Correspondence-Driven Hyperfunction in Bilevel Optimization

• The paper introduces correspondence-driven hyperfunction, replacing the classic optimistic hyperfunction with an algorithm-dependent model for bounded-rational followers.
• Gaussian smoothing and bifurcation geometry are used to handle discontinuities and abrupt changes in the follower’s response.
• The SCiNBiO algorithm implements this framework, achieving convergence and oracle-complexity guarantees in nonconvex bilevel optimization.

Searching arXiv for the cited papers and closely related work to ground the article. {"query":"arXiv (Jiang et al., 1 Sep 2025) Correspondence-Driven Approach for Bilevel Decision-making with Nonconvex Lower-Level Problems", "max_results": 5} Correspondence-driven hyperfunction is a bilevel-optimization construct introduced for decision-making with general nonconvex lower-level objectives. It replaces the classical optimistic hyperfunction, which assumes that the follower attains a global minimizer of the lower-level problem, with an algorithm-dependent object determined by what a bounded-rational follower can actually reach by running a fixed lower-level method from a specified initialization and step-size schedule. In this formulation, discontinuity of the leader’s induced objective is treated explicitly through Gaussian smoothing, while abrupt changes in the follower’s response are analyzed through bifurcation geometry; the resulting framework is coupled to the algorithm SCiNBiO and to convergence and oracle-complexity guarantees for the smoothed upper-level problem (Jiang et al., 1 Sep 2025).

1. Classical hyperfunction and the modeling shift

The classical optimistic bilevel hyperfunction is defined by

$$\phi(x):=\min_{y^\ast(x)\in \mathcal{S}(x)} f(x,y^\ast(x)), \qquad \mathcal{S}(x):=\arg\min_{y\in\mathbb{R}^m} g(x,y).$$

This formulation presumes that the follower can solve the lower-level problem globally. For general nonconvex $g(x,\cdot)$, the framework identifies this assumption as both computationally problematic and conceptually problematic: globally computing $\mathcal S(x)$ is typically intractable, and endowing the follower with the power to attain a global minimum makes the follower a “fully rational agent” with unrealistic capabilities (Jiang et al., 1 Sep 2025).

The correspondence-driven reformulation replaces that idealization with an algorithmic response model. The follower is specified by an upper-level decision $x\in\mathcal X$, an initialization $y_0\in\mathcal Y$, a step-size schedule $\boldsymbol{\eta}=\{\eta_t\}$, and a lower-level method $\mathcal M$. The induced upper-level objective is then determined not by the set of global minimizers of $g(x,\cdot)$, but by the subset of algorithmically reachable accumulation points that attain the smallest lower-level objective value among those reachable points.

This shift has direct modeling consequences. The follower is treated as a bounded-rational algorithmic agent, and the leader’s objective becomes explicitly dependent on the lower-level algorithm, initialization, and step sizes. A plausible implication is that the framework changes bilevel semantics from “best global response” to “best reachable response under a prescribed procedure,” which is materially different in the nonconvex regime.

2. Formal definition and follower correspondence

Let $\hat S(x,y_0,\boldsymbol{\eta},\mathcal M)$ denote the subset of accumulation points generated by running $\mathcal M$ on $g(x,\cdot)$0 from $g(x,\cdot)$1 with step sizes $g(x,\cdot)$2 that attain the smallest lower-level objective value. The follower’s selected response is defined by

$g(x,\cdot)$3

and the correspondence-driven hyperfunction is

$g(x,\cdot)$4

The paper also introduces an averaging variant over initializations,

$g(x,\cdot)$5

These definitions make the lower-level response a correspondence generated by algorithmic dynamics rather than by a global argmin operator (Jiang et al., 1 Sep 2025).

Two assumptions organize the analysis. The first is a method assumption: all accumulation points of the generated sequence are stationary points of $g(x,\cdot)$6. This is described as mild and as covering standard methods such as gradient descent. The second is a descent assumption: $g(x,\cdot)$7 Under this condition,

$g(x,\cdot)$8

so the follower’s response remains in a compact lower-level level set. The significance of this point is analytic as well as algorithmic: boundedness of the lower-level trajectory is used to control the induced upper-level objective.

3. Gaussian smoothing and regularized upper-level analysis

Because $g(x,\cdot)$9 is generally discontinuous in $\mathcal S(x)$0, the framework introduces Gaussian smoothing. With

$\mathcal S(x)$1

the smoothed objective is

$\mathcal S(x)$2

with gradient

$\mathcal S(x)$3

Using the Gaussian kernel identity, the gradient can be written as

$\mathcal S(x)$4

This representation is the basis for the stochastic outer-loop estimator used later in SCiNBiO (Jiang et al., 1 Sep 2025).

For every $\mathcal S(x)$5, $\mathcal S(x)$6 is $\mathcal S(x)$7. If $\mathcal S(x)$8 is continuous at $\mathcal S(x)$9, then

$x\in\mathcal X$0

If $x\in\mathcal X$1 is differentiable at $x\in\mathcal X$2, then

$x\in\mathcal X$3

The same asymptotic correspondence is established for the projected-gradient mapping

$x\in\mathcal X$4

The paper also provides quantitative bounds: $x\in\mathcal X$5 Thus the smoothing parameter $x\in\mathcal X$6 controls both regularization and gradient growth. This suggests a familiar trade-off: smaller $x\in\mathcal X$7 makes the smoothed objective closer to the original discontinuous one, while at the same time worsening the smoothness constants that govern the outer-level method.

4. Bifurcation structure and discontinuity of the follower response

The central geometric difficulty is bifurcation in the family $x\in\mathcal X$8. A point $x\in\mathcal X$9 is defined to be a bifurcation point if $y_0\in\mathcal Y$0 is not Morse in $y_0\in\mathcal Y$1, meaning that it has a degenerate stationary point. The bifurcation set is

$y_0\in\mathcal Y$2

equivalently,

$y_0\in\mathcal Y$3

The set $y_0\in\mathcal Y$4 is closed because Morse-ness is an open property (Jiang et al., 1 Sep 2025).

Away from $y_0\in\mathcal Y$5, stationary points are isolated and nondegenerate, and local solution curves behave smoothly by the implicit function theorem. Near $y_0\in\mathcal Y$6, stationary points may disappear, appear, merge, split, or change stability. The framework identifies this as the mechanism by which $y_0\in\mathcal Y$7 can become discontinuous. In practical terms, a small perturbation of $y_0\in\mathcal Y$8 can move the lower-level dynamics from one stable branch to another.

To sharpen this picture, the paper imports bifurcation theory from dynamical systems and defines fold bifurcation points. If $y_0\in\mathcal Y$9 is $\boldsymbol{\eta}=\{\eta_t\}$0, and $\boldsymbol{\eta}=\{\eta_t\}$1 is stationary in $\boldsymbol{\eta}=\{\eta_t\}$2, then $\boldsymbol{\eta}=\{\eta_t\}$3 is a fold bifurcation point if:

1. $\boldsymbol{\eta}=\{\eta_t\}$4 has exactly one zero eigenvalue and all others are nonzero;
2. if $\boldsymbol{\eta}=\{\eta_t\}$5 is the unit eigenvector of that zero eigenvalue, then

$\boldsymbol{\eta}=\{\eta_t\}$6

1. with $\boldsymbol{\eta}=\{\eta_t\}$7 the third derivative tensor,

$\boldsymbol{\eta}=\{\eta_t\}$8

Condition (2) is equivalent to the block Jacobian

$\boldsymbol{\eta}=\{\eta_t\}$9

having full row rank at stationary points.

Under the weaker assumption that $\mathcal M$0 is Morse for almost every $\mathcal M$1, the bifurcation set is measure zero. If $\mathcal M$2 is semi-algebraic, then $\mathcal M$3 is semi-algebraic and admits a stratified manifold decomposition

$\mathcal M$4

with each $\mathcal M$5 diffeomorphic to an open hypercube and each $\mathcal M$6, hence

$\mathcal M$7

This geometric control underlies the neighborhood estimates used in the complexity analysis.

5. SCiNBiO and the coupled inner–outer optimization scheme

The algorithm proposed for the smoothed problem is SCiNBiO, abbreviated from Smooth Correspondence-driven Nonconvex lower-level Bilevel Optimization. It solves

$\mathcal M$8

with projected SGD at the upper level and cubic-regularized Newton at the lower level (Jiang et al., 1 Sep 2025).

For a fixed $\mathcal M$9, the inner solver computes

$g(x,\cdot)$0

and then

$g(x,\cdot)$1

with $g(x,\cdot)$2. It returns the best iterate according to the stationarity metric

$g(x,\cdot)$3

and sets $g(x,\cdot)$4.

At outer iteration $g(x,\cdot)$5, SCiNBiO performs the following operations in sequence. It samples $g(x,\cdot)$6, forms perturbed decisions $g(x,\cdot)$7, assigns $g(x,\cdot)$8 whenever $g(x,\cdot)$9, and otherwise runs $\hat S(x,y_0,\boldsymbol{\eta},\mathcal M)$0 steps of cubic Newton on $\hat S(x,y_0,\boldsymbol{\eta},\mathcal M)$1 from $\hat S(x,y_0,\boldsymbol{\eta},\mathcal M)$2 to obtain $\hat S(x,y_0,\boldsymbol{\eta},\mathcal M)$3, then sets

$\hat S(x,y_0,\boldsymbol{\eta},\mathcal M)$4

The Monte Carlo gradient estimator is

$\hat S(x,y_0,\boldsymbol{\eta},\mathcal M)$5

and the projected update is

$\hat S(x,y_0,\boldsymbol{\eta},\mathcal M)$6

The method is explicitly described as biased stochastic gradient descent, because the quantities $\hat S(x,y_0,\boldsymbol{\eta},\mathcal M)$7 are evaluated at approximate lower-level solutions rather than exact ones. The significance of that bias is not treated as incidental; it is central to the theory, which decomposes the resulting error into lower-level approximation, bifurcation-neighborhood, and sampling components.

6. Convergence, oracle complexity, and conceptual placement

The analysis separates three sources of gradient error: lower-level approximation error $\hat S(x,y_0,\boldsymbol{\eta},\mathcal M)$8, bifurcation-neighborhood error controlled by $\hat S(x,y_0,\boldsymbol{\eta},\mathcal M)$9, and sampling variance controlled by $\mathcal M$0. If the bifurcation set has Minkowski dimension $\mathcal M$1, then the measure of its $\mathcal M$2-neighborhood satisfies

$\mathcal M$3

and this yields

$\mathcal M$4

The Monte Carlo variance obeys

$\mathcal M$5

With $\mathcal M$6 and a random output index $\mathcal M$7 chosen with probability proportional to $\mathcal M$8, the projected-gradient residual satisfies a nonasymptotic bound that yields

$\mathcal M$9

under the schedule

$g(x,\cdot)$00

To reach $g(x,\cdot)$01, the total upper-level function-value oracle complexity is

$g(x,\cdot)$02

Under the additional assumption that all degenerate stationary points are fold bifurcation points, the paper gives

$g(x,\cdot)$03

and states that, in this case, the bifurcation set is actually an $g(x,\cdot)$04-dimensional manifold (Jiang et al., 1 Sep 2025).

Conceptually, the framework places nonconvex bilevel optimization within an explicitly algorithmic and geometric semantics. The paper’s main message is that bilevel nonconvexity should be understood through algorithmic correspondences and bifurcation geometry, not through idealized global minimizers. This suggests that the correspondence-driven hyperfunction is not merely a technical variant of the classical hyperfunction, but a different modeling primitive for the lower-level agent.

The terminology is also specific to this bilevel-optimization setting. The word “hyperfunction” appears in distinct arXiv contexts, including the hyperfunction formulation of many body Green’s functions and the hyperfunction theory of aging (Smit, 2022, Aronoff et al., 11 Apr 2025). This suggests that “correspondence-driven hyperfunction” is a local term of art tied to bilevel response modeling rather than a continuation of those other hyperfunction literatures.