Edge modes as reference frames and boundary actions from

post-selection
Sylvain Carrozzaa,1 and Philipp A. Höhnb,c,2 .
a Institute for Mathematics, Astrophysics and Particle Physics, Radboud University, Heyendaalseweg
135, 6525 AJ Nijmegen, The Netherlands.
b Okinawa Institute of Science and Technology Graduate University, Onna, Okinawa 904 0495, Japan.
c Department of Physics and Astronomy, University College London, Gower Street, London, WC1E
6BT, United Kingdom.

Abstract
We introduce a general framework realizing edge modes in (classical) gauge field theory as dynamical
arXiv:2109.06184v4 [hep-th] 3 Feb 2022

reference frames, an often suggested interpretation that we make entirely explicit. We focus on a
bounded region M with a co-dimension one time-like boundary Γ, which we embed in a global
spacetime. Taking as input a variational principle at the global level, we develop a systematic
formalism inducing consistent variational principles (and in particular, boundary actions) for the
subregion M . This relies on a post-selection procedure on Γ, which isolates the subsector of the
global theory compatible with a general choice of gauge-invariant boundary conditions for the
dynamics in M . Crucially, the latter relate the configuration fields on Γ to a dynamical frame field
carrying information about the spacetime complement of M ; as such, they may be equivalently
interpreted as frame-dressed or relational observables. Generically, the external frame field keeps
an imprint on the ensuing dynamics for subregion M , where it materializes itself as a local field on
the time-like boundary Γ; in other words, an edge mode. We identify boundary symmetries as frame
reorientations and show that they divide into three types, depending on the boundary conditions,
that affect the physical status of the edge modes. Our construction relies on the covariant phase
space formalism, and is in principle applicable to any gauge (field) theory. We illustrate it on
three standard examples: Maxwell, Abelian Chern-Simons and non-Abelian Yang-Mills theories.
In complement, we also analyze a mechanical toy-model to connect our work with recent efforts on
(quantum) reference frames.

Contents
1 Introduction 2

Main notations 5

2 Edge modes and post-selection in a mechanical system* 6

2.1 Covariant phase space in mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Edge modes as “internalized” external reference frames . . . . . . . . . . . . . . . . . . 8
2.3 Symmetries as frame reorientations vs. gauge transformations . . . . . . . . . . . . . . 12
2.4 Reference frame changes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5 Subsystem phase space from post-selection on the global space of solutions . . . . . . . 16
2.6 Three types of symmetry transformations (frame reorientations) . . . . . . . . . . . . . 18
2.7 Reference frame changes as changes of foliation of the solution space . . . . . . . . . . 20

3 Covariant phase space method in field theory 21

3.1 Presymplectic structure of global field space . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Relation to canonical phase space methods* . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 Towards a regional covariant phase space: main questions to be addressed . . . . . . . 23
1
[email protected]
2
[email protected]

1
4 Edge modes as “internalized” external frames and gauge-invariant observables 24
4.1 Dynamical reference frames in gauge field theory . . . . . . . . . . . . . . . . . . . . . 24
4.2 Frame-dressed observables are relational observables . . . . . . . . . . . . . . . . . . . 26
4.3 Edge modes as dynamical reference frame fields . . . . . . . . . . . . . . . . . . . . . . 29
4.4 Concrete realization of group-valued edge reference frames . . . . . . . . . . . . . . . . 31
4.5 Symmetries as edge frame reorientations . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.6 Edge frame changes and transformations of relational observables . . . . . . . . . . . . 34
4.7 Frame-reorientation-invariant field-space form . . . . . . . . . . . . . . . . . . . . . . . 36

5 Symplectic geometry of splitting post-selection 36

5.1 Radiative presymplectic structure from a dressing flow in field space . . . . . . . . . . 36
5.2 Foliation with respect to field configurations on the time-like boundary . . . . . . . . . 38
5.3 Presymplectic structure of the subregion field space . . . . . . . . . . . . . . . . . . . . 40

6 Boundary actions from splitting post-selection 44

6.1 Splitting a global variational problem into local ones via post-selection . . . . . . . . . 45
6.2 Local Darboux coordinates, appearance of edge modes and non-uniqueness of the bound-
ary action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6.3 Summary of the algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

7 Illustration of splitting post-selection in example field theories 52

7.1 Scalar field theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
7.2 Maxwell theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
7.3 Abelian Chern-Simons theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
7.4 Yang-Mills theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

8 Conclusion and outlook 69

References 70

1 Introduction
It is well known that gauge field theories defined on manifolds with boundaries can support the emer-
gence of dynamical edge modes. This is unquestionable in the quantum theory, and in condensed
matter physics in particular, where such emergent degrees of freedom can be related to a wealth of
interesting phenomena (famously, chiral edge states are known to arise from the topological order un-
derlying the quantum hall effect [1, 2]). In a broader context, and already at the classical level, the
emergence of edge modes was argued to be a necessary consequence of gauge symmetry in the presence
of boundaries by Donnelly and Freidel [3]. This has led to a series of efforts to better understand the
notion of edge modes in gauge theories [4–15], gravity [5, 16–37] and string theory [38–40]. However,
the genericity of this particular argument for edge modes, its interpretation and physical relevance, all
remain debated in the literature. In particular, it is not a priori obvious how the two notions of edge
modes just mentioned — physical edge degrees of freedom on the one hand, and Donnelly–Freidel edge
modes on the other hand — are related to one another.
To various degrees, it was already suggested in the literature that Donnelly–Freidel edge modes
might be best understood as some kind of dynamical reference frames (see e.g. [8, 31] and also [17]).
Part of the purpose of the present work is to realize this idea in an entirely explicit manner. This will
allow us to rederive the structure of Donnelly–Freidel edge modes from first principles, while making
their interpretation as reference frames completely transparent; in particular, we will show that they
are dynamical reference frames in the same sense as they appear in the recent quantum reference
frame literature [41–54]. By investigating their interplay with generic boundary conditions, we will
also elucidate how they can sometimes (but not always) support a non-trivial algebra of boundary

2
symmetries acting on the physical phase space. At least in the classical set-up we are considering, this
is a necessary condition for Donnelly–Freidel edge modes to support the emergence of physical degrees
of freedom, and therefore, to reveal themselves in an experiment. We will show that different types
of boundary conditions lead to a refinement of the notion of boundary symmetry into three distinct
types. Since our primary focus will be on Donnelly–Freidel edge modes, from now on we will simply
refer to them as “edge modes”.
Beyond the question of edge modes per se, our work establishes new connections between subjects
that were previously discussed in the literature, which we believe to be of broader interest.
From a technical standpoint, we will employ covariant phase space methods [55–63] (see [64, 65]
for reviews) to describe the dynamics of fields in a (bounded) spacetime region M . For definiteness,
we will take M to have the topology of a cylinder, with time-like boundary Γ (but generalizing the
construction to the case of a null boundary should be reasonably straightforward). This problem has
received renewed interest in recent years [11,66–68], in part because of its relevance to holography, e.g.
see [37,69–78], gravitational observables [79], and more generally, to the characterization of asymptotic
symmetries in gauge and gravitational subsystems, see e.g. [28, 35, 68, 80–87]. Our work revisits such
constructions, from the vantage point of a global field space of solutions in a spacetime M ∪ M̄ ,
where M̄ is the complementary spacetime region of M . In that, it provides a direct link between the
global methods of [56] and the regional analysis of [11, 66]. In more concrete terms: given a consistent
variational principle for M ∪ M̄ , we will provide a general algorithm inducing consistent variational
principles, and specifically boundary actions, for fields in M .
By identifying edge modes as reference frames, our work establishes a bridge between field the-
ory and the recent literature on (quantum) reference frames. This permits us to identify the above
mentioned boundary symmetries as edge frame reorientations and to prove that frame-dressed observ-
ables — a systematic generalization of the examples considered in [3] — can be equivalently understood
as relational observables in the sense of [88–95]. It furthermore enables us to consider different systems
of edge frame fields and to show how to translate from the description relative to one frame field to
that relative to another, i.e. to establish the transformation between the relational observables relative
to different edge frames. Given that edge modes are only defined on the boundary Γ of M , rather than
its full bulk, they provide a particularly promising avenue for extending the efforts on quantum frame
covariance of physical properties [41–54] into the quantum field theory setting. Conversely, the dis-
tinction between (i) gauge transformations and symmetries (i.e. frame reorientations) or, equivalently,
constraints and charges, as originally proposed in [3], and (ii) different types of symmetries depending
on the boundary conditions as established below, has so far not been explored in the quantum reference
frame literature and warrants further investigation there.
A basic idea we will invoke in our construction is that of post-selection of a certain subspace of field
configurations in a global space of solutions. We view post-selection as a unifying concept allowing one
to incorporate operational constraints into a given overarching theory to select the subsector of that
theory consistent with the constraints. These are associated to physical events of a contingent nature.
Such constraints may encode the boundary conditions describing the context of a controlled experiment
(e.g. a scattering experiment), or more broadly speaking, the state of the world we happen to observe
as agents. This broad understanding seems particularly relevant in the gravitational context, where
controlled experiments are more the exception than the norm. In (quantum) general relativity, our
control/engineering abilities are essentially non-existent, but what we observe around us is nonetheless
post-selected from within the set of all possibilities admitted by the global theory.
In this paper, it is post-selection of field configurations on Γ that will provide a link between
the global covariant phase space for M ∪ M̄ , e.g. in the sense of [56], and the local one for M , e.g.
in the sense of [11, 66, 68]. The two additional guiding principles we will rely on in this endeavour
are: 1) that post-selection should operate on physical (as opposed to gauge) degrees of freedom, and
thus, on gauge-invariant functionals in the global solution space for M ∪ M̄ ; 2) that it should lead
to a consistent factorization of the global variational problem into two independent ones, for M and
its complement. Up to more minor and technical assumptions that we will make along the way,

3
this will be sufficient to derive the structure of edge modes on Γ, in a self-contained and conceptually
illuminating way. In particular, edge modes assume the role of “internalized” external dynamical frames
for subregion M and do not have to be added to the theory; they are always part of the global theory
and induced onto the regional one, encoding how the subregion M of interest relates to its spacetime
complement M̄ . Furthermore, the fact that edge modes can ultimately restore the invariance of the
regional presymplectic structure under field-dependent gauge transformations becomes secondary in
our approach: it is a direct consequence of our construction, not a main postulate (as it was in [3]).
Post-selection naturally prompts us to examine the interplay between edge modes and different
boundary conditions on Γ. Incidentally, we will investigate general boundary conditions, that can
be obtained by application of a linear canonical transformation of some set of gauge-invariant local
canonical Darboux coordinates for the presymplectic current on Γ. This will allow us to observe that,
within the same global theory space, the physical role of edge modes for the post-selected dynamics
in subregion M can in fact be contingent on the choice of boundary conditions. For instance, in
Maxwell theory, edge modes drop out of the regional presymplectic structure (and hence the regional
physical phase space) if one decides to impose Neumann boundary conditions, i.e. boundary conditions
on the flux ?F Γ . This suggests that, in this specific example, the concept of edge mode could be
dispensed with altogether. This is consistent with the interpretation of edge modes as external frames
for subregion M (originating in the complement M̄ ) and the fact that Neumann boundary conditions
correspond to reflecting boundary conditions and thus an essentially independent dynamics of M and
M̄ . Indeed, it was demonstrated explicitly in [6–9] that edge mode degrees of freedom are not strictly
necessary in order to construct a consistent phase space for Maxwell theory with boundary condition
on the flux. However, our work also demonstrates that they do have physical significance for other
(e.g., Dirichlet and Robin) boundary conditions in Maxwell, Chern-Simons and Yang-Mills theories.3
This ultimately leads to a refinement of the notion of boundary symmetries, as proposed in [3],
into three types depending on the boundary conditions, that we call symmetries, meta-symmetries
and boundary gauge symmetries. Symmetries are physical transformations in the regional phase space,
while meta-symmetries are symplectomorphisms between different regional phase spaces (i.e. different
subregion theories), and boundary gauge symmetries arise when the charges turn into additional first-
class constraints on the boundary (which will always be the case for Neumann boundary conditions).
Even though we are mainly motivated by (quantum) gravity, our focus in the present article will
be on non-gravitational gauge theories, in which the metric is non-dynamical (though can be curved).
We however anticipate our interpretation of edge modes as dynamical reference frames to extend (with
suitable adaptations) to gravitational theories, a problem we plan to return to in the future. In the long
run, our hope is that this viewpoint will help to further establish the notion of gravitational edge mode
as a valuable concept for quantum gravity, in line with recent efforts in this direction [21–32]. Given
the holographic nature of gravitational systems (see for instance [96]), and the fact that the quantum
reference frame literature provides a promising framework in which to define quantum generalizations
of the equivalence principle (see e.g. [53, 97] for initial explorations), we believe this hope is warranted.

Organization of the paper. We start in Sec. 2 by analyzing a Newtonian mechanical model featuring
a local gauge symmetry generated by a single constraint. This serves as a simple illustration of the
general formalism to be developed later, in a context where conceptual connections with the existing
literature on dynamical reference frames are not obscured by field theory technicalities. Readers who are
unfamiliar with the covariant phase space formalism will also find in this section a gentle introduction
to its philosophy. A more substantial introduction to this formalism is provided in Sec. 3, this time
in the context of field theory. This is where we fix some of our basic notations for the global field
space, and introduce additional background material on the relation between covariant and canonical
3
For Yang-Mills (and in particular Maxwell) theories, the imprint of edge modes on the reduced phase space remains
arguably mild: once boundary conditions have been imposed, at most a finite number of global charges survive at
the physical level. This is dramatically less than in Chern-Simons theory, which generates an algebra of symmetries
parametrized by a space of functions on Γ (hence, infinite-dimensional, see Sec. 7).

4
methods. In Sec. 4, we start by formalizing the notion of dynamical reference frame in the context of
field theory. This allows us, in a second step, to explicitly realize gauge field theory edge modes as
dynamical reference frames. We conclude that section by analyzing two types of transformations on
edge frames: boundary symmetries (following the nomenclature of [3]), which we identify as physical
frame reorientations; and changes of frames, which can be formalized as field-dependent generalizations
of the former. Along the way, we establish that the frame-dressed observables entering the construction
of Donnelly and Freidel can be generalized and understood as a covariant incarnation of the relational
observables originally defined in a canonical set-up by Dittrich [94, 95] (extending earlier work by
Rovelli [88–93]). We finally proceed with post-selection itself, which for didactic reasons, we describe
in two steps. In Sec. 5, we work exclusively on-shell of the bulk equations of motion, which already
allows us to induce a conserved presymplectic structure for the subregion M . Moving on to an off-shell
description of the same procedure in Sec. 6, we construct a general and systematic algorithm inducing
consistent variational principles for the subregion M , that is, to deduce an appropriate boundary action
imposing both the bulk equations of motion and the boundary conditions as dynamical equations on
Γ. Finally, we illustrate our results on standard examples of field theories in Sec. 7: scalar field theory,
which does not necessitate the introduction of edge modes, but in which the general algorithm of
Sec. 6 is still operational; Maxwell theory, which we take as a main basic example to illustrate our
formalism throughout the paper, and only work out in full detail in that section; Abelian Chern-
Simons theory, a particularly interesting example in which the edge modes construction gives rise to
an infinite-dimensional algebra of boundary symmetries; and finally, Yang-Mills theory, as a way to
illustrate non-Abelian features of the general formalism.

Main notations
For the reader’s convenience, we collect here our main field space notations; they will be of particular
relevance in Secs. 3 to 7.

Spacetime structures

M Spacetime region with the topology of S d−1 × R

M̄ Spacetime complement of M

M̊ Interior of region M

Γ Time-like boundary of M

Σ ∪ Σ̄ Space-like Cauchy surface in M ∪ M̄ , such that Σ ⊂ M and Σ̄ ⊂ M̄

 Spacetime volume form

Γ Volume form on Γ induced by 

d Spacetime exterior derivative

∧ Spacetime wedge product

Γ
Pullback to Γ

Field space structures

F Space of global field configurations

S Subspace of solutions (S ⊂ F)

P Physical phase space of gauge orbits (P = S/G, where G is the gauge group)

5
FM , SM , PM Analogously defined field spaces for the spacetime subregion M
X0
SM Subspace of solutions in SM obeying the boundary condition X = X0

δ Field-space exterior derivative

Field-space wedge product (kept implicit)

· Field-space interior product

Weak equalities

= Equality on F (resp. FM )

≈ Equality on S (resp. SM )
X0
≈
ˆ Equality on SM

Presymplectic structure

Θ Presymplectic potential (defined on F)

ω Presymplectic current ω := δΘ

Presymplectic form on S (Ω :=
R
Ω Σ∪Σ̄ ω)

ω∂ Boundary presymplectic current ((2, d − 2)-form on Γ)

X0
Presymplectic form on SM (ΩM := Σ ω + Γ ω∂ )
R R
ΩM

Gauge transformations and symmetries

Xα Field-space vector field generating the gauge transformation δα := Xα · δ

C[α] Gauge constraint associated to Xα (δC[α] = Xα · ΩM )

Yρ Vector field generating the symmetry (i.e. frame reorientation) ∆ρ := Yρ · δ

Q[ρ] Charge associated to Yρ (δQ[ρ] = Yρ · ΩM )

2 Edge modes and post-selection in a mechanical system*

*This section may be skipped by a quick reader only interested in our field theory constructions. It may
serve as an introduction to the covariant phase space formalism for the uninitiated.
As a warm-up exercise, and for the sake of conceptual clarity, we shall illustrate a few illuminating
observations in a mechanical toy model before later exploiting them in the field theory context. This
model mimics finite region gauge theories through groups of particles subject to translation invariance.
Specifically, this model serves to show

• that edge modes constitute “internalized” external reference frames for a subregion and are dy-
namical frames in the same sense in which they have been discussed in the recent quantum
reference frame literature [41–54];

• how to change from the description of the subregion relative to one choice of edge mode frame
to another one;

• how to obtain a subregion phase space structure from a boundary condition induced foliation of
the global space of solutions through a splitting post-selection procedure;

6
 • that gauge transformations and symmetries need to be distinguished [3], where we identify the
latter as frame reorientations. These frame reorientations have to be further distinguished into
three types: symmetries, meta-symmetries and boundary gauge symmetries. Symmetries leave a
subregion theory defined by given boundary conditions invariant, while a meta-symmetry maps
one such theory into another one defined by different boundary conditions. Boundary gauge
symmetries arise in addition to the already present gauge symmetry and only for Neumann
boundary conditions.

While the mechanical toy model captures many of the qualitative features we shall later encounter
in gauge field theory, it also leads to some differences which we try to highlight. For example, owing to
the discrete nature of the degrees of freedom, it does not give rise to corner terms. Furthermore, gauge
constraints arise as identities already off-shell and, due to the simplicity of the model, are non-local in
contrast to our field theory examples.

2.1 Covariant phase space in mechanics

Suppose we are given
R t a group of N particles in 1D Newtonian space, collectively called M , subject to
an action SM = t12 LM with Lagrangian form

LM = L(qi , q̇i )dt .

Our configuration space is Q = RN which is the mechanical analog of the space of instantaneous field
configurations, i.e. space of field configurations on some Cauchy slice. By contrast, the mechanical
analog of the space of field configurations here is the space of spatiotemporal configurations of M , in
other words, the space of histories for the time interval [t1 , t2 ], which we take to be FM̊ := {c : [t1 , t2 ] →
Q | c is a C 2 curve} and for simplicity often refer to as “field space”. M̊ denotes here internal or “bulk”
degrees of freedom in distinction of the situation when we include external frame particles below. The
action is a function SM : FM̊ → R, while the Lagrangian can be thought of as a functional L[c, t] =
L(qi (t), q̇i (t)) on FM̊ × [t1 , t2 ] because a history c = {qi (t), t ∈ [t1 , t2 ]} and an instant t determine the
positions and velocities (qi (t), q̇i (t)). For the moment it is irrelevant whether the Lagrangian L features
a gauge symmetry.
FM̊ is itself a smooth infinite-dimensional manifold, allowing us to define standard differential
geometric notions on it. In what follows, we will be using the covariant phase space formalism,
constructing a phase space from the subspace SM̊ ⊂ FM̊ of solutions to the equations of motion. In
line with standard convention for the covariant phase space in field theory, which we adopt throughout
this work, we denote the exterior derivative on field space FM̊ by δ, while reserving the notation d for
the exterior derivative of spacetime forms. In the mechanical case, these are forms on the time manifold
[t1 , t2 ] underlying the action SM . As such, we will be dealing with (r, s)-forms µ, which means µ is
an r-form on FM̊ and an s-form on spacetime. For example, the Lagrangian form LM is a (0, 1)-form.
The two exterior derivatives commute with each other ([d, δ] = 0), and both square to zero (d2 = 0
and δ 2 = 0). Rt
The variation of the action on field space δSM = t12 δLM is determined through the exterior
derivative of the Lagrangian (0, 1)-form
N  
X ∂L d ∂L
δLM = − δqi dt + dΘM̊ , (1)
∂qi dt ∂ q̇i
i=1

where the Lagrangian symplectic potential (a (1, 0) form) takes the form
N
X ∂L
ΘM̊ = δqi (2)
∂ q̇i
i=1

7
and gives rise to the Lagrangian presymplectic structure (a (2, 0) form)
N N N
X X ∂2L X ∂2L
ΩM̊ = δΘM̊ = δqj δqi + δ q̇j δqi . (3)
∂ q̇i ∂qj ∂ q̇i ∂ q̇j
i,j=1 i,j=1 i,j=1

Again in line with standard covariant phase space convention, we keep the field space wedge product
between field space one-forms δqj and δ q̇i implicit, keeping in mind that δqj δ q̇i = −δ q̇i δqj ; the notation
∧ is reserved for the spacetime wedge product.
The presymplectic structure is conserved on-shell, which follows from
dΩM̊
dΩM̊ = dt ≈ 0 , (4)
dt
where ≈ denotes evaluation on SM̊ ⊂ FM̊ , i.e. on solutions to the Euler-Lagrange equations of motion
∂qi − dt ∂ q̇i = 0. Indeed, this follows immediately from applying δ to δLM ≈ dΘM̊ , which is implied
∂L d ∂L

by equation (1).
q̇i , we find the potential and presymplectic structure in the simple forms
Defining 4 pi := ∂∂L

N
X N
X
ΘM̊ = pi δqi and ΩM̊ = δpi δqi , (5)
i=1 i=1

which will become useful shortly.

2.2 Edge modes as “internalized” external reference frames

In this subsection, we will illustrate the mechanical analog of edge modes. To this end, we have to
introduce a gauge symmetry. As the simplest possibility, suppose the Lagrangian is of the form

L(qi , q̇j ) = L {qi − qj , q̇i − q̇j }N (6)

i,j=1 ,

so that it features an invariance under translations qi (t) → qi (t) + α(t) and q̇i (t) → q̇i (t) + α̇(t), for an
arbitrary function of time α(t).5 In the mechanical context, the time manifold R assumes the role of
spacetime and since α depends on t, we can view this as a gauge-transformation of the corresponding
action.6 The gauge group is thus the translation group (R, +).
We denote by Xα the vector field on FM̊ which generates the above translation. It satisfies Xα (qi ) =
α = Xα · δqi and Xα (q̇i ) = α̇ = Xα · δ q̇i , where Xα · µ denotes the interior product of the vector field
Xα and the form µ on FM̊ × [t1 , t2 ]. Let g(α) ∈ R denote the translation group element amounting
to translation by α. Let f (qi (t), q̇i (t)) be an arbitrary functional on FM̊ × [t1 , t2 ] which depends
analytically on positions and velocities. It transforms under this group action as

g(α) . f (qi , q̇i ) = eXα f = f (qi + α, q̇i + α̇) , (7)

which clearly satisfies g(α + β) . f = g(α) . g(β) . f = g(β) . g(α) . f .

4
We are not performing a Legendre transformation and so pi is here simply a function on FM̊ × [t1 , t2 ].
5 1
PN 2
For example, the Lagrangian could read 2N i,j=1 (q̇i − q̇j ) − V ({qi − qi+1 }), which amounts to subtracting the
center-of-mass kinetic energy from the total kinetic energy and a translation-invariant potential.
6
For simplicity, we choose this spatially global gauge symmetry. In order to better mimic the
local P gauge symmetries in field theory, one could for instance choose a Lagrangian of the form
L = 21 i (q̇2i − q̇2i+1 )2 − V ({q2i+1 − q2i − q2i−1 + q2i−2 }). This would lead to a “local” gauge symmetry
(q2i (t), q2i+1 (t)) → (q2i (t) + αi (t), q2i+1 (t) + αi (t)), in term of a gauge parameter αi (t) which can be explicitly
“space” dependent. The exposition would become somewhat more convoluted, however.

8
2.2.1 Internal description of a group of particles
We can select any of the particles of M , say particle N , as a dynamical internal reference frame for
the translation group (R, +), relative to which we describe the remaining ones. Indeed, we can build
gauge-invariant quantities, such as relative distances qi − qN , using the frame degrees of freedom. More
formally, note that the frame position qN constitutes a (translation) group-valued dynamical reference
frame U = qN on field space, which transforms by group multiplication (here, the addition) under
translations, g(α) . U = g(α) + U = qN + α. The group being Abelian, left and right actions of
the group are of course the same. Using this frame, we can turn any functional f (qi (t), q̇i (t)) into a
gauge-invariant one by subjecting it to a frame-dressed gauge transformation defined by α = −qN + x,
for an arbitrary real function x(t):
Of,U (x) := g(−qN + x) . f = (U − x)−1 . f := g(α) . f α=−U +x
, (8)
where (U − x)−1 denotes the inverse group element.7 In coordinates, we thus have
Of,U (x, qi , q̇i ) = f (qi − qN + x, q̇i − q̇N + ẋ) , (9)
which clearly is translation-invariant. This frame-dressed observable is the (Lagrangian analog) of
a relational Dirac observable. They constitute so-called gauge-invariant extensions of gauge-fixed
quantities [46,58,94,98] and here encode the question “what is the value of f when the reference frame
(particle N ) is in orientation (position) x?” [88,89,91–94]. This is particularly transparent when setting
f = qi , giving the relative distance
Qi|N (x) := Oqi ,U (x) = qi − qN + x . (10)
Since the value of Oqi ,UN (x) is constant on any gauge orbit, this value is the same as on the gauge-fixing
surface qN = x, in this sense measuring the position of particle i when frame particle N is in position
x.
Next, we observe that Xα constitutes a null direction of the presymplectic structure. To this end,
we first note that, owing to our assumption on the Lagrangian in equation (6), pi = ∂∂L q̇i is translation-
invariant, so that Xα (pi ) = Xα · δpi = 0. Furthermore, the total momentum vanishes identically on all
of FM̊ × [t1 , t2 ]8
XN
PM = pi = 0 . (11)
i=1
This yields
Xα · ΩM̊ = −αδPM = 0 . (12)
In particular, Xα is also off-shell a degenerate direction of the presymplectic structure. By contrast, in
the field theory examples later, the constraints will be related to the equations of motion in such a way
that gauge directions in field space will only be on-shell null directions of the presymplectic structure.
We can use the relational observables to split the symplectic potential and presymplectic structure
into gauge-invariant and gauge parts, that is ΘM̊ = Θ̊inv + Θ̊gauge and ΩM̊ = Ω̊inv + Ω̊gauge , respectively,
where
N
X −1
Θ̊inv := pj δQj|N (x) Θ̊gauge := PM δqN = 0 , (13)
j=1
N
X −1
Ω̊inv := δpj δQj|N (x) Ω̊gauge := δPM δqN = 0 . (14)
j=1
7
For the translation group, we of course have (U − x)−1 = −(U − x). The reason we also use the notation g −1 for
the inverse group element here is for facilitating the comparison with the corresponding construction of gauge-invariant
observables for general groups in gauge field theory, which will be the focus of Sec. 4.
8
This is because PM constitutes a primary constraint (upon Legendre transformation) in the language of constrained
systems. This is yet another difference to the field theory examples later, where the gauge constraints are not identities
on field space, but only vanish on the subspace of solutions. The reason is that they are secondary constraints.

9
 For distinction from the case with external frame particles below, we have equipped the forms with a
˚. A characteristic feature of the gauge contributions Θ̊gauge and Ω̊gauge (here, but also in field theory)
is that they involve the constraints. However, in contrast to the field theory case later, the gauge part
vanishes identically here.
Let us now consider the construction of a phase space for the particle group M . Owing to the
degeneracy of the presymplectic form ΩM̊ , the field space FM̊ is not a phase space. In order to
construct a phase space PM̊ from it, we have to factor out the null directions of ΩM̊ : PM̊ := FM̊ / ∼,
where c ∼ c0 if the histories c, c0 differ by such a null direction. Clearly, Xα is a null direction, but
there are more: recalling that ΩM̊ is a (2, 0)-form, we have to fix a time t ∈ [t1 , t2 ] in order to turn it
into a genuine two-form ΩM̊ (t) on F. Note that outside of the space of solutions SM̊ ⊂ FM̊ , ΩM̊ (t)
will depend on the choice of t. It is clear that any non-zero vector field X which vanishes at t will be a
null direction of ΩM̊ (t), simply because it only changes qi (t0 ), q̇i (t0 ) at times t0 6= t. Taking both types
of degeneracies together, the equivalence class [c] consists of all histories that have the same gauge-
N −1
invariant (Qj|N (x, t), Q̇j|N (x, t))j=1 at t as c. The phase space PM̊ can thus be parametrized by these
gauge-invariant variables and is isomorphic to R2N −2 , hence finite-dimensional in contrast to FM̊ . We
−1
can, however, equivalently use (Qj|N , pj )N j=1 as a coordinate system for PM̊ . To see this, note that we
   
∂ Q̇ N −1
can write pj = ∂ Q̇∂L since ∂ q̇j|N j
= 1 and L {q i − q ,
j iq̇ − q̇ } N
j i,j=1 = L {Q j|N , Q̇j|N j=1 , so pj is
}
j|N
−1 N −1 N −1
determined by (Qj|N , Q̇j|N )N
j=1 . Conversely, (Qj|N , pj )j=1 determine (Qj|N , Q̇j|N )j=1 by the implicit
∂pj ∂2L
function theorem because ∂ Q̇j 0 |N
= ∂ Q̇j|N ∂ Q̇j 0 |N
is non-degenerate owing to the gauge-invariance of the
involved variables. Hence, PM̊ is also equipped with the non-degenerate symplectic form Ω̊inv (t).
In the present class of models, it turns out that the so constructed phase space, in fact, does not
depend on t. The reason is that it coincides with the phase space constructed directly from the space
of solutions SM̊ ⊂ FM̊ on which ΩM̊ (t) is independent of t thanks to equation (4). To see this, note
that Xα is also a degenerate direction of the pullback of ΩM̊ (t) to SM̊ , and that the set of Xα contains
any non-zero vector field X tangential to SM̊ which vanishes at t. The ensuing equivalence relation
s ∼ s0 of solutions s, s0 ∈ SM̊ is thus simply the restriction of the field space equivalence relation c ∼ c0
to the subspace of solutions. First consider the null directions defined by non-zero vector fields which
vanish at t (which are comprised of gauge directions Xα for which α, α̇ vanish at t). Factoring out these
null directions from SM̊ yields a finite-dimensional space isomorphic to T Q ' R2N , parametrized by
initial data (qi (t), q̇i (t)) at some time t ∈ [t1 , t2 ]. Further factoring out all remaining gauge directions
Xα , which amounts to identifying the initial data sets (qi (t), q̇i (t)) and (qi (t) + α(t), q̇i (t) + α̇(t)) for
arbitrary α(t), produces a space which is conveniently parametrized by the relational observables
N −1
(Qj|N , Q̇j|N )j=1 . This is, of course, the same phase space as above, i.e. altogether
PM̊ = SM̊ / ∼ = FM̊ / ∼ . (15)
This is also equivalent to the reduced phase space one would obtain if, instead of the covariant
phase space method, one were to first perform a Legendre transformation, followed by a constraint
reduction. Indeed, here the Legendre transformation, in coordinates defined by pi = ∂∂L q̇i (viewing L
at fixed t as a function L : T Q → R), would map the space T Q of positions and velocities (qi , q̇i ) to
the kinematical phase space Pkin := T ∗ Q. On Pkin , PM = 0 no longer is an identity, but a non-trivial
primary constraint defining a constraint submanifold C ⊂ Pkin . Factoring out the gauge orbits of PM
from C then yields the reduced phase space PM̊ , coinciding with equation (15). However, within the
covariant phase space formulation, the Legendre transformation is not a natural map to consider as it
relies on fixed t.
The ultimate reason for the coincidence of the two phase spaces, constructed from the field space
and space of solutions, respectively, is that Xα constitutes a degenerate direction of ΩM̊ both on- and
off-shell. This will be different in Maxwell, Chern-Simons and Yang-Mills theory later where con-
straints will only hold on-shell (they are secondary), such that gauge directions Xα are only degenerate
directions of the pullback of ΩM̊ to SM̊ . In those cases, the equivalence relation ∼ does not encom-
pass gauge directions and so Pkin := FM̊ /∼, C := SM̊ /∼ and P := SM̊ / ∼ /G, where G denotes the

10
group of spacetime gauge transformations, are distinct sets, constituting what are usually known as
the kinematical phase space, constraint surface and reduced or physical phase space, respectively; see
e.g. [56] (as well as Sec. 3.2). A field theory example where the kinematical and physical phase spaces
coincide as here is parametrized field theory [56].
Altogether, this constitutes a purely internal description of the group of particles M . At times it
will be interesting though to explore how a subsystem relates to its environment.

2.2.2 External description of the particle group: “extending the Heisenberg cut”
While the internal description of the particle group M , e.g. relative to particle N , thus lives in the
physical phase space PM̊ = SM̊ / ∼, i.e. on the set of equivalence classes of solutions, the kinematical
phase space Pkin , as we shall now elucidate, can be associated with the description of M relative to
an external reference frame. As just seen, the kinematical phase space is not a natural object to
consider in the covariant phase space formulation of the present mechanical model because it is here
not obtained through the natural reduction methods of the covariant formulation as in equation (15),
but through a Legendre transformation. We nevertheless discuss this physical interpretation of Pkin
here since it is useful for understanding edge modes conceptually and as we are ultimately interested
in field theory where no Legendre transformation is needed in order to construct Pkin (see Sec. 3.2).
From the point of view of the covariant method, the following interpretation of Pkin will thus be more
natural to entertain in field theory.
PM̊ comprises the physical states of M that are distinguishable relative to an internal frame of M at
any time t ∈ [t1 , t2 ], while Pkin can, in a sense to be made more precise below, be viewed as comprising
those states that are distinguishable relative to a frame external to M . This frame could be fictitious
or a physical one that we have thus far ignored. In the sequel, we consider the latter situation and
“internalize” such an external frame through a field space and corresponding phase space extension.
The set of physical states of the ensuing internal description of M relative to the new reference frame
will then be equivalent to Pkin . Although we are not in the quantum theory and neither considering
measurement interactions, this is somewhat reminiscent of the extension of the Heisenberg cut. In
the quantum theory, the below extension is related to what is known as the ‘paradox of the third
particle’ [48, 99] (see also [49] for a summary).
Suppose the group of particles M is not alone in the world and there is a further particle that
we call R1 (see Fig. 1). It will assume the role of the “internalized” external reference frame and will
be the mechanical analog of an edge mode for M . To accommodate it, we will have to extend the
translation-invariant Lagrangian in equation (6) as well as the underlying field space accordingly. In
line with the subsequent field theory discussion, let us denote this field space by FM , which is defined
in complete analogy to FM̊ above, incorporating all histories of the now N + 1 particles (group M and
frame R1 ). For this (N + 1)-body problem, we can now repeat the same exercise as before. Specifically,
the symplectic potential and presymplectic structure on FM (which depend on t) read
N
X N
X
ΘΣt = pi δqi + pR1 δqR1 and ΩM = δpi δqi + δpR1 δqR1 , (16)
i=1 i=1

where, in preparation for the field theory case, Σt here highlights that the potential depends on the
“time slice” t both on- and off-shell, in contrast to ΩM which on SM will be t-independent due to
equation (4). In contrast to equation (11), we now have the identity

PM R1 := PM + pR1 = 0 (17)

on all of FM × [t1 , t2 ], so PM need no longer vanish.

All the relational observables of the previous subsection remain invariant observables on FM . How-
ever, thanks to the additional frame particle R1 we can now turn all the kinematical degrees of freedom
associated with M into gauge-invariant ones. Specifically, we can now also relate the old internal frame,

11
particle N , to the “edge mode” R1 , yielding new relational observables, such as

QN |R1 (y) := OqN ,UR1 (y) = qN − qR1 + y . (18)

This permits us to decompose the symplectic potential and symplectic structure once more into gauge-
invariant and gauge parts, ΘΣt = Θinv + Θgauge and ΩM = Ωinv + Ωgauge , taking equation (13) into
account, where:
N
X −1
Θinv := pj δQj|N (x) + PM δQN |R1 (y) , Θgauge := PM R1 δqR1
j=1
N
X −1
Ωinv := δpj δQj|N (x) + δPM δQN |R1 (y) , Ωgauge := δPM R1 δqR1 . (19)
j=1

In particular, observe that only the last term in the invariant parts are dressed by the edge mode R1 .
These are the mechanical analogs of the radiative contributions to the presymplectic potential and
presymplectic structure on the boundary that we shall later see in field theory, while the first terms
are the mechanical analog of the bulk contribution in field theory. In contrast to the field theory case,
no corner term can arise in this discrete model and the gauge parts vanish identically.
It is clear that the physical phase space

PM := SM / ∼ = FM / ∼ (20)

of M and R1 together is isomorphic to R2N , equipped with the non-degenerate symplectic form Ωinv ,
−1
and parametrized by the pairs (Qj|N , pj ; QN |R1 , PM )N j=1 . Specifically, we can describe PM equivalently
through a gauge-fixing since, e.g., qR1 = y for some y ∈ R is a globally valid gauge fixing. On the
corresponding gauge-fixing surface we have QN |R1 (y) = qN such that the coordinates on PM become
N −1
(Qj|N , pj ; qN , PM )j=1 . This is equivalent to the Legendre transformation induced natural coordinates
(qi , pi )i=1 on the kinematical phase space Pkin := T ∗ Q ' R2N ; in particular, the invariant
N
PN part of the
symplectic structure can (in this gauge-fixing) be written equivalently as Ωinv = i=1 δpi δqi . In this
sense, the set of gauge-invariant states of M relative to R1 is equivalent (even symplectomorphic) to
the kinematical phase space associated with M only, i.e. PM ' Pkin . This equivalence is rooted in
the above observation that we can turn all kinematical degrees of freedom associated with M into
gauge-invariant ones by relating them to the edge mode R1 .

2.3 Symmetries as frame reorientations vs. gauge transformations

Given the internalized external frame R1 , we now have to distinguish two types of transformations
which are indistinguishable in a purely internal description of M on PM̊ , as they leave all the relations
within M invariant (see Fig. 1):
Gauge transformations: A global translation

qi → qi + α , i = 1, . . . , N , and qR1 → qR1 + α (21)

of M and R1 leaves the relation between the new frame R1 and M invariant and thus describes
the same physical situation. Denote the vector field on FM generating this transformation once
more by Xα , where α may be “field-dependent”, i.e. depend on where it is evaluated on field space
FM . It constitutes a null-direction of the presymplectic structure

Xα · ΩM = δC[α] = 0 , (22)

with “constraint” 9
C[α] = −αPM R1 = 0 . (23)
9
We write it in quotation marks as it vanishes identically everywhere on FM × [t1 , t2 ].

12
Symmetries (frame reorientations): A relative translation of M and R1 by a distance ρ. While
we could equivalently formulate such a transformation as a transformation of only M , we shall
write it as a transformation purely of the new frame R1 :

qi → qi , i = 1, . . . , N , and qR 1 → qR 1 + ρ . (24)

This transformation does not leave the relation between R1 and M invariant and thereby changes
the physical situation. Since the position qR1 constitutes the “orientation” of the frame, this
transformation amounts to a frame reorientation. Denote the vector field on FM generating this
transformation by Yρ , where ρ must now be field-space-independent.
For concreteness, let us assume that thePonly dependence of the Lagrangian for M, R1 on the
velocities is of the form Lkin = 4(N1+1) N +1
I,J=1 (q̇a − q̇b ) , where I, J runs over all particles in
2

M and the edge mode R1 , which is tantamount to assuming unit mass for the particles and
subtracting the center-of-mass kinetic energy from the total kinetic energy, e.g. see [42, 43]. This
yields pI = q̇I − N 1+1 J q̇J and therefore the variations
P

ρ̇
∆ρ Qj|N = Yρ · δQj|N = 0, ∆ρ QN |R1 = Yρ · δQN |R1 = −ρ, and ∆ρ pi := Yρ · δpi = − ,
N +1
where i = 1, . . . , N runs over the particles of M . Contraction with the presymplectic structure
gives
Yρ · ΩM = Yρ · Ωinv = δQ[ρ] , (25)
with gauge-invariant charge
N −1
1 X N
Q[ρ] = ρPM − ρ̇ Qj|N − ρ̇ QN |R1 . (26)
N +1 N +1
j=1

which does not in general vanish even on-shell. Specifically, for an edge-mode-independent func-
tional on FM of the form f (qi , q̇i ), we have, thanks to the analog of equation (9) for when R1 is
used as frame,
Of,UR1 +ρ (y) = Of,UR1 (y − ρ) (27)
and
∂Of,UR1 (y) ∂Of,UR1 (y)
(28)

ρ + ρ̇ = Q[ρ], Of,UR1 (y) = ZO · Yρ · Ωinv ,
∂y ∂ ẏ
where ZO is the Hamiltonian vector field on PM associated with the relational observable
Of,UR1 (y). In line with the interpretation of frame reorientations, we thus see that the charge
generates changes in the frame orientation label y of the edge mode for such observables.
Just like the group action . in equation (7) corresponds to gauge transformations acting on all
dynamical variables, we can now define the group action corresponding to symmetries, i.e.

g(ρ) f (qi , qR1 , q̇i , q̇R1 ) = eYρ f = f (qi , qR1 + ρ, q̇i , q̇R1 + ρ̇) . (29)

This permits us to rewrite equation (27) as

g(ρ) Of,UR1 (y) = Of,g(ρ) UR1 (y) = Of,UR1 (y − ρ) . (30)

This will become useful to describe frame changes shortly.

This distinction between gauge transformations and symmetries is the mechanical analog of the
distinction in gauge theories put forth in [3]. What is new here is the connection with dynamical
reference frames and especially the identification of symmetries as frame reorientations of the “edge
mode”. We shall see in Sec. 2.6, that we have to introduce a further distinction of symmetries into
three types.

13
 α α α α M M̄ α
Xα ···
1 1 2 2 3 3 N N R1 R1

M M̄ ρ
Yρ ···
1 2 3 N R1 R1

Figure 1: Two types of transformations which are indistinguishable from their action on particles in
region M , but distinguishable relative to the dynamical frame R1 . The vector field Xα (top) moves
both the particles in M and the frame R1 by the same distance, therefore does not affect relative
distances and corresponds to a gauge direction. The symmetry (frame reorientation) generator Yρ
(bottom) only acts on R1 , thereby affects relative distances to the reference frame: it generates a
physical transformation.

We finally note that an infinitesimal frame reorientation Yρ , which we have so far defined on FM ,
does not necessarily translate into an actual symmetry of the solution space SM . The translation
parameter ρ needs to verify some non-trivial set of differential equations in order for Yρ to act tangen-
tially to SM , that is, in order for ∆ρ to map solutions to solutions. In the present model, requiring
those equations to have non-trivial solutions turns out to impose severe constraints on the shape of
the translation-invariant potential V . As a result, for the purpose of illustrating the physical action
of symmetries on solutions, it will be convenient to simply assume a vanishing potential V = 0. In
that case, it is clear that Yρ leaves the space of solutions invariant (in particular ṗR1 = 0, where
pR1 = q̇R1 − N 1+1 ( Ni=1 q̇i + q̇R1 )), if and only if ρ̈(t) = 0. That is, we must require ρ(t) to be a linear
P
function of time.

2.4 Reference frame changes

The choice of dynamical reference frame is not unique. For instance, we have already used the internal
frame particle N and the “edge mode” R1 . As in the field theory context later, there could be additional
external frames that we can internalize as further edge modes. It is therefore natural to inquire about
a further type of transformation we have so far not considered: changes of reference frame. How
can we change from the description relative to one dynamical frame to that relative to another?
Plenty of recent efforts in the context of quantum reference frames have focused on developing a
framework for dynamical frame covariance [41–54]. For example, the classical discussion of [42, 43]
could be directly applied to the present mechanical scenario. Here, we shall restrict our attention to
the transformation of relational observables under changes of dynamical frame, i.e. we consider an active
change from the relational observables relative to one frame to those relative to another. The following
discussion can thereby be viewed as an extension of the exploration in some of the above references
where relational observables relative to different frame choices where considered (e.g., see [50]), but the
explicit transformations between them were not given.10 It also bears some resemblance to the “clock
transformations” in [100, Sec. 5].
Suppose we are given a second “edge mode” R2 in addition to R1 and the original group of particles
M . R2 can be accommodated by extending accordingly the translation-invariant Lagrangian, as well
as the field space (and thus physical phase space), in the same manner in which these structures were
extended in the previous subsection to internalize the first “edge mode” R1 . We denote the group-
valued orientation variables of the two dynamical external frames to M by URk = qRk , k = 1, 2. See
10
The frame changes in these references rather encompass how one and the same invariant observable can be described
relative to different internal frame choices. This involves an additional reduction using a gauge fixing procedure compared
to the present treatment.

14
Fig. 2.

QR2 |R1
Q1|R1 Q2|R1

···
1 2 N R1 R2
M M̄

R1 → R2

Q1|R2 Q2|R2
QR1 |R2

···
1 2 N R1 R2
M M̄

Figure 2: Pictorial illustration of the frame change mapping the relational observables {Qi|R1 } to
{Qi|R2 }. It is implemented by a field-dependent reference frame transformation, namely a translation
by the relative distance UR2 − UR1 .

Consider an arbitrary functional f (qI (t), q̇I (t)) which depends analytically on positions and veloci-
ties, where I runs over the N particles in M , as well as R1 , R2 . Invoking equation (8), the corresponding
relational observable relative to frame Rk reads
Of,URk (yk ) = (URk − yk )−1 . f := g(α) . f α=−URk +yk
. (31)

This implies that the transformation from Of,UR1 to Of,UR2 is given by a relation-conditional translation
by the gauge-invariant relative distance between the two frames (which is a dynamical element of the
translation group (R, +)):
g21 (UR1 , UR2 ) := QR1 |R2 (0) := OUR1 ,UR2 (0) = UR1 − UR2 . (32)
This implies that the transformation from the description relative to frame R1 to the one relative to
frame R2 is given by the relation-conditional translation of the frame R1 :
−1
UR1 7−→ UR2 = g21 UR1 := UR1 − g21 , (33)
−1
Of,UR1 (y1 ) 7−→ Of,UR2 (y2 ) = (−y1 + y2 + g21 ) Of,UR1 (y1 ) := Of,(−y1 +y2 +g21 )−1 UR1 (y1 ) .

Note that this constitutes an extension of the definition of in equation (29) to field-dependent frame
translations, in complete analogy to how equation (8) extends field-independent to field-dependent
gauge transformations. In particular, just like equation (8) gives rise to observables invariant under
gauge transformations, equation (33) generates observables invariant under symmetries (reorientations
of the frame R1 ). Indeed,

−1
Of,UR2 (y2 ) = (UR1 − y1 ) − (UR1 − y1 ) + (UR2 − y2 ) .f

−1
−1 
= − g21 − y2 + y1 . UR1 − y1 .f

−1
= − g21 − y2 + y1 Of,UR1 (y1 ) (34)
= g(β) . f β=(−y1 +y2 +g21 )−1 α(UR1 )
,

where (−y1 + y2 + g21 )−1 α(UR1 ) = −UR2 + y2 and α(UR1 ) = −UR1 + y1 is the field-dependent
group parameter used in equation (31).11 We can therefore interpret the frame change transformation
11
Note that we cannot write the third line in (34) in terms of the global translation group action as

−1
g21 − y2 + y1 . Of,UR1 (y1 ) since g . Of,UR1 (y1 ) = Of,UR1 (y1 ) for all g ∈ R.

15
as a relation-conditional frame reorientation. Specifically, it is equivalent to the relation-conditional
translation of the frame orientation label

y1 7→ y1 + (−y1 + y2 + g21 ) = QR1 |R2 (y2 ) (35)

i.e.
−1
(36)


− g21 − y2 + y1 Of,UR1 (y1 ) = Of,UR1 QR1 |R2 (y2 ) .
This is especially transparent in the coordinate expression, cf. equation (9), for relational observables
Of,UR1 (y1 ) = f (qi − UR1 + y1 , q̇i − U̇R1 + ẏ1 ).
It is instructive to illustrate this on a generating set of observables. Since the momenta are
translation-invariant and therefore trivially relational observables relative to both R1 , R2 , we only
need to look at the relational observables associated with the positions f = qI :

QI|R1 (y1 ) = qI − UR1 + y1 7−→ qI − UR1 + QR1 |R2 (y2 ) = QI|R2 (y2 ) . (37)

This includes the tautological observables (e.g., encoding “what is the position of R1 when R1 is in
position y1 ?”)

QR1 |R1 (y1 ) = y1 7−→ QR1 |R2 (y2 )

QR2 |R1 (y1 ) = UR2 − UR1 − y1 7−→ y2 = QR2 |R2 (y2 ) .

Including the tautological observables and the identity PM + pR1 + pR2 = 0, the change of frame thus
amounts to a canonical transformation on the physical phase space

N
N
Qi|R1 , pi ; QR2 |R1 , pR2 i=1
7−→ Qi|R2 , pi ; QR1 |R2 , pR1 i=1
. (38)

This can also be viewed as a “field redefinition” on the physical phase space.
It is important to note that, despite the change of frame transformation admitting the interpretation
of a relation-conditional (and hence configuration-dependent) frame reorientation, it is not a symmetry;
as seen in the previous subsection, symmetries are configuration-independent translations of the frame
R1 . In particular, the change of frame transformation is not generated by the charges Q[ρ]. Note also
that, in contrast to a symmetry, it does not change the relations between the N + 2 particles, only the
description of them.

2.5 Subsystem phase space from post-selection on the global space of solutions
Thus far, we have considered the covariant phase space structures associated with the group of particles
M with and without the two “edge modes” R1 , R2 . Accommodating the edge modes required us to
extend the field and phase spaces, embedding a setup with fewer into a setup with more particles. Let
us now consider the opposite process: how can we recover the covariant phase space structure of fewer
particles from that with more particles? This serves as a toy scenario for recovering local subregion
phase space structures from those associated with a global spacetime region in field theory, which we
shall elaborate on later. This can be achieved through a post-selection on the global space of solutions.
Suppose that R1 , R2 are part of a second group M̄ of N̄ particles, the complement of M . We shall
now illustrate how to recover the phase space structure of the subsystem M from the global one for
M ∪ M̄ . To this end, let us assume for concreteness and simplicity that we are dealing with free particle
dynamics subject to a translation-invariant Lagrangian of the form LM ∪M̄ = 4(N1+N̄ ) N
P +N̄
I,J=1 (q̇I − q̇J ) ,
2

where I, J now runs over all particles in M ∪ M̄ . Then pI = q̇I − N +1 N̄ J q̇J , so that pi = q̇i −
P

1 PN 1 PN +N̄
j=1 q̇j − N +N̄ q̇C̄ , where i = 1, . . . , N runs over the particles of M . Here, qC̄ := N̄ j̄=N +1 qj̄ ,
N̄
N +N̄
where j̄ labels the N̄ particles in M̄ , is the center-of-mass of M̄ . In other words, the momenta in
M now only depend on M̄ through its center-of-mass. We denote the global field space on which
the Lagrangian is defined by F and the subspace of solutions by S. Let us choose particle N as the

16
internal frame for group M , as before, while we now select the center-of-mass C̄ of M̄ as an external
frame for M and an internal frame for M̄ . We could in principle choose any of the configuration
degrees of freedom in M̄ as the edge mode, however, to ensure a suitable dynamical decoupling, it
will be convenient to choose qC̄ as the edge mode. Repeating the previous exercises on F, we find the
presymplectic structure in the form
N
X −1 NX
+N̄
Ω= δpj δQj|N + δPM δQN |C̄ + (δPM + δPM̄ ) δqC̄ + δpj̄ δQj̄|C̄ , (39)
j=1 j̄=N +1

where j labels the N − 1 particles in M , except particle number N , and PM̄ = PC̄ = j̄ pj̄ is the
P

total momentum of M̄ . The relational observables QN |C̄ , Qj̄|C̄ are constructed as before according to
equation (10).12 We now have the identity

PM + PM̄ = 0 . (40)

The third contribution to Ω is the pure gauge part and thus vanishes identically. The first and
last contributions pertain solely to M and M̄ , respectively, and are mechanical analogs of the bulk
contribution of a subregion. By contrast, the second term Ωrad := δPM δQN |C̄ is the mechanical analog
of the edge-mode-dressed radiative contribution on the boundary between adjacent subregions M and
M̄ in field theory. We will consider the particle N as the boundary of M and (QN |C̄ , PM ) as the
“boundary variables”.
In splitting post-selection, we shall now post-select on the global space of solutions S those solutions
that are compatible with a choice of gauge-invariant boundary conditions on particle N . This will split
the global solution space into subgroup solution spaces for M and M̄ . The “boundary contribution“
Ωrad will play an important role in this process. In order to encompass different types of boundary
conditions (Dirichlet, Neumann, Robin and mixed), let us consider the linear combinations of the
translation-invariant data

X = a QN |C̄ + b PM , Y = c QN |C̄ + d PM , (41)

where a, b, c, d are real functions of t subject only to ad − bc = 1, in which case they define a symplectic
transformation
Ωrad = δPM δQN |C̄ = δY δX . (42)
 
The space of solutions S can be parametrized by solutions Qj|N (t), QN |C̄ (t), qC̄ (t), Qj̄|C̄ (t); t ∈ [t1 , t2 ]
 
and so also by Qj|N (t), X(t), qC̄ (t), Qj̄|C̄ (t); t ∈ [t1 , t2 ] . In particular, we can thus use the configura-
tion variable X to foliate S
G
S= SX0 , SX0 := {all histories c(t) ∈ S|X(t) = X0 (t)} , (43)
X0

where X0 (t) is some non-dynamical background function specifying the gauge-invariant boundary
condition for particle N (which, of course, have to be consistent with the equations of motion of the
relevant Lagrangian). For example, we could have a(t1 ) = d(t1 ) = 1 and b(t1 ) = c(t1 ) = 0 initially and
a(t2 ) = d(t2 ) = 0 and b(t2 ) = 1 = −c(t2 ) at the end, so that the boundary conditions are initially on
X(t1 ) = QN |C̄ (t1 ), hence Dirichlet, and finally on X(t2 ) = PM (t2 ), thus Neumann, and so altogether
of mixed type. Clearly, different choices of the functions a, b, c, d will lead to distinct foliations of S.
Post-selection now proceeds by selecting a leaf SX0 from the global solution space S, i.e. restricting
to all solutions consistent with the history X0 (t) of the variable X. Pulling back to any of these leaves,
we thus have δX S = δX0 = 0 and so
X0

Ωrad SX0
= 0. (44)

12 PN̄
There is a redundancy in the last term of equation (39) as j̄=1 Qj̄|C̄ = 0, but this will be of no relevance below.

17
Henceforth, we also write the weak equality ≈ ˆ for identities that only hold on SX0 (i.e. after the
boundary conditions have been imposed), so that Ωrad ≈ ˆ 0. In other words, post-selection guarantees
that there is no (physical) symplectic flux between M and M̄ since Ωrad is the only gauge-invariant
contribution to the global presymplectic structure Ω in (39) that involves degrees of freedom from M
and M̄ that are generically non-constant (the momenta, which depend on all particles, satisfy ṗI ≈ 0).
This also splits the phase space structures associated with M and M̄ . In particular, we can consider
the restriction of any leaf to M . As can be easily checked, the equations of motion for the qi and edge
mode qC̄ decouple from the remaining ones for M̄ and can thereby be solved independently; we can
therefore perform post-selection on this subset. This suggests to define the restriction of the leaf SX0
to M by including the edge mode in the form
n   o
X0
SM := c(t) M = Qj|N (t), QN |C̄ (t), qC̄ (t) , t ∈ [t1 , t2 ] c ∈ SX0 . (45)

Accordingly, we equip this “regional” space of solutions for M with the presymplectic structure (cf.
equation (39))
N
X −1
δpj δQj|N + δPM δQN |C̄ + (δPM + δPM̄ ) δqC̄ = ΩM̊ + Ωrad ≈
ˆ ΩM̊ , (46)
j=1

recalling that the second term on the left hand side vanishes on-shell of the boundary conditions, while
the last term vanishes
P identically. On-shell, the surviving piece of the presymplectic structure for M is
thus ΩM̊ = Ω̊inv = j δpj δQj|N given in (13), which is the “bulk” non-degenerate symplectic form for
M on the physical phase space PM̊ discussed at the end of Sec. 2.2.1. Indeed, using the equations of
motion in the form ṗI ≈ 0 and noting that Q̇j|N = pj −pN and Q̇N |C̄ = pN + N̄1 PM , it is straightforward
to check that dtd
ΩM̊ + Ωrad ≈ 0, i.e. the restricted presymplectic structure for M is conserved on-

shell. It is also gauge-invariant and any Xα constitutes a degenerate direction of it.13 Hence, upon
X0
factoring out the degenerate directions of the presymplectic structure from SM , we recover the phase
space PM̊ of internally distinguishable states for the subsystem M from splitting post-selection on
the global solution space for M ∪ M̄ . As discussed shortly, this does not mean, however, that the
X0
dynamics on PM̊ induced by SM will coincide with the isolated dynamics for M defined by the pure
subsystem Lagrangian L implicitly considered in Sec. 2.2.1; depending on the boundary condition X0 ,
the Hamiltonian may be distinct.
Splitting post-selection can also be performed at the level of the action, however, we shall only
discuss this in the field theory case later. This will lead to a systematic algorithm for generating
boundary actions.

2.6 Three types of symmetry transformations (frame reorientations)

Each leaf SX0 in the global space of solutions S can be regarded as (the space of solutions of) a
specific subsystem theory for M (resp. M̄ ), defined by the boundary conditions X0 . Indeed, since the
“edge-mode-dressed” variables X, Y contain degrees of freedom from both M and M̄ , these boundary
conditions specify how M couples to its complement M̄ . Different choices of X0 for X will lead
to distinct subsystem Hamiltonians (e.g., different parameters or explicit time dependence) for the
remaining N − 1 particles in M . This will be discussed in more detail in [101]. That is to say, different
X0
“subleaves” SM yield the same physical phase space PM̊ , but will correspond to distinct subsystem
dynamics and thereby contain different sets of solutions for Qj|N (t). We are therefore justified to think
X0
of the different SM as different theories for the subsystem M and, consequently, of the global space
of solutions S as a “meta-theory” containing all the different subsystem theories.
In Sec. 2.2.2, we discussed the distinction between gauge and symmetry transformations, originally
put forth in [3], and interpreted the latter as frame reorientations. Since our subsystem theory is now
13
Neither conservation nor invariance would be true had we only included the generally non-vanishing δPM δqC̄ for the
edge mode contribution.

18
subjected to boundary conditions when considered as a post-selection of the global theory, we have to
explore the consequences for symmetries in the presence of these boundary conditions. In particular,
we have to investigate whether the frame reorientations preserve the boundary conditions or whether
they give rise to changes of leaves SX0 in the foliations of S.
For concreteness, let us once more assum a free particle dynamics as in the previous subsection.
The vector field Yρ on S generating the frame reorientation (24) (for the center-of-mass) then leads to
the variations ∆ρ pi := Yρ · δpi = − N N̄
+N̄
ρ̇, where i = 1, . . . , N runs over the particles of M . This has
two consequences: first, it leads to the following variation of the variable fixed through the boundary
conditions
N N̄
∆ρ X = Yρ · δX = −ρa − ρ̇b . (47)
N + N̄
Hence, only symmetries with ρ such that ∆ρ X = 0 will leave the boundary conditions invariant.
X0
Second, restricting Yρ to the subleaf SM , we can contract it with its presymplectic structure ΩM̊ +Ωrad .
X0
Note that, while the pullback of Ωrad to SM vanishes, the contraction with Yρ , which may not be
tangential to SX0 in S, will not in general vanish. This leads to

(48)


Yρ · ΩM̊ + Ωrad = δQ[ρ]

with charges in the form (cf. (26))

N −1
N̄ X N N̄
Q[ρ] = ρPM − ρ̇ Qj|N − ρ̇ QN |C̄ . (49)
N + N̄ j=1 N + N̄

Given that they constitute generators of an Abelian transformation group

(translations of R1 ), they

generate an Abelian Poisson algebra, {Q[ρ], Q[σ]} = Yσ · Yρ · ΩM̊ + Ωrad = 0.
This leads to a further distinction of three types of symmetry transformations (frame reorienta-
tions).

Symmetries. These are frame reorientations Yρ which preserve the boundary condition X = X0 ,
hence leave SX0 in S, and feature an unconstrained charge. These are symmetries with time-
dependent ρ(t) such that ∆ρ X = 0 and δQ[ρ] ≈ 6 ˆ 0, by which we mean that the pullback of
X0
δQ[ρ] to the leaf SM does not vanish (e.g., Q[ρ] cannot be proportional to X). The above
implies that this is only possible for a, b 6= 0 and thus requires Robin boundary conditions.
Solving ∆ρ X = 0 in this case and recalling from the end of Sec. 2.3 that for free particles frame
reorientations must be of the form ρ(t) = c1 t + c2 with c1 , c2 = const to be determined, yields

c2 N + N̄ b
= −t − ,
c1 N N̄ a
which only has a solution for special choices of a, b. When it has a solution, only the ratio c2 /c1
is specified, which means there will then be a one-parameter family of solutions. The Abelian
X0
Poisson algebra of the non-vanishing charges on SM ,
 
N −1
N̄ c1  N X
Q[ρ] = − X0 + Qj|N  , (50)
N + N̄ a
j=1

is then of dimension dim R = 1, in line with generating a translation group symmetry.

Meta-symmetries. These are frame reorientations Yρ which do not preserve the boundary condition,
∆ρ X 6= 0, and so change leaf SX0 (or subsystem theory) in the foliation of the meta-theory S.
For example, in the case of Dirichlet boundary conditions, i.e. a = 1 and b = 0 so that
X = QN |C̄ = Q0N |C̄ , we have ∆ρ QN |C̄ = −ρ so that any frame reorientation will lead to a
change of subsystem theory. In fact, Dirichlet boundary conditions are the only ones which

19
 imply that every frame reorientation is a meta-symmetry. By contrast, for some fixed Robin
boundary conditions, hence a, b 6= 0, there will typically exist some ρ satisfying ∆ρ X 6= 0 and
others satisfying ∆ρ X = 0, so Robin boundary conditions will generally admit meta-symmetries,
as well as symmetries.
While Yρ thus leads to a change of leaf SX0 , note that it is a symplectomorphism in the meta-
theory S. Indeed, using Cartan’s magic formula LY Ω = δ(Y · Ω) + Y · (δΩ), we find

(51)


LYρ ΩM̊ + Ωrad = δ (δQ[ρ]) = 0 ,

which is the infinitesimal form of a symplectomorphism.

Boundary gauge symmetries. These are frame reorientations Yρ preserving the boundary condi-
tions, ∆ρ X = 0, and generating a constrained charge δQ[ρ] ≈ ˆ 0. For instance, we could have
Q[ρ] = g X for some background function g(t). In the present model, this is only possible for Neu-
mann boundary conditions, i.e. for a = 0 so that X = bPM = X0 , and ρ time-independent,
which leads to the new “boundary constraint”
 
1
C∂ [ρ] = ρ PM − X0 ≈ ˆ 0. (52)
b

Since Yρ defines a null-direction, Yρ · ΩM̊ + Ωrad = δC∂ [ρ] = 0 on this subsystem theory, it

constitutes a gauge transformation. Indeed, all physical variables that are to survive on the sub-
system phase space have to leave the conditions defining the theory invariant and must therefore
be invariant under Yρ . While (Qj|N , pj ) are invariant (recall that ρ̇ = 0), we have ∆ρ QN |C̄ = −ρ.
Hence, while the edge-mode-dressed QN |C̄ is gauge-invariant under global translations Xα , it
fails to be invariant under the boundary gauge symmetries appearing for Neumann conditions.
However, since Yρ therefore only varies a term in Ωrad (through Y ) which vanishes on-shell of
X0
the boundary conditions, it constitutes a direction that in any case is factored out from SM in
the reduced phase space construction. It therefore does not lead to an additional reduction of
the dimension of the reduced phase space, in contrast to the analogous situation in field theory
later. A special case of this constraint is PM ≈ ˆ 0 in which case one obtains precisely the purely
internal subsystem theory on PM̊ discussed in Sec. 2.2.1, in which no edge mode appeared.

In conjunction with the previous subsection, this implies that the edge-mode-dressed configuration
observables of M drop out from the ensuing phase space PM̊ for the subsystem M ,14 regardless of
the boundary conditions. However, the edge mode C̄ plays a crucial role in the formulation of the
boundary conditions giving rise to the foliation of the global solution space into subsystem theories
for M . Through the boundary conditions, the edge mode will also determine the dynamics on PM̊
induced by the global Lagrangian on M ∪ M̄ [101]. Altogether, the edge mode is therefore relevant for
describing in an invariant manner how a subsystem couples to its complement.
Below (especially in Sec. 7), we shall elaborate on these observations in the field theory setting,
where the situation is similar, but also more subtle.

2.7 Reference frame changes as changes of foliation of the solution space

Symmetries as frame reorientations thus divide into those that are tangential or transversal to a leaf
SX0 in a foliation of S, but they always preserve the foliation. What kind of transformations map
between distinct foliations of S and so between different classes of boundary conditions, i.e. subsystem
theories?
For a fixed external frame R1 , different foliations arise from different choices for the functions
a, b, c, d that define the “boundary variables” X, Y in equation (41). These different foliations are
therefore related by the corresponding canonical transformations in S.
14
The momentum variables pi , which depend on q̇C̄ , however, survive on PM̊ .

20
 Similarly, the second external frame R2 will give rise do different edge-mode-dressed “boundary
variables” X 0 , Y 0 , not encompassed by X, Y in equation (41). These lead to a distinct family of
foliations of S. It is the reference frame transformations of Sec. 2.4 that link the family of foliations
relative to frame R1 with the one relative to frame R2 . This bears some analogy to spacetime, where
different foliations are related by (spacetime) reference frame transformations too.

3 Covariant phase space method in field theory

We now begin our systematic investigation of edge modes in (gauge) field theory. In the remainder of the
paper, we will consider relativistic and Lagrangian field theories defined on d-dimensional Lorentzian
manifolds of the form M ∪ M̄ , where M is the spacetime subregion of primary interest and M̄ its
complement. For definiteness, the interface between M and M̄ , which we denote by Γ, is assumed to
be time-like with the topology of S d−2 × R; see Fig. 3 from Sec. 4. The main purpose of this section
is twofold: first, to summarize the key ingredients of the covariant phase space formalism we will need
at the global level, that is in M ∪ M̄ ; and second, to anticipate the main questions we will have to
address in going from the global level to the regional covariant phase space for subregion M .

3.1 Presymplectic structure of global field space

We start from the space of off-shell field configurations in M ∪ M̄ , which we denote by F. Given a
Lagrangian L, one can define a presymplectic potential Θ on F via

δL = Ea δΦa + dΘ , (53)

where Φa are local coordinates on F, Ea ≈ 0 are the bulk equations motion associated to L and
ω := δΘ is the presymplectic current. We note that L, Θ and ω are (0, d)-, (1, d − 1)- and (2, d − 1)-
forms, respectively, where we recall that an (r, s)-form is an r-form on F and an s-form on M ∪ M̄ .
By construction, ω is spacetime closed on-shell, since:

dω = dδΘ = δdΘ = δ (δL − Ea δΦa ) = −δEa δΦa ≈ 0 . (54)

We can then introduce a canonical presymplectic two-form

Z
Ω= ω (55)
Σ∪Σ̄

on the subspace of on-shell configurations S ⊂ F, where Σ ∪ Σ̄ is some arbitrary Cauchy surface in

M ∪ M̄ , such that Σ ⊂ M and Σ̄ ⊂ M̄ . In case M ∪ M̄ has spatial boundaries (whether asymptotic
or finite), we assume that suitable boundary (or fall-off) conditions have been imposed at the global
level so that ω ≈ 0 there. Together with (54), this in turn ensures that Ω, seen as a two-form on the
solution space S, is independent from the choice of Cauchy surface Σ ∪ Σ̄.
Vector fields Xα on F which on-shell lie in the kernel of the presymplectic form,

Xα · Ω ≈ 0 , (56)

are infinitesimal gauge transformations. They form a Lie algebra [56] and generate the gauge group G
(group of spacetime gauge transformations). Any two solutions related by an element of G lie on the
same gauge orbit and must be considered as physically equivalent. Ignoring technical subtleties (such
as the Gribov problem), one obtains a symplectic form Ω̂ on the space of gauge orbits P := S/G by
quotienting Ω with respect to G. We shall refer to P as the physical phase space or, equivalently, as
the reduced phase space.
Our goal in the rest of the paper will be to elucidate how to induce a consistent dynamics for
fields in the subregion M via a post-selection procedure on the solution space of the global dynamical
system in M ∪ M̄ . In the presence of gauge symmetries, post-selection can in principle be performed at

21
two levels. As a first option, one can post-select in the physical phase space of gauge orbits equipped
with the symplectic form Ω̂, to directly obtain a phase space PM for subregion M , equipped with a
symplectic form Ω̂M . Alternatively, one can implement post-selection prior to gauge reduction. Given
S and Ω as input, one then looks for a presymplectic form ΩM for the post-selected solution space
SM in region M . Technical obstructions aside, one might subsequently expect to recover (PM , Ω̂M )
from (SM , ΩM ) by a suitable gauge reduction at the level of the subregion M . This is captured in the
following (and tentative) commutative diagram:

/G
(S, Ω) (P, Ω̂)
post−selection post−selection
/G
(SM , ΩM ) (PM , Ω̂M )

However, the process of gauge reduction being in general highly non-trivial, horizontal arrows are not
always readily available in practice. Moreover, even in theories in which the process of moding out
by gauge transformations is reasonably well-understood, it can only be implemented at the expense
of locality.15 This should not be too surprising, since in gauge theories, working with a redundant
description of physical states is precisely the price one pays to preserve the locality of the dynamcis.
Given our focus on dynamical questions in the present work (as well as in follow-ups [101]), it is highly
desirable for our purpose to preserve locality as much as possible. We will therefore implement post-
selection at the level of the presymplectic structure. Of course, our construction will take the gauge
equivalence relation induced by the gauge group G into account, but it will not directly rely on the
existence of (P, Ω̂).

3.2 Relation to canonical phase space methods*

*This section may be skipped by a quick reader. It is only relevant for Sec. 4.2, where it will be estab-
lished that frame-dressed observables formulated with covariant methods (generalizing the construction
proposed in e.g. [3]) can be re-expressed as relational observables in the standard canonical setting (as
defined in e.g. [94]).
As in the mechanical example of Sec. 2, the field space F is not a phase space; the presymplectic
structure features degenerate directions which must be factored out. Indeed, the presymplectic current
density ω = δΘ is a (2, d − 1) form. Hence, in order to turn it into a genuine 2-form on F, we have
to choose a Cauchy slice Σ ∪ Σ̄ in the spacetime M ∪ M̄ and integrate to obtain the presymplectic
structure Ω = Σ∪Σ̄ ω. Outside (but not on) the subspace of solutions S ⊂ F this Ω depends on
R

the choice of Cauchy slice (see Sec. 3.1 and [56]). In interesting field theories, such as Maxwell and
Yang-Mills theory, or general relativity, the degenerate directions of Ω consist of vector fields X on
F with vanishing support on Σ ∪ Σ̄. Such vector fields only generate changes in field configurations
away from the Cauchy slice; thus, they do not encompass gauge transformations Xα with α such
that it and its first derivatives feature non-trivial support on the Cauchy slice [56]. This is because,
in contrast to the mechanical model of Sec. 2, these theories yield only on-shell constraints (see also
Sec. 7). For this reason, factoring out the degenerate directions X of the presymplectic structure, which
define an equivalence relation ∼ for spacetime field configurations, results in a kinematical phase space
Pkin = F/ ∼ with non-degenerate symplectic form Ωkin . This is the standard phase space of the field
theory of interest, parametrized by the pullback of field configurations to Σ ∪ Σ̄, as well as their normal
derivatives on the Cauchy slice. Furthermore, for this reason, the restriction of the equivalence classes
to the subspace of solutions S/ ∼, in contrast to the mechanical toy model, does not coincide with
Pkin , but rather resides as a constraint surface C in Pkin , on which the pullback of Ωkin is degenerate
(for details on these statements, we refer the reader to [56]).
15
For a lucid and pedagogical illustration of the interplay between locality and gauge redundancy, see [9].

22
 Next, we can ask what sort of functionals descend from the field space F to the quotient space
Pkin = F/ ∼. It is clear that these must be functionals which are constant along the degeneracy orbits
in F. These are functionals Ψ on F which can be written in the pullback form Ψ̃ ◦ π = Ψ, where π
denotes the natural projection from F to Pkin and Ψ̃
is a functional on Pkin . Suppose we are given
a local functional Φ, i.e. a functional on F × M ∪ M̄ as the relational observables in equation (70).
Given the above observations, we can turn it into a functional on F that descends to Pkin if we pull
it back and smear it appropriately across the chosen Cauchy slice Σ ∪ Σ̄ so as to remove non-trivial
spacetime dependence away from the Cauchy slice. In particular, we can choose a Dirac delta function
as a smearing function in order to obtain a local functional on Pkin , which now depends on x ∈ Σ ∪ Σ̄.
Similarly, we may ask what sort of vector fields Z project down from F to Pkin as π∗ Z, where
π∗ is the natural push forward from the tangent space Tf F to Tπ(f ) Pkin defined by the projection
π : F → Pkin , where f ∈ F. As explained in [56], these vector fields are characterized by the conditition
that LX Z · Ω = 0 for all degeneracy vector fields X; i.e. the change of Z along the degeneracy orbits is
a degenerate direction itself. In this sense, the transversal part of Z is then constant on the degeneracy
orbit and thereby admits a well-defined projection. In order to yield a non-vanishing vector field on
Pkin , Z has to feature non-vanishing transversal components to the degeneracy orbits. For gauge
theories such as Maxwell or Yang-Mills theory, as discussed later in this work, this is the case for any
gauge symmetry directions Xα with α such that it features non-trivial support on the chosen Cauchy
slice Σ ∪ Σ̄ [56].16 We will henceforth denote the so projected gauge directions on Pkin by X̃α := π∗ Xα
and restrict to field-independent α for now.
This permits us to consider the constraint functionals C̃[α], associated with the Lie-algebra-valued
spacetime functions α, defining the constraint surface C := S/ ∼ inside Pkin . These are (non-uniquely)
defined by
X̃α = Ω−1
kin δ C̃[α] , (57)
where here δ denotes the exterior derivative on Pkin [56].17 The X̃α are degenerate directions of the
pullback of Ωkin to C. Note also that the C̃[α] will be local constraint functionals that are smeared
across the Cauchy slice Σ ∪ Σ̄; they are related to the smeared field space constraints C[α] we shall
later encounter in Sec. 7 by C[α] = C̃[α] ◦ π.

3.3 Towards a regional covariant phase space: main questions to be addressed

We conclude this section by listing the main questions to be addressed in the remainder of the paper.

1. How can we account for gauge-invariant regional degrees of freedom, i.e. degrees of freedom in
M or M̄ , in a minimally non-local way?

2. After post-selecting global solutions with respect to suitable gauge-invariant boundary conditions
on Γ, what type of presymplectic structure descends from the global to the local field space?

3. Given a variational principle for M ∪ M̄ together with a suitable set of boundary conditions on
Γ, is there a systematic way of constructing a variational principle for subregion M ?

In Sec. 4, we will explain how to gauge-invariantly dress boundary degrees of freedom on Γ by suitable
reference frames drawn from M̄ , hence providing an answer to the first problem. This will allow us
in Sec. 5 to focus on a post-selected subset of solutions, defined by gauge-invariant conditions at the
interface Γ. In this subspace, we will be able to canonically induce a presymplectic structure for the
subregion M from the global one, thus providing a resolution to the second problem. Finally, the
third question will be answered in a systematic and algorithmic manner in Sec. 6. In conjunction,
this post-selection procedure can be viewed as deriving the covariant phase space formulation for finite
16
In general relativity, one needs to restrict to the space of solutions in order to also obtain a well-defined projection
of non-spatial diffeomorphisms [56].
17
As emphasized in [56], non-degeneracy of Ωkin does not imply its invertibility, given that Pkin is (assumed to be) an
infinite-dimensional manifold. Invertibility of Ωkin is thus an additional assumption.

23
regions with boundaries, e.g. as established in [11, 66, 68], from the covariant phase space construction
for global spacetimes, e.g. as established in [56]. We illustrate some of the constructions in Maxwell
and Yang-Mills theory along the way, and provide an in-depth exposition of example theories in Sec. 7.

4 Edge modes as “internalized” external frames and gauge-invariant

observables
Gauge symmetries can be understood as a manifestation of the intrinsically relational nature of certain
observables [102, 103]. A valuable concept for dealing with such observables is that of a dynamical
reference frame [41–54]: it is constituted by a collection of dynamical degrees of freedom which can be
used to keep track of the evolution of the rest of the system in a relational and gauge-invariant way. We
will make use of a particular class of such reference frames to correctly account for the dynamics of gauge
fields in a finite subregion M relative to its complement M̄ . These frames are the edge modes which,
as in the mechanical setting of the previous section, we identify as “internalized” external frames for
M . As such, they are not additional degrees of freedom one has to add to the theory; they are already
ingredients of the global theory for M ∪ M̄ one starts with. After discarding any additional dynamical
information from M̄ , the reference frame turns into an independent dynamical field living on the time-
like boundary of M . This observation explains why, in gauge theories, dynamical edge modes can
emerge at finite boundaries. As we will see, another advantage of the reference frame formalism is that
it clarifies under which operational assumptions such edge modes are to be considered as physically
relevant: namely, when we have the operational means to measure gauge-variant observables in M
relative to a reference frame in M̄ , which jointly, give rise to a gauge-invariant relational or dressed
observable with support on both M and M̄ . These relational observables will in later sections become
crucial for formulating the boundary conditions inducing foliations of the global space of solutions for
M ∪ M̄ into distinct subregion theories for M (compare also with the mechanical case in Secs. 2.5
and 2.6).

4.1 Dynamical reference frames in gauge field theory

Let us first specify what we mean by a dynamical reference frame for some local (gauge) symmetry
Lie group G acting on some system S. Loosely speaking, it should be a dynamical subsystem R
internal to the total system S that is as asymmetric under G-transformations as possible. Its purpose
in life is to constitute a dynamical coordinate system for parametrizing G-orbits in the configuration
space of the total system S. As such, R constitutes a dynamical reference system relative to which
we can describe the remaining degrees of freedom of S as they transform under G in such a way that
the relations translate into gauge-invariant observables of the theory. Subsystems that are already
invariant therefore constitute the worst possible choices of frames. In a nutshell, this is the philosophy
underlying much of the recent efforts on quantum reference frames [41–54]. In gauge field theory, we
will take G to be the local structure group rather than the (global) gauge group, the latter coinciding
in this context with the group of gauge transformations (which can be described locally as G-valued
functions). Our construction will therefore be spacetime local by nature. In particular, the system S
and reference frame R will consist of local degrees of freedom at (or in a neighborhhod of) a spacetime
point x.
More precisely, in order for R to parametrize G-orbits, it must (locally in spacetime) feature a set of
at least dim G configuration degrees of freedom that transform freely under G. In other words, R must
be such that G acts freely on its configuration space R (i.e. the isotropy group for each configuration
is trivial). If G acts furthermore regularly on R (i.e. it also acts transitively, connecting any two
configurations), then R is a principal homogeneous space for G — so dim R = dim G — and R features
as many configurations as there are group elements. In the language of [104], such a frame R is called a
complete reference frame. If, by contrast, G acts freely, but not regularly on R, the latter contains more
configurations than there are group elements, in which case we may call R an overcomplete reference

24
frame. It is also possible to consider partial reference frames characterized by the action of G on R not
being free, e.g. see [105], but we shall not explore them here much further. Henceforth, we shall refer
to the configurations of R as its frame orientations. For example, in Sec. 2, we considered particles as
complete reference frames for the translation group, whose positions were their frame orientations.
Let us now consider the general case of a gauge theory with local symmetry governed by a connected
Lie group G (which may be non-Abelian), and for definiteness we assume it to be a compact matrix
group (which could be weakened). The group of gauge transformations G is a space of G-valued
functions on spacetime, x 7→ g(x), with suitable fall-off conditions at asymptotic infinity. The group law
is the pointwise multiplication in G: given two gauge transformations g1 and g2 , g1 g2 (x) = g1 (x)g2 (x).
The connection is a g-valued one-form A, transforming under the group action:

g . A = gAg −1 − dgg −1 . (58)

Indeed, we can check that (g1 g2 ) . A = g1 . g2 . A. The generators of such gauge transformations are
vector fields Xα on field configuration space F, labeled by g-valued spacetime functions α:18

Xα (A) = [α, A] − dα , (59)

where [·, ·] denotes the Lie bracket. In particular, etXα (·) = etα . (·) for any t ∈ R.
Let us suppose that the gauge theory contains other fields, which are dynamically independent
from A, and that we collectively denote by Ψ. We can now consider the three types of reference frames
R alluded to above for the field theory (the system S) parametrized by A, Ψ. The following applies
both to the bulk and boundary of spacetime regions, so for the moment we do not restrict to edge
modes. The simplest possible type of frame R is a G-valued reference frame: its configurations are
described by a functional U [Ψ, A] with values in the space of gauge transformations, which transforms
by left-multiplication under .. Namely, at least locally in field space F, U [Ψ, A] assigns a group element
U [Ψ, A](x) ∈ G to every point x in spacetime and:

g . U [Ψ, A] = gU [Ψ, A] . (60)

The frame orientation space R is therefore parametrized by these G-valued functions on spacetime.
When such a construction is possible, U [Ψ, A] provides a field-space-local and complete reference
frame for the local structure group G at every point in spacetime, since the action of G on itself by left
multiplication is regular. Owing to the Gribov obstruction, this will in general not be possible globally
in field space F; for if it was possible, one could use U [Ψ, A] to define global gauge fixing conditions
for the G-orbits in F.
A group-valued reference frame is characterized by a particular choice of free group action of
the local structure group G on the configuration space R: namely, the action of G on itself by left
multiplication, which enters into the definition (60) (so that, in particular, R = G and the action is also
regular). It is straightforward to generalize this construction to other types of reference frames, with
R= 6 G; the free character of the G-action on R is the only necessary condition we need to preserve
in order to obtain a reference frame whose configurations can be used to parametrize G-orbits in field
space. Indeed, suppose we are given a free group action of the local symmetry group G on R, that is,
a (smooth) map D : G × R → R , (g, R) → g · R such that: for any (g, R) ∈ G × R, g · R = R implies
that g = e (where e is the identity element in G). A local reference frame for the structure group G
can then be obtained by constructing a R-valued functional R[Ψ, A] transforming as:

(g . R[Ψ, A])(x) = g(x) · R[Ψ, A](x) . (61)

If the group action D is regular on R, then the latter defines a complete reference frame for G at
every point x in spacetime; if it is only free (hence not transitive), the resulting local reference frame
18
This is true at least locally, and relative to a choice of trivialization. Gauge transformations are more generally
understood as G-valued sections, but this will not play any significant role in the present paper.

25
is overcomplete. Again, on account of the Gribov obstruction, the frame will generally have these
properties only locally in field space F.19
Finally, we could imagine relying on group actions that fail to be free to construct partial reference
frames. Such structures may be of interest to describe situations in which the observer has incomplete
knowledge of her frame of reference. For simplicity, we will ignore this possibility here, and instead
adopt the view that, at least in principle, it is always possible to construct (field-space-local) complete
reference frames for the gauge degrees of freedom of interest. Concretely, this may require the (explicit
or implicit) introduction of a sufficiently rich set of ancilla systems into our description.20 For instance,
in pure Maxwell theory, we could imagine adding test charges to make operational sense of holonomies
along open paths, which can in turn be used to construct group-valued reference frames at their open
ends. In most of the remainder of this manuscript, we shall work with group-valued frames.

4.2 Frame-dressed observables are relational observables

Suppose now that we have access to a group-valued — and hence complete — reference frame U [Ψ],
which for simplicity does not depend on the connection A. This permits us to decompose any local
functional Φ[A] of the connection canonically into:

Φ = Φinv + Φgauge , (62)

where Φinv is the gauge-invariant frame-dressed field

Φinv := U [Ψ]−1 . Φ , (63)

and
Φgauge := Φ − U [Ψ]−1 . Φ (64)
carries the gauge dependence of Φ. Such a frame associated decomposition is similar to the splitting
into vertical and horizontal components relative to a field-space connection, as advocated by Gomes and
Riello [106,107]. While the latter splitting was done for a spatial slice, here we take a covariant picture,
applicable to spacetime regions. As an illustration, we find for the connection and field strength:
X 1
Ainv = U [Ψ]−1 AU [Ψ] − dU [Ψ]−1 U [Ψ] , Agauge = −U [Ψ]−1 dU [Ψ] − [A, u]n , (65)
n!
n≥1
X 1
Finv = U [Ψ]−1 F U [Ψ] , Fgauge = − [F, u]n , (66)
n!
n≥1

where we have defined the Lie algebra generator u via U [Ψ](x) := eu(x) , used the general formula
AdeX = eadX , and introduced the short-hand [A, u]n := [[[A, . . .], u], u]. Owing to the Gribov obstruc-
tion, this gauge-invariant dressing only needs to hold locally in field space F. When G = U(1), a special
19
A mechanical toy model for dynamical reference frames in the presence of the Gribov problem is the N -body problem
in 3D subject to rotation and translation symmetry, where globally valid gauge-fixing conditions likewise do not exist
on the constraint surface [43]. This leads to an absence of globally valid internal frame perspectives, in some analogy to
the absence of globally valid coordinate systems on general spacetimes.
20
This is best illustrated with an example based on the natural action of G = SO(3) on the unit two-sphere S 2 of R3 .
This group action is not free since a point on the sphere is left invariant by a U(1) subgroup of G. For the same reason,
the action of SO(3) on three copies of S 2 fails to be free. However, by restricting the latter to the orbit of a normal
basis (e1 , e2 , e3 ), we do obtain a free group action of G. In other words, while a single unit vector does not constitute
a complete local reference frame in R3 , a basis does (for a related discussion
 R in quantum theory, see [51]). Likewise,

in gauge field theory, a matter-dressed Wilson line R[A](x) := P exp − γx A ψ(γ(0)), where γ : [0, 1] → M ∪ M̄ is
a path in spacetime and ψ denotes a charged spinor field, transforms as (g . R[A])(x) = g(x) · R[A](x) under gauge
transformations, where · is a group action on spinors. This action is in general not free, so that R will only constitute
a partial frame for G at x. Nonetheless, it is always possible to construct a complete reference frame from a suitable
collection of such matter-dressed Wilson lines, in a conceptually analogous manner to the previous example. Anticipating
Sec. 4.3, this shows that edge reference frames (or edge modes) which are not group-valued are in principle relevant in
gauge field theory, even though, for simplicity, we will focus exclusively on group-valued ones from Sec. 4.4 onwards.

26
case we will restrict to in some examples, Ainv = A − dϕ and Agauge = dϕ, where U [Ψ](x) = eiϕ(x)
(and we have absorbed a factor of i in A so that Aµ (x) ∈ R).
While the choice of dynamical reference frame U is not a choice of gauge, standard gauge-fixing con-
ditions can be recovered from this formalism and translate into specific restrictions on the orientation
of U . For instance, in a U(1) gauge theory where U (x) = eiϕ(x) , we can view the Lorenz gauge-
fixing condition ∇µ Aµ = 0 as an equation for ϕ: ∇µ ∇µ ϕ = −∇µ Aµrad , with the gauge-independent
contribution on the right-hand side playing the role of source.
This construction extends to the case of a frame R[Ψ] not valued in G that, however, transforms
freely under a G-action D (as specified in equation (61)). The only extra ingredient we need in order
to make sense of frame-dressings in this context is an equivariant isomorphism between G and R; that
is, a (smooth) invertible map E : G → R such that:

∀g, g 0 ∈ G , E(gg 0 ) = g · E(g 0 ) . (67)

Since E is one-to-one, we can use its inverse E −1 : R → G to introduce the decomposition:

Φinv (x) := [E −1 (R[Ψ](x))]−1 . Φ(x) , Φgauge (x) = Φ(x) − [E −1 (R[Ψ](x))]−1 . Φ(x) . (68)

Owing to equation (61), together with the fact that E −1 is itself equivariant,21 the frame-dressed field
Φinv is again gauge-invariant. What we have left to do is to justify the existence of an equivariant
and invertible map E. But it suffices to define E(g) := g · R0 , where R0 is some arbitrary (and fixed)
configuration in R. As can be checked explicitly, this is an equivariant map, and given that G acts
freely on R, it is also injective. To make sure it is surjective, all we need to do is to restrict the
configuration space to its image E(G), namely, define R := G · R0 .
In fact, recalling the mechanical example from Secs. 2.2.1 and 2.2.2, we can slightly generalize
the above construction of the gauge-invariant frame-dressing. Restricting to group-valued frames for
simplicity, we can introduce a dependence on the frame orientation g ∈ G into the frame-dressed
observable via right multiplication on the frame (cf. equation (8)):

−1
OΦ,U (g; x) := U [Ψ](x) g −1 (x) . Φ[A](x) (70)

for any local functional Φ[A] of the connection. We have here written the dependence on the spacetime
point x to emphasize that these observables are functionals on F × M ∪ M̄ . This observable is



−1
−1 0−1
clearly gauge-invariant, g 0 . OΦ,U (g) = g 0 U g −1 . (g 0 . Φ) = U g −1 g . (g 0 . Φ) = OΦ,U (g).
It is a relational observable, describing in a gauge-invariant manner how the functional Φ transforms
locally under G relative to the frame U ; it encodes the question “what is the value of Φ when the frame
U is in orientation g ∈ G?”. Indeed, this can be seen as follows: since OΦ,U (g) is gauge-invariant,
we can evaluate it on a gauge-fixing surface in field space without it changing value. In particular,
we can evaluate it on the local gauge-fixing condition U [Ψ] = g, in which case the observable simply
reduces to Φ. Relational observables are therefore so-called gauge-invariant extensions of gauge-fixed
quantities [58, 94]. As a special case, we note for later use that “Φ when the frame is in the origin e”
coincides with the invariant observables above

OΦ,U (e) ≡ Φinv . (71)

In the remainder of this subsection, we shall demonstrate that the frame-dressed observables in
equation (70) indeed constitute the covariant phase space version of the relational observables originally
constructed in [95], using the standard Legendre transformation based approach to phase spaces (and
generalizing earlier proposals [88, 89, 92, 93]). To this end, the reader is invited to remember how the
standard phase space formulation in field theory is obtained from the covariant one, which is discussed
21
That is, for any (g, R) ∈ G × R:
gE −1 (R) = E −1 (g · R) , (69)
which is a direct consequence of equation (67), evaluated for g 0 = E −1 (R).

27
in Sec. 3.2. One elementary aspect to keep in mind is that, in their canonical form, and in contrast
to their covariant incarnation above, the relational observables depend on a point on a Cauchy slice,
rather than a spacetime point. To map from the covariant to the canonical picture, we therefore need
the canonical projection map π : F → Pkin , where Pkin denotes the standard kinematical phase space,
the construction of which is outlined in Sec. 3.2.
With this key ingredient at hand, we are in a good position to discuss the “projection” of the
frame-dressed observables in equation (70) from F to Pkin . Recalling, eXα (·) = eα . (·) and setting
0
U g −1 = eu e−α =: eu , for u0 ∈ g, we can also write them as

OΦ,U (g; x) = eXα Φ[A](x) α(x)=−u0 (x)

. (72)

Provided x ∈ Σ ∪ Σ̄ and the reference frame functional U [Ψ](x), which may depend non-locally on
dynamical fields, only has support in this chosen Cauchy slice, it will “project” to a group valued
local reference frame functional on Pkin via U = Ũ ◦ π. We shall henceforth assume that this is
the case (which may require adjusting the Cauchy slice accordingly).22 In this case, also the field-
dependent Lie-algebra-valued spacetime functions u0 descend to the kinematical phase space as u0 =
ũ0 ◦π since pointwise addition and multiplication of “projectable” functions is preserved by the projection
π. Altogether, if Φ[A](x) is a local field functional and x ∈ Σ ∪ Σ̄, OΦ,U (g; x) constitutes a functional
which only depends on the degeneracy orbits in F and therefore descends to a functional on Pkin , i.e.

OΦ,U (g; x) = ÕΦ̃,Ũ (g; x) ◦ π . (73)

Specifically, the assumption on u0 restricts us to α with non-trivial support in the Cauchy slice. As
noted above, the corresponding Xα contain transversal directions to the degeneracy orbits in F. Since
Φ[A](x) is constant along the degeneracy orbits for x ∈ Σ ∪ Σ̄, we have that only the transversal
directions contribute to the derivative Xα (Φ). This implies Xα (Φ) = X̃α (Φ̃) ◦ π (transversal derivatives
are preserved by the projection π) and thus ultimately

eX̃α Φ̃ ◦ π = eXα Φ α=−u0

. (74)
α=−ũ0

Recalling equation (57), this yields for x ∈ Σ ∪ Σ̄

∞
X 1
ÕΦ̃,Ũ (g; x) = eX̃α Φ̃ = e{C̃α ,Φ̃} = {C̃α , Φ̃}n , (75)
α=−ũ0 α=−ũ0 n!
n=0 α=−ũ0

where {f, g}n = {f, {f, g}n−1 } denotes the nth -nested Poisson bracket with convention {f, g}0 = g.
We emphasize that the Poisson bracket is here the one defined in terms of the kinematical symplectic
form Ωkin on Pkin . This is precisely the power-series representation of relational observables in field
theory originally constructed in [95, Eqs. (4–6)] using the notion of a kinematical phase space.
Equation (70) therefore constitutes the covariant phase space formulation of relational observables
and we conclude that all frame-dressed observables are in fact relational observables. We also note
that it is more convenient to construct relational observables covariantly as done here; this does not
necessitate to choose a Cauchy surface and neither to choose a (possibly non-locally defined) reference
frame geared to this Cauchy slice.
So far we have considered relational observables OΦ,U on the full field space, but ultimately one is
interested in their restriction to the space of solutions S, which in the kinematical phase space formu-
lation means restricting equation (75) to the constraint surface C. Note that, while the construction
of Pkin depends on the choice of Cauchy slice, the construction of C in fact is independent of it since
Ω does not depend on it on solutions. This permits one to also dynamically reconstruct relational
observables on spacetime points x not lying on the chosen Cauchy slice by solving the equations of
motion. By contrast, off-shell, one would have to construct a new kinematical phase space.
22
For example, in the below subregion problems, this will be the case for holonomy constructed frames whose holonomies
do not have support outside Σ ∪ Σ̄.

28
4.3 Edge modes as dynamical reference frame fields
We now apply this general formalism to a subregion M , delimited by initial and final partial Cauchy
surfaces Σ1 and Σ2 , and a time-like boundary Γ. In particular, we shall henceforth only consider
dynamical frames and relational observables on Γ: that is, we should content ourselves with a realization
of formula (61) in which x spans the time-like boundary Γ rather that the whole spacetime. For
simplicity, we assume that the hypersurface Γ has the topology of Sd−2 times an interval. We can
arbitrarily extend Σ1 and Σ2 into complete Cauchy surfaces Σ1 ∪ Σ̄1 and Σ2 ∪ Σ̄2 . We then denote by
M̄ the complement of M in the spacetime slab delimited by Σ1 ∪ Σ̄1 and Σ2 ∪ Σ̄2 . See Fig. 3.

Γ0
Σ2

Σ̄2
y
Γ M W U [Ā](y) sy
x

sx
M̄ U [Ā](x)
Σ1

Σ̄1
Figure 3: The spacetime region M and its complement. From a system of Wilson lines {γx } anchored at
asymptotic source {sx }, we construct the dynamical edge field U [Ā], which provides a reference frame
for the time-like boundary Γ. It allows us to decompose gauge-invariant observables with support on
both M and M̄ , such as the Wilson loop W , into composites of regional gauge-invariant observables.

We collectively denote local field configurations in M (resp. M̄ ) by Φ (resp. Φ̄). In the presence of
gauge symmetries, there might exist gauge-invariant observables with support in M ∪ M̄ , which cannot
be reconstructed from the knowledge of gauge-invariant observables which are solely supported in M
or M̄ . Let us consider two examples, we shall revisit below:23

(a) Consider a Wilson loop W with support in both Mand M̄ , that intersects Γ at two points x
and y, see Fig. 3. Calling H̄xy [Ā] := P exp − γxy Ā (resp. Hxy [A]) the holonomy from x to y
R

along the portion of the loop that lies inside M̄ (resp. M ),24 we have

W := Tr H̄xy [Ā]−1 Hxy [A] (76)

and neither Hxy [A] nor H̄xy [Ā] are gauge-invariant. If G is Abelian this can be decomposed into
gauge-invariant contributions from M and M̄, but not otherwise. Indeed, consider the closed
H 
Wilson line Hxx [A] = Hxy [A] Hyx ∂ [A] = P exp −
γ A , where γ is the closed path consisting of
the previous path from x to y through M and some path from y to x that passes solely through
the interface Γ. Similarly, we can define H̄xx [Ā] = H̄xy [Ā] Hyx
∂ [A], so that the Wilson loop can

be written as
W = Tr H̄xx [Ā]−1 Hxx [A] , (77)

see Fig. 4. When G is Abelian, both Hxx [A] and H̄xx [Ā] are gauge-invariant. In the case that
G = U(1), the trace is trivial, so that the Wilson loop W becomes the product of the purely
23
In the mechanical toy model of sec. 2.5, examples would be intergroup relative distances, such as, say, Qj|j̄ between
particle j of group M and particle j̄ of group M̄ .  R 
24
Note that in the special case G = U(1) our convention to have real angles entails that H̄xy [Ā] = P exp −i γxy Ā .

29
 regional Wilson loops Hxx [A] and H̄xx [Ā] [9]. In the non-Abelian case, neither Hxx [A], nor
H̄xx [Ā] are gauge-invariant. The following example shows that the decomposition of W into
regional gauge-invariant quantities in the Abelian case is a degenerate feature that, regardless of
the group, does not hold anymore as soon as dynamical coupling to matter is included into the
picture.
(b) Consider a Wilson line passing through the point y in the interface Γ

ψ † (x) Hxy [A] H̄yx̄ [Ā] ψ̄(x̄) (78)

and connecting a charged antiparticle ψ † (x) in M to a charged particle ψ̄(x̄) in M̄ .25 Regardless
of the group, neither ψ † (x)Hxy nor H̄yx̄ ψ̄(x̄) are gauge-invariant.

Hxy [A] ∂
Hyx [A]
H̄xy [Ā]

M M̄

Figure 4: Wilson loop W = Tr H̄xy [Ā]−1 Hxy [A] , with support on both M and M̄ .

More abstractly, in the quantum theory, the failure to decompose gauge-invariant observables with
support in both M ∪ M̄ into gauge-invariant observables from M and M̄ is rooted in the appearance
of a constraint-induced non-trivial common center in the algebras of gauge-invariant observables of
M and M̄ , e.g. see [108]; given any Cauchy surface, this center does not commute with arbitrary
gauge-invariant cross-boundary observables.26
In such a situation, two natural options can be considered when projecting the global dynamics
down to the subregion M :
• if one is only interested in gauge-invariant observables which are fully supported in M , it is
sufficient to construct a variational principle for the regional configuration variables Φ, and
completely forget about the configuration variables Φ̄;27
• by contrast, in order to be able to reconstruct all gauge-invariant observables with support in
M ∪M̄ (including those which cannot be reconstructed from regional gauge-invariant observables),
it is necessary to keep track of a sufficient amount of relational information between M and M̄ .
In the present work, we are interested in the second option because it is more general, but we empha-
size that there is nothing wrong with following the first route: it is more restrictive, but completely
appropriate in a physical situation where one has no operational access to the region M̄ .
The main idea is to construct a reference frame R[Φ̄, Φ] on the time-like boundary Γ, with a non-
trivial dependence on degrees of freedom Φ̄ (in a sense we will make more precise below), and possibly
on Φ as well. Namely, we assume we can construct a functional:

R[Φ̄, Φ] : Γ → R , ∀x ∈ Γ, (g . R[Φ̄, Φ])(x) = g(x) · R[Φ̄, Φ](x) , (79)

25
To avoid confusion with our convention to equip fields in M̄ with a¯, we denote the antiparticle field with a †.
26
For instance, the algebra generated by AΣ ∪ AΣ̄ , where AΣ (resp. AΣ̄ ) is the algebra of gauge-invariant observables
supported on the partial Cauchy slice Σ (resp. Σ̄) is obtained by taking the bicommutant (AΣ ∪ AΣ̄ )00 . Since both the
commutant and bicommutant contain the center, the result cannot include arbitrary gauge-invariant cross-boundary
observables; that is, (AΣ ∪ AΣ̄ )00 ( AΣ∪Σ̄ .
27
Sec. 2.2.1 considers the analogous situation in mechanics.

30
that is kinematically independent (i.e. can be varied independently) from the pull-backs φ := Φ Γ , as
we will elucidate in more detail in the following subsection.28 This functional R[Φ̄, Φ] constitutes an
edge mode field on Γ, which we can use as a reference frame to decompose any φ on Γ as:

φ = φrad + φgauge (80)

where

φrad (x) := [E −1 (R[Φ̄, Φ](x))]−1 . φ(x) , φgauge (x) := φ(x) − [E −1 (R[Φ̄, Φ](x))]−1 . φ(x) , (81)

and E is an equivariant map as defined in (67). For a group-valued edge mode frame field, which
for distinction we will continue to denote by the letter U , we also consider more generally the edge
relational observables (cf. equation (70))

−1
Oφ,U (g) := U [Φ̄, Φ] g −1 . φ. (82)

Because this decomposition operates solely on the time-like boundary, we call the gauge-invariant
dressed field φrad the radiative component of φ relative to R[Φ̄, Φ].29 We see that the joint data
(φ, R[Φ̄, Φ]) allows us to define the gauge-invariant observable φrad , which cannot in general be recon-
structed from φ alone (unless gauge symmetries act trivially on φ). It is therefore crucial to keep track
of R[Φ̄, Φ] when restricting to the subregion M . A natural way of doing so is to extend the configura-
tion space for M spanned by Φ by an R-valued field on Γ, that we will simply denote by R. We will
therefore construct a regional variational principle for the pair of bulk and edge fields (Φ, R). Such
an extension (without variational principle) was originally introduced in [3], but we emphasize that
from the point of view of the post-selection procedure discussed in the next sections (see also Sec. 2.5
for the mechanical analog), R can be directly interpreted as a projection of R[Φ̄, Φ] from the global
variational problem for M ∪ M̄ to the regional one for M . This clarifies why R must be dynamical,
since R[Φ̄, Φ] is kinematically independent from Φ in the global field space associated with M ∪ M̄ .
Likewise, this point of view naturally explains why, when gluing back the two regions M and M̄ , R
will serve as a placeholder for R[Φ̄, Φ] [101].
Hence, R[Φ̄, Φ] does not constitute a set of degrees of freedom that has to be added to the theory: it
is part of the global theory to start with, however, only becomes relevant when we restrict our attention
to the subregion M under the premise that we would like to know how it relates to its complement M̄ .
We can therefore view the edge mode field R, in the sense of Sec. 2.2.2, as an “internalized” external
reference frame for M . In particular, the relational observables (82) encode how M relates to its
complement M̄ . As we will illustrate in the next subsection, once we explicitly include the edge frame
into the picture, we can recover the cross-boundary observables in examples (a) and (b) above from
regional relational observables.

4.4 Concrete realization of group-valued edge reference frames

As an illustration of the general framework outlined in the previous subsection, let us discuss how
group-valued edge reference frames can be constructed from a gauge connection A.
A natural idea is to construct a system of paths {γx |x ∈ Γ} such that the path γx ends at x and
originates from the interior of the complementary region M̄ .30 One can then define U [Ā, A](x) as the
holonomy of (Ā, A) along that path. In order for U [Ā, A](x) to transform as in (60), the Wilson line
must originate from a hypersurface where gauge transformations vanish. We can for instance imagine
that M̄ has an asymptotic boundary where gauge transformations are constrained to fall-off sufficiently
28
We emphasize that Φ Γ is our notation for the pull-back of Φ to Γ. To avoid any potential confusion, we will never
invoke the distinct notion of restriction of a field (for instance, the restriction of a normal vector field to Γ).
29
We are here adopting the suggestive nomenclature introduced in the related work [7–9], even though our set-up is
not identical.
30
γx may or may not intersect M , but it must have some support in M̄ .

31
rapidly, so that we only consider bulk-supported gauge transformations in the global field space.31 Such
a boundary provides a non-dynamical anchor for the Wilson lines, which can be understood as a proxy
for background matter degrees of freedom which are not explicitly included in the description (such as
heavy degrees of freedom constituting measurement apparata). We can for example assume that the
path γx originates from a reference point sx lying on an asymptotic boundary, and that the latter is
equipped with a background reference frame U0 (sx ) that no longer transforms under G, as illustrated
in Fig. 3.
To be more precise, we will assume that we have a continuous map γ : [0, 1] × Γ → M ∪ M̄ ,
(t, x) 7→ γx (t), such that γ(·, x) = γx for any x ∈ Γ. In particular, the reference set Γ0 := γ(0, Γ) =
{sx , x ∈ Γ} is homotopic to Γ = γ(1, Γ), which will prove important later on. We also assume sufficient
regularity of γ (such as, for instance, differentiablity) to ensure that U [Ā, A] : Γ → G is well-defined
and differentiable. Lastly, in order for U [Ā, A] to qualify as a maximally rich and interesting external
reference frame for the region M , we will require it to be a non-trivial functional of the fields in M̄ ,
namely:32
• for any x ∈ Γ, U [Ā, A](x) can be varied independently from A in M ;
• for any x, y ∈ Γ, U [Ā, A](x) and U [Ā, A](y) can be varied independenly from each other whenever
x 6= y.
We will assume that those conditions hold at the dynamical level, that is to say in S, in which case
we will say that U [Ā, A](x) is dynamically independent from A M . This may be hard to check in full
generality,33 but this is a reasonable assumption that we can expect to hold true for many systems of
paths. Furthermore, we can easily justify that the conditions do hold in F, a weaker statement we may
call kinematical independence. Indeed, at the kinematical level, the first condition is fullfilled by our
assumption that Γ0 ⊂ M̄ ˚ , where˚denotes the interior of a region: it is then always possible to change
˚ . Similarly, one can enforce
the value of U [Ā, A](x) by a variation of A with support in γx ([0, 1]) ∩ M̄
the second condition by requiring, for any x 6= y, the existence of tx and ty such that: γx (tx ) ∈ M̄ ˚,
˚ , and γ (t ) 6= γ (t ). In words, any two distinct paths are assumed to be non-overlapping
γy (ty ) ∈ M̄ x x y y
somewhere inside M̄ . For conceptual clarity, and without significant loss of generality, we will from
now on assume that the paths γx are entirely supported in M̄ , as illustrated in Fig. 3, so that the
reference frame is a functional of Ā alone and reads explicitly
 Z 
U [Ā] := P exp − Ā . (83)
γx

Once we have selected a dynamical reference frame for the boundary Γ, it is possible to decom-
pose any gauge-invariant observable as a composite of regional gauge-invariant observables, relative
to that particular frame. For instance, we can equivalently write the cross-boundary Wilson loop in
equation (76) as:
(84)


W = Tr H̄s [Ā]Hxy [A]rad ,
where
Hxy [A]rad := U [Ā]−1 . Hxy [A] = U [Ā](y)−1 Hxy [A]U [Ā](x) , (85)
−1 −1
H̄s [Ā] := U [Ā](x) Hxy [Ā] U [Ā](y) , (86)
31
In doing so, we are ignoring potentially interesting effects associated to asymptotic symmetries and soft modes.
Understanding the interplay between our general construction and asymptotic limits is an important objective that we
leave for future work.
32
Even though these conditions make the formalism physically more transparent, they are not strictly necessary for
our general construction to go through. One could for instance consider a situation in which, because of a non-generic
choice of system of paths, U [Ā, A] happens to be determined uniquely across Γ once we know its value at one point. In
such a situation, U [Ā, A] would effectively generate a reference frame valued in a single copy of the structure group G,
rather than in a space of group-valued functions. But this would not preclude its use as a dressing field.
33
For example, this assumption holds in the mechanical toy model of Sec. 2.

32
and both of these quantities are gauge-invariant. Similarly, the dressed Wilson line in equation (78)
can be split into
ψ † (x) Hxy [A] Hyx̄ [Ā] ψ̄(x̄) = (ψ † (x) Hxy [A])rad ψ̄s [Ā](x̄) , (87)
where now

(ψ † (x) Hxy [A])rad := ψ † (x) Hxy [A] U [Ā](y) , ψ̄s [Ā](x̄) := U [Ā](y)−1 Hyx̄ [Ā] ψ̄(x̄) . (88)

4.5 Symmetries as edge frame reorientations

Now that we have set up an edge frame field, we are in the position to consider two types of frame
transformations as in the mechanical toy setting in Secs. 2.2.2 and 2.4: frame reorientations and
frame changes. Both types of transformations act on the right of the frame U [Φ̄], in contrast to
gauge transformations (79) which act on the left. The essential difference between the two is that
frame reorientations are field-independent, while frame changes depend on the gauge-invariant relation
between the old and new frames. We will begin with the former and discuss the latter in the next
subsection.
These observations are particularly clear in the example of the group-valued reference frame U [Ā]
we have just introduced. A reorientation of the edge mode U [Ā] can be interpreted as a change
of background reference frame, parametrized by a group-valued function g0 (sx ) on the asymptotic
boundary Γ0 where the Wilson lines are anchored.34 The effect of this change of asymptotic background
frame on the holonomy along γx can be written as
 Z   Z 
P exp − Ā → P exp − Ā g0−1 (sx ) (89)
γx γx

and constitutes an asymptotic change of trivialization. Since γx (0) = sx , this induces a group-valued
map g : x 7→ g0 (sx ) on Γ, which acts on U [Ā] in equation (83) from the right as:

U [Ā](x) → U [Ā](x)g(x)−1 , (90)

where we emphasized that g, corresponding to a change of background frame, is independent from

dynamical fields. In particular, it is generated by a vector field Yρ on field space F where ρ is some
non-dynamical Lie-algebra-valued field on Γ, corresponding to the group element g.
Transformations of this kind that act “on the other side” of the edge mode from gauge transfor-
mations have been identified in [3] as symmetries.35 Our construction elucidates what they mean
physically: they are edge frame reorientations. In [3, 11], these symmetries have been associated with
charges Q[ρ] obtained via δQ[ρ] = Yρ · ΩM , where ΩM is the regional presymplectic structure for M . In
the mechanical toy model, we have shown the analog in Sec. 2.2.2 and demonstrated in Sec. 2.6 that,
in the presence of boundary conditions, one has to distinguish symmetries into three further types. In
order to expand on this using our construction, we will need more structure. In particular, we will
explain in Sec. 5 how to obtain ΩM from the global presymplectic structure on M ∪ M̄ via a post-
selection procedure involving the edge modes. This will then permit us to explore the charges in Secs. 5
and 7 in detail in Maxwell, Chern-Simons and Yang-Mills theories, including the further distinction of
symmetries into genuine symmetries, meta-symmetries and boundary gauge transformations.
For now, we observe that frame reorientations change the physical situation, as they change the
gauge-invariant relation between the system of interest (here the fields φ on Γ) and the frame. This
34
By contrast, in the case when we use an open ended Wilson line dressed byR a charge in the bulk or on the asymptotic
boundary of M̄ and running solely through M̄ to Γ, i.e. Ũ [Ā](x) = P exp(− γx Ā)ψ̄(sx ), the reorientation of the edge
frame corresponds to a physical transformation of the charge ψ̄(sx ) 7→ g0 (sx )−1 ψ̄(sx ).
35
In the convention of [3], gauge transformations act on the right of the edge mode, while symmetries act on the left, in
contrast to here. We would obtain the convention of [3] if we chose the opposite orientation for the holonomies defining
our edge mode frame.

33
is clear from the associated transformations of the relational observables in (82) (see also (27) for the
mechanical analog):
Oφ,U g−1 (g 0 ) = Oφ,U (g 0 g) . (91)
This constitutes a change of frame orientation label g(x), x ∈ Γ, in the family of gauge-invariant
observables Oφ,U (g). Hence, this amounts to a change of observable within this family and therefore a
change of relation since the family encodes the value of φ when the frame U is in orientation g.
In analogy to the left group action . corresponding to field-independent gauge transformations
eXα (·) = eα . (·), we can define the right group action corresponding to symmetries on functionals f
depending on boundary fields φ, U , as well as their derivatives by (cf. equation (29) for the mechanical
analog)
g(ρ) f [φ, U ] := eYρ f = f [φ, U g(ρ)] . (92)
In particular, frame reorientations take the form g −1 U = U g −1 , leaving other degrees of freedom
invariant, and we can rewrite the relational observable transformation in equation (91) as (cf. equa-
tion (30) for the mechanical analog)

g −1 Oφ,U (g 0 ) = Oφ,g−1 U (g
0
) = Oφ,U (g 0 g) . (93)

4.6 Edge frame changes and transformations of relational observables

We have provided a prescription for how to construct an edge mode frame field on Γ using a system
of paths {γx |x ∈ Γ}. However, this construction is by no means unique, as illustrated in Fig. 5. The
global theory in M ∪ M̄ supports a whole plethora of different edge mode fields on the interface Γ, each
corresponding to a different system of paths, e.g. from the asymptotic boundary Γ0 to Γ. Each ensuing
such edge mode field constitutes a different external frame field for the subregion M of interest, which
we could then “internalize” by extending the field space associated with M correspondingly, as sketched
at the end of Sec. 4.3 (and expanded on in Sec. 5 below).36 For each such edge mode field, we can
construct relational observables on Γ according to equation (82) and the relational observables relative
to different edge frame choices will encompass the different ways in which M relates to its complement
M̄ .

Γ0
M Γ M̄ s(2)
y
(2)
sx
γy(2)

y γx(2)

x
γy(1) s(1)
y

γx(1)
s(1)
x

Figure 5: Two systems of paths γ (1) and γ (2) , supporting distinct families of Wilson lines. They
generate two distinct group-valued reference frames U1 [Ā] and U2 [Ā] for the boundary Γ.
36
As mentioned in Sec. 2.2.2, this is related to an extension of the Heisenberg cut in the quantum theory. In fact, the
analogy is stronger in the field theory case here because of the local nature of the gauge constraints: extending the field
space with additional edge mode fields does not change the form of the gauge constraints. As such, the gauge-invariant
physics on Γ intrinsic to M (i.e. not involving the edge modes) is contained in the same form in any of such field space
extensions. By contrast, in the mechanical setting of Sec. 2, the total momentum “constraint” as well as the action
contribution involving the particles of group M change with every addition of a new particle to M̄ (see also Sec. 2.5).

34
 Since there is therefore no distinguished choice of edge frame, we better explain how to relate
different frame choices, especially how to translate the relational observables relative to one frame
into relational observables relative to another. The following is essentially a translation of the refer-
ence frame changes discussed in the mechanical toy model in Sec. 2.4 into the field theory context.
Consider two edge frame fields U1 , U2 and the corresponding relational observables (cf. equation (82))

−1
Oφ,Uk (gk ) = Uk [Ā] gk−1 . φ, k = 1, 2. The relation between the two sets of observables depends on
the relation between the two frames “when the second frame U2 is in the origin e ∈ G”, encoded in the
gauge-invariant group-valued function

x 7→ g21 [Ā](x) := OU1 ,U2 (e, x) = U2 [Ā](x)−1 . U1 [Ā](x) (94)

−1
= U2 [Ā] (x)U1 [Ā](x) , (95)

which is now dynamical. We can then write the transformation from frame U1 to frame U2 as the
relation-conditional right action on the frame U1 (cf. the mechanical counterpart in equation (33)):

U1 [Ā] 7−→ U2 [Ā] = g21 [Ā]−1 U1 [Ā] := U1 [Ā] g21 [Ā]−1 (96)

−1
Oφ,U1 (g1 ) 7−→ Oφ,U2 (g2 ) = g1−1 g2 g21 Oφ,U1 (g1 ) := Oφ,(g−1 g2 g21 )−1 U1 (g1 ) . (97)
1

Note that this amounts to an extension of the right action defined in equation (92) to field-dependent
frame reorientations, in analogy to how equation (72) extends field-independent gauge-transformations
to field-dependent ones. Similarly to how equation (72) generates gauge-invariant observables, equa-
tion (97) produces observables invariant under reorientations of frame R1 . Indeed, we have:
 −1

−1 
Oφ,U2 (g2 ) = U1 [Ā] U1 [Ā]−1 U2 [Ā]g2−1 g1 g1−1 . φ = U1 [Ā]−1 U2 [Ā]g2−1 g1 U1 [Ā]g1−1


.φ .

We note that this is equivalent to (cf. the field-independent frame reorientation in equation (93))

Oφ,U2 (g2 ) = Oφ,(g−1 g2 g21 )−1 U1 (g1 ) ≡ Oφ,U1 g1 (g1−1 g2 g21 [Ā]) = Oφ,U1 g2 g21 [Ā] (98)



1

and hence a relation-conditional right action on the frame orientation label of the relational observable.
This encompasses the tautological observables describing the edge frames relative to themselves

OU1 ,U1 (g1 ) = g1 7−→ OU1 ,U2 (g2 ) = g2 g21 [Ā] , (99)
−1
OU2 ,U1 (g1 ) = g1 g21 [Ā] 7−→ OU2 ,U2 (g2 ) = g2 . (100)

In view of the previous subsection, we can thus interpret frame changes as relation-conditional
frame reorientations. However, in contrast to the field-independent frame reorientations of the previous
subsection, frame changes are not symmetries (and not generated by charges). Symmetries change the
physical situation, while frame changes only change the description of it. In particular, equation (97)
constitutes a field redefinition for the gauge-invariant data on Γ. As such, a change of edge frame
simply amounts to a change of coordinates on the field space FM (which includes the edge modes).
Recall also that field-independent frame reorientations map observables from one family Oφ,U (g) of
relational observables into others within the same family (only changing the orientation label g of the
frame in a field-independent manner). By contrast, the frame changes here map the observables from
one family Oφ,U1 (g1 ) of relational observables (associated with frame U1 ) to a distinct family Oφ,U2 (g2 )
(associated with frame U2 ).
In order to say that this coordinate change is equivalent to a symplectomorphism, as in the me-
chanical case of Sec. 2.4, we first need control of the presymplectic structure. As argued in Sec. 2.7 for
the mechanical setting, different edge frame choices induce distinct families of foliations of the solution
space S which are crucial for the post-selection procedure; a frame change relates a foliation within
the family associated with the first frame to a foliation in the family associated with the second frame.
Accordingly, frame changes relate different implementations of the post-selection procedure. This will
also be true in the field theory case, as we will argue in Sec. 5.

35
4.7 Frame-reorientation-invariant field-space form
We finally introduce the field-space right-invariant Maurer-Cartan form δU U −1 which, being invariant
under frame reorientations, will play an important role in our construction. For notational convenience,
we will from now on keep the dependence of the frame U on Ā implicit. Its variation is given by
1
δ(δU U −1 ) = (δU U −1 )(δU U −1 ) = δU U −1 , δU U −1 , (101)

2
where [·, ·] is to be understood as the Lie-bracket on field-space forms.37 As in finite dimensions, one can
understand this equation as a flatness condition for a field-space gauge potential −δU U −1 = U δU −1 .38
The Maurer-Cartan form also transforms nicely under field-dependent gauge transformations, namely:

g . δU U −1 = gδU U −1 g −1 + δgg −1 , (102)

where we have kept the field-dependence of g implicit. At the infinitesimal level, this yields:

LXα δU U −1 = Xα · δ(δU U −1 ) + δXα · δU U −1 = [α, δU U −1 ] + δα , (103)

Finally, the differential of the field-space Maurer-Cartan form δU U −1 can be related to the variation
of the spacetime Maurer-Cartan form dU U −1 , as follows:

d(δU U −1 ) = δ(dU U −1 ) + [dU U −1 , δU U −1 ] . (104)

5 Symplectic geometry of splitting post-selection

We shall now explain how to extend the splitting post-selection procedure, illustrated in the mechanical
example in Sec. 2, to field theory. The underlying idea of our construction is to take the total space
of solutions S to the global equations of motion in M ∪ M̄ and post-select the subset of solutions on
it, which satisfy certain boundary conditions X = X0 on Γ. X will be some differential form, locally
constructed out of the pullbacks φa := Φa |Γ of the fields to Γ and their derivatives, as well as of U and
its derivatives, and X0 is a non-dynamical background field. We require the boundary conditions —
and thus X — to be gauge-invariant, so as to cleanly distinguish between physical boundary conditions
and gauge fixing conditions.
First implementing splitting post-selection on-shell, rather than at the level of the action, has
several advantages. From the point of view of M , this allows us to cleanly disentangle bulk dynamical
equations from boundary conditions (which, as we will explain in Sec. 6, will eventually be enforced
dynamically). More generally, it is useful to investigate the relation between the covariant phase spaces
of M and M̄ independently from a choice of variational principle. By focusing on the (pre-)symplectic
geometry underlying post-selection, we will be readily able to connect our work to previous literature
on edge modes.

5.1 Radiative presymplectic structure from a dressing flow in field space

Let us first proceed to show that the group-valued dressing of edge fields introduced in Sec. 4.4 induces
a natural and unambiguous splitting of the presymplectic structure on Γ into radiative and gauge con-
tributions. This is important, as only the radiative component will be subject to boundary conditions,
while, on-shell, the gauge component will essentially constitute the corner term.
Dressing being implemented by a field-dependent gauge transformation φ 7→ U −1 . φ, this operation
does not commute with the field-space exterior derivative δ. As a result, it is not completely obvious
37
That is, [ω1 , ω2 ] := ω1 ω2 − (−1)pq ω2 ω1 if ω1 is a p-form and ω2 a q-form (as before, the field-space wedge product is
kept implicit).
38
Following e.g. [6, 106, 107], one could thus define a flat covariant differential D := δ − δU U −1 for fields on Γ. We will
not rely on such a formalism in the present paper, but it would be interesting to explore it further.

36
how to extend a prescription like (81) to field forms such as θ := Θ Γ and ω Γ .39 To circumvent this
difficulty, we will implement dressing by means of a particular flow in field space. The advantage of this
point of view is that it will readily generalize to field forms by means of the concept of Lie derivative.
Given our working hypotheses for the system of paths {γx , x ∈ Γ}, we can find a homotopy between
U (x) and the identity frame field 1l : Γ → G, x 7→ e (where e is the identity element in G). Indeed, for
any t ∈ [0, 1] and x ∈ Γ, we can define Ũt (x) as the holonomy between the points sx and γx (t), along
the path γx . We then have Ũ1 = U , Ũ0 = 1l, and given that (t, x) 7→ Ũt (x) is continuous, we conclude
that U is homotopic to the identity frame field 1l. As a result, U is in the image of the exponential on
the Lie algebra of the gauge group. In particular, there exists a continuous g-valued function u on Γ,
such that: for any x ∈ Γ, U (x) = eu(x) .40 As anticipated, it is therefore legitimate to view the dressing
φ 7→ φrad = U −1 . φ as a field-dependent gauge transformation (in this work, we do not consider large
gauge transformations as part of the gauge group).
In order to extend this dressing to field-space forms, it is convenient to realize it as a flow. To this
effect, and given a fixed field-configuration A, we consider a path {At := gt .A|t ∈ [0, T ]} in field-space,
where {gt |t ∈ [0, T ]} is a one-parameter family of (global) field-dependent gauge transformations. All
we demand from gt is that g0 = Id, and gT (x) = U −1 [A](x) for any x ∈ Γ.41 Given those definitions,
φrad (where we again collectively denote pull-backs of fields on Γ by φ) can be recovered as the edge
configuration induced by AT :
φrad = φT . (105)
Likewise, the frame field flows from its initial value U to the identity frame field U −1 . U = 1l.42 In
other words, the two defining properties of the path At are that: 1) it is confined to the gauge orbit of
A0 = A, and 2) U [ĀT ] = 1l. In particular, we can find a vector field Xα such that:

d
At = Xα (At ) = α . At , (106)
dt
where the infinitesimal gauge parameter α is a priori field-dependent.
The advantage of this somewhat contrived definition of the dressing is that it immediately general-
izes to field forms, upon replacement of the vector field Xα by the Lie derivative LXα . In more detail,
given a form p on the edge field-space, we introduce pt as the solution to the flow
d
pt = LXα pt = Xα · δpt + δXα · pt , (107)
dt
with initial condition p0 = p. We can then define the radiative component of the form p by

prad := pT . (108)

Owing to [LXα , δ] = 0, which directly follows from Cartan’s magic formula, an immediate consequence
of this definition is that:
(δp)rad = δ(prad ) . (109)
In particular, θrad is a potential for ωrad :

ωrad = δθrad . (110)

Furthermore, using the Leibniz property of LXα , we infer that

(pq)rad = prad qrad , (111)

39
In the following, we will simply denote ω Γ by ω, keeping the pull-back to Γ implicit.
40
u is not unique, but at least locally in field-space, one can choose a smooth map U 7→ u from group- to algebra-valued
fields, defining local logarithmic coordinates.
41
Such a path necessarily exists since U , seen as a field on Γ, is connected to the identity. At is actually highly
non-unique, but we will see that the resulting dressing of edge modes is independent from such ambiguity.
42
It is crucial at this level that U transforms nicely under field-dependent gauge transformations.

37
for any two forms p and q. Crucially, this observation is sufficient to show that our dressing prescription
is independent from the choice of vector field representative Xα , and in particular, of the choice of path
At . Indeed, φrad = U −1 . φ is itself independent from Xα , and therefore the statement holds for
0-forms. Invoking (109) and (111), we can then extend this unicity to forms of arbitrary degree.43
Considering that φrad is invariant under arbitrary field-dependent gauge transformations, a similar
reasoning invoking (109) and (111), allows us to extend this invariance to any dressed form prad . In
particular, we have
LXα θrad = 0 and LXα ωrad = 0 , (112)
for any gauge parameter α (including field-dependent ones).
Finally, we define the gauge contributions of the presymplectic potential and the presymplectic
current as:
θgauge = θ − θrad and ωgauge = ω − ωrad . (113)

For completeness, let us illustrate our construction with a concrete choice of vector field Xα . Since U
is in the image of the exponential, it is natural to choose suitable logarithmic coordinates U (x) = eu(x) ,
for instance by demanding u to be minimal relative to some well-suited norm.44 We can then require
the field-dependent gauge parameter α to verify: α[A](x) = −u[A](x), for any x ∈ Γ. Given such a
generator, it follows that:
Xα φ = −u . φ , (114)
for any local field on Γ. We can then observe that the following flow, which implements a homotopy
between the edge mode and identity frame field, is well-defined and unique:45
d
(φt , Ut ) = Xα (φt , Ut ) , φ0 = φ , U0 = U . (115)
dt
Its solution is:
(e(−1+τ (t))u . φ, eτ (t)u ) , τ (t) = e−t , t ∈ [0, +∞[ , (116)
which one can check by direct computation. Using that τ 0 (t) = −τ (t) in conjuction with the transfor-
mation rule α . U = uU for the reference frame, we indeed find that
d (−1+τ (t))u
(e . φ, eτ (t)u ) = ((τ 0 (t)ue(−1+τ (t))u ) . φ, τ 0 (t)ueτ (t)u ) = (−τ (t)u) . (e(−1+τ (t))u . φ, eτ (t)u )
dt
(117)
= Xα (e(−1+τ (t))u . φ, eτ (t)u ) . (118)

Consequently, φrad = U −1 . φ can be recovered as the limit:

φrad = lim φt , (119)

t→+∞

and, as result, any field-space form can be dressed by following the induced Lie flow.

5.2 Foliation with respect to field configurations on the time-like boundary

Splitting post-selection proceeds in three steps (see Sec. 2.5 for this procedure in mechanics): one

1. foliates the global space of solutions S with respect to suitable field configurations on Γ; then
43
Alternatively, we could have directly defined the dressing of forms by equations (81), (109) and (111).
44
One could for instance minimize the L2 norm. That is, given a group-valued U , we look for an algebra-valued u such

1/2
that U = eu and kuk := Γ (u(x) · u(x))dx
R
is minimal (where · denotes here the Killing form on g). The solution
will be unique, except on a set of field configurations of measure zero. For those special configurations, we choose one
solution in an arbitrary way.
45
The norm kuk is strictly decreasing along the flow, so that one never encounters field configurations where u is not
uniquely defined, except possibly at t = 0.

38
 2. restricts to a particular leave in this foliation; and finally

3. discards the complementary region M̄ to obtain the looked for presymplectic form ΩM .
We call this process splitting post-selection as the second step “post-selects” those solutions in S, which
satisfy a specific choice of boundary conditions on the interface Γ, thereby effectively splitting the
global space of solutions S for M ∪ M̄ into two, one for each spacetime subregion. Upon the third step,
one then focuses on the dynamical problem for the subregion M of interest, defined by the selected
boundary conditions. Since the latter define the dynamical theory for M (and, likewise, for M̄ ), we
can think of each leaf in the foliation of S as a particular subregion theory. As such, the global solution
space S assumes the role of a space of subregion theories, in this sense constituting a meta-theory for
the local subregions. This will lead us to distinguish symmetries (frame reorientations) on Γ into (i)
genuine symmetries that leave the subregion theory invariant, (ii) meta-symmetries, which change leaf,
i.e. subregion theory within S, and (iii) symmetries that become gauge transformations on particular
boundary conditions. We stress that case (ii) only constitutes meta-symmetries for the subregion
theories, not, however, for the global dynamics in M ∪ M̄ , since it maps one solution into another one
of the same global theory.46
In more detail, we look for a decomposition of S of the form
G
S= SX0 , SX0 := {Φ ∈ S|X [φrad ] = X0 } (120)
X0

for some suitable choice of X and X0 , where X0 is a (set of) background field(s) that spans a function
space on Γ appropriate for a well-posed boundary value problem, and X is a map from field configu-
rations on Γ to field configurations on Γ (including the reference frame U , which enters the definition
of φrad ). The post-selected leaf SX0 can be thought of as (the space of solutions to) a specific sub-
region theory. The boundary condition X [φrad ] = X0 must fix sufficiently many degrees of freedom
to give rise to a well-posed dynamical problem in M , and not too many so that partial Cauchy data
in the interior of Σ remains freely specifiable. In particular, we will require that there is no physical
symplectic flux through Γ after pulling-back Ω to SX0 . A natural and sufficient requirement seems at
first to be that ω Γ,S = 0. However, in the presence of gauge symmetries such a condition may be
X0

overly restrictive, in the sense that it is not always gauge-invariant. In such a situation, ω Γ,S = 0
X0
would amount to imposing a vanishing flux, as well as a non-trivial gauge-fixing condition. To make
sure that we only constrain physical degrees of freedom, we instead impose the weaker condition:

ωrad SX0
= 0. (121)

In the following, we will denote equalities that only hold on the leaf SX0 , that is on-shell of both the
bulk equations of motion and the boundary conditions, by the symbol ≈. ˆ In particular, ωrad ≈ ˆ 0. This
ensures that the presymplectic current ω is gauge-equivalent to zero when restricted to Γ and a specific
leaf SX0 , but not necessarily identically zero.
To make sure that we do not constrain ωrad more than necessary, we will impose boundary conditions
on a Lagrangian submanifold of S. In practice, we will choose local and gauge-invariant Darboux
coordinates X and Y (which are fields on Γ) such that ωrad = δXδY . To implement (121), it is then
sufficient to impose e.g. δX ≈ˆ 0, while leaving Y dynamical.
Let us look at a simple but illuminating example in Maxwell theory (which we will discuss again
in more depth in Sec. 7.2). The presymplectic current can be expressed in the canonical Darboux
coordinates as ω = δA ∧ δ ? F . Given a U(1) reference frame with phase ϕ(x) one can define the
radiative data (see equations (65) and (66)):

Arad = A − dϕ , (?F )rad = ?F , ωrad = δArad ∧ δ(?F )rad ≈ δA ∧ δ ? F − d(δϕδ ? F ) . (122)

46
This global theory S too is to be defined by some asymptotic boundary conditions that are unaffected by the
meta-symmetries (ii).

39
These coordinates induce two natural foliations. We can foliate the solution space S with respect to the
pull-back of ?F , which amounts to defining X[Arad , (?F )rad ] := (?F )rad . Since ?F is gauge-invariant to
start with, it follows that ω Γ ≈ ˆ 0. Relative to this particular choice of foliation, the dressing by U can
be safely ignored. By contrast, it becomes relevant again if one decides to impose a boundary condition
on Arad , that is, if we foliate the solution space with respect to X[Arad , (?F )rad ] := Arad . In that case,
ωrad P ≈ ˆ 0 while ω Γ ≈ ˆ d(δϕδ ? F ) does not vanish in general. Note that the boundary condition
f
on Arad fixes the pull-back of F to Γ completely, but is stronger in that it also contains relational
information about A Γ and U . Finally, one can impose boundary conditions on linear combinations of
Arad and (?F )rad , which we will investigate thoroughly in Sec. 7.
Let us now briefly discuss the two types of transformations one can perform on the canonical
splitting of A relative to the reference frame U = eiϕ :47

A Γ
= Arad + dϕ . (123)

The gauge contribution Agauge = dϕ is unconstrained by purely local gauge-invariant observables, such
as F Γ , but is fixed unambiguously by the reference frame data. Gauge transformations act on both
A and ϕ in a way that leaves Arad invariant:

δα A = dα , δα ϕ = α ⇒ δα Arad = 0 . (124)

By contrast, frame reorientations take the form of a symmetry transformation ∆ρ which acts solely on
ϕ, and thereby changes the radiative part of the connection:

∆ρ A = 0 , ∆ρ ϕ = −ρ ⇒ ∆ρ Arad = dρ . (125)

Depending on the choice of boundary condition, the solutions might not be stable under this transfor-
mation, in which case ∆ρ generates a meta-symmetry, i.e. a change of post-selected theory SX0 within
the global solution space S. For instance, when foliating the solution space with respect to Arad , ∆ρ
acts transversally to the leaves and thereby changes the boundary condition (that is, the value of X0 in
(120)), unless ρ is constant on Γ (in which case ∆ρ does generate a symmetry). As we will see below,
this meta-symmetry is a symplectomorphism within S. The definition (125) is an Abelian incarnation
of the symmetry transformation originally introduced by [3] in Yang-Mills theory. Our construction
clarifies its relation to reference frames, and its interplay with boundary conditions.
Our construction has so far been implicit about the choice of edge frame U [Ā] that goes into the
construction of the radiative data φrad (cf. equation (81)) defining the boundary conditions X0 . For
a fixed frame choice, we can choose different maps X, each corresponding to a distinct foliation of S.
We can consider them as part of one family of foliations, associated with the frame U . They will be
related by the transformations relating the different choices of X (e.g., in the mechanical setting of
Sec. 2.7, these were canonical transformations). As argued in Sec. 4.6, there is a plethora of distinct
choices of edge frame fields. It is clear that each of them will lead to a distinct family of foliations (120)
of the global solution space S. Any choice of boundary conditions X0 fixes a set of edge relational
observables associated with some frame. Since the edge frame transformations of Sec. 4.6 map the fixed
set of relational observables into relational observables associated with a different frame, they also relate
the different frame associated families of foliations of S.48 The frame changes, as specific coordinate
changes on S, can thereby be invoked to link physically distinct implementations of post-selection.

5.3 Presymplectic structure of the subregion field space

Let us now introduce a presymplectic structure ΩM on the space of dynamical field configurations in
M , which includes both bulk fields and the boundary reference frame and which is consistent with
47
We focus exclusively on the connection because its conjugate ?F is gauge-invariant in the present Abelian context.
48
Since the frame changes are field-dependent, the new set of relational observables will generally not be fixed via the
boundary conditions X = X0 .

40
the global presymplectic structure Ω for fields in M ∪ M̄ . In the pure gauge-theory examples we are
interested in, this is defined as:49
X0
(126)

SM := (A M , U [Ā])|(A, Ā) ∈ SX0 .

We first assume that ΩM is locally identical to Ω in the bulk of M (resp. M̄ ), and that the regional
presymplectic forms are additive, in line with the idea that the global degrees of freedom are constituted
by the sum of the local degrees of freedom:

ΩM + ΩM̄ ≈
ˆ Ω. (127)

We are then naturally led to the following Ansatz:

Z Z Z Z
ΩM = ω+ ω∂ , and ΩM̄ = ω− ω∂ . (128)
Σ ∂Σ Σ̄ ∂Σ

In this equation, ω∂ is a (2, d − 2)-form on Γ that may depend on both bulk and boundary fields and
constitutes a corner term for ΩM . To ensure that δΩM = 0, we will furthermore assume that ω∂ = δθ∂
for some (1, d − 2)-form θ∂ , so that for every Cauchy surface Σ,
Z Z
ΘΣ := Θ+ θ∂ (129)
Σ ∂Σ

defines a potential (which may depend on Σ) for ΩM . We call ω∂ (resp. θ∂ ) the boundary presymplectic
current (resp. boundary presymplectic potential ). We can take advantage of this splitting ambiguity to
enforce two additional desirable properties for ΩM , namely: its independence from the choice of partial
Cauchy surface, and its invariance under field-dependent gauge transformations. As we shall see, the
latter property follows from the former.

Σ2

Γ12 M12

Σ1

Figure 6: We now allow Σ1 and Σ2 to be arbitrary surfaces anchored on Γ; we call Γ12 the portion of
Γ delimited by ∂Σ1 and ∂Σ2 , while M12 designates the portion of M delimited by Σ1 and Σ2 .

Asking for (128) to be conserved across distinct Cauchy surfaces is quite constraining. Let us
consider an initial and final surface Σ1 and Σ2 , which we now permit to vary; see Fig. 6. We call Γ12
the portion of Γ whose boundary is ∂Σ1 ∪ ∂Σ2 , and M12 the subregion of M delimited by Σ1 and Σ2 .
From dω ≈ 0, it follows by Stokes’ theorem that:
Z Z Z Z
ω− ω+ ω= dω ≈ 0 . (130)
Σ2 Σ1 Γ12 M12
49
Under the assumptions we made in Sec. 4.4 — namely, that U [Ā] can be varied independently from A M at the
X0
dynamical level —, we can conclude that SM factorizes as a solution space for A M times a space of G-valued functions
X0
on Γ. As a result, SM can be understood as a boundary extended functional space for region M , in the spirit of [3].
X0
If dynamical independence were not to hold, SM could still be embedded in the same extended functional space, hence
our construction could straightforwardly be generalized to include this case. To avoid introducing further notation that
would only cloud the main conclusions of the paper, we refrain from demonstrating this explicitly.

41
Inserting this relation into the consistency condition
Z Z Z Z
ω+ ω∂ ≈ˆ ω+ ω∂ , (131)
Σ1 ∂Σ1 Σ2 ∂Σ2

we find Z Z Z
ω∂ − ω∂ ≈
ˆ ω. (132)
∂Σ2 ∂Σ1 Γ12

For this relation to hold for any choice of Σ1 and Σ2 , we must have:

ω Γ
≈
ˆ −dω∂ . (133)

Irrespective of the choice of gauge-invariant boundary condition, equations (113) and (121) entail that
ωΓ≈ ˆ ωgauge on SX0 , and therefore:
ωgauge ≈
ˆ −dω∂ . (134)
This fixes ω∂ up to a closed form, provided that ωgauge is itself exact (at least when pulled-back to the
leaf SX0 ). We will find out that this consistency requirement is fulfilled in Maxwell, Chern-Simons and
Yang-Mills theory. Finally, equation (134) implies that

θgauge ≈
ˆ −dθ∂ − δ`∂ . (135)

for some (0, d − 1)-form `∂ .

In Maxwell theory, this leads to:

dω∂ = −δdϕ ∧ δ ? F = −d(δϕδ ? F ) + δϕδd ? F ≈ −d(δϕδ ? F ) , (136)

where we have used the equation of motion d ? F Γ = 0 in the last equality.50 Henceforth, we can
define ω∂ as
ω∂ = −δϕδ ? F + c , (137)
where c is some closed form on Γ. If ∂Σ is simply-connected and has no boundary, c is also exact and
does not contribute to the presymplectic form. We will assume we are in this situation and choose
c = 0.51 If we decide to impose a boundary condition on ?F , we find out that dω∂ ≈ ˆ 0; ω∂ = 0
is then a consistent choice, as expected. By contrast, as soon as the boundary condition constrains
Arad ,52 thereby leaving at least some components of ?F dynamical on the boundary, the boundary
presymplectic current contributes non-trivially.
Let us now consider the question of gauge invariance. In Yang-Mills theories, the presymplectic
current ω and presymplectic potential Θ are gauge-invariant to start with, but are not invariant under
field-dependent gauge transformations. In [3], the authors postulated such an extended invariance,
which in turn motivated the introduction of abstract edge modes and led to the extended presymplectic
structure ΩM . In the present work, we are somewhat reversing the logic. Edge modes are physical
from the start and already present in the global phase space. The way they must contribute to
the regional presymplectic structure ΩM has then been determined by the physical requirement of
a vanishing symplectic flux through Γ. As we will see now, invariance under field-dependent gauge
transformations is a necessary by-product of this construction, but conceptually secondary.
Field-dependent gauge transformations Xα[A] verify two interesting properties:
50
Note that it is also possible to consistently include boundary matter sources: it is indeed sufficient to assume that
d?F Γ ≈ ˆ j where j is some background field for the consistency condition dω∂ ≈ ˆ ω Γ to hold.
51 1
Given our working hypotheses, ∂Σ is homeomorphic R to S in 2 + 1 dimensions, and therefore fails to be simply-
connected. In that case, including a closed form c with ∂Σ c 6= 0 might be relevant to consistently
R account for non-trivial
winding modes of U (x). However, in our set-up, U (x) is always homotopic to the identity ( ∂Σ dϕ = 0), and c = 0
remains a consistent choice.
52
That is, for Dirichlet and Robin boundary conditions, as discussed in Sec. 7.2.

42
 X0
(a) Irrespective of the choice of gauge-invariant boundary condition X[φrad ] = X0 , SM is stable
under the flow of Xα[A] , since the latter maps solutions to solutions and leaves φrad invariant.

(b) Xα[A] generate local gauge symmetries, which can be tuned independently in disjoint neighbor-
hoods of M .
Given a Cauchy surface Σ, we can always pick up a second Cauchy surface Σ0 such that Σ and Σ0 are
disjoint. By (131) and property (a), we know that:
Z Z Z Z
Xα[A] · ω + Xα[A] · ω∂ ≈ Xα[A] · ΩM ≈ Xα[A] · ω + Xα[A] · ω∂ . (138)
Σ ∂Σ Σ0 ∂Σ0

Furthermore, thanks to property (b), we can always find a gauge transformation α̃[A] that agrees with
α[A] on Σ but vanishes on Σ0 :

∀x ∈ Σ , α[A](x) = α̃[A](x) , and ∀y ∈ Σ0 , α̃[A](y) = 0 . (139)

It follows from (138) that

Xα[A] · ΩM ≈ 0 . (140)
Using Cartan’s magic formula, this implies gauge-invariance of the regional presymplectic structure:

LXα[A] ΩM ≈ 0 , (141)

which can be equivalently written as

LXα[A] ΘΣ ≈ δ(Xα[A] · ΘΣ ) . (142)

In our main examples of Sec. 7, we will see that Xα[A] · ΘΣ is a constraint (a fact we will discuss
explicitly for Maxwell theory in Sec. 7.2), so that LXα[A] ΘΣ ≈ 0. Such a formula played the role of a
postulate in the construction of the extended phase space for Yang-Mills theory first outlined in [3].
Our starting point was different, but we recover it in a second step. One can directly check that (140)
holds in Maxwell’s theory:
Z Z Z Z
Xα[A] · ΩM = dα[A] ∧ δ ? F − α[A]δ ? F ≈ d(α[A]δ ? F ) − α[A]δ ? F = 0 . (143)
Σ ∂Σ Σ ∂Σ

It is interesting to contrast the invariance of ΩM under field-dependent gauge transformations

with its behaviour under the reference frame symmetries U 7→ U g −1 . For definiteness and simplicity,
let us focus on Maxwell theory (Chern-Simons and Yang-Mills theory will be discussed in Sec. 7,
accompanied with a more in-depth discussion of Maxwell theory). The reference frame symmetries are
then generated by the transformations ∆ρ defined in (125); let us denote the associated field-dependent
vector fields by Yρ[A] . The status of Yρ[A] depends on the type of boundary condition one is imposing. If
X0
X[Arad , (?F )rad ] = (?F )rad , any Yρ[A] leaves SM invariant, so that condition (a) holds. Since condition
(b) also holds, we are led to the conclusion that the vector fields Yρ[A] generate gauge transformations
rather than symmetries. Indeed, we find that the charge vanishes by virtue of the boundary condition
?F Γ = X0 Z
Yρ[A] · ΩM = ρ[A]δ ? F ≈
ˆ 0. (144)
∂Σ
For
R field-independent ρ, we can thus define, in addition to the standard U(1) bulk constraints C[α] =
Σ αd ? F (see Sec. 7.2 for further details), the edge constraints
Z
C∂ [ρ] := ρ(?F − X0 ) , (145)
∂Σ

which, by virtue of ∆ρ ? F = 0, are first-class

{C∂ [σ], C∂ [ρ]} = Yρ · Yσ · ΩM = 0 . (146)

43
Note that, by (137), the edge constraints are conjugate to the edge mode ϕ of Maxwell theory. Hence,
for Neumann boundary conditions in Maxwell theory the edge modes assume a somewhat redundant
role. The phase space has been extended, but also subjected to another set of first-class momentum-
type constraints so that the phase space extension does not amount to an increase in the number of
independent gauge-invariant degrees of freedom on Γ, despite the presence of edge modes as reference
frames.
By contrast, when X[Arad , (?F )rad ] = Arad , only the vector fields Yρ[A] with dρ[A] = 0 leave
X0
the solution space SM invariant. That is, ρ[A] must be a constant on Γ. For such transformations,
condition (a) still holds but (b) does not. As a consequence, they escape our general argument. Indeed,
we can explicitly verify that such transformations do generate non-trivial symmetries, recorded by
integrable charges Q[ρ] when ρ is furthermore assumed to be field-independent:53
Z
Yρ · ΩM = δQ[ρ] with Q[ρ] := ρ?F . (147)
∂Σ

Finally, transformations Yρ with ρ field-independent but not constant on Γ are neither symmetries nor
gauge: since they affect the boundary conditions themselves, they can be understood as non-trivial
X0 X0
transformations of background fields, which map one solution space SM to another SM0 , both descend-
ing from the same foliation of the global solution space S in (120). They are symplectomorphisms with
non-vanishing charges, but between different phase spaces. Indeed, Cartan’s magic formula implies
LYρ ΩM = δ (Yρ · ΩM ) = δ 2 Q[ρ] = 0; they therefore constitute meta-symmetries.

6 Boundary actions from splitting post-selection

After exploring splitting post-selection directly at the level of solution spaces, let us extend this discus-
sion by constructing a suitable variational principle for it. Post-selecting the total space of solutions
S of M ∪ M̄ on certain boundary conditions X = X0 on Γ will have the effect of turning the global
variational problem for M ∪ M̄ into two separate variational problems for M and M̄ . We will begin
by considering a fictitious boundary Γ between the two subregions M and M̄ , i.e. a boundary which
is not physically distinguished through sources or currents residing on it. In other words, the regular
bulk equations of motion of the theory have to hold on the interface Γ. We shall address the question
of physical boundaries in a follow-up work [101].
In order to have well-defined equations of motion everywhere on the space of field configurations
F, we shall assume the latter to only contain field configurations Φa which are C q across M ∪ M̄ (incl.
Γ) if the equations of motion are of order q in derivatives.
A variational problem for the subregion M of interest can be formulated in two equivalent ways.
(1) If the field configuration space FM associated with M only contains configurations which satisfy
the desired boundary conditions, the role of the boundary action will be limited to canceling the
presymplectic potential on the interface Γ on-shell. This is necessary in order to make the boundary
value problem well-defined and consistent with the global variational problem. (2) If instead FM
contains all C q field configurations in M , it accordingly requires a second type of contribution to the
boundary action, with the purpose of imposing the desired boundary conditions as boundary equations
of motion. Post-selecting the global solution space on X = X0 will first yield the former formulation,
which we however then use to construct the equivalent one in terms of boundary actions.
Once translated into the covariant phase space language, post-selection will result in a simple but
general algorithm which, given a set of admissible (Dirichlet, Neumann, Robin or mixed) boundary
conditions, produces a family of appropriate boundary actions implementing these boundary conditions.
53
In particular, note that the Dirichlet boundary condition Arad = X0 , in contrast to the case of Neumann boundary
conditions, is now a holonomic constraint (devoid of momentum degrees of freedom) and such a constraint does not
constitute a generator of gauge transformations. Accordingly, the extension of the phase space through edge modes does
in this case lead to additional independent gauge-invariant degrees of freedom.

44
6.1 Splitting a global variational problem into local ones via post-selection
Suppose we are given some field theory and would like to formulate its variational principle for the
subregion M subject to the boundary conditions X a = X0a on the interface Γ, where the X a are some
functionals locally constructed out of the pullbacks of the radiative fields φrad and the X0a denote a
corresponding choice of background fields.
A natural way to proceed is to start with the space of solutions S ⊂ F associated with the global
action Z
SM ∪M̄ = L = SM + SM̄ (148)
M ∪M̄
and post-select (i.e. restrict to) the subset SX0 ⊂ S of solutions satisfying the desired conditions on Γ.
Noting that the boundary conditions are integrable, δX = 0,54 this can be achieved by imposing them
as a holonomic constraint, yielding the total action
Z
Stot := SM ∪M̄ + λa (X a − X0a ) , (149)
Γ

where λa denote a set of Lagrange multiplier densities defined on Γ.55 In order to maintain gauge-
invariance of the total action when there is a gauge symmetry, we shall require the Lagrange multipliers
λa to transform trivially under the relevant gauge transformations. Furthermore, since we are dealing
with an interface between two regions, we will have to choose a convention as to which orientation of
Γ we tacitly invoke in the integral of (149); henceforth, we shall always choose the ‘perspective’ of our
subregion of interest M , i.e. orient Γ relative to the normal pointing outward from M .
The action (149) constitutes a well-defined variational problem since (assuming asymptotic fall-off
conditions of Φa in M̄ )
Z Z
δStot = (Ea δΦ + dΘ) + (δλa (X a − X0a ) + λa δX a )
a

ZM ∪M̄ Z Γ Z
= a
Ea δΦ + Θ + (δλa (X a − X0a ) + λa δX a ) , (150)
M ∪M̄ Σ2 ∪Σ̄2 −Σ1 ∪Σ̄1 Γ

where we used the shorthand notation A−B :=R A − B . Note that there is no contribution Rof the
R R R

presymplectic potential Θ on Γ; the contribution Γ θ coming from M cancels the corresponding Γ̄ θ =

− Γ θ coming from its complement M̄ because their integrals along Γ feature opposite orientation.56
R

In other words, around field configurations satisfying the bulk equations of motion Ea = 0 and the
interface equations of motion X a = X0a and λa = 0, the total variation is stationary up to terms on
the future and past Cauchy surface
Z
δStot ≈
ˆ Θ. (151)
Σ2 ∪Σ̄2 −Σ1 ∪Σ̄1

We would like to turn this global variational problem into two independent variational problems,
one for M and one for its complement M̄ , which together are equivalent to the global one. This can
be achieved by noting that the two action contributions for M and M̄ in (148) and (149) are only
uniquely defined up to the addition and subtraction of a local integral over their interface Γ. More
precisely, we are free to replace SM and SM̄ in (149) with
Z Z
S̃M := SM + `corr , S̃M̄ := SM̄ − `corr (152)
Γ Γ
54
Since X0 is a background field, the condition X − X0 = 0 entails the differential constraint δX = 0, which thus is
integrable and, in turn, holonomic.
55
We emphasize that X a = X0a does not constitute a gluing condition as will be discussed in our follow-up work [101]
(see also [10, 11]) since X0a is a non-dynamical background field and we are here not identifying a priori independent
degrees of freedom from M and M̄ on Γ.
56
Γ̄ here denotes the interface equipped with the canonical orientation induced by M̄ .

45
without changing Stot , where `corr is any density locally constructed out of φa , the pullback of the
derivatives of Φa to Γ, as well as the reference frame U .57 This argument is completely analogous
to the one used in Sec. 5 to motivate the Ansatz (128), now transposed onto the level of the action.
In fact, we can understand what follows as an off-shell extension of our previous construction, with
respect to both equations of motion and boundary conditions.
In particular, we can exploit this ambiguity and choose `corr such that
Z Z Z Z
δ S̃M ≈
ˆ Θ+ C, δ S̃M̄ ≈
ˆ Θ+ C. (153)
Σ2 −Σ1 ∂Σ2 −∂Σ1 Σ̄2 −Σ̄1 ∂ Σ̄2 −∂ Σ̄1

We allow for a possible corner contribution C, which we will view below as arising from a (1, d−2)-form
intrinsic to Γ and locally constructed out of the fields, their variations and derivatives. Note that we
then have
δStot ≈
ˆ δ S̃M + δ S̃M̄ (154)
since the two corner contributions will cancel one another, seeing as they come with opposite orientation
on ∂Σi and ∂ Σ̄i , i = 1, 2. Hence, upon invoking the boundary conditions through post-selection, the
global variational problem in (150) splits into two independent subregion variational problems in (153),
which together are equivalent to the global one.
We can render the condition on `corr in (153) more explicit by noting that the first equality, which
is associated to the subregion of interest M , is equivalent to
Z Z
(θ + δ`corr ) ≈ − dC .
ˆ (155)
Γ Γ

Using Stokes’ theorem, the corner terms have been reexpressed as an integral over Γ, where the extra
minus sign is explained by the fact that the orientations of the integrals over ∂Σ2 and ∂Σ1 induced
by Σ2 and Σ1 are opposite to the orientation of the integral over ∂Γ which is induced by Γ. If we
require the condition in (155) to hold for any choice of initial and final time slice Σ1 and Σ2 (which is
tantamount to varying Γ), it reduces to a local one:

θcorr := θ + δ`corr ≈
ˆ −dC . (156)

That is to say, we have a shifted presymplectic potential θcorr on Γ which is such that it vanishes up
to a corner term on field configurations satisfying the boundary conditions. Since, however, ωcorr :=
δθcorr = δθ = ω Γ , we can already see that this will not affect the symplectic flux through the interface
and we shall discuss this further below. A sufficient condition for equation (156) to hold is that

θcorr ≈ −dC + Ya δX a (157)

for some suitable set of densities Ya constructed locally out of the φa and the pullbacks of the derivatives
of Φa to Γ, as well as of U and its derivatives.
The conditions (156) and (157) are similar to the sufficient condition for a well-defined subregion
variational principle put forth in [66], however, are also somewhat different. Here we arrive at them
from a global variational principle using post-selection, and `corr is a boundary Lagrangian whose role
is not to impose the desired boundary conditions, which are already satisfied in (156). `corr rather
assumes two different roles for the subregion M that are otherwise played by the complement M̄ and
which are necessary in order to render the subregion variational problem well-defined and consistent
with the global one:

(i) Varying the fields in M generates a presymplectic potential θ on Γ which is cancelled by the
corresponding contribution −θ from varying the fields in M̄ . This is important for the global
variational problem as θ does not in general vanish (or reduce to a corner term) individually
57
In both modified actions we oriented Γ according to our convention above and, accordingly, the volume form Γ and
any normal derivatives in `corr are defined relative to M .

46
 when the boundary conditions are fulfilled. From the perspective of M the interface Lagrangian
`corr mimics the effect of the complementary region M̄ in such a way that θ gets cancelled up
to a corner term once the boundary conditions are satisfied. This enables us to split the global
variational problem into two independent ones. As we shall see below in theories admitting
Darboux coordinates, `corr constitutes a generating function of a canonical transformation that
implements a change of polarization on field space. This in turn ensures that the boundary
conditions X a = X0a become truly holonomic constraints that only depend on field configurations,
but not their derivatives.58

(ii) In some topological theories, such as Chern-Simons theory (see Sec. 7.3), the action SM asso-
ciated with a finite subregion may not be gauge-invariant, with its non-invariance coming from
a boundary piece. This boundary piece has to be cancelled by the complement SM̄ in order
to ensure invariance of the global action in the bulk of M ∪ M̄ . For such theories, `corr will
also mimic the role of the complement M̄ by cancelling any non-invariance of SM and thereby
establishing gauge-invariance of the subregion variational principle. In other words, `corr is only
gauge-invariant if SM is.

So far we have split the global variational problem into independent subregion problems after
globally imposing the boundary conditions. The ensuing variational problem for subregion M is defined
by the action S̃M on the post-selected field configuration space FM X=X0 , containing all C q field
configurations in M which satisfy the boundary conditions X = X0 on Γ. As in [67,109] the boundary
conditions are thus directly built into the field configuration space. Clearly, we can equivalently
formulate this subregion variational problem on the larger space FM of all C q field configurations
in M if we dynamically impose the boundary conditions through Lagrange multipliers. On FM the
variational principle for the subregion is therefore defined by
Z
S̃M ∪Γ := S̃M + λa (X a − X0a ) = SM + Sps , (158)
Γ

where Z
Sps := (`corr + λa (X a − X0a )) (159)
Γ
is the post-selection boundary action. Condition (157) implements a polarization of field configuration
space in which the X a are configuration degrees of freedom. This guarantees that we can consistently
impose the boundary conditions X a = X0a as genuine holonomic constraints. In problems with holo-
nomic constraints the Lagrange multipliers have physical significance: they can be solved for as the
forces imposing the constraints and this solution can be reinserted into the action, defining the vari-
ational principle for the remaining degrees of freedom. In our case, (157) entails that the constraint
forces read λa = −Ya which yields the final subregion action on FM

SM ∪Γ := SM + SΓ , (160)

with boundary action Z

SΓ := Sps λ=−Y
= (`corr − Ya (X a − X0a )) . (161)
Γ
As we will illustrate below, this procedure will provide a systematic algorithm for generating boundary
actions, encompassing various well-known ones in the literature as well as producing new ones. This
will also elucidate the appearance of edge modes in boundary actions.
58
Note that, since we are considering canonical transformations of δθ, the relevant field space here is the space of fields
pulled-back to Γ. As a result, constraints involving normal derivatives of the fields but no intrinsic derivative on Γ are
holomomic.

47
6.2 Local Darboux coordinates, appearance of edge modes and non-uniqueness of
the boundary action
In this subsection, we shall render the above general construction more concrete, using the edge mode
dressed observables and field space forms of Secs. 4 and 5. We shall also discuss the non-uniqueness of
the boundary action implementing a specific boundary value problem.

6.2.1 General algorithm

In order to render this construction more explicit, let us assume we are given a theory admitting the
use of Darboux coordinates on F, so that the presymplectic potential reads

Θ = Πa δΦa , (162)

for some density Πa and where Φa denote the fields in the Lagrangian. Its pullback to Γ will be denoted
by
θ = πa δφa . (163)
For example, this assumption is true in all theories whose Lagrangian depends on the fields Φa and
their first, but not higher order derivatives [56]; this is the case e.g. in Maxwell, Chern-Simons and
Yang-Mills theory as discussed later. In this case, Πa typically (but not always) depends on only the
first derivatives of Φa . This assumption can, however, also be realized in theories with Lagrangians
featuring higher orders of derivatives, which might require either a relabeling of field variables or taking
care of boundary terms.59
When dealing with a gauge theory, it will be convenient to split the field degrees of freedom φa and
the momenta πa on Γ into pure gauge and radiative parts, as explained in Sec. 4:

φa = φarad + φagauge and πa = πarad + πagauge . (164)

Owing to the symmetry induced redundancy, this can always be done in many different ways in the
global covariant phase space, but it is uniquely defined once we have specified a reference frame.
Next, we shall make one further assumption, namely that we can use the split into pure gauge
and radiative degrees of freedom to decompose the presymplectic potential on solutions to the bulk
equations of motion on Γ as

θ = θrad + θgauge , θrad = πarad δφarad θgauge ≈ −dθ∂ − δ`∂ , (165)

where θ∂ and `∂ have already been introduced in equation (135). The extra assumption we are making
is that equation (134) (from which (135) follows) can be extended off-shell of the boundary conditions.
In other words, the non-invariant part of θ is in the kernel of δd on-shell of the equations of motion.
This assumption holds true in many gauge theories and we shall illustrate it in Maxwell, Chern-Simons
and Yang-Mills theory below. The assumption is essentially that we can find local Darboux coordinates
for the degrees of freedom on Γ that split into gauge-invariant and -variant pairs, where the ‘momenta’
of the latter pair are simply the constraints of the theory which vanish on-shell. Such a split is related
to a so-called constraint abelianization [58].
The gauge-invariant pairs (φarad , πarad ) constitute the set of degrees of freedom on which boundary
conditions will be imposed. Typically, one imposes boundary conditions either on the fields (Dirichlet),
their normal derivatives (Neumann), or linear combinations (Robin). Since the latter contain the former
two as special cases, we shall directly formulate the variational problem for Robin boundary conditions.
In fact, we shall be even slightly more general and permit the nature of the boundary conditions to
change as ∂Σ evolves through Γ, so that we encompass mixed boundary conditions, i.e. time-dependent
59
For example, also the presymplectic potential of general relativity can be put into the form of equation (162) by
taking care of boundary terms appearing in the Gauss-Codazzi equations [21, 56].

48
Robin boundary conditions. To this end, it will suffice to consider linear combinations of ‘position’
and ‘momentum’ variables, which for simplicity we write in matrix notation
     
X φrad a b
=M , M= , (166)
y prad c d

where a, b, c, d are real n×n-matrix background fields and prad is the n-vector with scalar components
a , defined by πa = pa Γ , where Γ is the volume form on Γ, and similarly y is the n-vector whose
prad rad rad

components are defined by a new ‘momentum’ density Ya = ya Γ . In order for this to define a symplectic
transformation, we shall require

ωrad = δπ rad .δφrad = δY.δX , (167)

where . denotes the dot product between n-vectors, which is equivalent to

d> a − b> c = 1, c> a = a> c ,

a> d − c> b = 1 , d> b = b> d , (168)

so that M ∈ Sp(2n, R).

We are now in a position to determine the shape of the interface action `corr ensuring consistency
of the subregion variational problem from the condition in (157)

δ`corr ≈ Y.δX − θ − dC
≈ Y.δX − π rad .δφrad + d(θ∂ − C) + δ`∂ (169)
1  
= δ Y.X − π rad .φrad + 2`∂ + d(θ∂ − C) ,
2
where we have invoked (168) in arriving at the last line. To solve this consistency relation, we henceforth
define C and `corr on FM by:60
1 
C := θ∂ , `corr := `∂ + Y.X − π rad .φrad . (170)
2
The above observations permit us to shed new light on the so-called extended presympletic potential,
as considered, for instance, in [3, 11]:61

θe := θ + dθ∂ + δ`∂ . (171)

The corner term θ∂ has been added to θ in [3,11] to render the potential gauge-invariant even for gauge
transformations with support on Γ, and independently of any choice of boundary conditions. Noting
the decomposition (165), on-shell the extended potential becomes the pure radiative contribution

θe ≈ π rad .δφrad = θrad . (172)

This is a consequence of our variable split (164) which renders the meaning of the extended potential
especially transparent. Similarly, the extended symplectic current ωe := δθe of [3, 11] then coincides
on-shell with the radiative contribution constructed in equation (110),

ωe ≈ ωrad . (173)

It is thus the extended symplectic structure which ultimately is post-selected to vanish boundary
on-shell in equation (121), ωrad ≈
ˆ 0, so as to eliminate gauge-invariant symplectic flux through the
interface Γ (see Sec. 5).
60
It is implicitly assumed here that quantities such as θ∂ and `∂ can be naturally extended from forms on SM to forms
on FM .
61
The field-space exact term δ`∂ was not included in those works, as it happens to be irrelevant in many situations of
interest; we keep it for full generality.

49
 The boundary condition X = X0 in (159) is clearly the integrated version of the differential
constraint δX = 0, which thus is holonomic. Holonomic constraints can be written as functions of
only the configuration degrees of freedom. This is not the case in the old variables (φrad , π rad ), unless
b = c = 0, but clearly in the new boundary variables (X, Y). The interface action `corr thus implements
a change of polarization of the field configuration space, as can already be inferred from (157). Indeed,
`corr − `∂ can be interpreted as a generating function of a canonical transformation in multiple ways.
For example, if b is invertible and `∂ = 0, one can use (166) to rewrite `corr as a type I generating
function of a canonical transformation
1
X.(db−1 X) − 2(b−1 X).φrad + (b−1 aφrad ).φrad Γ ,


`corr =
2
which can be checked to satisfy the functional derivative relations

δ`corr (x) δ`corr (x)

= −πarad (x)δ (d−1) (x, y) , = Ya (x)δ (d−1) (x, y) .
δφarad (y) δX a (y)

Similarly, when d is invertible, one can write `corr in terms of a type II generating function, etc.
Finally, we can now also write the boundary action for the subregion variational principle on FM
in (161) explicitly
Z
1 
SΓ = `, ` := `corr − Y.(X − X0 ) = Y.X0 − X.Y + π rad .φrad + `∂ . (174)
Γ 2

In more detail, we find that

1 1
` = `∂ + (aφrad + bπ rad ).X0 − φrad .(a> cφrad ) − π rad .(b> dπ rad ) − π rad .(d> aφrad ) . (175)
2 2
When `∂ vanishes, this action is gauge-invariant by construction. This is for instance the case in Yang-
Mills theories. By contrast, in the case that SM is not gauge-invariant, as for example in Chern-Simons
theory, one still has the freedom to add the corresponding non-invariant piece on Γ to `corr (via `∂ ) as
mentioned under (ii) in Sec. 6.1. Such a requirement may be imposed in addition to our main on-shell
consistency condition (169).
This concludes the general procedure for post-selecting on the gauge-invariant variables X.

6.2.2 Non-unicity of the boundary action

The boundary action (174) was derived from a particular choice of polarization (X, Y) for the symplec-
tic current ωrad = δπ rad .δφrad = δY.δX. Since we are imposing boundary conditions on X, any other
polarization (X0 , Y0 ) such that X0 is related to X by an invertible transformation is also admissible.
More precisely, any two such polarizations are related by an Sp(2n, R) transformation of the form
 0     
X X a 0
=N , N= , (176)
y0 y c d

where d> a = 1 and c> a = a> c. It is convenient to parametrize such transformations by the invertible
matrix a and the symmetric matrix s = a> c. They generate the subgroup of Sp(2n, R)
  
a 0
N = |a ∈ GL(n, R) , s ∈ Sym(n, R) . (177)
(a> )−1 s (a> )−1

We can view this set as the matrix product N = DT , where

  
a 0
D= |a ∈ GL(n, R) ' GL(n, R) (178)
0 (a> )−1

50
and   
1 0
T = |s ∈ Sym(n, R) ' Sym(n, R) . (179)
s 1
As a group, N is in fact
  a semidirect product of D and T . Indeed, introducing the notation N(a, s) :=
a 0
, we are led to the semidirect group multiplication law:
(a> )−1 s (a> )−1

N(a1 , s1 )N(a2 , s2 ) = N(a1 a2 , a>

2 s1 a2 + s2 ) . (180)

We therefore conclude that

N ' GL(n, R) n Sym(n, R) . (181)
Among the admissible changes of polarization encoded in N , which transformations affect the
functional form of the boundary action `? From the construction of the previous subsection, and in
particular equation (169), it is clear that any canonical transformation that preserves the flux E := YδX
will also preserve δ` (and thereby `, up to an irrelevant additive ambiguity). Such transformations
generate the (normal) stabilizer subgroup D ⊂ N . Ambiguities in the definition of ` are therefore
parametrized by the Abelian subgroup T (which is not normal). Acting with such a transformation
on the polarization (X, y) yields new coordinates

X0 = X , y0 = sX + y , (182)

where s is a real symmetric matrix. By application of (174), the boundary Lagrangian produced by
our algorithm in the new coordinates is
1  1
`0 = Y0 .X0 − X.Y0 + π rad .φrad + `∂ = ` + (sX).X0 − X.(sX) (183)
2 2
1 1
= ` − (X − X0 ). (s(X − X0 )) + X0 . (sX0 ) , (184)
2 2
where we have used the fact that s is symmetric in the last line.
We see that `0 only differs from ` by:

• a non-trivial quadratic term − 21 (X − X0 ). (s(X − X0 ));

• and a field-independent additive contribution 21 X0 . (sX0 ).

The second term can be ignored since, by construction, ` is only defined up to such additive contribu-
tions. By contrast, the first term affects the functional form of δ` and is a priori non-trivial. However,
because it is quadratic in (X − X0 ), we have:

δ`0 = δ` − (X − X0 ). (sδX) , (185)

which implies that δ` and δ`0 coincide on-shell of the boundary conditions. As a result, this ambiguity
only affects the functional form of the boundary action, not the presymplectic structure that results
from it.62 It is then clear that we are not changing anything of substance when we change polarization.
Likewise, we will show in an upcoming paper that work functionals can be consistently associated to
the boundary Lagrangian `, and are not affected by such quadratic (or higher order) ambiguities.
62
In fact, this ambiguity in δ` is of the general form (X i −X0i )δZi , and such contributions have been implicitly discarded
in the definition (157).

51
6.3 Summary of the algorithm
In summary, given a set of boundary conditions X = X0 to be imposed dynamically, a suitable
boundary Lagrangian (0, d − 1)-form ` can be obtained by implementation of the following steps:
1. Decompose the presymplectic potential on Γ into radiative and gauge contributions, relative to
a frame U : θ = θrad + θgauge .
2. Find local Darboux coordinates for the radiative degrees of freedom, so that θrad = πarad δφarad ,
where πarad = prad
a Γ is the momentum density dual to the frame-dressed field φrad = (U
a −1 . φ)a .

3. Find a (1, d − 2)-form θ∂ and a (0, d − 1)-form `∂ such that, on-shell: θgauge + dθ∂ + δ`∂ ≈ 0.
4. Find a linear transformation to new Darboux coordinates (Y, X) such that:63 δθrad = δY.δX.
5. Define the boundary Lagrangian as:
1 
` = Y.X0 − X.Y + π rad .φrad + `∂ . (186)
2

Then, as we have explained in the main text, the stationary points of the action SM ∪Γ =
R R
M L+ Γ`
X0
in the configuration space FM , span the solution space SM .

7 Illustration of splitting post-selection in example field theories

We showcase the splitting post-selection algorithm for generating boundary actions that impose Robin,
Dirichlet and Neumann boundary conditions in the examples of scalar field, Maxwell, Abelian Chern-
Simons, and Yang-Mills theories. The latter three examples will more specifically illustrate the ap-
pearance and role of edge modes on the boundary Γ. We also emphasize that, given that the matrix
M implementing the change of polarization in (166) is a background field, it allows us in principle to
continuously interpolate between different types of boundary conditions along Γ. For clarity of the
exposition, we will not consider such mixed boundary conditions in the following examples.
We emphasize that throughout this section, we consider general Lorentzian spacetimes, which in
particular can be curved and, except in the Chern-Simons theory case, of arbitrary dimension d ≥ 2.

7.1 Scalar field theory

Varying the Lagrangian  
1
L=− µ
∇µ Φ∇ Φ + V (Φ)  (187)
2
of the bulk scalar field theory action SM = M L[Φ] yields
R

δL = ∇µ ∇µ Φ − V 0 (Φ) δΦ + dΘ , (188)

where  is the spacetime volume form and the presymplectic potential reads

Θ = θ · , with θµ = −∇µ ΦδΦ . (189)

The pullback of Θ to Γ takes the form

θ=Θ Γ
= −∂n φδφΓ , (190)

where φ := Φ Γ , ∂n φ := nµ ∇µ Φ Γ and nµ is the outward pointing normal to Γ. Since there is no gauge

symmetry in this model, we directly have

φrad = φ , prad = −∂n φ , θrad = θ . (191)

63
In particular, X a must be linear in φarad and prad
a , as we have assumed in our construction.

52
The coefficient matrices defining X and y in (166) are here real functions a, b, c, d ∈ R satisfying
ad − bc = 1, so that M is an SL(2, R) matrix field. The generating function of this canonical
transformation reads `corr = 2 Y X − π φrad , where π rad = prad Γ .
1 rad

Equation (174) implies that the variational principle forR subregionR M subject to Robin boundary
conditions X = X0 is thus given by SM ∪Γ = SM + SΓ = M L[Φ] + Γ `[φ] with boundary Lagrangian
 
1
` = yX0 − (yX − φ∂n φ) Γ . (192)
2
Indeed, varying this action yields the shifted symplectic potential θcorr = θ + δ`corr = Y δX and
Z Z Z
0
µ
(193)


δSM ∪Γ = ∇µ ∇ Φ − V (Φ) δΦ + Θ − (X − X0 ) Γ δy ,
M Σ2 −Σ1 Γ

thus dynamically imposing the desired boundary conditions as boundary equations of motion.
In the special case that a = d = 1 and b = 0, we post-select on X = φ, giving Dirichlet boundary
conditions φ = φ0 , therefore a vanishing generating function `corr = 0, and a boundary Lagrangian
(setting for simplicity also c = 0)
` = ∂n φ (φ − φ0 ) Γ . (194)
The dual case a = 0 and b = −1 = −c amounts to post-selection on the normal derivative X = ∂n φ
and thus generates Neumann boundary conditions ∂n φ = N0 which are imposed by the simple
boundary action (setting d = 0)
` = φN0 Γ . (195)
This leads to a change of polarization compared to (190)

θcorr = φδ∂n φΓ . (196)

7.2 Maxwell theory

While Maxwell theory is the restriction of the Yang-Mills theory construction in Sec. 7.4 to the special
U(1) case, we illustrate it here for clarity and its standalone physical interest. We start from the bulk
Lagrangian in d spacetime dimensions:
1
L[A] = − F ∧ ?F , (197)
2
where F := dA is the curvature of a U(1) connection. The variation of L gives

δL = −dδA ∧ ?F = −d(δA ∧ ?F ) − δA ∧ d ? F , (198)

from which we recover the bulk equation of motion, d ? F = 0, and the presymplectic potential

Θ = −δA ∧ ?F , (199)

so A and ?F are conjugated.

On the boundary Γ, we invoke the split of A into a radiative and a pure gauge part

A Γ
= Arad + dϕ , (200)

where Arad is defined relative to the edge mode frame in equations (65) and (122), which splits the
presymplectic potential on Γ into θ = θrad + θgauge , where

θrad = −δArad ∧ ?F , θgauge = (d ? F ) δϕ − d (?F δϕ) ≈ −d (?F δϕ) . (201)

53
In order to formulate the decomposition (165) at the form level, we have to invoke the Hodge star
operator ∗ on Γ. Using that ∗ ∗ v = (−1)(d−1−q)q+1 v for a q-form v on a (d − 1)-dimensional Lorentzian
manifold, we can write the radiative piece of the presymplectic potential as

θrad = −δArad ∧ ?F = (−1)d δArad ∧ ∗ ∗ ?F = (−1)d hδArad , ∗ ? F iΓ

= (−1)d h∗ ? F, δArad iΓ = (−1)d ∗ ?F ∧ δ ∗ Arad , (202)

where h·, ·i denotes the inner product of 1-forms on Γ. Accordingly, the assumption in (165) is realized
with θ∂ = ?F δϕ, `∂ = 0, and gauge-invariant momentum one-form π rad = (−1)d ∗ ?F .64 Here,
ϕ is the phase of a U(1) edge reference frame U (x) = eiϕ(x) , see Sec. 4. Clearly, it transforms as
δα ϕ := Xα · δϕ = α under the infinitesimal gauge transformation Xα acting on F (see equation (124)).
The momentum density conjugated to the edge mode is a constraint: the
R bulk equation of motion
pulled back to Γ. Likewise, we find that the contraction of ΘΣ = Σ Θ + ∂Σ θ∂ with Xα is an on-shell
R

constraint, as was already anticipated after equation (142):

Z
C[α] := Xα · ΘΣ = − αd ? F ≈ 0 . (203)
Σ

The splitting of the presymplectic current into ω = ωrad + ωgauge can be deduced from (201) by
acting with the exterior derivative δ, leading to:

ωrad = δArad ∧ δ ? F , ωgauge = δ(d ? F )δϕ − d(δ ? F δϕ) ≈ −d(δ ? F δϕ) . (204)

Following our construction of Sec. 5, we obtain a presymplectic structure which is conserved on-shell,
and gauge-invariant independently of any choice of boundary conditions (see equation (141) which here
can be directly checked):
Z Z Z Z
ΩM = ω+ δθ∂ = δA ∧ δ ? F + δ ? F δϕ . (205)
Σ ∂Σ Σ ∂Σ

This expression coincides with the extended presymplectic structure used in [11], but is here a result
of post-selection. As an aside, we note that considering the extended potential as in [11], but using
our variable split (200), yields the purely radiative piece

θe = θ + dC = −δArad ∧ ?F + (d ? F )δϕ ≈ −δArad ∧ ?F = θrad , (206)

since the gauge part θgauge of θ is cancelled by C on shell, exemplifying (172).

The split in (200) permits a transparent understanding of the distinction between gauge transfor-
mations and symmetries, as originally proposed in [3] (see also [11]). Gauge transformations act as

δα A = dα , δα Arad = 0 , δϕ = α , (207)
while symmetry transformations can be understood as leaving the vector potential on Γ invariant, but
changing its split into radiative and pure gauge parts:

∆ρ A = 0 , ∆ρ Arad = dρ , ∆ρ ϕ = −ρ . (208)

As frame reorientations, symmetries transform the edge mode and thereby indirectly transform the
frame-dressed gauge-invariant observables, such as Arad = A − dϕ (see Sec. 4.5), while gauge trans-
formations by construction leave all gauge-invariant data invariant. In the present case, a symmetry
transformation exploits the fact that the split into radiative and gauge parts is only unique up to
64
We emphasize that in all expressions defined on Γ, ?F denotes the pullback ?F Γ which for simplicity of notation we
shall not write explicitly. Moreover, we slightly depart from the notations introduced in (165): we distribute the volume
form Γ differently, so that π rad is a one-form rather than a density. This is for mere convenience due to our reliance on
the form language, which otherwise does not affect the applicability of the formalism laid out in Secs. 5 and 6.

54
gauge transformations and changes the physical part Arad by a total derivative. This has no impact
on local gauge-invariant quantities such as F or ?F in M ∪ Γ, but it does affect the relation between
the subregion M and the (non-locally defined) reference frame U , originating in the complement M̄ .
Let us now illustrate post-selection on Robin boundary conditions, from which we obtain Dirichlet
and Neumann conditions as special cases. If we wanted to proceed in the general manner of Sec. 6.2.1
we would have to choose a basis field in the cotangent space of Γ to decompose A Γ into scalar
components φa = Aa with a = 0, . . . , d − 2. This can clearly be done, however, here we choose to
work with form language instead, which provides a more elegant split (200) into radiative and gauge
parts and a more concise formulation of the boundary value problem. Since we will now build linear
combinations directly at the form (rather than component) level in order to define Robin boundary
conditions, the matrix M in (166) will be a 2 × 2 (rather than 2(d − 1) × 2(d − 1)) matrix. The Robin
boundary conditions thereby encompassed will thus only be a subset of all the possible Robin boundary
conditions one can formulate in component language. Besides offering a more concise formulation, the
advantage is that our boundary value problem will be independent of the choice of spacetime frame
field on Γ. More precisely, while the 1-form X = Xa dxa ≈ ˆ X0 , which we will use below for imposing
boundary conditions, is invariant under local frame (coordinate) changes, the components Xa transform
covariantly under the latter. The types of boundary conditions we are excluding by focusing on an
SL(2, R) subgroup of all canonical transformations are those that, in component form, are not even
covariant under local frame transformations.
In order to formulate the linear combinations (166) at the form level, we recall equation (202). This
suggests to define the fields for Robin post-selection in the form

X := aArad + (−1)d−1 b ∗ ?F , Y := cArad + (−1)d−1 d ∗ ?F , (209)

for a, b, c, d real functions (d ? F should not be confused with the equation of motion d ? F ), so that
X and Y are both (0, 1)-forms. Imposing the SL(2, R) constraint ad − bc = 1 ensures that (209) is a
canonical transformation of the radiative presymplectic current:65

ωrad = δArad ∧ δ ? F = δX ∧ ∗δY . (210)

Let us now determine the generating function `corr . Condition (169) translates here into

δ`corr ≈ −δX ∧ ∗Y − θ − dC
≈ −δX ∧ ∗Y + δArad ∧ ?F − d(C − ?F δϕ)
1
= δ (Arad ∧ ?F − X ∧ ∗Y ) − d(C − ?F δϕ) . (211)
2
In arriving at the last line, we have made use of the various identities above. Henceforth, we define
1
`corr := (Arad ∧ ?F − X ∧ ∗Y ) , C := ?F δϕ , (212)
2
in line with (170). In turn, we find that

θcorr := θ + δ`corr = −δX ∧ ∗Y + (d ? F )δϕ − d(?F δϕ) , (213)

to be contrasted with (201). The post-selection action (159) on FM thus reads

Z  
1
Sps = (Arad ∧ ?F − X ∧ ∗Y ) + λ ∧ (X − X0 ) , (214)
Γ 2

with the gauge-invariant (0, d − 2)-form λ playing the role of Lagrange multiplier.
65
We use the above identity for applying the Hodge operator twice, and note that hδv, δui = −hδu, δvi since δv, δu are
1-forms on field space.

55
 In conjunction, the subregion variational problem defined by post-selection on Robin boundary
conditions X = X0 is given by the action SM ∪Γ = M L[A] + Γ ` with gauge-invariant boundary
R R

Lagrangian (see equation (174))

1
` = −X0 ∧ ∗Y + (Arad ∧ ?F + X ∧ ∗Y ) , (215)
2
in analogous form to (192) of the scalar field case. Variation of this action yields
Z Z Z  
δSM ∪Γ = d ? F ∧ δA − δA ∧ ?F + (d ? F )δϕ + (X − X0 ) ∧ δ ∗ Y − d (?F δϕ) , (216)
M Σ2 −Σ1 Γ

hence implementing the Robin boundary conditions as boundary equations of motion and, in line with
our assumption that we are dealing with a fictitious boundary Γ, also imposing the bulk equation of
motion d ? F = 0 on Γ.
Dirichlet boundary conditions are obtained in the special case a = d = 1 and b = 0, which
amounts to post-selection on the radiative part of the connection X = Arad = A0rad . The generating
function `corr vanishes and the boundary Lagrangian becomes (when setting also c = 0 for simplicity)

` = (Arad − A0rad ) ∧ ?F = A − dϕ − A0rad ∧ ?F , (217)

where we have restored the standard connection A on Γ and the associated edge mode via (200). This is
a generalization of the boundary action for Maxwell theory proposed in [11, 13], which we thus recover
through our post-selection algorithm. In particular, our procedure demonstrates that the edge mode
in the action does not have to be postulated, but actually is part of the global theory space from the
start, and appears thanks to our variable split into radiative and gauge parts relative to the reference
frame U .66
By contrast, post-selection on ?F can be implemented by setting b = 1 and a = 0, so that
X = (−1)d−1 ∗ ?F = (−1)d−1 ∗ f0 , for some background (d − 2)-form f0 . This yields Neumann
boundary conditions. If we also set c = −1 and d = 0 for simplicity, we have the ‘momentum’
Y = −Arad and the boundary action reads

` = Arad ∧ f0 = (A − dϕ) ∧ f0 . (218)

Accordingly, the shifted presymplectic potential on Γ reads on-shell

θcorr ≈ δ ? F ∧ Arad − d(?F δϕ) , (219)

which amounts to the dual polarization of equation (201), i.e. with the roles of radiative configuration
and momentum degrees of freedom interchanged.
Finally, let us revisit the distinction between gauge transformations and symmetries. The former
yield constraints, while the latter give rise to non-vanishing charges. Indeed, let Xα denote the vector
field on FM inducing the gauge transformation in (207). On the space of solution SM , this defines a
degenerate direction of ΩM given in (205)
Z
δC[α] := Xα · ΩM = − αδd ? F ≈ 0 , (220)
Σ
66
More precisely, the action in (4.28) of [11] is given for d = 4 in the form `Γ = j ∧ (da + A), where a is a postulated
edge mode transforming as δα a = −α and j is a gauge-invariant dynamical 2-form that can be viewed as a Lagrange
multiplier. One can solve for the Lagrange multiplier, giving j = ?F , and the equations of motion also yield F Γ = 0.
In our action, we get the edge mode ϕ ≡ −a with correct transformation property directly from the reference frame
construction of Sec. 4, which allows us to split the boundary degrees of freedom into radiative and gauge parts as in (200).
Furthermore, the action we obtained generates the more general boundary value F Γ = dA0rad through the background
field introduced via post-selection. Hence, (217) is a generalization of the action in [11], evaluated on the solution for
the Lagrange multiplier j = ?F .

56
so that the constraint takes the expected form (203). Since we are dealing with a U(1) gauge theory,
the constraints yield an Abelian Poisson algebra

{C[α], C[β]} = Xβ · (Xα · ΩM ) = 0 . (221)

On the other hand, let Yρ with field-independent ρ denote the vector field on FM generating the
symmetry transformation in (208). In SM (that is, before boundary conditions are imposed), this does
not define a degenerate direction of the presymplectic form, since
Z
δQ[ρ] := Yρ · ΩM = ρδ ? F (222)
∂Σ

does not vanish in general. Instead, this yields the charge

Z
Q[ρ] = ρ?F . (223)
∂Σ

This expression is in agreement with the one in [11]. Constituting the generators of a U(1) frame
transformation, the charges likewise form an Abelian Poisson algebra

{Q[ρ], Q[σ]} = Yσ · (Yρ · ΩM ) = 0 (224)

and they are clearly gauge-invariant boundary observables, {C[α], Q[ρ]} = Yρ · (Xα · Ω) = 0. Finally,
X0
the interpretation of those generators in the regional solution space SM is contingent on the type of
boundary conditions being considered. As it has been discussed in Sec. 5.2 for gauge field theories, as
well as in Sec. 2.3 for a mechanical toy-model, we have three possibilities.

Symmetries. These are transformations which leave the boundary conditions invariant, but allow for
non-trivial variations of the charge, that is:

∆ρ X = Yρ · δX = a dρ ≈ ˆ 0, (225)
Z
δQ[ρ] = ρδ ? F ≈
6ˆ 0. (226)
∂Σ

Since δ?F — and thereby δQ[ρ] — vanishes for Neumann boundary conditions, this is only possible
for Dirichlet or Robin boundary conditions (a 6= 0), and requires ρ to be constant on Γ. Up
to
R multiplication, there is therefore a unique charge, which records variations of the total flux
∂Σ ?F .

Meta-symmetries. This corresponds to the situation in which Yρ fails to leave the boundary condi-
tion invariant (i.e. when adρ ≈6 ˆ 0). In this case, the transformation changes leaf (or subregion
theory) SXM in the foliation (120) of the global space of solutions S. This can only happen for
0
a 6= 0, that is for Dirichlet or Robin boundary conditions. In that case, meta-symmetries
include any vector field Yρ with ρ not constant on Γ.

Boundary gauge symmetries. Finally, one can have both ∆ρ X ≈ ˆ 0 and δQ[ρ] ≈ 0, in which case
the generator Yρ is transmuted into a gauge direction due to the boundary condition. This
requires δ ? F ≈ ˆ 0, hence a Neumann boundary condition, and in particular a = 0. This
is also a sufficient condition, since a = 0 automatically implies ∆ρ X ≈
ˆ 0, so that any generator
Yρ constitutes a boundary gauge symmetry. This case leads to the additional edge constraints
ˆ 0 discussed in equation (145).
C∂ [ρ] ≈

Lastly, let us comment on the status of edge modes and the edge-mode-dressed observables in
the presence of these various types of symmetries and boundary conditions, especially as far as the
reduced phase space is concerned. Let us firstly focus on the case of Dirichlet and Robin boundary
X0
conditions. Note that in this case, for any choice of post-selected subsystem theory SM defined by

57
X = X0 , there will exist both symmetries and meta-symmetries (but not boundary gauge symmetries),
the former acting tangentially to the leaf, while the latter act transversally. In order to construct the
X0 X0
regional reduced phase space PM from SM , one has to factor out the degenerate directions of the
X0
pullback of ΩM to SM . The symmetry transformation — unique up to multiplication by a constant
(see above) — does not constitute a degenerate direction and meta-symmetries are not even tangential.
The symmetry transformation, which ultimately originates in the edge mode, thus corresponds to a
X0
physical direction and thereby a non-trivial transformation on PM ; specifically, since Q[ρ] is conjugate
to the edge mode variable ϕ by equation (205), the symmetry will transform any non-trivially frame-
X0
dressed observable. At least some such frame-dressed observables must therefore survive on PM . By
contrast, the standard gauge generators Xα will constitute degenerate directions of ΩM . This entails
that a physical solution in M ∪ Γ is specified by the gauge-invariant boundary condition X = X0 ,
as well as suitable gauge-invariant initial data on some Cauchy slice Σ (cf. the related discussion in
Sec. 3.2) which must have some non-trivial dependence on the edge mode frame U ; the reduced phase
X0
space PM is thus labeled by the boundary conditions and parametrized by the invariant Cauchy data.
The invariant Cauchy data will thus not only contain regional gauge-invariant data (i.e. only de-
pending on fields in M ), but also non-local observables involving the edge mode frame (which originates
in the complement M̄ ). For example, it may contain observables relating the connection A at some
bulk point σ of Σ to the edge frame at x ∈ ∂Σ, obtained by shooting a Wilson line in from x to σ
and contracting it with the frame U , similarly to the constructions in Sec. 4. Note that such non-local
observables, which ultimately relate bulk properties in M to the edge frame U (and thus M̄ ), cannot
all be reconstructed from purely local gauge-invariant quantities (that is, from the radiative data X, Y
on ∂Σ together with gauge-invariant bulk variables); they thus amount to independent initial data. In
other words, in the case of Dirichlet and Robin boundary conditions, the regional reduced phase space
X0
PM contains frame-dressed observables and therefore more than the purely regional gauge-invariant
data one would find when ignoring edge modes in the first place. In this case, edge modes are therefore
of direct physical significance, besides also being a prerequisite for deriving the boundary value problem
from the global variational problem through post-selection.
The situation is, however, quite different for Neumann boundary conditions, in which case all
X0
symmetry generators Yρ constitute degenerate directions of the pullback of ΩM to SM , and correspond
to additional boundary gauge symmetries encoded in the additional edge constraints C∂ [ρ]. Since the
edge constraints are essentially what would be the charges in the case of symmetries, they too are
conjugate to the edge mode variable ϕ by equation (205). However, in this case they correspond to
X0
gauge directions in SM that are factored out when constructing the regional reduced phase space
X0
PM . The latter will thus not contain any frame-dressed observables, but exclusively the regional
gauge-invariant data one would find when ignoring edge modes. Intuitively, this can be understood
from the fact that Neumann boundary conditions correspond to reflective boundary conditions, so that
the subregions M and M̄ evolve essentially independently. Note, however, that even in this case, the
edge mode is crucial in order to construct the foliation of the space of solutions in equation (120) and
thereby to consistently formulate the boundary value problem through post-selection. Through its
X0
appearance in the boundary action `, it also affects the reduced dynamics on PM .

7.3 Abelian Chern-Simons theory

The edge mode construction of [3] was first extended to Chern-Simons theories in [5] (see also [11]). As
we will now illustrate in our post-selection formalism, an interesting feature of Abelian Chern-Simons
theory is that it can support an infinite-dimensional set of boundary symmetries, even after appropriate
boundary conditions have been imposed on Γ. This is a crucial structural difference between Maxwell
(or more generally Yang-Mills) and Chern-Simons edge modes, but it is important to note that our
general construction and its interpretation apply equally well to both types of theories. In this regard,
our formalism differs from that advocated in [6–9] (see in particular [9] for a nice summary of this
series of works): by relying on specific features of Yang-Mills theories (for instance, the functional
form of the Gauss constraint), this work is not directly exportable to Chern-Simons theories. We also

58
note that Chern-Simons symmetry algebras were previously studied in the context of boundaries, e.g.
in [110, 111], though in a different formalism and without an explicit edge mode construction.
The bulk action SM = M L[A] of Abelian Chern-Simons theory, a topological field theory in d = 3
R

spacetime dimensions, features the Lagrangian

L=A∧F , (227)
where F = dA is the field strength of the U(1) connection 1-form A, which under gauge transformations
transforms as usual: δα A = dα. The variation of the bulk Lagrangian gives
δL = 2δA ∧ F − d(A ∧ δA) , (228)
yielding the bulk equation of motion F = 0 and the presymplectic potential
Θ = δA ∧ A . (229)
Its pullback to Γ can be written as
θ = hδA, ∗AiΓ , (230)
where ∗ is the Hodge dual on Γ. The presymplectic current
ω = −δA ∧ δA (231)
pulls back to Γ in the form
ω = hδA, −δ ∗ AiΓ . (232)
In local coordinates, this means that the conjugate of a component Aµ is (∗A)µ . A has two components
once pulled-back to Γ; imposing boundary conditions on a Lagrangian submanifold amounts to fixing
just one of them.67 In local inertial coordinates (t, x), where x is space-like and t time-like, we could
(0) (0)
for instance impose a boundary condition Ax = Ax for some background field Ax . However, such a
condition is neither invariant nor covariant under local Lorentz transformations. A better option is to
impose a boundary condition on light-cone coordinates A+ = At + Ax or A− = At − Ax . The reason is
that A± only changes by a multiplicative constant under a Lorentz transformation, so that the type of
boundary condition is invariant: two local inertial observers can agree on the fact that one is imposing
(0)
a boundary condition on A± , even if they might disagree about the value of the background field A± .
Furthermore, the boundary condition A± = 0 is fully invariant. Therefore, boundary conditions of the
(0)
form A± = A± fall in the same class as those already investigated for Maxwell theory. We shall focus
exclusively on such boundary conditions.
In coordinate-independent language, we can simply define:
A± := A ± ∗A (∗A± = ±A± ) (233)
as the (anti)self-dual part of A. Indeed, in the coordinates (t, x) we used above, one can check that
∗dt = dx, ∗dx = dt, and therefore:
A± = (At ± Ax )(dt ± dx) = A± (dt ± dx) . (234)
The frame-independent boundary condition A± = A±
at the form level is thus equivalent to fixing the
0
component A± in a local inertial frame. We can decompose θ and ω as:68
1
1
δA+ ∧ A− + δA− ∧ A+ = −hδA+ , A− i + hδA− , A+ i Γ , (236)


θ=
4 4
1 1
ω = − δA+ ∧ δA− = hδA+ , δA− iΓ . (237)
2 2
67
While we will ultimately be imposing boundary conditions on the components of the edge-frame-dressed connection
Arad , we will here briefly explore the frame-(in)dependence of boundary conditions for the gauge-dependent A, as the
discussion applies in the same manner to Arad , but is notation wise simpler.
68
We make use of the following useful fact:
1 1 1
δA± ∧ A± = hδA± , ±A± iΓ = ± δhA± , A± iΓ = δ(A± ∧ A± ) = δA± ∧ A± + A± ∧ δA± = 0 .


(235)
2 2 2

59
 On Γ we shall once more invoke the split of the connection into a gauge-invariant radiative part
and a pure gauge part (see equation (65))

A Γ
= Arad + dϕ . (238)

As in the Maxwell case, ϕ is the edge mode, which is part of the global theory space from the start
and clearly has to transform as δα ϕ = α (see the discussion in Secs. 4.1 and 4.2). This permits us to
rewrite the pullback of the presymplectic potential to Γ in the form

θ = θrad + θgauge , (239)

θrad = δArad ∧ Arad , (240)
θgauge = δdϕ ∧ Arad + δArad ∧ dϕ + δdϕ ∧ dϕ . (241)

In order to implement the post-selection procedure of Sec. 5.2, we now have to impose boundary
−
conditions on the radiative data. It is clear that one can make ωrad = − 12 δA+
rad ∧ δArad to vanish by
imposing
δA+ rad ≈ 0
ˆ or δA−
rad ≈ 0 ,
ˆ (242)
where
A± ±
rad = A − (dϕ ± ∗dϕ) . (243)
To determine the correction term `corr that will need to be included in order for the shifted presymplectic
potential θcorr to reduce to a corner term on-shell (see equation (156)) across Γ, we firstly rewrite the
radiative part as:
1 1
θrad = δA∓ ∧ A± + δ A± ∧ A∓ (244)


rad rad rad rad .
2 4
The second term will have to be cancelled by a contribution − 41 A± ∓
rad ∧ Arad to `corr , for otherwise

equation (156) cannot be realized.

Next, let us turn to θgauge . As in Maxwell theory, this contains the edge mode contributions
that will allow us to impose the gauge constraint F = 0 dynamically on Γ. However, the situation
is slightly less trivial here as it requires a change of polarization. In Maxwell theory, we readily
had θgauge = d ? F δϕ − d(δϕ ? F ); that is, the variation of the reference frame times a constraint,
supplemented with a corner term. This is not so in Chern-Simons theory, as

θgauge = −2F δϕ + d(δϕArad − ϕδArad + δϕdϕ) + δ(ϕF ) . (245)

The first term includes the constraint F ≈ 0, the second is a corner contribution, and the third — the
difference to the Maxwell case — an exact field space form resulting from the change of polarization
to radiative and edge mode degrees of freedom. Given that this last term vanishes on-shell, defining
θ∂ = −δϕArad + ϕδArad − δϕdϕ and `∂ = 0 would be in compliance with the general algorithm
summarized in Sec. 6.3. However, recalling the discussion at the end of Secs. 6.1 and 6.2.1, we will find
it advantageous to cancel the exact form δ(ϕF ) even off-shell, by means of a non-vanishing `∂ := −ϕF ,
which in turn will contribute to `corr . All in all, we are thus led to defining:
1
`corr := −ϕF − A∓ ∧ A±
rad . (246)
4 rad
Recalling item (ii) from Sec. 6.1, the term −ϕF can be equivalently motivated by the desire to
restore gauge-invariance of the subregion variational problem. Indeed, an important feature of Chern-
Simons theory as compared to Maxwell theory is that SM fails to be gauge-invariant on its own. More
precisely, for fictitious boundaries SM is kinematically not gauge-invariant, but on-shell it is because
Z
δα SM = αF . (247)
Γ∪Σ2 ∪Σ1

60
In principle, we thus do not need to add a piece to the action in order to restore gauge-invariance for
a subregion delimited by fictitious boundaries.69 However, it is convenient to have gauge-invariance
built-in at the kinematical level, and therefore require the action to be gauge-invariant even off-shell.
Including a term −ϕF into `corr achieves just that.
Note that the decomposition (245) is not unique as one can shift ϕF in the third term by an exact
form dk and compensate this change by a shift −dδk of the second term. However, this translates into
nothing else but the usual ambiguity in the definition of the boundary Lagrangian ` (which is unique
up to integration by parts).
Altogether, we are led to introducing the post-selection action (see equation (159)):
Z
1
Sps = −ϕF − A∓ ∧ A± ± ±
rad + λ ∧ (Arad − A0 ) (248)
Γ 4 rad
Z
1
= −ϕF ± ∗ Arad ∧ Arad + λ ∧ (Arad ± ∗Arad − A± 0 ), (249)
Γ 2

where A±0 is a background self-dual (resp. antiself-dual) form entering the boundary condition A ≈
± ˆ
±
A0 , and λ is a gauge-invariant 1-form that we can furthermore assume to be antiself-dual (resp. self-
dual). After evaluating the Lagrange multiplier on-shell (λ ≈ 12 A∓
rad ), we obtain the following boundary
Lagrangian (see equation (174)):
1 1 ±
`± = −ϕF + A± ∓ ∓
0 ∧ Arad − Arad ∧ Arad (250)
2 4
1 1
= −ϕF + A± ∧ (Arad ∓ ∗Arad ) ± Arad ∧ ∗Arad . (251)
2 0 2
In terms of the original variables A and ϕ, this is simply:
1 1 1
`± = −ϕdA + A±
0 ∧ (A ∓ ∗A − dϕ ± ∗dϕ) ± A ∧ ∗A ± A ∧ ∗dϕ ± dϕ ∧ ∗dϕ , (252)
2 2 2
from which we deduce that the frame field ϕ inherits a non-trivial kinetic term ± 12 dϕ ∧ ∗dϕ. By
construction, this boundary Lagrangian allows us to impose the constraint F = 0 on Γ together with
± ±
the boundary R ± Arad = A0 as dynamical equations of motions. Indeed, varying the action
condition
SM ∪Γ = M L[A] + Γ `Γ gives
R

Z Z
δSM ∪Γ = 2δA ∧ F + δA ∧ A (253)
M Σ2 −Σ1
Z 
1 ∓ 
+ δArad ∧ (A±rad − A ±
0 ) − 2F δϕ + d(δϕArad − ϕδArad + δϕdϕ) , (254)
Γ 2

which is stationary up to variation on Σ1 and Σ2 provided that: F = 0 in the bulk as well as on Γ,

and A± ±
rad = A0 .
We note that our final expression (252) coincides with the boundary Lagrangian proposed in [13],
up to an irrelevant integration by parts and in the special case of vanishing background field A± 0 . In
this, our derivation justifies and generalizes this previous proposal. It is also closely related to the
boundary action previously introduced in [11]. Translated in our own notations, the authors of this
paper proposed the boundary Lagrangian
1
`˜± = −ϕF + j ∧ Arad ± j ∧ ∗j , (255)
2
in which an extra gauge-invariant and dynamical boundary 1-form j has been introduced. The equation
of motion resulting from varying j is ∗j = ∓Arad , and once coupled to the bulk action, the equation
69
In subsequent work [101], we shall also deal with physical boundaries on which the bulk equation of motion need not
hold in the form F Γ = 0 on account of possible sources, in which case SM will not be gauge-invariant by itself on-shell.

61
of motion resulting from varying A is j = Arad . Arad is therefore (anti)self-dual on-shell, and it turns
out that we recover the same equations of motion for (A, ϕ) as with `± , again provided that A± 0 = 0.
±
In this sense, our proposal similarly extends the results of [11] to non-vanishing A0 (at least at the
classical level, which is the main focus of the present work). In order to accommodate non-vanishing
boundary conditions, the action (255) can be augmented to:
Z
± 1 1
˜
`Γ = −ϕF + j ∧ Arad + A± 0 ∧ (Arad ∓ ∗Arad ) ± j ∧ ∗j , (256)
Γ 2 2

which remains linear in Arad . By varying A we now obtain j = Arad − A±

0 , while the variation of j still
± ±
yields ∗j = ∓Arad . It follows that Arad = A0 , as claimed.
Let us conclude this subsection by disentangling gauge transformations from boundary symmetries
and their charges. Computing ωgauge = δθgauge , we find:

ωgauge = −δdϕ ∧ δdϕ − 2δdϕ ∧ δArad ≈ −d (δϕδdϕ + 2δϕδArad ) . (257)

The (on-shell) consistency relation (134), dω∂ ≈ˆ −ωgauge , can be solved by postulating a boundary
presymplectic current
ω∂ := δϕδdϕ + 2δϕδArad = 2δϕδA − δϕδdϕ . (258)
Vector fields Xα act on ΩM = Σ ω + ∂Σ ω∂ as
R R

Z Z
Xα · ΩM = −2 dα ∧ δA + (2αδA − 2δϕdα − αδdϕ + δϕdα) (259)
Z Σ ∂Σ Z
= (−2d(αδA) + 2αδF ) + (2αδA − d(αδϕ)) (260)
ΣZ ∂Σ

= 2 αδF ≈ 0 , (261)
Σ

and are therefore gauge symmetries. Defining the constraints

Z
C[α] = 2 αF (262)
Σ

for field-independent α, we find that they form an Abelian Poisson algebra

{C[α], C[β]} := Xβ · Xα · ΩM = 0 . (263)

Next, we can introduce vector fields Yρ that only act on the edge reference frame as frame reorien-
tations (see Sec. 4.5 for the corresponding general discussion):

∆ρ A := Yρ (A) = 0 and ∆ρ ϕ := Yρ (ϕ) = −ρ , (264)

so that, in particular, the frame-dressed observables vary according to ∆ρ Arad := Yρ (Arad ) = dρ and
∆ρ A± ±
rad := Yρ (Arad ) = dρ ± ∗dρ. We then have Yρ · ω = 0, while

Yρ · ω∂ = −2ρδA + ρδdϕ − δϕdρ = −2ρδArad − d(ρδϕ) . (265)

For a field-independent ρ it follows that:

Z
Yρ · ΩM = δQ[ρ] , where Q[ρ] := −2 ρArad , (266)
∂Σ

and the charges Q obey a Kač-Moody algebra:

Z
{Q[ρ], Q[σ]} := Yσ · Yρ · ΩM = −2 ρdσ . (267)
∂Σ

62
 We would find trivial charges (δQ[ρ] ≈ 0) if one were to naively impose a boundary condition on
the full radiative connection Arad . By contrast, with the weaker (anti)-selfdual boundary condition
A± ± ∓
rad = A0 , Arad is free to fluctuate, so that we can have non-trivial charges. More precisely, the
regional covariant phase space is stable under Yρ provided that A±
rad is left invariant. Such vector fields
are labelled by functions ρ on Γ, such that

dρ ± ∗dρ = 0 . (268)

In the local coordinates (t, x), this is nothing but ∂t ρ ± ∂x ρ = 0, with general solution of the form
ρ = ρ(t ∓ x). Crucially, and in sharp contrast with Maxwell theory, we therefore obtain an infinite-
X0
dimensional space of symmetries.70 This entails a non-trivial regional reduced phase space PM with
propagating edge-mode-dressed degrees of freedom. For example, as in the case of Dirichlet and Robin
boundary conditions in Maxwell theory, these may include observables relating A in the bulk of M
to the edge frame via Wilson lines. Owing to the topological nature of the theory, such quasi-local
observables would have been absent from the regional reduced phase space had one ignored edge modes
in the first place.
Given that δA± rad = 0, we can define the charges by
Z
±
Q [ρ] = − ρA∓
rad , (269)
∂Σ

so that Q[ρ] = Q+ [ρ]+Q− [ρ], where we have added an explicit superscript ± to distinguish our two sets
of boundary conditions. On solutions to the boundary conditions A± ±
rad = A0 , the charges Q [ρ] are
∓

no longer dynamical and thereby drop out of the Poisson-bracket relations in Eq. (267). The surviving
charges Q± [ρ] satisfy the Kač-Moody commutation relations. Indeed, owing to the (anti)-selfduality
condition (268):
Z Z
± ±
{Q [ρ], Q [σ]} := Yσ · Yρ · ΩM = − ρ(dσ ∓ ∗dσ) = −2 ρdσ . (270)
∂Σ ∂Σ

Finally, the vector fields Yρ parametrized by functions ρ which fail to satisfy the stability condition (268)
are to be interpreted as meta-symmetries: owing to (266), they do generate symplectomorphisms, but
between distinct regional covariant phase spaces. Note that, as a result, boundary gauge symmetries
are excluded in this example.

7.4 Yang-Mills theory

As our last example, and in complement to Sec. 7.2, let us consider Yang-Mills theory for a general
(compact and connected) matrix Lie group G. This will serve as a non-Abelian illustration of the
post-selection and edge frame formalism.
Yang-Mills theory is also one of the two main models originally studied in the pioneering work of
Donnelly and Freidel on edge modes [3]. Where comparable, our technical results will be consistent
with theirs. At the same time, our work extends their construction and analysis. For example, 1) by
establishing the interpretation of edge modes as internalized external reference frames (see Sec. 4), we
are clarifying their overall physical meaning and that they are already included in the global theory
to start with; 2) we are focusing on a spacetime region with a non-trivial time-like boundary, rather
than a spatial Cauchy slice, which provides an alternative starting point for the construction of the
regional presymplectic structure ΩM (see Sec. 5.3); 3) we demonstrate how to obtain the regional
phase space and variational principle from the global ones via post-selection on edge-mode-dressed
boundary conditions and clarify the status of edge modes in the context of different types of boundary
conditions. In this regard, our work also complements the discussion in [6–9], where the fate of edge
70
This difference holds more generally in comparison to Yang-Mills theories. As we will see below, an additional
specificity of non-Abelian Yang-Mills theories is that the analogue of (268) explicitly depends on the background fields
defining the boundary condition. As a result, the dimension of the solution space itself depends on this background data.

63
modes in Yang-Mills theory in causal diamonds subject to Neumann boundary conditions was studied
in a different formalism and we comment on that below.
The Lagrangian is defined as:
1
L = − Tr [F ∧ ?F ] (271)
2
where F := dA+A∧A is the curvature two-form of a G-connection A. We also recall that the covariant
exterior derivative dA acts on a Lie-algebra valued form η as dA η := dη + [A, η].71 The variation of
the curvature two-form can be nicely expressed as δF = dA δA, from which we deduce:

δL = −Tr [dA δA ∧ ?F ] = −Tr [dA (δA ∧ ?F )] − Tr [δA ∧ dA ? F ] = −dTr [δA ∧ ?F ] − Tr [δA ∧ dA ? F ] .

(272)
The bulk equations of motion and the presymplectic potential are therefore

dA ? F ≈ 0 and Θ = −Tr [δA ∧ ?F ] , (273)

from which we also read out the presymplectic current ω = Tr [δA ∧ δ ? F ].

The radiative data relative to the reference frame U on Γ is (cf. equations (65) and (66)):

Arad = U −1 AU + U −1 dU and (?F )rad = U −1 ? F U . (274)

To determine the radiative presymplectic potential, we first remark that

U δ(U −1 dU )U −1 = d(δU U −1 ) , (275)

U δ(U −1 AU )U −1 = δA + A, δU U −1 , (276)
 

and therefore
U δArad U −1 = δA + dA (δU U −1 ) . (277)
It follows that

θrad := −Tr [δArad ∧ (?F )rad ] = −Tr [δA ∧ ?F ] − Tr dA (δU U −1 ) ∧ ?F , (278)

 

or, in other words,

θgauge = Tr dA (δU U −1 ) ∧ ?F = dTr δU U −1 ? F − Tr δU U −1 dA ? F . (279)

     

On-shell, the second term vanishes and θgauge is therefore exact. We see in particular, as anticipated
in Sec. 4.7, that the (field-space) right-invariant Maurer-Cartan form δU U −1 is the non-commutative
analogue of the variation δϕ that was entering the gauge presymplectic structure of Maxwell and
Abelian Chern-Simons theories. The main difference is that δU U −1 is not field-space exact when G
is not Abelian, which leads to extra contributions to the gauge part of the presymplectic current (as
compared to formula (204) in Maxwell theory):

ωgauge = δθgauge = dTr (δU U −1 )(δU U −1 ) ? F − dTr δU U −1 δ ? F

   

− Tr (δU U −1 )(δU U −1 )dA ? F + Tr δU U −1 δdA ? F . (280)

   

When the equations of motion are satisfied on the boundary, dA ? F Γ

≈ 0, ωgauge becomes exact:

ωgauge ≈ dTr (δU U −1 )(δU U −1 ) ? F − dTr δU U −1 δ ? F . (281)

   

Our consistency condition (134) can therefore be solved by the following choice of boundary presym-
plectic current:
ω∂ = Tr δU U −1 δ ? F − Tr (δU U −1 )(δU U −1 ) ? F . (282)
   

71
Here, [·, ·] is the Lie bracket on Lie-algebra valued forms, which can be defined in terms of the wedge product as
[η1 , η2 ] = η1 ∧ η2 − (−1)pq η2 ∧ η1 , where η1 is a p-form and η2 is a q-form.

64
 Let us explicitly check that the regional presymplectic structure ΩM = Σ ω + ∂Σ ω∂ is invariant
R R

under field-dependent gauge transformations, as argued around equation (140). Remember that the
vector field Xα acts as Xα (A) = [α, A]−dα = −dA α and Xα (F ) = [α, F ], where α is Lie-algebra-valued
and possibly field-dependent. We can first compute
Xα · ω = −Tr [dA α ∧ δ ? F ] + Tr [δA ∧ [?F, α]] (283)
= −dTr [αδ ? F ] + Tr [αdA δ ? F ] + Tr [δA ∧ [?F, α]]
= −dTr [αδ ? F ] + Tr [αδdA ? F ] − Tr [α[δA, ?F ]] + Tr [δA ∧ [?F, α]]
= −dTr [αδ ? F ] + Tr [αδdA ? F ] ≈ −dTr [αδ ? F ] ,
where we have used Leibniz’s rule for dA in going from the first to the second line, and the commutation
relation [δ, dA ]η = [δA, η] (for any Lie-algebra-valued form η) in going from the second to the third.
Owing to the constraint δdA ? F ≈ 0, Xα · ω is exact on-shell and reduces to a boundary term upon
integration, as expected in a local gauge theory. Given that Xα · δU U −1 = α, we also have
Xα · ω∂ = Tr [αδ ? F ] + Tr δU U −1 [?F, α] − Tr [α, δU U −1 ] ? F = Tr [αδ ? F ] . (284)
   

As anticipated, we conclude that:

Z
Xα · ΩM = Tr [αδdA ? F ] ≈ 0 , (285)
Σ

meaning that the vector field Xα is a gauge direction, even for a field-dependent α. Specifically, invoking
Cartan’s magic formula, this entails invariance of the regional presymplectic structure, LXα ΩM ≈ 0.
In [3], such an invariance under field-dependent gauge transformations was the primary postulate that
led to the introduction of the edge field U , together with the boundary presymplectic structure ω∂ .
From a conceptual point of view, our construction proceeds differently (see Sec. 5.3); in particular,
invariance under field-dependent gauge transformation is an output, not an input of our procedure,
but we recover the same results. For a field-independent α, we can furthermore define the constraint
generators Z
C[α] := Tr [αdA ? F ] , δC[α] = Xα · ΩM . (286)
Σ
They obey the following Poisson algebra:72
Z Z
{C[α], C[β]} := Xβ · Xα · ΩM = Tr [α[β, dA ? F ]] = Tr [[α, β]dA ? F ] = C[[α, β]] . (287)
Σ Σ

Hence, the constraints are first-class and define a homomorphism from g to the constraint Poisson
algebra.
On the other hand, the edge reference frame reorientations U → U g −1 (see Sec. 4.5) are generated
by vector fields Yρ , defined as:
∆ρ A := Yρ (A) = 0 and ∆ρ U U −1 := Yρ · δU U −1 = −U ρU −1 . (288)
They act on the frame-dressed observables Arad and ?Frad like gauge transformations intrinsic to Γ, in
the sense that:73
∆ρ Arad := Yρ (Arad ) = [ρ, Arad ] − dρ = −dArad ρ , ∆ρ (?F )rad := Yρ (?Frad ) = [ρ, ?Frad ] . (289)
Note that these expressions are examples of the frame-reorientation induced relational observable
transformations in equations (91) and (93), however, in their infinitesimal form. We then find that
Yρ · ω = 0, while
Yρ · ω∂ = −Tr[U ρU −1 δ ? F ] + Tr[[U ρU −1 , δU U −1 ] ? F ] = −Tr[ρδ ? Frad ] (290)
Here, it is useful to remember that g . (dA ? F ) = gdA ? F g −1 .
72
73
Although being gauge-invariant, we can view Arad as a vector potential for fields on Γ, which justifies the slight
abuse of notation dArad η := dη + [Arad , η], for any Lie algebra valued form on Γ.

65
Restricting to field-independent ρ we can define the charges
Z
Q[ρ] := − Tr [ρ ? Frad ] , δQ[ρ] = Yρ · ΩM , (291)
∂Σ

which verify the algebra

Z
{Q[ρ], Q[σ]} := Yσ · Yρ · ΩM = − Tr [[ρ, σ] ? Frad ] = Q[[ρ, σ]] . (292)
∂Σ

It is also clear that the charges are gauge-invariant and Poisson-commute with the constraints

{Q[ρ], C[α]} = Xα · Yρ · ΩM = 0 . (293)

Even though the charges Q and the constraints C obey similar-looking algebras, it is important
to realize that they are not isomorphic. Gauge transformations act in the whole spacetime region M ,
and are consequently labelled by Lie-algebra-valued fields (α, β,...) on M . By contrast, the charges
act solely on the edge field U , and are therefore labelled by Lie-algebra-valued fields (ρ, σ,...) on Γ.
Let us now turn to the construction of boundary Lagrangians, following the algorithm put forward
in Sec. 6. As before, we only consider boundary conditions which are covariant under local changes
of inertial frames and linear in the canonical local Darboux coordinates of ωrad . Similarly to (209),
the resulting generalized Robin boundary conditions can be parametrized by the following change of
polarization:

X := aArad + (−1)d−1 b ∗ ?Frad , Y := cArad + (−1)d−1 d ∗ ?Frad , (294)

where the SL(2, R) condition ad − bc = 1 ensures that

ωrad = Tr [δArad ∧ δ ? Frad ] = Tr [δX ∧ ∗δY ] . (295)

In order to impose the boundary condition X = X0 , one should first reduce the on-shell presym-
plectic potential on Γ to a corner term as in equation (156). This is the role of the correction term
`corr , which by construction must verify (cf. equations (169) and (211)):

δ`corr ≈ −Tr [δX ∧ ∗Y ] − θrad − θgauge − dC (296)

≈ −Tr [δX ∧ ∗Y ] + Tr [δArad ∧ ?Frad ] − d C + Tr δU U −1 ? F (297)
 

1
= δ (Tr [Arad ∧ ?Frad ] − Tr [X ∧ ∗Y ]) − d C + Tr δU U −1 ? F (298)
 

2
This can be satisfied by defining
1
C := −Tr δU U −1 ? F . (299)
 
`corr := (Tr [Arad ∧ ?Frad ] − Tr [X ∧ ∗Y ]) and
2
The second step in our algorithm consists in a dynamical imposition of the boundary condition X = X0 ,
by means of the post-selection action (159):
Z
Sps = `corr + Tr [λ ∧ (X − X0 )] . (300)
Γ

After solving for λ, we obtain the following boundary Lagrangian for Robin boundary conditions:
1
` = −Tr [X0 ∧ ∗Y ] + (Tr [Arad ∧ ?Frad ] + Tr [X ∧ ∗Y ]) . (301)
2
This is a straightforward generalization of the boundary Lagrangian (215) of Maxwell theory, the
important difference lying in the fact that ?F is not gauge invariant in the non-Abelian case and
therefore needs to be approppriately dressed. As in the Maxwell case, variation of this Lagrangian

66
yields the boundary conditions as boundary equations of motion, as well as the bulk equations of
motions (incl. on Γ).
We can impose a Dirichlet boundary condition Arad = A0 by specializing to a = d = 1 and
b = c = 0, which leads to a vanishing generating function `corr = 0 and (cf. equation (217) in Maxwell
theory)

` = Tr [(Arad − A0 ) ∧ ?Frad ] (302)

= Tr (A − U A0 U −1 ) ∧ ?F + Tr dU U −1 ∧ ?F . (303)
   

Choosing a = d = 0 and b = −c = 1 instead, yields the Neumann boundary condition ?Frad =

f0 := ∗X0 (cf. equation (218) in Maxwell theory):
1 h i
` = Tr [X0 ∧ ∗Arad ] + Tr [Arad ∧ ?Frad ] + Tr (−1)d ∗ ?Frad ∧ ∗Arad (304)
2
= Tr [Arad ∧ f0 ] = Tr U −1 AU ∧ f0 + Tr U −1 dU ∧ f0 . (305)
  

We conclude by analyzing the physical status of the frame reorientations (symmetries) Yρ and
edge modes, which varies depending on the type of boundary condition being imposed, similar to the
Maxwell case at the end of Sec. 7.2. Ignoring mixed types of boundary conditions for simplicity, we
can distinguish two cases.
Suppose first that a = 0 everywhere on Γ. The boundary condition is then of the Neumann type
and reduces to ?Frad = f0 for some background function f0 . Owing to equation (289), the vector field
Yρ leaves the boundary condition invariant if and only if [ρ, f0 ] = 0. However, by virtue of (291), any
such vector field corresponds to a gauge direction of ΩM . Indeed, in analogy to the edge constraints
(145) of Maxwell theory, we now obtain the first-class edge constraints
Z
C∂ [ρ] := − Tr [ρ (?Frad − f0 )] (306)
∂Σ

in addition to the standard bulk constraints in equation (286). The condition [ρ, f0 ] = 0 ensures that
X0
the Poisson brackets of such constraints does vanish on SM :
Z Z Z
{C∂ [ρ], C∂ [σ]} = − Tr[[ρ, σ] ? Frad ] = Tr[σ[ρ, ?Frad ]] ≈
ˆ Tr[σ[ρ, f0 ]] = 0 , (307)
∂Σ ∂Σ ∂Σ

that is, they constitute a first-class algebra. By contrast, any C∂ [ρ] with [ρ, f0 ] 6= 0 is second-class,74
and therefore generates a frame reorientation Yρ that does not preserve the boundary conditions. Such
X0
transformations constitute meta-symmetries that induce a change of subregion theory SM within the
foliation (120) of the global solution space S. This is in contrast to the case of Neumann boundary
conditions in Maxwell theory, where, owing to the Abelian character of the group, no such meta-
symmetry exists. As in the Maxwell case, however, Neumann boundary conditions are not compatible
X0
with proper symmetries of the regional solution space SM . We conclude that edge modes are somewhat
redundant with such boundary conditions: even though they play a conceptually enlightening role at
intermediate steps of our construction and are crucial in the formulation of the boundary value problem
through post-selection, they do not manifest themselves at the level of the regional physical phase
X0
space PM . Indeed, as emphasized in the Maxwell case, Neumann boundary conditions correspond to
reflecting boundary conditions, in which case M and its complement M̄ give rise to largely independent
regional dynamics. Incidentally, alternative constructions without edge modes have been proposed in
the literature for this particular class of boundary conditions (see again [6–9], and in particular [9] for
a nice summary).
74
This can be observed by computing its Poisson bracket with other constraints C∂ [σ] with [σ, f0 ] 6= 0.

67
 Suppose instead that a 6= 0 everywhere on Γ, so that we are dealing either with Dirichlet or
Robin boundary conditions. We can then choose a = 1 without any loss of generality. In order for
the vector field Yρ to leave the boundary condition X = X0 invariant, the function ρ must verify:

∆ρ X = [ρ, Arad ] − dρ + (−1)d−1 b ∗ [ρ, ?Frad ] = −dX0 ρ = 0 . (308)

This generalizes the condition dρ = 0 we already came across for Maxwell theory with boundary
condition on Arad (see Sec. 5.3 and the end of Sec. 7.2). In view of (291), δQ[ρ] does not generally vanish
since ?Frad is free to fluctuate. We conclude that the charges Q[ρ] such that dX0 ρ = 0 generate a non-
trivial algebra of symmetries, with commutation relations (292). However, equation (308) being a first
order partial differential equation for ρ, the dimension of this algebra can at most be equal to dim g,75
and for non-Abelian groups, it depends on the choice of background field X0 .76 For fixed Dirichlet
or Robin boundary conditions, the remaining frame reorientations Yρ will violate equation (308) and
thereby correspond to meta-symmetries.
The situation in non-Abelian Yang-Mills theory is therefore very similar to the Abelian Maxwell
theory case, the only difference being that here the dimension of the symmetry algebra can be anything
between zero and dim g, while in Maxwell theory it is always of dimension dim u(1) = 1. When there
are solutions to equation (308) and the algebra is of non-trivial dimension, we can invoke the same
argumentation as at the end of Sec. 7.2 to conclude that edge modes are of direct physical significance in
X0 77
the regional reduced phase space PM . This phase space will contain edge-frame-dressed observables,
including non-local ones that relate bulk properties of M to the edge frame U via Wilson lines. By
contrast, when the symmetry algebra happens to be zero-dimensional, edge modes do not manifest
themselves at the level of the regional reduced phase space; the frame-dressed observables rather
X0
parametrize transversal directions to the regional solution space SM in the foliation of the global
space of solutions S.
Our conclusion that edge modes are physically significant in the presence of Dirichlet or Robin
boundary conditions verifying equation (308) is consistent with the findings of [7, 112]. There, it
was shown without invoking edge modes that an equation analogous to (308) can be understood as
encoding a fundamental ambiguity in the gluing properties of the system. That is, whenever this
equation admits non-trivial solutions, it turns out not to be possible to reconstruct a unique state
for Σ ∪ Σ̄ from the knowledge of the local states characterizing the gauge-invariant data with support
respectively in Σ and Σ̄. The physical symmetries generated by the analog of Q[ρ] in [7, 112] capture
the missing holistic properties of the global system, and once specified, restore unicity of the gluing
procedure. In our construction of edge modes as reference frames, this holistic nature of the symmetry
charges Q[ρ] is explicit from the outset: U is non-locally defined and is precisely introduced to account
for the gauge-invariant observables of the system which have support on both M and M̄ . However, it
is enlightening to note that the authors of [7, 112] arrived at a similar conclusion without introducing
edge modes per se, but instead relied on a careful analysis of the gluing properties of partial Cauchy
slices.
Altogether, this shows that edge modes, in the sense of [3], do not always manifest themselves at
the physical level in Yang-Mills theory; and when they do, they add a finite number of parameters
to an otherwise infinite-dimensional regional reduced phase space. In this regard, Yang-Mills theory
is dramatically different from (Abelian) Chern-Simons theory: in the latter case, frame reorientations
were giving rise to an infinite-dimensional algebra of edge symmetries (270), owing to the fact that the
analog of (308), equation (268), admits infinitely many solutions (irrespectively of the background).
In this respect, our conclusions seem consistent with at least some of the views on Yang-Mills edge
modes laid out in [6–9] (see also [112,113] for a philosophical discussion), and it would be interesting to
compare the two frameworks more closely. Notwithstanding, also in Yang-Mills and Maxwell theory,
75
A more formal justification of this statement would invoke the Cauchy-Lipschitz theorem.
76
For a generic X0 , one actually expects dX0 ρ = 0 to have no solution.
77
In the language of [7], the existence of stabilizers (that is, of transformations Yρ such that (308) holds) leads to a
stratification of the reduced phase space generated by the symmetry charges.

68
edge modes play a conceptually illuminating role and are crucial for consistently formulating the
regional boundary value problem, as well as the foliation of distinct subregion theories. Ultimately,
they provide the proper tools to relate M to its complement M̄ .

8 Conclusion and outlook

In this paper, we have provided a self-contained and detailed analysis of edge modes in gauge field
theory, which we believe clarifies their physical nature. By embedding a spacetime region M with time-
like boundary Γ into a global spacetime M ∪ M̄ , we were able to view the regional dynamics in M as
the result of a post-selection procedure on the global solution space. This also constitutes a connection
between covariant phase space constructions for global spacetimes [56] and bounded subregions as e.g.
in [11, 66–68]. This global perspective allowed us to make precise sense of edge modes as dynamical
reference frames, and thereby elucidate their interpretation. Our main conclusions are the following.
• Edge modes can be systematically and consistently interpreted as dynamical reference frames,
in the same sense in which they have appeared in the recent quantum reference frame literature
[41–54]. Here, they are in general necessary to specify gauge-invariant boundary conditions on
the time-like boundary Γ of a finite region M . They are non-local functionals of the global
configuration fields (defined e.g. via systems of Wilson lines with support in the complementary
region M̄ ), which however materialize themselves as local fields on Γ. See Sec. 4.
• Hence, edge modes are a direct manifestation of the relational character of gauge field degrees of
freedom [102,103]. Indeed, frame-dressed observables on Γ are gauge-invariant functionals which
relate the region M to its complement M̄ , and that can be precisely interpreted as relational
observables [89–95] in a covariant phase space setting (see Sec. 4.2).
• Given a suitable set of gauge-invariant boundary conditions on Γ (they can be of Dirichlet,
Neumann, Robin or mixed type), we have developed a systematic algorithm allowing one to
induce a consistent variational principle for subregion MR fromR the global one on M ∪ M̄ . This
includes the definition of a presymplectic form Ω M = Σ ω + ∂Σ ω∂ , and more generally, of an
action SM ∪Γ = M L + Γ `. The presymplectic structure is made independent from the choice of
R R

partial Cauchy slice Σ thanks to the inclusion of a corner term verifying the on-shell consistency
condition ω∂ + dωgauge ≈ ˆ 0 (see Sec. 5). The boundary Lagrangian form ` can be motivated by
similar arguments at the off-shell level (see Sec. 6).
• While edge modes are always useful concepts in the construction of the variational problem for
subregion M , they may sometimes drop out from the resulting on-shell presymplectic structure.
However, this only happens for very specific choices of boundary conditions (such as a Neu-
mann boundary condition on ?F in Maxwell theory). For generic boundary conditions, the edge
reference frame does contribute to the regional presymplectic structure.
• As was originally emphasized in [3], symmetries need to be carefully distinguished from gauge
transformations. As we have shown, symmetry transformations of edge degrees of freedom can
be interpreted as reference frame reorientations, which makes their physical character all the
more clear. At the level of the subregion M , the inclusion of boundary conditions necessitates
a refinement of this dichotomy. Frame reorientations then split into three further subcategories:
proper boundary symmetries, which leave the regional field space (in particular, the boundary
conditions) invariant and carry non-trivial charges; boundary gauge symmetries, which also leave
the regional field space invariant, but whose charges become trivial on-shell of the boundary
conditions; and finally meta-symmetries, which, by changing the boundary conditions themselves,
give rise to symplectomorphisms between distinct regional field spaces.
• Our formalism is in principle applicable to any gauge field theory, as examplified in Sec. 7 (see also
Sec. 2 for a mechanical model featuring both gauge symmetries and edge modes). In particular, we

69
 do not see any major difference in the physical interpretation of Yang-Mills, Maxwell and Chern-
Simons edge modes: they can all be derived in a systematic way by application of our algorithm.
However, in contrast to Yang-Mills theories (including Maxwell), Chern-Simons theories do have
the peculiarity of supporting an infinite-dimensional algebra of boundary symmetries, even after
suitable boundary conditions have been imposed on Γ (see Sec. 7.3 for an illustration of this
claim in the Abelian context). In that, the dynamical role of edge modes is greatly enhanced in
such theories.

The present article has been conceived as a first in a series of upcoming publications [101]. In
complement to splitting post-selection, we plan to examine the reverse process of gluing in the language
of dynamical reference frames. It will also be important to test our formalism on other interesting
gauge systems. An obvious example to consider, that was left out of the present work for simplicity, is
non-Abelian Chern-Simons theory. Finally, beyond the definition of consistent variational principles,
our main objective is to explore the statistical (and, ultimately, quantum) properties of bounded
subregions in field theory, from the point of view of universal fluctuation relations in non-equilibrium
thermodynamics.
Another and related objective we leave for the future is to investigate the nature of edge modes
in gravitational systems from the reference frame perspective advocated in the present paper. In
the context of full Einstein gravity, it will be particularly interesting to revisit recent proposals such
as [20, 21, 26, 30–32, 68]. Finally, in order to test our construction in the realm of quantum gravity, it
will also be enlightening to investigate symmetry-reduced models which are reasonably well understood
at the quantum level (such as Jackiw-Teitelboim gravity).

Acknowledgements
We thank Laurent Freidel, Marc Geiller, Henrique Gomes, Aldo Riello, and Nicholas Teh for helpful
discussions, and Josh Kirklin for comments on the manuscript. SC is supported by a Radboud Excel-
lence Fellowship from Radboud University in Nijmegen, the Netherlands. PH is grateful for support
through an ‘It from Qubit’ fellowship of the Simons Foundation in the initial stages of this work and
from the Foundational Questions Institute under grant number FQXi-RFP-1801A. He furthermore
thanks Perimeter Institute for support and hospitality during a collaborative visit for the project.
This project was initiated at Perimeter Institute and University College London. Research at Perimeter
Institute is supported, in part, by the Government of Canada through the Department of Innovation,
Science and Economic Development Canada, and by the Province of Ontario through the Ministry of
Colleges and Universities. This work was also supported in part by funding from Okinawa Institute of
Science and Technology Graduate University.

References
[1] B. I. Halperin, Quantized hall conductance, current-carrying edge states, and the existence of
extended states in a two-dimensional disordered potential, Physical Review B 25 (1982) 2185.

[2] X.-G. Wen, Theory of the edge states in fractional quantum hall effects, International journal of
modern physics B 6 (1992) 1711.

[3] W. Donnelly and L. Freidel, Local subsystems in gauge theory and gravity, JHEP 09 (2016) 102
[1601.04744].

[4] W. Donnelly and A. C. Wall, Entanglement entropy of electromagnetic edge modes, Phys. Rev.
Lett. 114 (2015) 111603 [1412.1895].

[5] M. Geiller, Edge modes and corner ambiguities in 3d Chern–Simons theory and gravity, Nucl.
Phys. B 924 (2017) 312 [1703.04748].

70
 [6] H. Gomes, F. Hopfmüller and A. Riello, A unified geometric framework for boundary charges
and dressings: non-Abelian theory and matter, Nucl. Phys. B 941 (2019) 249 [1808.02074].

[7] H. Gomes and A. Riello, The quasilocal degrees of freedom of Yang-Mills theory, SciPost Phys.
10 (2021) 130 [1910.04222].

[8] A. Riello, Symplectic reduction of Yang-Mills theory with boundaries: from superselection
sectors to edge modes, and back, SciPost Phys. 10 (2021) 125 [2010.15894].

[9] A. Riello, Edge modes without edge modes, 2104.10182.

[10] A. Blommaert, T. G. Mertens and H. Verschelde, Edge dynamics from the path integral —
Maxwell and Yang-Mills, JHEP 11 (2018) 080 [1804.07585].

[11] M. Geiller and P. Jai-akson, Extended actions, dynamics of edge modes, and entanglement
entropy, JHEP 09 (2020) 134 [1912.06025].

[12] G. Wong, A note on entanglement edge modes in Chern Simons theory, JHEP 08 (2018) 020
[1706.04666].

[13] P. Mathieu, L. Murray, A. Schenkel and N. J. Teh, Homological perspective on edge modes in
linear Yang-Mills and Chern-Simons theory, Lett. Math. Phys. 110 (2020) 1559 [1907.10651].

[14] P. Mathieu and N. J. Teh, Boundary electromagnetic duality from homological edge modes,
JHEP 21 (2020) 192 [2102.06799].

[15] J. François, Bundle geometry of the connection space, covariant Hamiltonian formalism, the
problem of boundaries in gauge theories, and the dressing field method, JHEP 03 (2021) 225
[2010.01597].

[16] L. Freidel and A. Perez, Quantum gravity at the corner, Universe 4 (2018) 107 [1507.02573].

[17] W. Donnelly and S. B. Giddings, Observables, gravitational dressing, and obstructions to

locality and subsystems, Phys. Rev. D 94 (2016) 104038 [1607.01025].

[18] A. J. Speranza, Local phase space and edge modes for diffeomorphism-invariant theories, JHEP
02 (2018) 021 [1706.05061].

[19] M. Geiller, Lorentz-diffeomorphism edge modes in 3d gravity, JHEP 02 (2018) 029

[1712.05269].

[20] L. Freidel, E. R. Livine and D. Pranzetti, Gravitational edge modes: from Kac–Moody charges
to Poincaré networks, Class. Quant. Grav. 36 (2019) 195014 [1906.07876].

[21] L. Freidel, M. Geiller and D. Pranzetti, Edge modes of gravity. Part I. Corner potentials and
charges, JHEP 11 (2020) 026 [2006.12527].

[22] L. Freidel, M. Geiller and D. Pranzetti, Edge modes of gravity. Part II. Corner metric and
Lorentz charges, JHEP 11 (2020) 027 [2007.03563].

[23] L. Freidel, M. Geiller and D. Pranzetti, Edge modes of gravity. Part III. Corner simplicity
constraints, JHEP 01 (2021) 100 [2007.12635].

[24] M. Dupuis, L. Freidel, F. Girelli, A. Osumanu and J. Rennert, On the origin of the quantum
group symmetry in 3d quantum gravity, 2006.10105.

[25] M. Geiller, E. R. Livine and F. Sartini, Symmetries of the black hole interior and singularity
regularization, SciPost Phys. 10 (2021) 022 [2010.07059].

71
[26] W. Donnelly, L. Freidel, S. F. Moosavian and A. J. Speranza, Gravitational Edge Modes,
Coadjoint Orbits, and Hydrodynamics, 2012.10367.

[27] E. R. Livine, Loop quantum gravity boundary dynamics and gauge theory, Class. Quant. Grav.
38 (2021) 135031 [2101.07565].

[28] L. Freidel, R. Oliveri, D. Pranzetti and S. Speziale, Extended corner symmetry, charge bracket
and Einstein’s equations, 2104.12881.

[29] W. Wieland, Fock representation of gravitational boundary modes and the discreteness of the
area spectrum, Annales Henri Poincare 18 (2017) 3695 [1706.00479].

[30] W. Wieland, New boundary variables for classical and quantum gravity on a null surface, Class.
Quant. Grav. 34 (2017) 215008 [1704.07391].

[31] W. Wieland, Null infinity as an open Hamiltonian system, JHEP 04 (2021) 095 [2012.01889].

[32] W. Wieland, Gravitational SL(2, R) algebra on the light cone, JHEP 07 (2021) 057
[2104.05803].

[33] T. Takayanagi and K. Tamaoka, Gravity Edges Modes and Hayward Term, JHEP 02 (2020)
167 [1912.01636].

[34] E. De Paoli and S. Speziale, A gauge-invariant symplectic potential for tetrad general relativity,
JHEP 07 (2018) 040 [1804.09685].

[35] M. Geiller and C. Goeller, Dual diffeomorphisms and finite distance asymptotic symmetries in
3d gravity, 2012.05263.

[36] L. Freidel, C. Goeller and E. R. Livine, The Quantum Gravity Disk: Discrete Current Algebra,
2103.13171.

[37] J. Kirklin, The Holographic Dual of the Entanglement Wedge Symplectic Form, JHEP 01
(2020) 071 [1910.00457].

[38] W. Donnelly and G. Wong, Entanglement branes in a two-dimensional string theory, JHEP 09
(2017) 097 [1610.01719].

[39] W. Donnelly, Y. Jiang, M. Kim and G. Wong, Entanglement entropy and edge modes in
topological string theory: I, 2010.15737.

[40] Y. Jiang, M. Kim and G. Wong, Entanglement entropy and edge modes in topological string
theory II: The dual gauge theory story, 2012.13397.

[41] F. Giacomini, E. Castro-Ruiz and v. Brukner, Quantum mechanics and the covariance of
physical laws in quantum reference frames, Nature Commun. 10 (2019) 494 [1712.07207].

[42] A. Vanrietvelde, P. A. Höhn, F. Giacomini and E. Castro-Ruiz, A change of perspective:

switching quantum reference frames via a perspective-neutral framework, Quantum 4 (2020) 225
[1809.00556].

[43] A. Vanrietvelde, P. A. Höhn and F. Giacomini, Switching quantum reference frames in the
N-body problem and the absence of global relational perspectives, 1809.05093.

[44] P. A. Höhn, Switching Internal Times and a New Perspective on the ‘Wave Function of the
Universe’, Universe 5 (2019) 116 [1811.00611].

[45] P. A. Höhn and A. Vanrietvelde, How to switch between relational quantum clocks, New J.
Phys. 22 (2020) 123048 [1810.04153].

72
[46] P. A. Höhn, A. R. H. Smith and M. P. E. Lock, Trinity of relational quantum dynamics, Phys.
Rev. D 104 (2021) 066001 [1912.00033].

[47] P. A. Höhn, A. R. H. Smith and M. P. E. Lock, Equivalence of Approaches to Relational

Quantum Dynamics in Relativistic Settings, Front. in Phys. 9 (2021) 181 [2007.00580].

[48] M. Krumm, P. A. Höhn and M. P. Müller, Quantum reference frame transformations as

symmetries and the paradox of the third particle, Quantum 5 (2021) 530 [2011.01951].

[49] P. A. Höhn, M. Krumm and M. P. Müller, Internal quantum reference frames for finite Abelian
groups, 2107.07545.

[50] P. A. Höhn, M. P. E. Lock, S. A. Ahmad, A. R. H. Smith and T. D. Galley, Quantum Relativity

of Subsystems, 2103.01232.

[51] A.-C. de la Hamette and T. D. Galley, Quantum reference frames for general symmetry groups,
Quantum 4 (2020) 367 [2004.14292].

[52] E. Castro-Ruiz, F. Giacomini, A. Belenchia and v. Brukner, Quantum clocks and the temporal
localisability of events in the presence of gravitating quantum systems, Nature Commun. 11
(2020) 2672 [1908.10165].

[53] F. Giacomini and v. Brukner, Einstein’s Equivalence principle for superpositions of

gravitational fields, 2012.13754.

[54] A. Ballesteros, F. Giacomini and G. Gubitosi, The group structure of dynamical

transformations between quantum reference frames, Quantum 5 (2021) 470 [2012.15769].

[55] C. Crnkovic and E. Witten, Covariant description of canonical formalism in geometrical

theories., in Three Hundred Years of Gravitation, pp. 676–684, (1987).

[56] J. Lee and R. M. Wald, Local symmetries and constraints, J. Math. Phys. 31 (1990) 725.

[57] A. Ashtekar, L. Bombelli and O. Reula, The covariant phase space of asymptotically flat
gravitational fields, in Mechanics, analysis and geometry: 200 years after Lagrange,
pp. 417–450, Elsevier, (1991).

[58] M. Henneaux and C. Teitelboim, Quantization of Gauge Systems. Princeton University Press,
Princeton, 1992.

[59] R. M. Wald, Black hole entropy is the Noether charge, Phys. Rev. D 48 (1993) R3427
[gr-qc/9307038].

[60] T. Jacobson, G. Kang and R. C. Myers, On black hole entropy, Phys. Rev. D 49 (1994) 6587
[gr-qc/9312023].

[61] R. M. Wald and A. Zoupas, A General definition of ’conserved quantities’ in general relativity
and other theories of gravity, Phys. Rev. D 61 (2000) 084027 [gr-qc/9911095].

[62] V. Iyer and R. M. Wald, Some properties of Noether charge and a proposal for dynamical black
hole entropy, Phys. Rev. D 50 (1994) 846 [gr-qc/9403028].

[63] V. Iyer and R. M. Wald, A Comparison of Noether charge and Euclidean methods for
computing the entropy of stationary black holes, Phys. Rev. D 52 (1995) 4430 [gr-qc/9503052].

[64] I. Khavkine, Covariant phase space, constraints, gauge and the Peierls formula, Int. J. Mod.
Phys. A 29 (2014) 1430009 [1402.1282].

73
[65] F. Gieres, Covariant canonical formulations of classical field theories, 2109.07330.

[66] D. Harlow and J.-Q. Wu, Covariant phase space with boundaries, JHEP 10 (2020) 146
[1906.08616].

[67] J. Margalef-Bentabol and E. J. S. Villaseñor, Geometric formulation of the Covariant Phase

Space methods with boundaries, Phys. Rev. D 103 (2021) 025011 [2008.01842].

[68] V. Chandrasekaran and A. J. Speranza, Anomalies in gravitational charge algebras of null

boundaries and black hole entropy, JHEP 01 (2021) 137 [2009.10739].

[69] G. Compere and D. Marolf, Setting the boundary free in AdS/CFT, Class. Quant. Grav. 25
(2008) 195014 [0805.1902].

[70] T. Faulkner, M. Guica, T. Hartman, R. C. Myers and M. Van Raamsdonk, Gravitation from
Entanglement in Holographic CFTs, JHEP 03 (2014) 051 [1312.7856].

[71] T. Faulkner, F. M. Haehl, E. Hijano, O. Parrikar, C. Rabideau and M. Van Raamsdonk,

Nonlinear Gravity from Entanglement in Conformal Field Theories, JHEP 08 (2017) 057
[1705.03026].

[72] D. L. Jafferis, A. Lewkowycz, J. Maldacena and S. J. Suh, Relative entropy equals bulk relative
entropy, JHEP 06 (2016) 004 [1512.06431].

[73] A. Belin, A. Lewkowycz and G. Sárosi, Complexity and the bulk volume, a new York time story,
JHEP 03 (2019) 044 [1811.03097].

[74] A. Belin, A. Lewkowycz and G. Sárosi, The boundary dual of the bulk symplectic form, Phys.
Lett. B 789 (2019) 71 [1806.10144].

[75] J. Kirklin, Unambiguous Phase Spaces for Subregions, JHEP 03 (2019) 116 [1901.09857].

[76] J. Kirklin, Islands and Uhlmann phase: Explicit recovery of classical information from
evaporating black holes, 2011.07086.

[77] J. F. Pedraza, A. Russo, A. Svesko and Z. Weller-Davies, Sewing spacetime with Lorentzian
threads: complexity and the emergence of time in quantum gravity, 2106.12585.

[78] J. F. Pedraza, A. Russo, A. Svesko and Z. Weller-Davies, Lorentzian threads as ’gatelines’ and
holographic complexity, 2105.12735.

[79] D. Harlow and J.-q. Wu, Algebra of diffeomorphism-invariant observables in Jackiw-Teitelboim

Gravity, 2108.04841.

[80] G. Compère and A. Fiorucci, Advanced Lectures on General Relativity, 1801.07064.

[81] G. Barnich and F. Brandt, Covariant theory of asymptotic symmetries, conservation laws and
central charges, Nucl. Phys. B 633 (2002) 3 [hep-th/0111246].

[82] G. Barnich and G. Compere, Surface charge algebra in gauge theories and thermodynamic
integrability, J. Math. Phys. 49 (2008) 042901 [0708.2378].

[83] G. Barnich and C. Troessaert, Symmetries of asymptotically flat 4 dimensional spacetimes at

null infinity revisited, Phys. Rev. Lett. 105 (2010) 111103 [0909.2617].

[84] G. Barnich and C. Troessaert, BMS charge algebra, JHEP 12 (2011) 105 [1106.0213].

[85] G. Compere and F. Dehouck, Relaxing the Parity Conditions of Asymptotically Flat Gravity,
Class. Quant. Grav. 28 (2011) 245016 [1106.4045].

74
 [86] M. Geiller, C. Goeller and N. Merino, Most general theory of 3d gravity: Covariant phase space,
dual diffeomorphisms, and more, JHEP 02 (2021) 120 [2011.09873].

[87] L. Freidel, R. Oliveri, D. Pranzetti and S. Speziale, The Weyl BMS group and Einstein’s
equations, JHEP 07 (2021) 170 [2104.05793].

[88] C. Rovelli, What is observable in classical and quantum gravity?, Class. Quantum Grav. 8
(1991) 297.

[89] C. Rovelli, Quantum reference systems, Class. Quantum Grav. 8 (1991) 317.

[90] C. Rovelli, Partial observables, Phys. Rev. D 65 (2002) 124013 [gr-qc/0110035].

[91] C. Rovelli, Quantum Gravity. Cambridge University Press, Cambridge, 2004.

[92] C. Rovelli, Time in quantum gravity: Physics beyond the Schrödinger regime, Phys. Rev. D 43
(1991) 442.

[93] C. Rovelli, Quantum mechanics without time: a model, Phys. Rev. D 42 (1990) 2638.

[94] B. Dittrich, Partial and complete observables for Hamiltonian constrained systems, Gen. Rel.
Grav. 39 (2007) 1891 [gr-qc/0411013].

[95] B. Dittrich, Partial and complete observables for canonical general relativity, Class. Quant.
Grav. 23 (2006) 6155 [gr-qc/0507106].

[96] D. Marolf, Unitarity and Holography in Gravitational Physics, Phys. Rev. D 79 (2009) 044010
[0808.2842].

[97] F. Giacomini and v. Brukner, Quantum superposition of spacetimes obeys Einstein’s

Equivalence Principle, 2109.01405.

[98] L. Chataignier, Construction of quantum Dirac observables and the emergence of WKB time,
Phys. Rev. D 101 (2020) 086001 [1910.02998].

[99] R. M. Angelo, N. Brunner, S. Popescu, A. J. Short and P. Skrzypczyk, Physics within a

quantum reference frame, Journal of Physics A: Mathematical and Theoretical 44 (2011)
145304.

[100] B. Dittrich and J. Tambornino, Gauge invariant perturbations around symmetry reduced sectors
of general relativity: Applications to cosmology, Class. Quant. Grav. 24 (2007) 4543
[gr-qc/0702093].

[101] S. Carrozza and P. A. Höhn, untitled, to appear .

[102] C. Rovelli, Why Gauge?, Found. Phys. 44 (2014) 91 [1308.5599].

[103] C. Rovelli, Gauge Is More Than Mathematical Redundancy, Fundam. Theor. Phys. 199 (2020)
107 [2009.10362].

[104] S. D. Bartlett, T. Rudolph and R. W. Spekkens, Reference frames, superselection rules, and
quantum information, Rev. Mod. Phys. 79 (2007) 555.

[105] L. Chataignier, P. A. Höhn and M. P. E. Lock, Relational dynamics with periodic clocks, to
appear .

[106] H. Gomes and A. Riello, The observer’s ghost: notes on a field space connection, JHEP 05
(2017) 017 [1608.08226].

75
[107] H. Gomes and A. Riello, Unified geometric framework for boundary charges and particle
dressings, Phys. Rev. D 98 (2018) 025013 [1804.01919].

[108] H. Casini, M. Huerta and J. A. Rosabal, Remarks on entanglement entropy for gauge fields,
Phys. Rev. D 89 (2014) 085012 [1312.1183].

[109] S. Carlip, Statistical mechanics and black hole entropy, gr-qc/9509024.

[110] S. Carlip, The Statistical mechanics of the (2+1)-dimensional black hole, Phys. Rev. D 51
(1995) 632 [gr-qc/9409052].

[111] M.-I. Park, Symmetry algebras in Chern-Simons theories with boundary: Canonical approach,
Nucl. Phys. B 544 (1999) 377 [hep-th/9811033].

[112] H. Gomes, Gauging the Boundary in Field-space, Stud. Hist. Phil. Sci. B 67 (2019) 89
[1902.09258].

[113] H. Gomes, Holism as the empirical significance of symmetries, Eur. J. Phil. Sci. 11 (2021) 87
[1910.05330].

76