Rovelli's World
Rovelli's World
Rovelli's World
DOI 10.1007/s10701-009-9326-5
Rovelli’s World
1 Rovelli [11]; I will refer by section numbers, since a revised version is available on the web.
I happily dedicate this paper to Jeffrey Bub, whose work has inspired me for a good quarter of a
century.
B.C. van Fraassen ()
Department of Philosophy, San Francisco State University, 1600 Holloway Avenue, San Francisco,
CA 94132, USA
e-mail: fraassen@sfsu.edu
Found Phys (2010) 40: 390–417 391
1 Placing Rovelli
Thus observer system in this paper is any possible physical system (with more
than one state). If there is any hope of understanding how a system may behave
as observer without renouncing the postulate that all systems are equivalent,
then the same kind of processes—“collapse”—that happens between an elec-
tron and a CERN machine, may also happen between an electron and another
electron. Observers are not “physically special systems” in any sense.
We must treat this with some delicacy, since the usual explanation of such correlations
or entanglements is in terms of states conceived of as observer-independent. The
standard quantum mechanical formalism is used here, but understood in a new way.
Given the comparative loss of popularity of the older ‘external observation’ ap-
proach, at least among those who work on foundations of physics, Rovelli’s return to
it at this date imparts his view with a stimulating sense of novelty.
Noting the emphasis Rovelli puts on information, it is also important to place Rov-
elli’s approach with respect to the information-theory approach. This is a very lively
new development. While there were beginnings and precedents, this has recently
taken a quite radical turn, and Rovelli’s work can be seen as involved in that turn.
Let’s look at the beginnings first and then at the radical agenda in such recent work
as that of Christopher Fuchs, Jeffrey Bub and their collaborators.
In the 1950s H.J. Groenewold advocated that we should regard quantum states
as just summaries of information obtained through measurement. There are some
striking similarities between Groenewold’s description of the quantum mechanical
situation and Rovelli’s.
Groenewold [6, 7] proposed a formulation of the theory that would contain all
its empirical content without referring to states in any essential way. He derided the
idea that quantum states are to be thought of on the model of states in classical me-
chanics. His formulation re-appears quite clearly in Rovelli’s article, though there in
a more general form. The idea is that a situation of interest is to be depicted as the
effect of a series of measurements, represented by a series of observables (the ones
being measured) interspersed with evolution operators (governing evolution between
measurements). The sole real problem to be addressed, according to Groenewold, is
this:
given the outcomes of preceding measurements, what are the probabilities for
outcomes of later measurements in the series?
The answer is formulated in terms of transition probabilities.2 In the exposition of
Rovelli’s specific version below I shall explain and illustrate how that goes.
Groenewold offers an argument to the effect that states are to be regarded as ‘sub-
jective’ or ‘observer-relative’, determined by information available. Imagine that each
measurement apparatus in the series records its outcome.3 After the entire series has
2 Groenewold was not the only one; see for example [13].
3 See Dicke [3] for an argument about how this is physically possible without disturbance; see further the
discussion in [15, pp. 257–258].
394 Found Phys (2010) 40: 390–417
been concluded, a physicist O inspects those recorded results in some order, and
assigns states to the system measured for the times of those outcomes using von Neu-
mann’s Projection Postulate recipe (which everyone agrees is fine for such narrowly
focused predictive tasks). To begin, O assumes some initial state. Groenewold sug-
gests that in absence of other information that could be the entirely uninformative
mixture represented by the identity operator on the space. For time t between times
t1 and t2 where the state ρ(t1 ) is assumed or known the calculation looks like this:
(with a correspondingly longer such series for a longer series of measurements be-
tween the initial and final time) where the Ks are transition operators, and the time-
indexed ρ is the ascribed state; the U s are the normal evolutions while no measure-
ment or other interference occurs.
But now what would happen if O (or one of his colleagues) decides on a dif-
ferent order for inspection of the recorded outcomes? For the same times, although
having started with the same initial knowledge or assumptions about the system, the
assignment of states will be quite different.
There is nothing contentious in this imagined scenario itself. The contentious part
is Groenewold’s insistence that no other significance is to be accorded to the as-
signment of states. They are nothing more than compendia of information assumed,
known, or gathered through measurements, and thus determined entirely by a specific
history, the ‘observer’s’ history. The truly empirically testable part of the theory, he
insists, is contained in the transition probabilities. When they are tested, the conve-
nient calculation starts with an assignment of an initial state, but coherence requires
only that some such initial assignment leads to the right predictions—the transition
probabilities are independent of the states, they are formulable in terms of the observ-
ables.4
This insistence, that the states be thought of as playing no other role, is at the
heart of the recent innovations in the information theoretic approach. Christopher
Fuchs presents the program in its most radical form in his much discussed “Quantum
Mechanics as Quantum Information (and only a little more)”:
This, I see as the line of attack we should pursue with relentless consistency:
The quantum system represents something real and independent of us; the
quantum state represents a collection of subjective degrees of belief about
something to do with that system (even if only in connection with our ex-
perimental kicks to it). The structure called quantum mechanics is about the
interplay of these two things—the subjective and the objective [5, p. 5].)
He submits that “the quantum state is solely an expression of subjective information—
the information one has about a quantum system. It has no objective reality in and
4 They are often presented as probabilities for transitions between states, because the Projection Postulate
is generally taken for granted. In [15] I explained them in an intermediate way: the probability is that of the
outcome 1 of a measurement of the observable represented by projection on the vector representing second
state, given that the system measured is in the first state. But this can easily be replaced by a formulation in
terms of the two observables, which are the projections on the two states. For a nice introductory treatment
of the theory entirely in this form we can look to [13].
Found Phys (2010) 40: 390–417 395
of itself.” When asked “information about what?” he replies “The answer is ‘the po-
tential consequences of our experimental interventions into nature’.” (ibid, p. 7) But
Fuchs also has a precise proposal about how to describe the information-updating
process in response to measurement. (See further [4].) Drawing on results, both his
own and others, he depicts it as a special case of Bayesian updating of opinion by
conditionalization. We have to think here, as in Groenewold’s scenario, of an epis-
temic agent with a pertinent prior state of opinion—a physicist who accepts at least
the bare minimum of the quantum theory—reacting to recorded measurement out-
comes. There is also, without explicit attention paid, for both Groenewold and Fuchs,
a presumed coordination, so that tangible physical operations can be univocally rep-
resented in terms of an algebra of observables of a certain sort.
This reliance on a fundamental representation of the physical situation—the
coordination—becomes clearest in the important paper by Robert Clifton, Jeffrey
Bub, and Hans Halvorson. The physical system is characterized by means of an alge-
bra of observables, taken to be a C* algebra.5 But states are just generalized probabil-
ity functions—more accurately, expectation value functions—defined on this algebra
of observables. So far that is similar to the approach in more “realistically” under-
stood foundational treatments. The difference comes in what is added now so as to
single out quantum theories. What is added is constraints on information transfer,
with the states thought of as information depositories. From the premise that those
constraints are satisfied, the basic principles of quantum theory are deduced. As re-
flection on this result, Bub then argued in his “Why the Quantum?” that
A quantum theory is best understood as a theory about the possibilities and im-
possibilities of information transfer, as opposed to a theory about the mechanics
of non-classical waves or particles. [1, p. 42]
“Information” is here understood as Groenewold specified, in the technical sense of
information theory, as measured classically by the Shannon entropy or by the von
Neumann entropy for quantum states. And in “Quantum Mechanics Is About Quan-
tum Information”, Bub argues that
Quantum mechanics represents the discovery that there are new sorts of infor-
mation sources and communication channels in nature (represented by quantum
states), and the theory is about the properties of these information sources and
communication channels. You can, if you like, tell a mechanical story about
quantum phenomena . . . but such a story, if constrained by the information-
theoretic principles, will have no excess empirical content over quantum me-
chanics. So the mechanical story for quantum phenomena is like an aether story
for electromagnetic fields. [2, p. 558]
Bub’s answer to the question “Information about what?” is just the same as Fuch’s—
though in phrasing that shows his special interest in encryption and decoding.
Note once again that some form of coordination is presumed given, without re-
ceiving explicit attention: the measurements and their results are assumed univocally
5 This is a very general framework, which allows for the formulation of many sorts of physical theories,
both classical and quantum.
396 Found Phys (2010) 40: 390–417
representable in terms of the observables that characterize the system. This points to
‘absolute’ characteristics of the system, which are not aspects of information gath-
ered about it, but pertain to the system itself. That the system is characterizable in
such a way is presupposed when certain operations are classified as, or taken to be,
means of gathering information about it. Thus here, as for Groenewold (and equally
for Rovelli, as we shall see) there is a divide as well as a link between ‘subjective’
and ‘objective’ features of the experimental situation.
At first sight Rovelli’s treatment of states is not exactly what either Groenewold,
Fuchs, or Bub appears to advocate.6 Rovelli does bring states into the discussion, but
as states that measured objects can have relative to the measuring system. At first
sight we seem to detect a tension between what Rovelli does and what he tells us it is
possible to do. What he calls his Main Observation, motivating the view, is similar to
Groenewold’s though:
In quantum mechanics different observers may give different accounts of the
same sequence of events.
Having rejected the idea of observer-independent states, there is no question of one
of those descriptions being the sole truth, with the other illusion or error. Here is an
example that Rovelli describes in intuitive terms. I will elaborate on it, in several
steps.
O registers the value 1, say, and thus assigns pure state |A, 1 to S, or in other words,
S is now in state |A, 1 relative to O.
6 Though Rovelli’s article was clearly inspirational to the later literature; cf. Fuchs, op. cit. p. 3.
Found Phys (2010) 40: 390–417 397
Meanwhile P has the information that this measurement is taking place (presum-
ably on the basis of earlier measurements made on S + O). So P describes O as in
an initial state |init and S + O coupled at the beginning. The state of S + O evolves:
(α|A, 1 + β|A, 0) ⊗ |init → (α|A, 1 ⊗ |B, 1) + (β|A, 0 ⊗ |B, 0).
Here |B, 1 and |B, 0 are the ‘pointer reading states’ that P uses to characterize
observer O when O registers a definite value of 1 or 0 as measurement outcome.
That is, the measurement interaction between S and O is such as to effect the requisite
correlation between A pertaining to S and B pertaining to O.
If P now wonders what state to assign to S, but does not make a measurement, then
he calculates it by the usual ‘reduction of the density matrix’. Thus P assigns to S a
mixed state, namely the mixture of |A, 1 and |A, 0 in proportions α 2 and β 2 . So we
see that O and P assign different states to S. To put it in other words, S has different
states relative to O and relative to P .
Rovelli also insists on the orthodox eigenvalue-eigenstate link, so that A takes a
value 1 relative to O, but not relative to P —observables have values only relative to
observers, and may not have the same value relative to different observers.
But is this description of the situation then observer-independent, one that is in
fact not relative to any observer? Shouldn’t we object that the rationale forbids this,
because by Rovelli’s lights we can only have descriptions relative to some observer
or other?
The answer is that there is no incoherence here, but we must carefully distinguish
what Rovelli gives us when he presents his view, even in such an example, and the
description of the same situation by a third observer. The Example can indeed be
elaborated so as to include a third observer, whom we might call ROV. We could
imagine that ROV has, on the basis of previous measurements, information that can be
summarized by assignments of initial states to O, S, P and their composites relative
to ROV, plus later states based on their unitary evolution. We’ll look later at how this
goes, when we will also have occasion to consider measurements that P can make
on O or S later on. But right now we can point out that ROV’s information is not to
be confused with what Rovelli tells us about this sort of situation. The tension that a
reader might feel could be expressed this way:
Rovelli seemingly purports to be giving us a description of the world that would
on the one hand be on the same level as a description of the rest of the world
relative to some given system ROV, and yet on the other hand not relative to
anything!
But that is not so at all. Rovelli, who can give these examples, is telling us only some-
thing about the general form that these observers’ descriptions (their information) can
take, given that certain measurement interactions have taken place. The resolution of
this sensed tension is this: Rovelli does not give any specific such description of the
world—he describes the form that any description which assigns states must take.
398 Found Phys (2010) 40: 390–417
Rovelli describes not the world, but the general form of information that one system
can have about another—namely as the assignment of states relative to a given system
on the basis of information available to that system:
• there is no implication of possible specific information about what there is which
is independent of any point of view, but
• there can be knowledge of the form that any such information, relative to a partic-
ular vantage point, must take.
So we have here a transcendental point of view. Rovelli offers us this knowledge of
the general form, the conditions of possibility. We must take very seriously the fact
that as he sees it, quantum mechanics is not a theory about physical states, but about
(‘about’?) information. The principles he sees at the basis of quantum mechanics are
principles constraining the general form that such information can take, not to be
assimilated to classical evolution-of-physical-state laws.
This form is constrained by the insistence that specific information, had by one sys-
tem about another system, can only be a record of actual measurement outcomes.
The only way in which there can be information for one observer of what has hap-
pened to another observer is through a physical measurement by the former on the
latter. Communication, i.e. exchange of information, is physical (cf. end sect. III of
the article).
Before aiming at greater precision, let’s briefly summarize how this happens ac-
cording to Rovelli’s account. A question is asked of a system or source only when
an appropriate physical interaction takes place. This interaction is a measurement de-
livering a value for some observable, but also serves as a preparation, so that the
value obtained has (relative to the theory) predictive content. The probabilities of
future measurement outcomes are affected by the outcome obtained—the measured
system has gone into a new state relative to the measurement set-up. Thus he accepts
(explicitly, in his rejection of the Bohm and modal interpretations) von Neumann’s
eigenstate-eigenvalue link:
the system to which the observable’s value pertains is (at that time) in an eigen-
state of that observable, corresponding to that value.
But there is a twist, which changes the meaning, so that this says something quite
different from its original. The reference is here not to a physical state of the system,
but to the state of the system relative to the observer (the measurement apparatus). So
the ‘collapse’ is in that observer’s information; the state assigned to the system is a
summary of that information.
As mentioned earlier, because of the eigenstate-eigenvalue link it follows that if
states are relative, so are values of observables. That an observable takes or has a
certain value at a certain moment, that too is observable-relative (cf. end of sect. 2
[10]). Because information can only be had by actual, physical measurement, the
states assigned will rarely be pure. It is not easy to obtain maximal information about
a system, even with respect to targeted observables. So in general the value of an
observable, relative to a given observer, will not be sharp.
Found Phys (2010) 40: 390–417 399
This information is the subject of two postulates. Let us introduce them in such
a way as to spell out what is and is not observer relative. Each physical system S
is characterized in the first place by means of a set W (S) = {Qi : i in I } of ques-
tions that can be asked of it. This association of W (S) with S is not relative to any
observer—we may call it the first ‘absolute’. Although the presentation differs, this
set of questions pertaining to S is essentially the specification of the family of observ-
ables that pertain to S. (Eventually, the algebra of observables is reconstructed from
this family of questions; for our purposes we need not distinguish the two.) When the
sets of questions are the same for two systems we may call them of the same type.
Secondly, an observer who has been in measurement interaction with a system
has a record of the questions that have been asked and the sequence of outcomes
thus obtained. That the observer has this is not relative to another observer.7 It is our
second ‘absolute’. At the same time we must be careful not to equate this fact about
the observer with a quantum mechanical state! For while we could try to describe a
state that ostensibly is the state that O has if and only if it has a particular sequence of
0s and 1s registered in a series of measurement interactions with S, that would have
to be the state of O relative to another observer P who has obtained that information
by means of a later measurement on S. We’ll see later on whether, or to what extent,
there could be a discrepancy, or even a meaningful comparison.
3 States as Observer-Information
7 By taking this not to be relative, we have in this sequence of 0s and 1s something analogous to Einstein’s
local coincidences, the ‘bed rock’ of the representation. Rovelli’s criticism of the ‘consistent histories’
interpretation suggest strongly that he does not allow any ambiguity in this respect.
400 Found Phys (2010) 40: 390–417
8 Rovelli introduces and uses the term “complete family s of information” for “maximally non-redundant
c
question-answer sequence”.
9 Compare: “In particular, I identify one element of quantum mechanics that I would not label a subjective
term in the theory; it is the integer parameter D traditionally ascribed to a quantum system via its Hilbert-
space dimension.” Chris Fuchs [5], Abstract.
Found Phys (2010) 40: 390–417 401
Notice the modal character of this assertion! In contrast, some questions Q and Q
are compatible: on a given occasion, after receiving Yes to Q, the observer has only
non-zero probabilities for both possible answers to Q , but if he then asks Q , he can
base more precise predictions on the fact that he has had these two answers.
New assumption: this indeterminism is not a chaotic randomness, but can be
characterized in terms of definite probabilities.
Suppose the first complex apparatus A asks a “complete” question, so it yields a
record that provides a maximally non-redundant question-answer sequence. Before
that question has been asked we have no non-trivial information. Suppose the second
apparatus B is equally complete, though the question family is very different. Rovelli
posits a definite transition probability p(B|A) that a Yes answer to B will follow a
Yes answer to A, which is both idempotent and symmetric.
Intuitive mnemonics: look at the scenario in which a single source sends many
systems of the same type into the series of measurement apparatus for two-valued
observables A, B, . . . that the observer has installed. The stream is diminished by
some factor q by the first measurement, then by the transition probability p. Suppose
we do A again, then once again the stream is diminished by that factor p. So the
number goes from qM to pqM to ppqM by the operations A, AB, ABA so we could
write:
ABA = pA
and this is what a sequence of 1-dimensional projections would do to a vector. It is a
way to identify the transition probability. This is numerically equal to the cos2 of the
angle between the two 1-eigenvectors, onto which they project, or in Hilbert space
the squared modulus of the scalar product, or equivalently the trace of the product of
the two projections.
After a maximally non-redundant question-answer sequence performed by mea-
surement A, the next question might only e.g. ask “is the system in subspace J ?”,
with J of higher dimension—but here there is a definite probability as well, which
can be derived (in accordance with the practical calculation suggested by von Neu-
mann’s Projection Postulate).
Suppose that observer O has put a series of questions to system S and has arrived
at the point of attributing |A, x to S, where x is an eigen-value of A. Imagine once
again a second observer P , whose knowledge (gained earlier through a physical trans-
mission process) was enough to attribute an initial state to S + O, and a Hamiltonian
to govern their interaction, enough for him to attribute the evolution in question. Then
as we noted above P has the usual ‘distant’ description of S + O:
initially it is in state βi (|A, ai ⊗ |init) (1)
i
this evolves into the final state βi (|A, ai ⊗ |B, ai ) (2)
i
402 Found Phys (2010) 40: 390–417
which is quite different from |A, 1 or |A, 0. According to Rovelli, this is all there
is to be said, so far: S has one state relative to O, and another state relative to P .
The phrase ‘S has state |A, 1 relative to O’ means only that the information O has
obtained can be summed up or represented by the vector |A, 1. But is the fact that O
has certain information a fact that is or is not observer-relative? We must answer this
question in the light of two points Rovelli insists on:
(i) There is no meaning to the state of a system except within the information of a
further observer.
(ii) There is no way a system P may get information about a system O without
physically interacting with it, and therefore without breaking down (at the time
of the interaction) the unitary evolution description of O.
‘Information’ has a minimal sense in this context, to say that O has information
about S means only that there is a certain correlation in the state of S + O. That
much P was able to predict already, and so he can predict something with certainty
if a measurement will be made to confirm this. Note that what he is able to predict
with certainty amounts to information he already has.
More formally, there is an operator M on the Hilbert space of the S + O system
whose physical interpretation is “Is the pointer correctly correlated to A?” If P mea-
sures M, then the outcome of this measurement would be yes with certainty, when
the state of the S + O system is as in the state described in (2). The operator M is
given by
where the eigenvalue 1 of M means “yes, the hand of O indicates the correct state
of S” and the eigenvalue 0 means “no, the hand of O does not indicate the correct
state of S”. At time t2, the S + O system is in an eigenstate of M with eigenvalue 1;
therefore P can predict with certainty that O “knows” the value of A.
Thus, it is meaningful to say, according to the P description of the events E, that
O “knows” the quantity A of S, or that O “has measured” the quantity A of S, and
the pointer variable embodies the information (cf. middle of section II-D). But of
course P had a choice, P could have measured a different observable, say K, to try
and find out which result O obtained:
Intuitively speaking, this is what P would measure to find out what O found. She
would get either result 1 or result 0, and would say “O found 1” or “O found 0”
accordingly. But can we understand that literally as referring to what O had as infor-
mation before P made this measurement? If P finds result 1, does that imply that O
had found 1 and that O had assigned state |A, 1 to S?
According to Rovelli’s rules, this makes no sense. An interpretation of quantum
measurement as revealing pre-existing values is untenable.
We are now in a position to examine and resolve some puzzles that tend to occur
to practically any reader in first acquaintance with this interpretation.
Suppose that P will make a measurement on O + S after this point, and later report
the result to O. In the meanwhile O makes a prediction with certainty about what P
will find. Is it possible that O will find his prediction contradicted by P ?
Example: P will measure (I ⊗ A) on O + S. P predicts that he will get value 1
with probability < 1, and value 0 with some probability > 0. Suppose he gets
value 0.
Meanwhile O knows that he has seen value 1, and has a record of that, so as-
signs himself state |B, 1, and assigns to S the state |A, 1, and therefore to
O + S the state |B, 1 ⊗ |A, 1. So O predicts with certainty that P ’s measure-
ment will have result 1. And so O is making a false prediction here, one that is
falsified by what P finds.
REPLY: The reasoning is questionable in several ways.
404 Found Phys (2010) 40: 390–417
10 My suggestion is that this should not be added as a possibility; there certainly seems to me to be no
warrant in Rovelli’s interpretation for doing so. For a contrary view and a recent ‘Wigner’s friend’ type
example presented to challenge information-theoretic approaches (specifically Jeffrey Bub’s recent work)
see [8].
Found Phys (2010) 40: 390–417 405
So does this mean that O and P have no way to find out what either of them saw
earlier, as opposed to what it seems now that they have seen? That would still seem
almost as puzzling. To answer this properly, we need to construct our puzzle situation
with more precision and care.
PUZZLE 3. Can an Observer Find out What was Observed Earlier on?
To see how we can get into a confusion here, I am going to present this third version of
the puzzle first of all in the ‘old’ style, assuming that states are observer-independent.
Then the puzzle will again be resolved by seeing how the understanding of this situ-
ation changes on Rovelli’s conception.
Let the measured system S start off in a superposition β i |A, i of eigenstates of
an observable A corresponding to distinct eigenvalues, and let us measure A twice,
using two measuring systems O and P .11 For simplicity I’ll take A to be time-
independent (we could put in evolution operators, as Groenewold and Rovelli indi-
cate, but it would not seriously affect the argument), and take the pointer observables
of both O and P to be the same observable B.
Then, under the familiar idealized assumptions of a von Neumann measurement, the
combined system S + O will be in dynamic state βi (|A, i ⊗ |B, i) at the end of
the first measurement. At the end of the second measurement the dynamic state of
S + O + P will be, ignoring phase factors, βi (|A, i ⊗ |B, i ⊗ |B, i).
By reduction, we have states also for parts of the total system. Write P [|A, i]
for the projection on the ray containing |A, i, etc. At the end of the first measure-
ment,
the individual systems S and O are in dynamic states [ |β
i | P [|A,
2 i] and
|βi | P [|B, i], respectively. The final dynamic state of O +P is |βi |2 P [|B, i⊗
2
11 Assume that I and O each evolve freely after their measurement interaction, that there is no interaction
between O and P , and that both A and the ‘pointer-reading observable’ B for O commute with the free
Hamiltonians for S and O respectively.
406 Found Phys (2010) 40: 390–417
thus has the value k as well, we will say that its pointer reads k. Likewise, at
the end of the second measurement, S ends up in some state |A, m with P in
the corresponding |B, m; its pointer the reads m. Moreover, m = k.
Suppose we want to check now whether that is so. Then we can have a third mea-
surement, of that ‘agreement observable’ that Rovelli describes, as follows:
Let M be an observable for the combined system O + P , which has eigenvalue 1
on the space spanned by all |B, i×|B, i, and which has value 0 on all |B, i×|B, j ,
for j not equal to i. Then if O and P are in pure dynamic states |B, i and |B, j
respectively (always ignoring phase factors), the value of M will be 1 if and only if
i = j . In the usual interpretation, this means that in the only case in which our pointer
readings can have definite values, M will have the value 1 just in case these values
agree. In the context of that interpretation, then, it is reasonable to speak of M as
the observable which is, or registers, agreement between the two pointer readings.
Even in the context of Rovelli’s interpretation, one can continue to speak of M as the
‘agreement’ observable. The question is whether here, the locution needs to be taken
with a grain of salt—may M take up the value 1 even though the pointer readings do
not agree?
Our present example provides an illustration. The final dynamic state of O + P
is |βi |2 P [|B, i × |B, i]. Since all summands of the mixture are eigenstates of
M with eigenvalue 1, so is the state itself. So M takes the value 1 on the system
O + P . To arrive at this conclusion however, we needed only to know the mixed state
here ascribed to O + P —we did not need any information about what states O, P
are in individually. That information is logically compatible with the equally valid
conclusion that O, P are in mixtures of the various states {|B, i}. So the conclusion
that M takes value 1 cannot possibly, by itself, guarantee the suggestion that the
pointer reading on P is equal to the pointer reading on O. But if we assume von
Neumann’s rather than Rovelli’s interpretation, we do have that guarantee, since O,
P collapsed into definite pointer states.
REPLY: Once again, we have drawn a puzzling consequence for Rovelli by think-
ing about the situation in ‘old’ terms, and then having too quick a look at how his
view differs. To really see whether there is a puzzle here, we have to retell the story
from the beginning, in Rovelli’s way. Here is the retelling, which we can now exhibit
as a more elaborate example of Rovelli’s view:
12 It is part of ROV’s knowledge, based on past measurements, that I and O each evolve freely after their
measurement interaction, that there is no interaction between O and P , and that both A and the ‘pointer-
reading observable’ B for O commute with the free Hamiltonians for S and O respectively.
Found Phys (2010) 40: 390–417 407
What we have seen is that the puzzles one might have at first sight of Rovelli’s ac-
count can be resolved. But the resolution leaves one still uneasy, for it hinges on the
point that an observer O can register a measurement outcome—e.g. the answer 1 to
question ?A—but this fact is not equivalent to O being in a particular physical state,
whether relative to itself or relative to any other observer.
In other words there are elements of Rovelli’s ‘meta’ description which may in
particular cases not correspond to any information had by any observer, and hence
apparently not describable in the language of quantum mechanics. One might be
tempted to introduce the fiction that there is a ‘universal observer’ who knows what
information is had (what answers have been registered) by each ‘ordinary’ observer.
But this fiction can certainly not be admitted without ruining the story.
408 Found Phys (2010) 40: 390–417
At the same time, in our reflections on what the observers register as measurement
outcomes, we are targeting the very basis of Rovelli’s understanding of quantum
mechanics, and the very basis of the description of Rovelli’s world:
Quantum mechanics is a theory about the physical description of physical sys-
tems relative to other systems, and this is a complete description of the world.
(Sect. II-C)
Drawing on Rovelli’s favorite illustration of different frames of reference in Ein-
stein’s world, we are clearly tempted to ask: but what relations are there between
the descriptions that different observers give when they observe the same system? Of
course there can be no clue at all to an answer if we assume that there are no interac-
tions at all between these distinct observers. But perhaps we can get a clue if we think
of those distinct observers as themselves subject to observation by a third observer!
Doing so need not be illegitimate if we recall that Rovelli is describing the general
form that any ascription of states or observable-values can take, and that this is the
form of information that an observer could have.
Let us take a look back at how, in his ‘meta’ description, Rovelli introduces a sym-
bolism to express the fact that a given system ‘has’ information about another one:
If there is a maximal amount of information that can be extracted from the sys-
tem, we may assume that one can select in W (S) an ensemble of N questions
Qi , which we denote as c = {Qi , i = 1, N}, that are independent from each
other. There is nothing canonical in this choice, so there may be many distinct
families c, b, d, . . . of N independent questions in W (S). If a system O asks
the N questions in the family c to a system S, then the answers obtained can be
represented as a string that we denote as
sc = [e1 , . . . , eN ]c (4)
The string sc represents the information that O has about S, as a result of the
interaction that allowed it to ask the questions in c. (Section III-C)
The idea of a state of S relative to O enters now, because on the basis of this in-
formation, O can locate S in a finite subspace of the pertinent Hilbert space—even
assign it a particular pure state represented by a vector in that space if the question-
answer sequence was a maximally compatible one. This is what we describe infor-
mally in:
(Form 1) O registers answer 1 to complete question ?A, so S has state |A, 1
relative to O.
We observe now that there is in effect a time order: the order in which the questions
are asked. (Only order in time will be regarded for now, not time metric.) The N
questions in numbered line (4) appear in the order 1, . . . , N so we can think of them
as time-points, and can suggestively take them to indicate times t1 , . . . , tN . But then
the less formal description of (Form 1) should be expanded to the form:
Found Phys (2010) 40: 390–417 409
a) The measured system S starts off in a state θ = βi |A, i relative to ROV, which
is a superposition of eigenstates of an observable A corresponding to distinct
eigenvalues.
b) A will be measured twice, by two measuring systems O and P . Each of O and P
will be in the ‘ready to measure’ state relative to ROV to start, |B, r. The indicator
states are |B, i for eigenvalues i of A (which do not include r).
We assume that I and O each evolve freely after their measurement interaction, that
there is no interaction between O and P , and that both A and the ‘pointer-reading
observable’ B for O commute with the free Hamiltonians for S and O respectively.
c) The combined system S + O will be in state βi (|A, i ⊗ |B, i) relative to ROV
at t1 , the end of the first measurement.
d) Similarly at that time, taking into account the as yet
unchanging P , the state of
S +O +P relative to ROV will be the superposition βi (|A, i⊗|B, i⊗|B, r).
end of the second measurement the state of S + O + P relative to ROV
e) At t2 , the
will be βi (|A, i ⊗ |B, i ⊗ |B, i).
410 Found Phys (2010) 40: 390–417
When we look at d) and e) above, we see that the state of S relative to ROV does
not change after t1 , because the coefficients in the superposition do not change, even
though the components do.
To show this, note that by reduction, we have states also for parts of the total sys-
tem, namely S, O, P , relative to ROV. As before we write P [|A, i] for the projection
on the ray containing |A, i, etc. We deduce
f) ROV assigns to S all by itself an evolving mixture ρ(S, ROV)(t) of the states
|ψ(t, i) such that:
for t < t1 the state |ψ(t, i) = θ = βi |A, i,
for t1 ≤ t the state |ψ(t, i) = |A, i
This mixture has as components the projections on these evolving vectors, one for
each value i such that the coefficient βi is not zero, and the weights are the ‘squares’
of those coefficients.
Note well, that there is no change in this relative state at the second measurement
time, since in the superposition for the entire system, the values of B in O and P are
the same in each component (that is, for every eigenvalue i such that βi is not zero)
from that moment on. Hence the weights in the mixture do not change from t1 on.
But there is more. From the above it follows that the interaction between S and
the total system O + P is also the correlate of a measurement—in fact of three dis-
tinguishable measurements. For example, if we take B ⊗ I and I ⊗ B respectively as
pointer observables on O + P , then the vN criterion is satisfied for times t1 and t2
respectively. So we have:
two measurements of A by O + P , ending at t1 and t2 respectively
To see this, in the story as told in terms of states relative to ROV, let us look at the
overall evolution of the system, relative to ROV.
At the final time t2 , the complete system S + O + P is in pure state βi (|A, i ⊗
|B, i ⊗ |B, i) relative to ROV. By reduction the other states relative to ROV are:
S is in |βi |2 P [|A, i]
O and P are both in |βi |2 P [|B, i]
S + O and S + P are both in |βi |2 P [|A, i ⊗ |B, i]
O + P is in |βi |2 P [|B, i ⊗ |B, i]
Inspection shows that the vN Criterion is satisfied for the interactions I mentioned.
But we can add one more: taking B ⊗ B as pointer observable, we also see O + P
engaged in a measurement that ends at the later time t2 . That is the fifth measurement
which appears in this story of the states of these various systems relative to ROV, and
their various evolutions.13
The reason it is important to note this is of course that observers gain information
about systems only by measurement, and it is only if they gain information about
systems that those systems have states relative to them. So now we can continue,
in accordance with the meta-description of Rovelli’s world, to see what states S has
relative to O, P , and O + P .
States Relative to O, P , O + P
We can find the states of S relative to O, to P , and to O + P for that interval, except
that there will be some unknowns in it, namely the eigenstates that these observers
assign to S on the basis of the measurements they make on it. (ROV makes no mea-
surements on S, after the interval begins, that is why there are no similar unknowns
in our calculation of ρ(S, ROV ).) So we arrive at:
g) Observer O assigns to S an evolving pure state ρ(S, O)(t):
for t < t1 the state ρ(S, O)(t) = θ = βi |A, i,
for t1 ≤ t the state ρ(S, O)(t) = |A, m
13 There can be no objection, it seems to me, to allow for trivial limiting cases: if O has absolutely no
interactions with S through which information is gained, it is only a matter of bookkeeping if we say that
then the state of S relative to O is the represented by the Identity operator—the ‘informationless’ statistical
operator. This convention may at times smoothen the presentation, even if it is not really needed.
412 Found Phys (2010) 40: 390–417
and here the value m is an unknown, it is the result that O registers as outcome of the
measurement.
h) For P it is only a little more complicated: P assigns to S a mixture ρ(S, P )(t) of
the evolving pure states λ(t, i) with weights |βi |2 :
for t < t1 the state λ(t, i) = θ = βi |A, i,
for t1 ≤ t < t2 the state λ(t, i) = |A, i,
for t2 ≤ t the state λ(t, i) = |A, k
So consider how the situation looks to ROV. When ROV contemplates measurements
on these systems, to see if the pointer observables of O, P , O + P could be in
disagreement with each other at the pertinent times, the calculation of the Born con-
ditional probability for this will be zero. So, to follow the above suggestion as to
how to conceive of the un-measured world, ROV will conceive of the relations be-
tween what the subsystems register accordingly. The idea that any assertion about
what happens in nature must have cash value in what we can expect to detect, mea-
sure, or observe is strong in the Copenhagen tradition, even if contradicted by hidden
variable enthusiasts. It seems to me that it echoes precisely the sort of inspiration that
both the Copenhagen theorists and Rovelli derive from Einstein’s reasoning when he
introduced relativity.
So how is this inspiration honored by our Additional Postulate? If we now look
back to our description of the evolving states of S, through the relevant time interval,
relative to these three observers, we see the following pure state assignments:
For different values of m, k, r, s, those vectors are mutually orthogonal, since they
are all eigenvectors of the same operator. So the second and third line immediately
tell us that k = s. But the first and third line tell us that m = r when we attend to t1 ,
and similarly that m = s, when we attend to t2 . So all these numbers are after all the
same.
Result: the evolving states of S relative to the observers O and P are not the
same to begin, but they are the same once P makes its A-measurement on S,
sometime after O did (with no disturbance of A intervening meanwhile).
414 Found Phys (2010) 40: 390–417
6 Relational EPR
Laudisa [9] and Smerlak and Rovelli [12] have examined how the Einstein–Podolski–
Rosen situation can be regarded or modeled within Relational Quantum Mechanics.
They do not entirely agree in their approach. Here I shall show how the situation fares
if my Additional Postulate is accepted. The result appears to be different from what
is favored by Rovelli, though it does not seem to affect the empirical content of the
resulting formulation of quantum mechanics.
Let S be a two-part system α + β (such as a photon pair in singlet state), in a
superposition of correlated states ↑ ⊗ ↓ and ↓ ⊗ ↑. The arrows are eigenvalues of
observable A.
Observers P 1 and P 2 respectively measure A ⊗ I and I ⊗ A with pointer observ-
able B. ROV has information on initial states and dynamic process
and
(|B, 2 ⊗ x|↓) ⊗ (|B, 1 ⊗ |↑)
This implies that ROV assigns to α + β a mixture of (|↑ ⊗ |↓) and (|↓ ⊗ |↑).
By the Additional Postulate it follows that the state of α + β relative to P 1 + P 2
must be one of these, thus ruling out two of the possibilities noted above. And then
P 1 + P 2 will assign to α and β separately either |↑ and |↓ respectively or |↓ and
|↑ respectively. But then, again by the Postulate, the states of α and β relative to P 1
and to P 2 respectively cannot be the same, on pain of orthogonality to what they are
relative to P 1 + P 2.
Have we arrived at ‘spooky’ non-locality? We need to be worried by possible
conflict with the sentiment so clearly expressed in [12]:
There is no operational definition of observer-independent comparison . . . of
different observers’ information . . .: the information of different observers can
be compared only by a physical exchange of information between the ob-
servers.
Can ROV, in our story (including the Additional Postulate), compare the two states
of α + β relative to P 1 and P 2 before measuring P 1 and P 2 at the end?
YES and NO!
ROV can know that P 1 and P 2 did not register the same value for A. But to know
anything about which values they did register, ROV would have to make measure-
ments. So ROV can predict no more than someone who has not heard the additional
Postulate, but only that if he measured both he would find different registered values,
which is predictable with no reliance on the Additional Postulate.
We can define a function of the outcomes registered by P 1 and P 2, which takes
value 1 if the outcomes are the same and value 0 if they are different. It seems then
that ROV can know the value of this defined quantity, without having measured it.
But the defined quantity has value 1 if and only if the states of α and β relative to
P 1 and P 2 respectively are either |↑ and |↓, or |↓ and |↑. That this is so I do not
think follows in the original Relational Quantum Mechanics. Therefore this could be
counted as running contrary to the above cited sentiment.
But I would like to suggest that it may count as a reason for the suggested Addi-
tional Postulate. For otherwise we leave open the possibility that the state of α + β
relative to P 1 + P 2 is (|↑ ⊗ |↓, for example, although the states of α and β relative
to P 1 and P 2 respectively are both |↑!
Even if we were to insist that “ S has state . . . relative to O” can only have a
truth value related to a further observer ROV (and not be true or false ‘absolutely’)
this same difficulty would appear when ROV is in the picture.14 But there is much to
14 The suggestion here would be, it seems to me, a radicalization of the original Relational Quantum
Mechanics, but perhaps closer to the initial intuitions than what I have worked with here. However,
416 Found Phys (2010) 40: 390–417
explore here yet, including the most radical view, namely that even what the states
relative to any observer are must itself be relative to an observer.
Acknowledgements This research was supported by NSF grant SES-0549002. I want to thank Carlo
Rovelli and Matteo Smerlak for much helpful discussion. Thanks also to the discussion of a first draft
of this paper in a seminar at Princeton with a. o. David John Baker, Jeffrey Bub, Christopher Fuchs,
Hans Halvorson, and Tim Maudlin.
Finally, solely as an aid to the imagination, we can add some auxiliary symbolism, as
follows. We note that O registers an answer in an entirely physical way, in that this
measurement involves—and requires—a correlation of the measured observable A
on S with a ‘pointer’ observable B on O. Hence, if we wish to mark that correlation,
we have a final variant on Form 1:
though worth exploring further, I see it as difficult to sustain, given its obvious danger of either regress or
circularity—well, perhaps worth exploring precisely because of such danger!
Found Phys (2010) 40: 390–417 417
and in the there cited section of my 1991, it is possible to see just how the last line in
the Additional Postulate (“and so forth for larger composite situations”) would need
to be elaborated in detail.
References
1. Bub, J.: Why the quantum? Stud. Hist. Philos. Modern Phys. 35, 241–266 (2004)
2. Bub, J.: Quantum mechanics is about quantum information. Found. Phys. 35, 541–560 (2005)
3. Dicke, R.H.: Quantum measurements, sequential and latent. Found. Phys. 19, 385–395 (1989)
4. Fuchs, C.A.: Information gain vs. state disturbance in quantum theory. Fortschritte der Physik 46(4,5),
535–565 (1998) (Reprinted in Quantum Computation: Where Do We Want to Go Tomorrow? edited
by S.L. Braunstein (Wiley–VCH Verlag, Weinheim, 1999), pp. 229–259.)
5. Fuchs, C.A.: Quantum mechanics as quantum information (and only a little more) (2004).
arXiv:quant-ph/0205039
6. Groenewold, H.J.: Information in quantum measurement. Koninklijke Nederlandse Akademie van
Wetenschappen B55, 219–227 (1952)
7. Groenewold, H.J.: Objective and subjective aspects of statistics in quantum description. In Körner S.
(ed.) Observation and Interpretation in the Philosophy of Physics Proceedings of the Ninth Symposium
of the Colston Research Society held in the University . . . April 1st–April 4th, 1957, pp. 197–203
8. Hagar, A., Hemmo, M.: Explaining the unobserved—why QM ain’t only about information. Found.
Phys. 36, 1295–1324 (2006)
9. Laudisa, F.: The EPR argument in a relational interpretation of quantum mechanics. Found. Phys.
Lett. 14, 119–132 (2001)
10. Laudisa, F., Rovelli, C.: Relational quantum mechanics. In The Stanford Encyclopedia of Philosophy
(2005). http://plato.stanford.edu/entries/qm-relational/
11. Rovelli, C.: Relational quantum mechanics. Int. J. Theor. Phys. 35, 1637–1678 (1996)
12. Smerlak, M., Rovelli, C.: Relational EPR. ms (2006)
13. Temple, G.: The General Principles of Quantum Theory. Methuen, London (1948)
14. van Fraassen, B.C.: Modal interpretation of repeated measurement: Reply to Leeds and Healey. Philos.
Sci. 64, 669–676 (1997)
15. van Fraassen, B.C.: Quantum Mechanics: An Empiricist View. Oxford University Press, Oxford
(1991)
16. Wheeler, J.: Assessment of Everett’s ‘Relative State’ formulation of quantum theory. Review Mod.
Phys. 29, 463–465 (1957)