Markovian Holonomy Fields in Gauge Theories

• Markovian holonomy fields are random multiplicative assignments of group elements to paths that incorporate gauge symmetry, area-preserving invariance, and spatial Markov properties.
• They underlie the construction and classification of planar Yang–Mills fields, with discrete and continuous formulations reflecting independence over disjoint regions.
• In the free probability framework, these fields converge in large-N limits, yielding free unitary Lévy processes and the generalized master field corresponding to operator-norm continuous free planar fields.

Searching arXiv for relevant papers on Markovian holonomy fields and generalized master fields. Markovian holonomy fields are random multiplicative assignments of group elements to paths, or in the free setting of unitaries to loops in a tracial non-commutative probability space, constrained by gauge symmetry, area-preserving invariance, and a spatial Markov property formulated through independence or freeness over disjoint regions. In the planar theory these axioms lead to a rigid structure: regular classical fields are planar Yang–Mills fields built from conjugation-invariant Lévy processes, while free planar fields admit a complete classification by characteristic triplets and include the two-dimensional Yang–Mills master field as a distinguished operator-norm continuous case (Gabriel, 2015, Cébron et al., 2016).

1. Algebraic and geometric framework

The basic object is a multiplicative function on paths. For a manifold or surface $M$, a parametrized path $p$ is a Lipschitz map $p:[0,1]\to M$ or a constant path, taken up to increasing bi-Lipschitz time-change. The resulting space is denoted $P(M)$. If $p_1$ ends where $p_2$ begins, concatenation is defined by

$$p_1\cdot p_2(t)= \begin{cases} p_1(2t), & t\le \tfrac12,\ p_2(2t-1), & t>\tfrac12, \end{cases}$$

and inversion by $p^{-1}(t)=p(1-t)$. Loops are paths $l$ with $l(0)=l(1)$, based loops at $p$0 form $p$1, and reduced loops form the group $p$2 (Gabriel, 2015).

A $p$3-valued holonomy assignment is a function $p$4 satisfying

$p$5

This reversal of order is the conventional multiplicativity for holonomy. Gauge transformations $p$6 acting on vertices or basepoints act by

$p$7

Accordingly, the gauge-invariant information is encoded by the $p$8-field generated by diagonal-conjugation-invariant functions of loop holonomies (Gabriel, 2015).

A common simplification is to treat the theory as a collection of Wilson loop variables. The formal definition is stronger: the primary random object is a multiplicative map on paths, and loop observables are extracted from that pathwise structure. This is essential for locality, surgery, and extension from graphs to continuous surfaces.

2. Planar axioms and regularity

For $p$9 endowed with an area measure $p:[0,1]\to M$0, a strong planar Markovian holonomy field is a family $p:[0,1]\to M$1 of gauge-invariant probability measures on $p:[0,1]\to M$2 subject to three structural requirements. First, there is invariance under orientation-preserving bi-Lipschitz homeomorphisms preserving area, together with restriction invariance on finite planar graphs. Second, if $p:[0,1]\to M$3 are simple loops with disjoint closed interiors, the invariant $p:[0,1]\to M$4-fields generated by holonomies of paths lying in the corresponding closures are independent. Third, locality holds: if two area measures agree inside a simple loop $p:[0,1]\to M$5, then the induced laws agree on paths contained in $p:[0,1]\to M$6 (Gabriel, 2015).

The weak planar version keeps the same geometric spirit but is formulated only for piecewise-affine paths and uses ordinary independence rather than invariant-$p:[0,1]\to M$7-field independence. Discrete analogues are defined on finite planar graphs $p:[0,1]\to M$8 through measures $p:[0,1]\to M$9 satisfying axioms $P(M)$0–$P(M)$1. Regularity consists of local $P(M)$2-Hölder continuity in the area dependence of loop expectations and continuity under small area-preserving homeomorphisms of graphs (Gabriel, 2015).

In the formulation specialized to $P(M)$3, the set of rectifiable oriented loops in $P(M)$4 based at $P(M)$5, a $P(M)$6-valued planar Markovian holonomy field is a random multiplicative map $P(M)$7 satisfying area-preserving Lipschitz invariance, independence for subfamilies supported in disjoint simple-loop interiors, and gauge invariance under conjugation. When $P(M)$8, stochastic continuity is added: if $P(M)$9 uniformly with lengths converging to the length of $p_1$0, then $p_1$1 in probability (Cébron et al., 2016).

In this literature, “Markovian” is therefore a spatial notion. It is expressed through independence over disjoint interiors and locality under restriction, not through a one-parameter transition kernel alone.

3. Planar Yang–Mills fields and classification

The main construction starts from a $p_1$2-valued Lévy process $p_1$3 with $p_1$4, invariant under conjugation $p_1$5, and convolution semigroup of laws $p_1$6. For a finite planar graph $p_1$7 and an area measure $p_1$8, one chooses a rooted spanning tree $p_1$9 and facial loops $p_2$0 around bounded faces $p_2$1, then declares the reduced facial holonomies $p_2$2 to be independent with law $p_2$3. Gauge invariance extends this prescription to all multiplicative functions on paths of $p_2$4 (Gabriel, 2015).

These discrete fields satisfy the planar graph axioms and regularity, and extension theorems yield a unique stochastically continuous strong planar Markovian holonomy field on $p_2$5. These are the planar Yang–Mills fields associated to the driving Lévy process. If $p_2$6 has density $p_2$7, then on any graph $p_2$8,

$p_2$9

which recovers the usual lattice Yang–Mills weights (Gabriel, 2015).

The classification theorem in the regular planar category is stringent: any regular Planar Markovian Holonomy Field is a planar Yang–Mills field (Gabriel, 2015). The family is then partitioned by symmetry and support properties. If the driving process is invariant under full $$p_1\cdot p_2(t)= \begin{cases} p_1(2t), & t\le \tfrac12,\ p_2(2t-1), & t>\tfrac12, \end{cases}$$0-conjugation, the field is called pure. Writing $$p_1\cdot p_2(t)= \begin{cases} p_1(2t), & t\le \tfrac12,\ p_2(2t-1), & t>\tfrac12, \end{cases}$$1, the field is non-degenerate when $$p_1\cdot p_2(t)= \begin{cases} p_1(2t), & t\le \tfrac12,\ p_2(2t-1), & t>\tfrac12, \end{cases}$$2, otherwise degenerate. For a pure non-degenerate Yang–Mills field, equivalent characterizations include the condition that $$p_1\cdot p_2(t)= \begin{cases} p_1(2t), & t\le \tfrac12,\ p_2(2t-1), & t>\tfrac12, \end{cases}$$3 for any simple loop $$p_1\cdot p_2(t)= \begin{cases} p_1(2t), & t\le \tfrac12,\ p_2(2t-1), & t>\tfrac12, \end{cases}$$4, and Wilson loop expectations converging to Haar. Mixed and degenerate cases are characterized similarly through support or limit laws of the one-parameter family $$p_1\cdot p_2(t)= \begin{cases} p_1(2t), & t\le \tfrac12,\ p_2(2t-1), & t>\tfrac12, \end{cases}$$5 (Gabriel, 2015).

A notable technical ingredient is the use of braid-group symmetry. The braid group $$p_1\cdot p_2(t)= \begin{cases} p_1(2t), & t\le \tfrac12,\ p_2(2t-1), & t>\tfrac12, \end{cases}$$6 acts on $$p_1\cdot p_2(t)= \begin{cases} p_1(2t), & t\le \tfrac12,\ p_2(2t-1), & t>\tfrac12, \end{cases}$$7 by

$$p_1\cdot p_2(t)= \begin{cases} p_1(2t), & t\le \tfrac12,\ p_2(2t-1), & t>\tfrac12, \end{cases}$$8

The corresponding de Finetti theorem states that an infinite sequence is braidable if and only if it is conditionally i.i.d. with respect to its tail field and each marginal law is invariant under conjugation by its support; with an additional diagonal-conjugation independence condition, the mixing measure becomes invariant under full conjugation by $$p_1\cdot p_2(t)= \begin{cases} p_1(2t), & t\le \tfrac12,\ p_2(2t-1), & t>\tfrac12, \end{cases}$$9 (Gabriel, 2015). This provides the probabilistic mechanism behind the facial-loop construction.

4. Surfaces, constraints, and free-boundary reduction to the plane

The non-planar theory is formulated on a measured marked surface with $p^{-1}(t)=p(1-t)$0-constraints, written $p^{-1}(t)=p(1-t)$1, where $p^{-1}(t)=p(1-t)$2 is a finite set of disjoint smooth cycles in the interior and $p^{-1}(t)=p(1-t)$3 assigns a conjugacy class in $p^{-1}(t)=p(1-t)$4 to each boundary or mark cycle. A Markovian holonomy field assigns to such data a finite gauge-invariant measure $p^{-1}(t)=p(1-t)$5 on $p^{-1}(t)=p(1-t)$6 satisfying seven axioms: constraint-forcing, measurability in the boundary data, disintegration or gluing of marks, invariance under orientation-preserving area-preserving homeomorphisms, factorization over disjoint unions, locality under surgery or splitting along a mark, and a normalization condition on the disk (Gabriel, 2015).

Regularity here means stochastic continuity together with a Feller property in the boundary constraint. The surface theory is thus not merely an extension of the planar one by continuity; it is organized by gluing and surgery identities. A plausible implication is that the formalism is designed to behave functorially under topological decomposition, rather than only under restriction to subdomains.

The bridge back to the plane is given by free-boundary expectation. For a finite disk $p^{-1}(t)=p(1-t)$7, one defines $p^{-1}(t)=p(1-t)$8 by integrating the field with boundary constraint $p^{-1}(t)=p(1-t)$9 over $l$0. As $l$1 increases to $l$2, these expectations form a projective family converging to a planar Markovian field $l$3 (Gabriel, 2015).

This reduction is exact for Yang–Mills fields. On any finite planar graph $l$4,

$l$5

Moreover, any regular Markovian holonomy field on surfaces yields, through its free-boundary planar part, a stochastically continuous strong planar field arising from a unique conjugation-invariant Lévy process $l$6, identified by its $l$7-hole partition functions $l$8. Conversely, any admissible $l$9 arises from a unique Yang–Mills field on all surfaces (Gabriel, 2015).

5. Free planar Markovian holonomy fields and generalized master fields

The free analogue replaces classical probability by a tracial non-commutative probability space $l(0)=l(1)$0. A free planar Markovian holonomy field is a multiplicative map

$l(0)=l(1)$1

to the unitaries of $l(0)=l(1)$2, satisfying the analogues of planar invariance and spatial independence, with freeness replacing ordinary independence, together with continuity in non-commutative distribution (Cébron et al., 2016). Explicitly, if $l(0)=l(1)$3 are disjoint simple loops, then the subfamilies supported in $l(0)=l(1)$4 and $l(0)=l(1)$5 are free.

The decisive observation is that if $l(0)=l(1)$6 is a one-parameter family of growing simple loops of area $l(0)=l(1)$7, then

$l(0)=l(1)$8

is a free-unitary Lévy process: $l(0)=l(1)$9 has the law of $p$00, and increments are free. Such processes are completely characterized by a real drift $p$01, a speed $p$02, and a Lévy measure $p$03 on the unit circle $p$04 satisfying

$p$05

Equivalently, if

$p$06

then

$p$07

The non-commutative law of a free planar Markovian holonomy field is therefore in bijection with the triplet $p$08 (Cébron et al., 2016).

Conversely, given such a triplet, one constructs a free unitary Lévy process $p$09, then builds a consistent multiplicative assignment on finite graphs by the “purely braidable stationary process” machinery, and finally extends by continuity to all loops (Cébron et al., 2016). The generalized master fields are thus not arbitrary free loop processes; they are exactly the free planar Markovian holonomy fields classified by these characteristic triplets.

The case $p$10 is distinguished. Then $p$11 is a free unitary Brownian motion with drift $p$12 and speed $p$13, yielding the true two-dimensional Yang–Mills master field up to normalization of area and drift. Among all free planar Markovian holonomy fields, the unique one continuous in operator norm, equivalently in the $p$14-seminorm $p$15, is precisely this $p$16 case (Cébron et al., 2016).

6. Large-$p$17 approximation and the master-field limit

Free planar Markovian holonomy fields arise as large-$p$18 limits of ordinary $p$19-valued fields. Starting from a sequence $p$20 of $p$21-valued Lévy processes that are stochastically continuous and conjugation-invariant, one obtains associated $p$22-valued planar Markovian holonomy fields. For any free planar field $p$23 with characteristic triplet $p$24, there exists for each $p$25 a $p$26-valued planar Markovian holonomy field $p$27 whose non-commutative law converges to that of $p$28 (Cébron et al., 2016).

For each fixed loop $p$29, this is expressed by the trace convergence

$p$30

The approximation first matches the free unitary Lévy process $p$31 with $p$32-valued Lévy processes $p$33 having the same characteristic triplet, obtained by embedding the scalar measure $p$34 into $p$35 through conjugation invariance. The associated planar field $p$36 then satisfies uniform $p$37-estimates, described in the source as Levy’s area-bound or the more general Proposition 6.4, which are sufficient for passage from finite-dimensional convergence to convergence for loop holonomies (Cébron et al., 2016).

The special case $p$38 and $p$39 recovers the classical two-dimensional $p$40 Yang–Mills field driven by $p$41-Brownian motion of speed $p$42. Its large-$p$43 limit is the free unitary Brownian holonomy field with no jumps. In this sense, the true master field—identified in the source with Marino’s/Levy’s master field—is the unique operator-norm continuous free planar Markovian holonomy field, corresponding to free unitary Brownian motion with characteristic triplet $p$44 (Cébron et al., 2016).

7. Related holonomy-based field identities on graphs

A related, though structurally distinct, use of holonomy appears in covariant Gaussian and annealed field theories on finite graphs. Here one considers a finite connected graph $p$45 with a distinguished well $p$46, conductances $p$47, vertex measures $p$48, and killing rates $p$49. A rank-$p$50 real or complex vector bundle over the graph is equipped with a unitary or orthogonal connection $p$51, and the holonomy along a discrete path $p$52 is

$p$53

The covariant Laplacian is $p$54, and with a Hermitian nonnegative potential $p$55 one sets $p$56 (Kassel et al., 2016).

The associated Feynman–Kac formula expresses the semigroup through random-walk holonomy: $p$57 Accordingly, the Green kernel has a path integral representation,

$p$58

and the Gaussian free vector field with covariance $p$59 satisfies

$p$60

(Kassel et al., 2016).

In this setting, the classical isomorphism theorems of Dynkin, Eisenbaum, Le Jan, and Sznitman extend to identities involving holonomy along random paths and loops, and one obtains a covariant Symanzik expansion for moments of annealed non-Gaussian fields. When $p$61 and the connection is trivial, these formulas reduce exactly to the classical scalar identities (Kassel et al., 2016). This broader context suggests that holonomy is not only an observable attached to gauge fields, but also a structural variable linking random paths, Gaussian fields, loop soups, and annealed field theories.