only very bad approximations to Q
since (Q
f
, Q), d(Q
f
, Q) and J
,
(Q
f
, Q) are all innite. One also loses the characteristic covariance properties of
position and momentum.
B. Uncertainty relations for covariant approximations of position and momentum
Next we will discuss approximate joint measurements of position and momentum based on smearings of these
observables by means of convolutions.
Let , be probability measures on R. We dene observables Q

, P

via
Q

(X) =
_
R
(X q) Q(dq), P

(Y ) =
_
R
(Y p) P(dp). (30)
These observables have the same characteristic covariance properties as Q, P and they are approximations in the sense
that they have nite error bar widths relative to Q, P. Hence we call them approximate position and momentum.
For given Q

, P

we ask under what conditions they are jointly measurable, that is, there is an observable M on
R
2
such that M
1
= Q

and M
2
= P

. In order to answer this question, we need to introduce the notion of covariant

phase space observables.
Covariance is dened with respect to a unitary (projective) representation of phase space translations in terms of
the Weyl operators, dened for any phase space point (q, p) R
2
via W(q, p) = e
i
2
qp
e

q

P
e
i

p

Q
. An observable M
on R
2
is a covariant phase space observable if
W(q, p)M(Z)W(q, p)

= M(Z + (q, p)) (31)

10
for all Z. It is known
5
that each such observable is of the form G
T
, where
G
T
(Z) =
1
2
_
Z
W(q, p)TW(q, p)

dqdp, (32)
and T is a unique positive operator of trace one. The marginal observables G
T
1
and G
T
2
are of the form (30) where
=
T
= p
Q
T
and =
T
= p
P
T
and is the parity operator, (x) = (x).
Our question is answered by the following fundamental theorem which was proven in the present form by Carmeli
et al. (2004, Proposition 7) as a direct development of the work of Werner (2004).
Theorem 6 An approximate position Q

and an approximate momentum P

are jointly measurable if and only if

they have a covariant joint observable G
T
. This is the case exactly when there is a positive operator T of trace equal
to 1 such that =
T
, =
T
.
We next quantify the necessary trade-o in the quality of the approximations of Q, P by G
T
1
, G
T
2
, using the three
measures of inaccuracy introduced above. First we state uncertainty relations for the measures of intrinsic unsharpness.
The probability distributions associated with these marginals of G
T
are the convolutions p
Q

T
and p
P

T
; this
indicates that statistically independent noise is added to the distributions of sharp position and momentum. In
accordance with this fact, the standard deviations are obtained via the sums of variances,
(G
T
1
, )
2
= (Q, )
2
+ (
T
)
2
, (G
T
2
, )
2
= (P, )
2
+ (
T
)
2
. (33)
The noise interpretation is conrmed by a determination of the intrinsic noise operators of G
T
1
, G
T
2
, which are well-
dened whenever the operator T is such that

Q
2

T and

P
2

T are Hilbert-Schmidt operators (Kiukas et al., 2006b,

Theorem 4); in that case one obtains
N
i
(G
T
1
) = (
T
)
2
I, N
i
(G
T
2
) = (
T
)
2
I. (34)
Since T, and with it T

, has the properties of a state, the uncertainty relation (7) applies to the probability
measures
T
and
T
, giving the following uncertainty relation for intrinsic noise, valid for any G
T
:

i
(G
T
1
)
i
(G
T
2
) = (
T
) (
T
)

2
. (35)
Equations (33) and (35) yield the following version of state-preparation uncertainty relation with respect to G
T
,
which also reects the presence of the intrinsic noise:
(G
T
1
, ) (G
T
2
, ) . (36)
The resolution widths of G
T
1
, G
T
2
are given by

1
(G
T
1
) = W

1
(
T
),

2
(G
T
2
) = W

2
(
T
), (37)
so that the uncertainty relation for overall widths then entails:

1
(G
T
1
)

2
(G
T
2
) = W

1
(
T
) W

2
(
T
) 2 (1
1
2)
2
. (38)
The standard errors are
(G
T
1
, Q; )
2
=
_

T
[1]
_
2
+ (
T
)
2
, (G
T
2
, P; )
2
=
_

T
[1]
_
2
+ (
T
)
2
, (39)
and the inequality
(G
T
1
, Q) (G
T
2
, P)

2
(40)
holds as an immediate consequence of the noise uncertainty relation (35).
5
This result is due to Holevo (1979) and Werner (1984). Alternative proofs with dierent techniques were recently given by Cassinelli
et al. (2003) and Kiukas et al. (2006a).
11
The distances of G
T
1
, G
T
2
from Q, P are
d(G
T
1
, Q) =
_
[q[
T
(dq), d(G
T
2
, P) =
_
[p[
T
(dp), (41)
and they satisfy the trade-o inequality
d(G
T
1
, Q) d(G
T
2
, P) C, (42)
where the value of the constant C can be numerically determined as C 0.3047 (Werner, 2004). There is a unique
covariant joint observable G
T
attaining the lower bound in (42), but the optimizing operator T = [ ) [ is not given
by the oscillator ground state (Werner, 2004, Section 3.2).
Finally considering the error bar widths of G
T
1
, G
T
2
relative to Q, P, one nds:
J

1
(G
T
1
, Q) W

1
(Q, T), J

2
(G
T
2
, P) W

2
(P, T). (43)
Therefore, (13) implies that
J

1
(G
T
1
, Q) J

2
(G
T
2
, P) 2 (1
1

2
)
2
. (44)
We note that error bar widths in this inequality are always nite, in contrast to the standard errors or distances,
which are innite for some G
T
.
The existence of covariant phase space observables G
T
establishes the positive complement to the no-go statement
(B). We have given Heisenberg uncertainty relations for the necessary inaccuracies in the approximations of Q, P by
means of the marginal observables G
T
1
, G
T
2
. For each pair of values of the inaccuracies allowed by these uncertainty
relations there exists a G
T
which realizes these values. This conrms the suciency of the inaccuracy relations for the
existence of an approximate joint measurement of position and momentum, in the form of a covariant joint observable.
There is a (perhaps unexpected) reward for the positive attitude that led to the search for approximate joint
measurements of position and momentum: the family of covariant phase space observables G
T
contains the important
class of informationally complete phase space observables. An example is given by the choice T = [
a,0
)
a,0
[.
C. Uncertainty relations for general approximate joint measurements
While the uncertainty relations are necessary for the inaccuracies inherent in jointly measurable covariant approx-
imations Q

and P

, there remains possibility that one can overcome the Heisenberg limit by some clever choice of
non-covariant approximations of Q and P. Here we show that this possibility is ruled out. It follows that covariant
phase space observables constitute the optimal class of approximate joint observables for position and momentum.
Let M be an observable on R
2
. It was shown by Werner (2004) that there is a covariant phase space observable G
T
associated with M with the following property: d(M
1
, Q) d(G
T
1
, Q) and d(M
2
, P) d(G
T
2
, P). The same kind of
argument can be carried out in the case of the global standard error and the error bar width, so that the inequalities
(40), (44) and (42) entail the universally valid Heisenberg uncertainty relations
(M
1
, Q) (M
2
, P)

2
, (45)
d(M
1
, Q) d(M
2
, P) C, (46)
J

1
(M
1
, Q) J

2
(M
2
, P) 2 (1
1

2
)
2
. (47)
The same considerations leading to these inaccuracy relations can also be applied to obtain general trade-o relations
for the intrinsic noise and resolution width of the marginals of an approximate joint observable of Q, P:

i
(M
1
)
i
(M
2
)

2
, (48)

1
(M
1
)

2
(M
2
) 2 (1
1
2)
2
. (49)
Considering now examples of noncovariant observables on phase space, we recall rst that the commutative observ-
able M on R
2
of subsection IV.A has marginals with innite error bars. Here we give an example of an observable
12
M on phase space which is not covariant but is still an approximate joint observable for Q, P. Let G
T
be a covariant
phase space observable and dene M := G
T

1
, where (q, p) := (
1
(q),
2
(p)). We assume that
1
,
2
are strictly
increasing continuous functions such that
1
(q) q and
2
(p) p are bounded functions. Then it follows that the
marginals M

1
= G
T
1

1
1
and M

2
= G
T
2

1
2
have nite error bars with respect to Q, P. If is a nonlinear function
then M will not be covariant.
V. FROM NO MEASUREMENT WITHOUT DISTURBANCE TO SEQUENTIAL JOINT MEASUREMENTS
As concluded in Subsection III.B there is no way to determine the (sharp) position and momentum observables in a
sequential measurement. We show now that there are sequential measurements which are approximate simultaneous
determinations of position and momentum. As discussed above, the inaccuracy of the second measurement denes
an operational measure of the disturbance of momentum due to the rst, approximate measurement of position. It
therefore follows that for any sequential joint observable M on R
2
the inaccuracies satisfy the trade-o relations
(45), (46) and (47) and, moreover, these relations constitute now the long-sought-for inaccuracy-disturbance trade-o
relations.
Insofar as there are sequential measurement schemes in which these error and disturbance measures are nite, we
have thus established the positive complement to the no-go statement (C): the associated sequential joint observable
constitutes an approximate joint measurement, so that it is indeed possible to limit the disturbance of the momentum
by allowing the position measurement to be only approximate.
The existence of sequential measurements of approximate position and momentum can be demonstrated by means
of the standard model of an unsharp position measurement introduced by von Neumann (1932). In this model, the
position of the object is measured by coupling it to the momentum P
p
of the probe system via U = e
(i/)

Q

P
p
,
and using the position Q
p
of the probe as the readout observable. If
p
is the initial probe state (assumed to be a
bounded function), then the instrument of the measurement can be written in the form
1(X)() =
_
X
K
q
K

q
dq, (50)
with K
q
denoting the multiplicative operator (K
q
)(x) =


p
((q x))(x). The approximate position realized
by this measurement is Q

, where is now the probability measure with distribution function [

p
(x)[
2
.
Suppose now that one is carrying out rst an approximate position measurement, with the instrument (50), and
then a sharp momentum measurement. As shown by Davies (1970), this denes a unique sequential joint observable
M, in fact, a covariant phase space observables with marginals
M
1
(X) = 1

(X)(P(R)) = Q

(X), M
2
(Y ) = 1

(R)(P(Y )) = P

(Y ). (51)
Here the distorted momentum is P

, where is the probability measure with the distribution

1

p
(
p

)[
2
. It is obvious
that =
T
and =
T
, where T = [
()
)
()
[ with
()
(q) =

p
(q). This makes it manifest that M obeys
the uncertainty relations (45), (46) and (47), here in their double role as accuracy-accuracy and accuracy-disturbance
trade-o relations.
6
VI. ILLUSTRATION: THE ARTHURS-KELLY MODEL
The best studied model of a joint measurement of position and momentum is that of Arthurs and Kelly (1965). In
this model, a quantum object is coupled with two probe systems which are then independently measured to obtain
information about the objects position and momentum respectively. Arthurs and Kelly showed that this constitutes
a simultaneous measurement of position and momentum in the sense that the distributions of the outputs reproduce
the quantum expectation values of the objects position and momentum. They also derived the uncertainty relation
for the spreads of the output statistics corresponding to our Eq. (36). As shown in (Busch, 1982), the model also
satises the more stringent condition of an approximate joint measurement, that the output statistics determine a
covariant phase space observable whose marginals are smeared versions of position and momentum. This work also
extended the model to a large class of probe input states (Arthurs and Kelly only considered Gaussian probe inputs),
6
It is interesting to note that any covariant phase space observable can be formed in this way as a sequential joint measurement of Q

and P (Carmeli et al., 2006).

13
which made it possible to analyze the origin of the uncertainty relation for the measurement accuracies and identify
the dierent relevant contributions to it. This will be described briey below. For a detailed derivation of the induced
observable and the state changes due to this measurement scheme, see (Busch, 1982) and (Busch et al., 1997, Chapter
6). Further illuminating investigations of the Arthurs-Kelly model can be found, for instance, in (Stenholm, 1992)
and (Raymer, 1994).
The Arthurs-Kelly model is based on the von Neumann model of an approximate position measurement introduced
in Sec. V. The position

Q and momentum

P of the object are coupled with the position

Q
1
and momentum

P
2
of
two probe systems, respectively, which serve as the readout observables. Neglecting the free evolutions of the three
systems the combined time evolution is described by the measurement coupling
U = exp
_

Q

P
1
I
2
+
i

P I
1


Q
2
_
. (52)
If is an arbitrary input (vector) state of the object, and
1
,
2
are the xed initial states of the probes (given by
suitable smooth functions, with zero expectations for the probes positions and momenta), the probabilities for values
of

Q
1
and

P
2
to lie in the intervals X and Y , respectively, determine a covariant phase space observable G
T
of the
form (32) via
[G
T
(X Y )[) := U
1

2
[I Q
1
(X) P
2
(Y )[U
1

2
). (53)
The variances of the accuracy measures , associated with the marginals Q

, P

of G
T
can readily be computed:
()
2
=
1

2
(

Q
1
,
1
)
2
+

2
4
(

Q
2
,
2
)
2
,
()
2
=
1

2
(

P
2
,
2
)
2
+

2
4
(

P
1
,
1
)
2
.
(54)
If the two measurements did not disturb each other, only the rst terms on the right hand sides would appear; the
second terms are manifestations of the presence of the other probe and its coupling to the object. Since the observable
dened in this measurement scheme is a covariant phase space observable, it follows immediately that the accuracy
measures satisfy the trade-o relation (35), ()() /2. It is nevertheless instructive to verify this explicitly by
evaluating the product of the above expressions:
()
2
()
2
= Q+T,
Q :=
1
4
(

Q
1
,
1
)
2
(

P
1
,
1
)
2
+
1
4
(

Q
2
,
2
)
2
(

P
2
,
2
)
2

2
8
T :=
1
()
2
(

Q
1
,
1
)
2
(

P
2
,
2
)
2
+
()
2
16
(

Q
2
,
2
)
2
(

P
1
,
1
)
2

2
16
_
x +
1
x
_


2
8
,
where x :=
16
()
2
(

Q
1
,
1
)
2
(

P
2
,
2
)
2
.
(55)
Here we have repeatedly used the uncertainty relations for the probe systems, (

Q
k
,
k
)(

P
k
,
k
) /2.
It is evident that there are two independent sources of inaccuracy in this joint measurement model. Indeed, each
of the terms Q and T alone would suce to guarantee an absolute positive lower bound for the inaccuracy product.
The rst term, Q, is composed of two independent terms which reect the quantum nature of the probe systems;
there is no trace of a mutual inuence of the two measurements being carried out simultaneously. This feature is in
accordance with Bohrs argument concerning the possibilities of measurement, which he considered limited due to the
quantum nature of parts of the measuring setup (the probe systems).
By contrast, the term T reects the mutual disturbance of the two measurements as it contains the coupling
parameters and product combinations of variances associated with both probe systems. This feature of the mutual
disturbance of measurements was frequently highlighted by Heisenberg in thought experiments aiming at joint or
sequential determinations of the values of position and momentum.
A suitable modication of the measurement coupling U leads to a model that can be interpreted as a sequential
determination of position and momentum. Consider the unitary operator, dependent on the additional real parameter
,
U
()
= exp
_

Q

P
1
I
2
+
i

P I
1


Q
2

i
2
I

P
1


Q
2
_
. (56)
14
The Baker-Campbell-Hausdor decomposition of this coupling yields
U
()
= exp
_
( + 1)
i
2
I

P
1


Q
2
_
exp
_

Q

P
1
I
2
_
exp
_
i

P I
1


Q
2
_
. (57)
It turns out that this coupling denes again a covariant phase space observable. The variances of the inaccuracy
measures

associated with the marginals are given as follows:

(

)
2
=
1

2
(

Q
1
,
1
)
2
+ ( 1)
2

2
4
(

Q
2
,
2
)
2
,
(

)
2
=
1

2
(

P
2
,
2
)
2
+ ( + 1)
2

2
4
(

P
1
,
1
)
2
.
(58)
These accuracies still satisfy the uncertainty relation (35), but this time the contributions corresponding to Q and T
will both depend on the coupling parameters unless = 0. In particular, it does not help to make the coupling look
like that of a sequential measurement, by putting = 1. In that case, (
1
) is the accuracy of an undisturbed
momentum measurement, and (
1
) contains a term which reects the disturbance of the subsequent position
measurement through the momentum measurement. The disturbance of the position measurement accuracy is now
given by (

Q
2
,
2
), and together with the momentum inaccuracy it satises the uncertainty relation
_
1

2
(

P
2
,
2
)
2
_ _

2
(

Q
2
,
2
)
2
_


2
4
. (59)
VII. ON EXPERIMENTAL IMPLEMENTATIONS AND TESTS OF THE UNCERTAINTY PRINCIPLE
Turning now to the question of the empirical support [for the uncertainty principle], we unhesitatingly
declare that rarely in the history of physics has there been a principle of such universal importance with
so few credentials of experimental tests. (Jammer, 1974, p. 81)
This assessment was written by the distinguished historian of physics Max Jammer at a time when studies of phase
space observables based on positive operator measures were just beginning. He qualies it with a survey of early
proposed and actual tests of the preparation uncertainty relation, and he refers to some early model studies of joint
measurements, the rst of which being that by Arthurs and Kelly (1965).
Jammers verdict still holds true today. There are surprisingly few publications that address the question of
experimental tests of the uncertainty principle. Some of these report conrmations of the uncertainty principle, while
a few others predict or suggest violations. We will briey comment on some of this work below.
A. Tests of preparation uncertainty relations
The most commonly cited version of uncertainty relation is the preparation relation, usually in the familiar version
in terms of standard deviations. Conrmations of this uncertainty relation have been reported by Shull (1969)
for a single-slit diraction experiment with neutrons, by Kaiser et al. (1983) and Klein et al. (1983) in neutron
interferometric experiments, and more recently by Nairz et al. (2002) in a slit experiment for fullerene molecules.
In these slit diraction and interferometric experiments, typical measures used for the width of the spatial wave
function are the slit width and slit separation, respectively. The width of the associated momentum wave function is
given in terms of the width at half height of the central peak. It must be noted that in the mathematical modeling of
single slit diraction, the standard deviation of the momentum distribution is innite. Hence it is indeed necessary
to use another, operationally signicant measure of the width of that distribution. There does not seem to be a
universally valid uncertainty relation involving width at half height (in short, half width), but the authors of these
experiments make use of a Gaussian shape approximation of the central peak, which is in agreement with the data
within the experimental accuracy. This allows them to relate the half widths to standard deviations and conrm the
correct lower bound for the uncertainty product.
A model independent and thus more direct conrmation of the uncertainty principle can be obtained if the widths
of the position and momentum distributions are measured in terms of the overall width dened in Eq. (4). It is likely
that the data collected in these experiments contain enough information to determine these overall widths for dierent
levels of total probability 1
1
and 1
2
. In the case of the neutron interference experiment, it was pointed out by
Unk (1985) that a more stringent relation is indeed at stake, namely, a trade-o relation, introduced by Unk and
15
Hilgevoord (1985), between the overall width of the position distribution and the ne structure width (mean peak
width) of the momentum distribution.
It should be noted that these experiments do not, strictly speaking, constitute direct tests of the uncertainty relations
for position and momentum observables. While the position uncertainty, or the width of the position distribution, is
determined as the width of the slit, the momentum distribution is inferred from the measured position distribution at
a later time, namely when the particles hit the detection screen. This inference is based on the approximate far-eld
description of the wave function (Fraunhofer diraction in optics), and is in accordance with the classical, geometric
interpretation of momentum as mass times velocity. Thus, what is being tested is the uncertainty relation along with
the free Schrodinger evolution and the Fourier-Plancherel connection between position and momentum.
An alternative interpretation can be given in the Heisenberg picture, noting that the operators

Q,

P

:= m

Q(t)/t
are canonically conjugate, given the free evolution

Q(t) =

Q +

Pt/m. Here m is the mass of the particle, and t is
the time of passage of the wave packet from the slit to the detection screen. (If the distance between the slit and the
detection screen is large compared to the longitudinal width of the wave packet, the time t is fairly well dened.) The
width of the distribution of

Q is determined by the preparation (passage through the slit), and the distribution of

P

is measured directly.
B. On implementations of joint and sequential measurements
To the best of our knowledge, and despite some claims to the contrary, there is presently no experimental realization
of a joint measurement of position and momentum. Thus there can as yet be no question of an experimental test
of the uncertainty relation for inaccuracies in joint measurements of these quantities. But there are reports on the
successful experimental implementation of joint measurements of canonically conjugate quadrature components of
quantum optical elds using multiport homodyne detection.
There seem to be several communities in quantum optics and optical communication where these implementations
were achieved independently. The experiment of Walker and Carrol (1986) is perhaps the rst realization, with
a theoretical analysis by Walker (1987) yielding the associated phase space observable. This seems to have been
anticipated theoretically by Yuen and Shapiro (1980). See also (Lai and Haus, 1989) for a review. A more recent
claim of a quantum optical realization of a joint measurement was made by Beck et al. (2001). It must be noted
that in these works it is not easily established (in some cases for lack of suciently detailed information) whether the
implementation criterion is merely that of reproducing the rst moments of the two quadrature component statistics,
or whether in fact the statistics of a joint observable have been measured.
By contrast, Freyberger et al. (1993), Freyberger and Schleich (1993) and Leonhardt and Paul (1993), and Leonhardt
et al. (1995) showed that the eight port homodyne schemes for phase dierence measurements carried out by Noh et al.
(1991), Noh et al. (1992) yield statistics that approach the Q-function of the input state for a suitable macroscopic
coherent state preparation of the local eld mode. This is manifestly a realization of a joint observable. A simple
analysis is given in (Busch et al., 1997, Sec. VII.3.7.).
Turning to the question of position and momentum proper, the Arthurs-Kelly model is particularly well suited to
elucidate the various aspects of the uncertainty principle for joint and, as we have seen, sequential joint measurements
of approximate position and momentum. However, it is not clear whether and how an experimental realization of this
scheme can be obtained. Apart from the quantum optical realizations of joint measurements of conjugate quadrature
components, there are a few proposals of realistic schemes for position and momentum, e.g., (Royer, 1985), (Torma
et al., 1995), and (Power et al., 1997) mainly in the context of atom optics. In the latter two models the probe
systems are electromagnetic eld modes, and the readout probe observables are suitable phase-sensitive quantities.
The measurement coupling diers from the Arthurs-Kelly coupling in accordance with the dierent choice of readout
observables.
The experimental situation regarding the inaccuracy-versus-disturbance relation is far less well developed. This is
probably because, as we have seen above, rigorous, operationally relevant formulations of such a relation had not been
found until recently. Apart from some model considerations of the kind considered here in Sec. III there seems to be
no experimental investigations of accuracy-disturbance trade-o relations.
C. On some alleged violations of the uncertainty principle
Throughout the history of quantum mechanics, the joint measurement uncertainty relation has been the subject
of repeated challenges. There are two lines of argument against it which start from logically contrary premises. The
conclusion is, in either case, that only the preparation uncertainty relation is tenable (as a statistical relation) within
quantum mechanics.
16
The rst argument against the joint measurement relation was based on the claim that there is no provision for a
notion of joint measurement within quantum mechanics. Based on a careful assessment of the attempts existing at
the time, Ballentine (1970) concludes that a description of joint measurements of position and momentum in terms of
joint probabilities could not be obtained without signicant modications or extensions of the existing theory. Here
we have shown that the required modication was the introduction of positive operator measures and specically
phase space observables, which is entirely within the spirit of the traditional formulation of quantum mechanics; it
amounts merely to a completion of the set of observables.
The second argument was based on the claim that joint measurements of position and momentum are in fact possible
with arbitrary accuracy, and its authors, among them Karl Popper and Henry Margenau, attempted to demonstrate
their claim by means of appropriate experimental schemes.
Popper (1934) conceived a joint measurement scheme that was based on measurements of entangled particle pairs.
That this proposal was awed and untenable was immediately noted by von Weizsacker (1934). While Popper later
accepted this criticism, he suggested (Popper, 1982, footnote on p. 15) that his example may nevertheless have inspired
Einstein et al. (1935) to conceive the famous thought experiment. In fact, this experiment can be construed as a
scheme for making a joint measurement of the position and momentum of a particle that is entangled with another
particle in a particular state: provided that Einstein, Podolsky and Rosens assumption of local realism is tenable, a
measurement of the position of the latter particle allows one to infer the position of the rst particle without disturbing
that particle in any way. At the same time, one can then also measure the position of the rst particle.
It would follow that the individual particle has denite values of position and momentum while quantum mechanics
provides only an incomplete, statistical description. However, it a well-known consequence of arguments such as the
Kochen-Specker-Bell theorem (Bell, 1966; Kochen and Specker, 1967) and Bells theorem (Bell, 1964) that such value
assignments are in contradiction with quantum mechanics. Moreover, this contradiction has been experimentally
conrmed in the case of Bell s inequalities, and these tests turned out in favor of quantum mechanics.
Another proposal of a joint determination of arbitrarily sharp values of the position and momentum of a quantum
particle was made by Park and Margenau (1968) who considered the time of ight determination of velocity. As shown
in a quantum mechanical analysis in (Busch and Lahti, 1984), this scheme is appropriately understood as a sequential
measurement of rst sharp position and then sharp momentum, and does therefore not constitute even an approximate
joint measurement of position and momentum. But Park and Margenau are only interested in demonstrating that it
is possible to ascribe arbitrarily sharp values of position and momentum to a single system at the same time.
An analogous situation arises in the slit experiment, where one could formally infer arbitrarily sharp values for the
transversal momentum component from the bundle of geometric paths from any location in the slit to the detection
point. This bundle is arbitrarily narrow if the separation between slit and detection screen is made suciently large.
Thus the width of the spot on the detection screen and the width of the possible range of the inferred momentum
value can be made small enough so that their product is well below the order of .
In both situations, the geometric reconstruction of a momentum value from the two position determinations at
dierent times, which is guided by classical reasoning, constitutes an inference for the time between the two measure-
ments and cannot be used to infer momentum distributions in the state before the measurement or to predict the
outcomes of future measurements. Hence such values are purely formal and of no operational signicance. One could
be inclined to follow Heisenberg who noted in his 1929 Chicago lectures (Heisenberg, 1930, p. 25) that he regarded it
as a matter of taste whether one considers such value assignments to past events as meaningful.
However, it has been shown, by an extension of the quantum mechanical language to incorporate propositions
about past events, that hypothetical value assignments to past events lead to Kochen-Specker type contradictions.
This result was obtained by Quadt (1988) in his diploma thesis written under P. Mittelstaedts supervision at the
University of Cologne in 1988; the argument is sketched in (Quadt, 1989).
Popper returned to the subject many years later (Popper, 1982, pp. 27-29) with a novel experimental proposal with
which he aimed at testing (and challenging) the Copenhagen interpretation. In a subsequent experimental realization
it is reported that the outcome seems to conrm Poppers prediction, thus amounting to an apparent violation of the
preparation uncertainty relation.
In Poppers new experiment, EPR-correlated pairs of quantum particles are emitted from a source in opposite
directions, and then each particle passes through a slit, a narrow one on one side, the one on the opposite side
of wide opening. The particles are then recorded on a screen on each side. Popper predicts that independent
diraction patterns should build up on each side, according to the appropriate slit width; according to Popper, the
Copenhagen interpretation should predict that the particle passing through the wider slit actually shows the same
diraction pattern as the other particle. In the extreme of no slit on one side, this would still be the case. Poppers
interpretation of his experiment as a test of the Copenhagen interpretation was criticized soon afterwards, see, e.g.,
the exchange in (Collet and Loudon, 1987a,b; Popper, 1987) or (Sudbery, 1985).
The experimental realization of Poppers experiment by Kim and Shih (1999) shows, perhaps at rst surprisingly,
a behavior in line with Poppers prediction. Moreover, taking the width of the ghost image of the rst, narrow slit
17
at the side of the second particle (conrmed by Pittman et al. (1995)) as a measure of the position uncertainty of
the second particle, then this value together with the inferred width of the momentum distribution form a product
smaller than allowed by the preparation uncertainty relation. Kim and Shih hasten to assert that this result does not
constitute a violation of the uncertainty principle but is in agreement with quantum mechanics; still, the experiment
has aroused some lively and controversial debate (e.g., (Bramon and Escribano, 2005; Peres, 2002; Sancho, 2002)).
As pointed out by Short (2001), Kim and Shih overlook the fact that the two width parameters in question should be
determined by the reduced quantum state of the particle and thus should, according to quantum mechanics, satisfy
the uncertainty relation. Short gives an explanation of the experimental outcome in terms of the imperfect imaging
process which leads to image blurring, showing that there is indeed no violation of the uncertainty relation.
Finally, it seems that papers with claims of actual or proposed experiments indicating violations of the uncertainty
relation hardly ever pass the threshold of the refereeing process in major journals. They appear occasionally as
contributions to conference proceedings dedicated to realistic (hidden variable) approaches to quantum mechanics.
VIII. CONCLUSION
In this exposition we have elucidated the positive role of the uncertainty principle as a necessary and sucient con-
dition for the possibility of approximately localizing position and momentum. We have also noted that approximate
position measurements can allow a control of the disturbance of the momentum. Uncertainty relations for posi-
tion and momentum thus come in three variants: for the widths of probability distributions, for accuracies of joint
measurements, and for the trade-o between the accuracy of a position measurement and the necessary momentum
disturbance (and vice versa).
In his seminal paper of 1927, Heisenberg gave intuitive formulations of all three forms of uncertainty relations, but
it was only the relation for state preparations that was made precise soon afterwards. It took several decades until
the conceptual tools required for a rigorous formulation of the two measurement-related uncertainty relations had
become available. Here we identied the following elements of such a rigorous formulation.
First, a theory of approximate joint measurements of position and momentum had to be developed; this possibility
was opened up by the representation of observables as positive operator measures. Second, a criterion of what
constitutes an approximate measurement of one observable by means of another must be based on operationally
signicant and experimentally relevant measures of inaccuracy or error. Here we discussed three candidate measures:
standard error, a distance of observables, and error bars. For each of these, a universal Heisenberg uncertainty relation
holds, showing that for any observable on phase space the marginal observables cannot both approximate position
and momentum arbitrarily well.
The proofs of these uncertainty relations are rst obtained for the distinguished class of covariant phase space ob-
servables, for which they follow mathematically from a form of uncertainty relation for state preparations. This formal
connection between the preparation and joint measurement uncertainty relations is in accordance with a postulate
formulated by Bohr (1928) in his famous Como lecture of 1927 which states that the possibilities of measurement
should not exceed the possibilities of preparation. The uncertainty relation for a general approximate joint observable
for position and momentum can then be deduced from that for some associated covariant joint observable.
Apart from the limitations on the accuracy of joint approximations of position and momentum, we have found
Heisenberg uncertainty relations which quantify the necessary intrinsic unsharpness of two observables that are jointly
measurable, provided they are to be approximations of position and momentum, respectively. Both limitations are
consequences of the noncommutativity of position and momentum.
Finally, the idea of a measurement of (say) position disturbing the momentum has been made precise by recog-
nizing that a sequential measurement of measuring rst position and then momentum constitutes an instance of a
joint measurement of some observables, of which the rst marginal is an (approximate) position and the second a
distorted momentum observable. The inaccuracy inherent in the second marginal gives a measure of the disturbance
of momentum. The joint measurement uncertainty relations can in this context be interpreted as a trade-o between
the accuracy of the rst position measurement against the extent of the necessary disturbance of the momentum due
to this measurement.
Last we have surveyed the current status of experimental implementations of joint measurements and the question
of experimental tests of the uncertainty principle. While there do not seem to exist any conrmed violations of the
uncertainty principle, there do exist several experimental tests of uncertainty relations which have shown agreement
with quantum mechanics.
Acknowledgement
Part of this work was carried out during mutual visits by Paul Busch at the University of Turku and by Pekka Lahti
at Perimeter Institute. Hospitality and nancial support by these host institutions are gratefully acknowledged. We
are indebted to M. Leifer and R. Werner for helpful comments on an earlier draft version of this paper.
18
References
Appleby, D., 1998a, Int. J. Theor. Phys. 37, 1491.
Appleby, D., 1998b, Int. J. Theor. Phys. 37, 2557.
Arthurs, E., and J. Kelly, 1965, Bell System Tech. J. 44, 725.
Ballentine, L., 1970, Rev. Mod. Phys. 42, 358.
Beck, M., C. Dorrer, and I. Walmsley, 2001, Phys. Rev. Lett. 87, 253601/1.
Bell, J., 1964, Physics 1, 195.
Bell, J., 1966, Rev. Mod. Phys. 38, 447.
Bohr, N., 1928, Nature 121, 580.
Bramon, A., and R. Escribano, 2005, quant-ph/0501134 .
Busch, P., 1982, Unbestimmtheitsrelation und simultane Messungen in der Quantentheorie, Ph.D. thesis, University of Cologne,
english translation: Indeterminacy relations and simultaneous measurements in quantum theory, Int. J. Theor. Phys. 24
(1985) 6392.
Busch, P., M. Grabowski, and P. Lahti, 1997, Operational Quantum Physics (Springer-Verlag, Berlin), second corrected printing.
Busch, P., T. Heinonen, and P. Lahti, 2004, Phys. Lett. A 320, 261.
Busch, P., and P. Lahti, 1984, Phys. Rev. D 29, 1634.
Busch, P., P. Lahti, and P. Mittelstaedt, 1996, The Quantum Theory of Measurement (Springer-Verlag, Berlin), second revised
edition.
Busch, P., and D. Pearson, 2006, Universal joint-measurement uncertainty relation for error bars, unpublished available at
quant-ph/0612XXX.
Busch, P., T. Schonbek, and F. Schroeck, Jr., 1987, J. Math. Phys. 28, 2866.
Carmeli, C., T. Heinonen, and A. Toigo, 2004, J. Math. Phys. 45, 2526.
Carmeli, C., T. Heinonen, and A. Toigo, 2006, Intrinsic unsharpness and approximate repeatability of quantum measurements,
unpublished available at quant-ph/0611286.
Cassinelli, G., E. De Vito, and A. Toigo, 2003, J. Math. Phys. 44, 4768.
Collet, M., and R. Loudon, 1987a, Nature 326, 671.
Collet, M., and R. Loudon, 1987b, Nature 328, 675.
Cowling, M., and J. Price, 1984, SIAM J. Math. Anal. 15, 151.
Davies, E., 1970, J. Funct. Anal. 6, 318.
Davies, E., 1976, Quantum Theory of Open Systems (Academic Press, London).
Dym, H., and H. McKean, 1972, Fourier Series and Integrals (Academic Press, New York).
Einstein, A., B. Podolsky, and N. Rosen, 1935, Phys. Rev. 47, 777.
Folland, G., and A. Sitaram, 1997, J. Fourier Anal. Appl. 3, 207.
Freyberger, M., and W. Schleich, 1993, Phys. Rev. A 47, R30.
Freyberger, M., K. Vogel, and W. Schleich, 1993, Phys. Lett. A 176, 41.
Hall, M., 2004, Phys. Rev. A 69, 052113/1.
Heisenberg, W., 1927, Z. Phys. 43, 172.
Heisenberg, W., 1930, The Physical Principles of the Quantum Theory (University of Chicago Press, Chicago).
Heisenberg, W., 1977, in The Uncertainty Principle and Foundations of Quantum Mechanics. A Fifty Years Survey, edited by
W. Price and S. Chissick (J. Wiley & Sons, London), pp. 36.
Holevo, A., 1979, Rep. Math. Phys. 16, 385.
Jammer, M., 1974, The Philosophy of Quantum Mechanics (John Wiley & Sons, New York).
Kaiser, H., S. Werner, and E. George, 1983, Phys. Rev. Lett. 50, 560.
Kim, Y.-H., and Y. Shih, 1999, Found. Phys. 29, 1849.
Kiukas, J., P. Lahti, and K. Ylinen, 2006a, J. Math. Anal. Appl. 319, 783.
Kiukas, J., P. Lahti, and K. Ylinen, 2006b, J. Math. Phys. 47, 072104.
Klein, A., G. Opat, and W. Hamilton, 1983, Phys. Rev. Lett. 50, 563.
Kochen, S., and E. Specker, 1967, J. Math. Mech. 17, 59.
Lahti, P., 1986, Rep. Math. Phys. 23, 289.
Lai, Y., and H. Haus, 1989, Quantum Opt. 1, 99.
Landau, H., and H. Pollak, 1961, Bell System Tech. J. 40, 65.
Lenard, A., 1972, J. Functional Analysis 10, 410.
Leonhardt, U., B. Bohmer, and H. Paul, 1995, Opt. Commun. 119, 296.
Leonhardt, U., and H. Paul, 1993, Phys. Rev. A 47, R2460.
Nairz, O., M. Arndt, and A. Zeilinger, 2002, Phys. Rev. A 65, 032109/1.
von Neumann, J., 1932, Mathematische Grundlagen der Quantenmechanik (Springer, Berlin), english translation: Mathematical
Foundations of Quantum Mechanics, Princeton University Press, Princeton, 1955.
Noh, J., A. Foug`eres, and L. Mandel, 1991, Phys. Rev. Lett. 67, 1426.
Noh, J., A. Foug`eres, and L. Mandel, 1992, Phys. Rev. A 45, 424.
Ozawa, M., 2004, Ann. Physics 311, 350.
Park, J., and H. Margenau, 1968, Int. J. Theor. Phys. 1, 211.
Peres, A., 2002, Stud. Hist. Philos. Modern Phys. 33, 23.
19
Pittman, T., Y. Shih, D. Strekalov, and A. Sergienko, 1995, Phys. Rev. A 52, R3429.
Popper, K., 1934, Naturwiss. 22, 807.
Popper, K., 1982, Quantum Theory and the Schism in Physics (Unwin Hyman, London).
Popper, K., 1987, Nature 328, 675.
Power, W., S. Tan, and M. Wilkens, 1997, J. Mod. Opt. 44, 2591.
Quadt, R., 1988,

Uber die Nicht-Objektivierbarkeit quantenmechanischer Ereignisse in der Vergangenheit, Diploma thesis,
University of Cologne.
Quadt, R., 1989, Found. Phys. 19, 1027.
Raymer, M., 1994, Am. J. Phys. 62, 986.
Royer, A., 1985, Phys. Rev. Lett. 55, 2745.
Sancho, P., 2002, Found. Phys. 32, 789.
Scully, M., B.-G. Englert, and H. Walther, 1991, Nature 351, 111.
Short, A., 2001, Found. Phys. Lett. 14, 275.
Shull, C., 1969, Phys. Rev. 179, 752.
Stenholm, S., 1992, Ann. Physics 218, 233.
Sudbery, A., 1985, Phil. Sci. 52, 470.
Torma, P., S. Stenholm, and I. Jex, 1995, Phys. Rev. A 52, 4812.
Unk, J., 1985, Phys. Lett. 108A, 59.
Unk, J., and J. Hilgevoord, 1985, Found. Phys. 15, 925.
Walker, N., 1987, J. Mod. Optics 34, 15.
Walker, N., and J. Carrol, 1986, Optical and Quantum Electronics 18, 355.
von Weizsacker, K., 1934, Naturwiss. 22, 808.
Werner, R., 1984, J. Math. Phys. 25, 1404.
Werner, R., 2004, in Quantum information, Statistics, Probability, edited by O. Hirota (Paramus, NJ:Rinton), pp. 153171,
also available: quant-ph/0405184.
Ylinen, K., 1989, J. Math. Anal. Appl. 137, 185.
Yuen, H., and J. Shapiro, 1980, IEEE Trans. Inform. Theory 26, 78.