Basic Adjoint Relationship (BAR)

• BAR is a concept that links forward operations to their corresponding duals or boundary corrections, underpinning stationary identities and adjoint rules across mathematics.
• In stochastic processes, BAR characterizes stationary distributions through generator-boundary couplings and has been applied to reflected diffusions and queueing networks.
• In operator theory and algorithmic differentiation, BAR formalizes transpose and integration-by-parts identities, enabling reverse-mode differentiation and rigorous adjoint analysis.

Searching arXiv for the cited BAR-related papers to ground the article in the current literature. The expression Basic Adjoint Relationship (BAR) is used in several technically distinct ways across contemporary mathematics, probability, operator theory, numerical analysis, and higher algebra. In stochastic-process and queueing theory, BAR denotes a stationary weak-form identity that couples an interior generator to boundary terms and serves as a characterization tool for stationary distributions (Braverman et al., 2015). In operator-theoretic and algorithmic-differentiation settings, BAR denotes an adjoint or transpose relation that moves derivatives or linear maps to a dual pairing, typically in the form of reversed adjoints or transposed Jacobians (Gustafson et al., 2013). In recent work on matrix functions and neural ordinary differential equations, BAR is a reverse-mode differentiation principle derived from resolvent calculus or operator-adjoint identities (Goloubentsev et al., 2021). In higher algebra, the same phrase is used for the adjoint-pair mechanism underlying derived and classical bar/cobar constructions (Hörmann, 20 Jul 2025). These usages are not interchangeable, but they share a common structural theme: a relation that transfers a forward operation to a dual, backward, or boundary-corrected object.

1. Terminological scope and structural pattern

A recurrent feature of BAR across the cited literature is that it identifies a dual object by requiring compatibility with a bilinear pairing, a stationary identity, or an adjunction. In stochastic models, the pairing is between test functions and stationary or boundary measures; in operator theory and algorithmic differentiation, it is an inner-product identity that defines the adjoint map; in higher algebra, it is a representability statement for a pairing that produces an adjoint functor (Braverman et al., 2015).

The operator-theoretic form is especially explicit in the statement

$(AB)^* \supset B^*A^*,$

for densely defined closable operators with dense product domain (Gustafson et al., 2013). The algorithmic-differentiation form is expressed by

$$\langle \nabla F \cdot v^{(1)},\, y_{(1)} \rangle = \langle v^{(1)},\, \nabla F^* \cdot y_{(1)} \rangle, \qquad \nabla F^*=\nabla F^T,$$

which identifies the adjoint map with the transpose of the Jacobian (Naumann, 2019). In reflected diffusions and queueing networks, BAR instead takes the form of a stationary weak equation such as

$$\int_E Lf\,d\pi+\sum_{i=1}^d\int_{F_i}D_i f\,d\nu_i=0, \qquad f\in C_b^2(E),$$

or its queueing analog

$E[\mathcal A f(X)] = 0$

with additional jump terms or Palm corrections (Lu et al., 3 Jul 2026).

This suggests a family resemblance rather than a single universal definition. The shared content is the transfer of a forward relation—evolution, differentiation, multiplication, or pairing—to a dual object that encodes stationary, reverse, or adjoint information.

2. BAR in stochastic processes: stationary identities for reflected diffusions and queueing networks

In semimartingale reflected Brownian motion (SRBM) on the orthant $E=\mathbb R_+^d$, with drift $\mu$, covariance $\Sigma$, and reflection matrix $R=(R_1,\dots,R_d)$, the interior generator is

$Lf=\mu\cdot\nabla f+Q:D^2f,\qquad Q=\Sigma/2,$

and the oblique boundary derivative on face $F_i=\{x:x_i=0\}$ is

$$\langle \nabla F \cdot v^{(1)},\, y_{(1)} \rangle = \langle v^{(1)},\, \nabla F^* \cdot y_{(1)} \rangle, \qquad \nabla F^*=\nabla F^T,$$0

The BAR is the weak stationarity identity

$$\langle \nabla F \cdot v^{(1)},\, y_{(1)} \rangle = \langle v^{(1)},\, \nabla F^* \cdot y_{(1)} \rangle, \qquad \nabla F^*=\nabla F^T,$$1

where $$\langle \nabla F \cdot v^{(1)},\, y_{(1)} \rangle = \langle v^{(1)},\, \nabla F^* \cdot y_{(1)} \rangle, \qquad \nabla F^*=\nabla F^T,$$2 is an interior measure and $$\langle \nabla F \cdot v^{(1)},\, y_{(1)} \rangle = \langle v^{(1)},\, \nabla F^* \cdot y_{(1)} \rangle, \qquad \nabla F^*=\nabla F^T,$$3 are boundary occupation measures (Lu et al., 3 Jul 2026). For the true stationary regime, this identity follows from Itô’s formula, and the converse question is whether the BAR determines the stationary distribution.

For generalized Jackson networks, the BAR is an exact stationary identity for the Markov process

$$\langle \nabla F \cdot v^{(1)},\, y_{(1)} \rangle = \langle v^{(1)},\, \nabla F^* \cdot y_{(1)} \rangle, \qquad \nabla F^*=\nabla F^T,$$4

with a drift operator

$$\langle \nabla F \cdot v^{(1)},\, y_{(1)} \rangle = \langle v^{(1)},\, \nabla F^* \cdot y_{(1)} \rangle, \qquad \nabla F^*=\nabla F^T,$$5

and a jump sum over arrival and service-completion events (Braverman et al., 2015). A key step is to choose exponential test functions so that the jump term disappears in expectation, reducing the BAR to a tractable derivative identity. After diffusion scaling, the stationary moment generating functions asymptotically satisfy the SRBM BAR

$$\langle \nabla F \cdot v^{(1)},\, y_{(1)} \rangle = \langle v^{(1)},\, \nabla F^* \cdot y_{(1)} \rangle, \qquad \nabla F^*=\nabla F^T,$$6

with

$$\langle \nabla F \cdot v^{(1)},\, y_{(1)} \rangle = \langle v^{(1)},\, \nabla F^* \cdot y_{(1)} \rangle, \qquad \nabla F^*=\nabla F^T,$$7

(Braverman et al., 2015).

The multiclass extension under static-buffer-priority (SBP) disciplines retains the same philosophy—derive the heavy-traffic limit directly from a stationary equation—but the state is a piecewise deterministic Markov process, and the BAR must incorporate Palm expectations of jump increments: $$\langle \nabla F \cdot v^{(1)},\, y_{(1)} \rangle = \langle v^{(1)},\, \nabla F^* \cdot y_{(1)} \rangle, \qquad \nabla F^*=\nabla F^T,$$8 The use of Palm distributions resolves a queue-length truncation difficulty that appears to be unavoidable in the multiclass setting (Braverman et al., 2023). Under stability, state space collapse, and a tight-matrix condition, the limiting stationary law is the stationary distribution of an SRBM with effective reflection matrix

$$\langle \nabla F \cdot v^{(1)},\, y_{(1)} \rangle = \langle v^{(1)},\, \nabla F^* \cdot y_{(1)} \rangle, \qquad \nabla F^*=\nabla F^T,$$9

(Braverman et al., 2023).

Within this probabilistic tradition, BAR is therefore both a characterization principle and a proof method. It replaces limit-interchange arguments by direct stationary analysis, and it converts asymptotic steady-state questions into transform equations and boundary identities (Braverman et al., 2015).

3. Signed BAR uniqueness and the Harrison–Reiman class

A central recent development is the signed-measure uniqueness problem for the BAR of multidimensional reflected diffusions. In the signed setting one allows finite signed measures $$\int_E Lf\,d\pi+\sum_{i=1}^d\int_{F_i}D_i f\,d\nu_i=0, \qquad f\in C_b^2(E),$$0, leading to the linear identity

$$\int_E Lf\,d\pi+\sum_{i=1}^d\int_{F_i}D_i f\,d\nu_i=0, \qquad f\in C_b^2(E),$$1

The question is whether every signed BAR tuple is a scalar multiple of the stationary one (Lu et al., 3 Jul 2026).

For stable Harrison–Reiman data with a nonsingular $$\int_E Lf\,d\pi+\sum_{i=1}^d\int_{F_i}D_i f\,d\nu_i=0, \qquad f\in C_b^2(E),$$2-matrix reflection matrix, the answer is affirmative. Under the assumptions

$$\int_E Lf\,d\pi+\sum_{i=1}^d\int_{F_i}D_i f\,d\nu_i=0, \qquad f\in C_b^2(E),$$3

every finite signed BAR tuple is a scalar multiple of the stationary BAR tuple: $$\int_E Lf\,d\pi+\sum_{i=1}^d\int_{F_i}D_i f\,d\nu_i=0, \qquad f\in C_b^2(E),$$4 Equivalently, the vector space of finite signed BAR tuples is one-dimensional (Lu et al., 3 Jul 2026).

The proof strategy proceeds through the resolvent identity

$$\int_E Lf\,d\pi+\sum_{i=1}^d\int_{F_i}D_i f\,d\nu_i=0, \qquad f\in C_b^2(E),$$5

where

$$\int_E Lf\,d\pi+\sum_{i=1}^d\int_{F_i}D_i f\,d\nu_i=0, \qquad f\in C_b^2(E),$$6

The technical obstacle is that $$\int_E Lf\,d\pi+\sum_{i=1}^d\int_{F_i}D_i f\,d\nu_i=0, \qquad f\in C_b^2(E),$$7 need not be $$\int_E Lf\,d\pi+\sum_{i=1}^d\int_{F_i}D_i f\,d\nu_i=0, \qquad f\in C_b^2(E),$$8 up to corners. The solution combines pathwise differentiability of the reflected diffusion, feasible directional differentiability of the probabilistic resolvent, tangent projections

$$\int_E Lf\,d\pi+\sum_{i=1}^d\int_{F_i}D_i f\,d\nu_i=0, \qquad f\in C_b^2(E),$$9

with

$E[\mathcal A f(X)] = 0$0

and a one-sided mollification

$E[\mathcal A f(X)] = 0$1

which remains strictly inside the orthant (Lu et al., 3 Jul 2026).

The same paper shows that the nonsingular $E[\mathcal A f(X)] = 0$2-matrix assumption is structural rather than technical. In the larger completely-$E[\mathcal A f(X)] = 0$3 class, a singular proper principal block $E[\mathcal A f(X)] = 0$4 permits lower-dimensional boundary gauges, and under exponential ergodicity together with a one-step regulator bound, these produce nonzero zero-mass signed BAR tuples: $E[\mathcal A f(X)] = 0$5 The zero-mass interior BAR coordinates contain an infinite-dimensional subspace (Lu et al., 3 Jul 2026). This yields a sharp dichotomy: finite signed uniqueness holds in the stable Harrison–Reiman $E[\mathcal A f(X)] = 0$6-matrix class and fails in a natural completely-$E[\mathcal A f(X)] = 0$7 extension.

4. BAR in operator theory and adjoint linear algebra

In unbounded-operator theory, BAR refers to the relation between the adjoint of a product and the product of the adjoints. For densely defined closable operators $E[\mathcal A f(X)] = 0$8 and $E[\mathcal A f(X)] = 0$9 with dense $E=\mathbb R_+^d$0, the basic inclusion is

$E=\mathbb R_+^d$1

This inclusion may be strict, and much of the theory concerns conditions under which equality holds: $E=\mathbb R_+^d$2 The analysis is tied to closures of products, dense domains for $E=\mathbb R_+^d$3, and closedness conditions on $E=\mathbb R_+^d$4, $E=\mathbb R_+^d$5, and $E=\mathbb R_+^d$6 (Gustafson et al., 2013).

One clean sufficient criterion is: if $E=\mathbb R_+^d$7 and $E=\mathbb R_+^d$8 are densely defined, $E=\mathbb R_+^d$9 is closed, and $\mu$0, then

$\mu$1

In particular, this holds when $\mu$2 is unitary (Gustafson et al., 2013). Another sufficient theorem states that if $\mu$3 are densely defined, $\mu$4 is closed, and $\mu$5, then

$\mu$6

These formulas are used to sharpen criteria for self-adjointness and normality of products and to clarify operator-product questions for Dirac operators (Gustafson et al., 2013).

In the linear-algebraic and algorithmic-differentiation setting, BAR is the inner-product identity that defines adjoints by transposition: $\mu$7 Applied to BLAS-level primitives, this yields the standard reverse-mode formulas. For the matrix-vector product

$\mu$8

the tangent relation

$\mu$9

induces the adjoint propagation

$\Sigma$0

For the matrix-matrix product

$\Sigma$1

one obtains

$\Sigma$2

For linear systems $\Sigma$3, the adjoint sensitivity of the solve is another transpose solve (Naumann, 2019).

These two literatures differ in emphasis. The unbounded-operator literature studies domain, closure, and self-adjointness subtleties, whereas the AD literature treats BAR as a constructive reverse-mode rule. The common feature is the reversal of operator order under adjunction.

5. BAR in reverse-mode differentiation, matrix functions, and dynamical systems

For generic matrix functions

$\Sigma$4

with $\Sigma$5 square and $\Sigma$6 holomorphic near every eigenvalue of $\Sigma$7, BAR is a reverse-mode differentiation formula that maps the incoming adjoint $\Sigma$8 to the adjoint $\Sigma$9 directly, without differentiating through a particular factorization component-by-component (Goloubentsev et al., 2021). Using the contour representation

$R=(R_1,\dots,R_d)$0

and differentiating the resolvent, the paper derives

$R=(R_1,\dots,R_d)$1

When $R=(R_1,\dots,R_d)$2 is diagonalizable, this reduces to the divided-difference formula

$R=(R_1,\dots,R_d)$3

where $R=(R_1,\dots,R_d)$4 is the matrix of divided differences of $R=(R_1,\dots,R_d)$5 on the eigenvalues (Goloubentsev et al., 2021). The same template yields closed-form adjoints for the positive-part map $R=(R_1,\dots,R_d)$6, the nearest correlation matrix routine, and a regularized regression construction (Goloubentsev et al., 2021).

In neural ordinary differential equations, BAR denotes the operator-adjoint or integration-by-parts identity relating the forward sensitivity equation

$R=(R_1,\dots,R_d)$7

to the adjoint equation

$R=(R_1,\dots,R_d)$8

For terminal-only loss and time-independent parameter $R=(R_1,\dots,R_d)$9, the gradient is

$Lf=\mu\cdot\nabla f+Q:D^2f,\qquad Q=\Sigma/2,$0

A central claim of the paper is that the loss gradient is not an ODE but an integral, and that the traditional continuous adjoint formulation is not generally equivalent to backpropagation through the actual discretized solver unless the backward discrete scheme uses the same discrete scheme as the forward solver (Hu, 2024).

In hybrid multibody dynamical systems, trajectories and sensitivities are piecewise smooth and discontinuous at events. There the BAR is the bilinear identity

$Lf=\mu\cdot\nabla f+Q:D^2f,\qquad Q=\Sigma/2,$1

which must remain valid across jumps (Corner et al., 2018). If the direct sensitivity jump is

$Lf=\mu\cdot\nabla f+Q:D^2f,\qquad Q=\Sigma/2,$2

then preservation of the pairing

$Lf=\mu\cdot\nabla f+Q:D^2f,\qquad Q=\Sigma/2,$3

forces the adjoint jump condition

$Lf=\mu\cdot\nabla f+Q:D^2f,\qquad Q=\Sigma/2,$4

The paper validates this framework on a five-bar mechanism and reports agreement of direct and adjoint sensitivities to within less than $Lf=\mu\cdot\nabla f+Q:D^2f,\qquad Q=\Sigma/2,$5 in the cost-function sensitivity (Corner et al., 2018).

Across these examples, BAR functions as a reverse-mode principle. It transports sensitivity information through analytic functional calculus, continuous-time dynamics, or event-driven jumps by an adjoint pairing.

6. Bar/cobar adjunctions and categorical BAR

In higher algebra, the phrase Basic Adjoint Relationship is used for the adjoint-pair mechanism behind bar and cobar constructions. The central idea is that a twisted-arrow-type object represents a pairing, and that representability yields an adjunction (Hörmann, 20 Jul 2025).

For an ordinary small category $Lf=\mu\cdot\nabla f+Q:D^2f,\qquad Q=\Sigma/2,$6, the paper recalls that for a complete and cocomplete $Lf=\mu\cdot\nabla f+Q:D^2f,\qquad Q=\Sigma/2,$7-category $Lf=\mu\cdot\nabla f+Q:D^2f,\qquad Q=\Sigma/2,$8 one obtains an adjunction

$Lf=\mu\cdot\nabla f+Q:D^2f,\qquad Q=\Sigma/2,$9

with $F_i=\{x:x_i=0\}$0 right adjoint (Hörmann, 20 Jul 2025). This is the derived bar/cobar pair in the sense of Lurie, and here bar is left adjoint while cobar is right adjoint.

The same work isolates a different adjunction, called the classical bar/cobar adjunction,

$F_i=\{x:x_i=0\}$1

where $F_i=\{x:x_i=0\}$2 is a fully faithful “bar” embedding and $F_i=\{x:x_i=0\}$3 is its left adjoint (Hörmann, 20 Jul 2025). Thus the variance is reversed relative to the derived case: in the classical setting, bar is the right-hand fully faithful functor and cobar is its left adjoint.

The abstract framework employs cofibrations of $F_i=\{x:x_i=0\}$4-operads, Day convolution, and relative operadic Kan extensions. The derived adjunction is expressed as a representability identity

$F_i=\{x:x_i=0\}$5

while the classical one takes the form

$F_i=\{x:x_i=0\}$6

Within this framework, the paper recovers classical comparison maps, including the Szczarba and Hess–Tonks maps, and relates Lurie’s constructions to the classical bar of Eilenberg–MacLane, Kan’s loop group, and Adams cobar (Hörmann, 20 Jul 2025).

This usage of BAR is categorically remote from stochastic and operator-theoretic BAR, but the formal pattern is recognizable: a pairing becomes represented by adjoint functors.

7. Disambiguation and common confusions

The multiplicity of meanings attached to BAR creates several recurrent confusions. The first is terminological: BAR in reflected diffusions and queueing theory is a stationary identity for measures, not an adjoint operator formula. Its variables are typically test functions, generators, and boundary measures, rather than Hilbert-space adjoints or Jacobian transposes (Lu et al., 3 Jul 2026).

A second confusion arises within adjoint-based computation. In neural ODEs, the cited paper argues that the loss gradient is an integral rather than an ODE, and that the continuous adjoint is not generally the adjoint of the discrete solver unless the same discrete scheme is used forward and backward (Hu, 2024). This is a stronger claim than the standard continuous-time formal derivation and should be read in the paper’s specific operator-adjoint sense.

A third confusion concerns the distinction between an adjoint representation and a basic adjoint relationship. In non-perturbative gauge theory, the “adjoint” or “octet” channel of the static $F_i=\{x:x_i=0\}$7 potential refers to the adjoint representation in the decomposition

$F_i=\{x:x_i=0\}$8

and the corresponding potential is extracted by projecting the temporal-gauge kernel onto the adjoint color sector (Rossi et al., 2013). This is not a BAR in the probabilistic, operator-theoretic, or categorical senses.

A plausible implication is that BAR is best treated as a context-dependent technical term rather than a single cross-disciplinary definition. What unifies the usages is a structural idea—transfer to a dual object, often through a pairing—but the actual meanings, hypotheses, and applications are domain-specific.