Computational Max-Divergence

• Computational max-divergence is a family of divergence measures that quantifies the maximal KL error in approximating a target distribution by a model, applicable in both classical and quantum contexts.
• It leverages geometric formulations such as information geometry, Voronoi polytopes, and toric models to transform complex divergence maximization into tractable optimization problems.
• Efficient measurement constraints in quantum settings lead to a computational formulation that bounds distinguishability and guides operational performance in state discrimination.

Computational max-divergence denotes a family of extremal divergence constructions rather than a single invariant. In statistical model approximation, it is the maximal information divergence from a model $M\subseteq \Delta(X)$, namely the largest Kullback–Leibler error incurred when approximating an arbitrary target distribution by $M$. In quantum information, closely related notions appear as maximal quantum $f$-divergences, defined through reverse tests and characterized as largest monotone quantum $f$-divergences, and as computational max-divergences induced by families of efficiently implementable binary measurements rather than the full positive semidefinite cone (Montufar et al., 2013, Matsumoto, 2013, Hiai, 2018, Yángüez et al., 25 Sep 2025). This multiplicity of meanings is not merely terminological: each setting fixes a different optimization domain, projection geometry, and operational interpretation.

1. Basic quantities and conceptual scope

For a finite state space $X$, the classical maximum-information-divergence problem starts from a model $M\subseteq \Delta(X)$ and the Kullback–Leibler divergence

$D(p\|q)=\sum_{x\in X} p(x)\log\frac{p(x)}{q(x)}.$

The divergence from a distribution $p$ to the model is

$D(p\|M):=\inf_{q\in M}D(p\|q),$

and the quantity of interest is the worst-case approximation error

$D_M:=\max_{p\in\Delta(X)}D(p\|M).$

A distribution $M$0 achieving the infimum is called the $M$1-projection of $M$2 onto $M$3 (Montufar et al., 2013).

In this setting, $M$4 measures the worst-case model misspecification error: one picks the distribution hardest for the model to approximate, computes the minimum KL divergence to the model, and maximizes over all targets. The same quantity admits a geometric reading. The model is regarded as a subset of the probability simplex, the $M$5-projection is the KL-closest point in the model closure, and the maximal divergence is a global “radius” of the model in information geometry (Montufar et al., 2013).

The phrase “maximal” is used differently in quantum work. There, a maximal quantum $M$6-divergence is defined either as the minimum classical $M$7-divergence over all reverse tests, or by an operator formula when $M$8 is operator convex, and it is maximal in the sense of being the largest quantum $M$9-divergence compatible with data processing and agreement with the classical divergence on commuting states (Matsumoto, 2013). In the computationally constrained quantum setting, the ordering $f$0 is replaced by a cone order induced by efficient binary measurements, so the “max” refers to distinguishability certified by efficient observers (Yángüez et al., 25 Sep 2025). This suggests that “computational max-divergence” is best understood as a class of extremal divergence problems indexed by the admissible model or test family.

2. Exponential families and optimization over $f$1

For an exponential family $f$2 on a finite set $f$3, specified by a strictly positive reference measure $f$4 and a sufficient-statistics matrix $f$5, the $f$6-projection $f$7 of a probability measure $f$8 is characterized by the moment-matching identity

$f$9

It satisfies the Pythagorean identity

$f$0

and therefore

$f$1

where

$f$2

Equivalently, $f$3 maximizes $f$4 among all probability measures satisfying $f$5 (0912.4660).

Local maximizers of $f$6 obey a strong projection property. If $f$7 is a local maximizer with support $f$8 and $f$9 is its $X$0-projection, then

$X$1

Thus the maximizer is obtained by restricting its projection point to its support and renormalizing. There is also a support bound,

$X$2

The paper further shows that $X$3 can be paired with a disjoint-support measure $X$4 satisfying $X$5, sharing the same $X$6-projection, with

$X$7

A useful identity is

$X$8

which couples the divergence values of the two disjoint-support maximizers (0912.4660).

A central computational reformulation replaces optimization over all probability measures by optimization over the kernel of the sufficient-statistics map. Every nonzero $X$9 decomposes uniquely as $M\subseteq \Delta(X)$0 with disjoint supports, and with normalization $M\subseteq \Delta(X)$1, one introduces

$M\subseteq \Delta(X)$2

Maximizing $M\subseteq \Delta(X)$3 over $M\subseteq \Delta(X)$4 is equivalent to maximizing $M\subseteq \Delta(X)$5: global maximizers correspond bijectively, and local maximizers of $M\subseteq \Delta(X)$6 yield local maximizers of $M\subseteq \Delta(X)$7 (0912.4660).

When $M\subseteq \Delta(X)$8 has integer entries, the first-order conditions become algebraic after fixing a sign vector $M\subseteq \Delta(X)$9. The resulting quasi-critical equations exponentiate to binomial equations, placing the problem in the setting of toric and lattice-ideal computation. Two algorithms are proposed. The first enumerates sign vectors of $D(p\|q)=\sum_{x\in X} p(x)\log\frac{p(x)}{q(x)}.$0, computes bases of $D(p\|q)=\sum_{x\in X} p(x)\log\frac{p(x)}{q(x)}.$1, forms binomial systems, and solves them via saturation and primary decomposition; TOPCOM, Singular, solve.lib, and 4ti2 are explicitly noted. The second uses the projection-point parameterization

$D(p\|q)=\sum_{x\in X} p(x)\log\frac{p(x)}{q(x)}.$2

and solves $D(p\|q)=\sum_{x\in X} p(x)\log\frac{p(x)}{q(x)}.$3 for the parameters $D(p\|q)=\sum_{x\in X} p(x)\log\frac{p(x)}{q(x)}.$4. The first method is easier when $D(p\|q)=\sum_{x\in X} p(x)\log\frac{p(x)}{q(x)}.$5 is small, whereas the second is better suited when the dimension of the exponential family is small (0912.4660).

These methods resolve nontrivial examples. For the binary independence model on $D(p\|q)=\sum_{x\in X} p(x)\log\frac{p(x)}{q(x)}.$6, the two global maximizers are

$D(p\|q)=\sum_{x\in X} p(x)\log\frac{p(x)}{q(x)}.$7

For the binary $D(p\|q)=\sum_{x\in X} p(x)\log\frac{p(x)}{q(x)}.$8-$D(p\|q)=\sum_{x\in X} p(x)\log\frac{p(x)}{q(x)}.$9 model, the maximal value is

$p$0

and for the independence model of cardinalities $p$1, the maximal value is

$p$2

with a unique global maximizer up to symmetry (0912.4660).

3. Linear and toric models: Voronoi geometry and chamber-complex algorithms

A later development treats maximum information divergence through logarithmic Voronoi polytopes. For a model $p$3, the divergence from $p$4 to the model is

$p$5

and the maximum information divergence is

$p$6

For fixed $p$7, the map $p$8 is strictly convex on the simplex (Alexandr et al., 2023).

For a $p$9-dimensional linear model $D(p\|M):=\inf_{q\in M}D(p\|q),$0, the principal theorem states that the maximum divergence is always achieved at a vertex of the logarithmic Voronoi polytope $D(p\|M):=\inf_{q\in M}D(p\|q),$1, where $D(p\|M):=\inf_{q\in M}D(p\|q),$2 itself is a vertex of $D(p\|M):=\inf_{q\in M}D(p\|q),$3: $D(p\|M):=\inf_{q\in M}D(p\|q),$4 The proof uses strict convexity on each convex polytope $D(p\|M):=\inf_{q\in M}D(p\|q),$5, together with a co-circuit description of the vertices. If $D(p\|M):=\inf_{q\in M}D(p\|q),$6 is a normalized co-circuit, then the corresponding vertex $D(p\|M):=\inf_{q\in M}D(p\|q),$7 satisfies

$D(p\|M):=\inf_{q\in M}D(p\|q),$8

which is linear in $D(p\|M):=\inf_{q\in M}D(p\|q),$9. Hence the global problem reduces to maximizing a linear function over a polytope (Alexandr et al., 2023).

For toric models, including discrete exponential families, the logarithmic Voronoi polytope at $D_M:=\max_{p\in\Delta(X)}D(p\|M).$0 is the affine fiber

$D_M:=\max_{p\in\Delta(X)}D(p\|M).$1

A local maximizer of $D_M:=\max_{p\in\Delta(X)}D(p\|M).$2 must be a projection point, and every such projection point is a complementary vertex of $D_M:=\max_{p\in\Delta(X)}D(p\|M).$3, where $D_M:=\max_{p\in\Delta(X)}D(p\|M).$4 is the maximum likelihood estimate of the maximizer. The support of any maximizer satisfies

$D_M:=\max_{p\in\Delta(X)}D(p\|M).$5

This sparsity statement aligns with the support bound for exponential-family maximizers obtained by kernel methods (Alexandr et al., 2023).

The combinatorial types of the fibers $D_M:=\max_{p\in\Delta(X)}D(p\|M).$6 are organized by the chamber complex $D_M:=\max_{p\in\Delta(X)}D(p\|M).$7. For $D_M:=\max_{p\in\Delta(X)}D(p\|M).$8 in the relative interior of a chamber, the combinatorial type of $D_M:=\max_{p\in\Delta(X)}D(p\|M).$9 is constant, and the supports of vertices and faces are constant on that chamber. This yields an explicit algorithm: compute equations of the toric variety, compute the chamber complex, choose one representative $M$00 per chamber, identify complementary vertex/face pairs, parametrize the corresponding line families, substitute into the equations of the toric variety, and solve the resulting systems using numerical algebraic geometry, with Bertini and HomotopyContinuation.jl given as examples (Alexandr et al., 2023).

The same framework produces explicit formulas in structured families. For twisted Veronese models $M$01,

$M$02

For box models $M$03,

$M$04

For the conditional independence model $M$05,

$M$06

The paper also emphasizes that reducible models admit decomposition results for Voronoi polytopes and divergence, but that compatibility of maximizers can fail, so naive additive upper bounds are not always attained (Alexandr et al., 2023).

4. Neural-network statistical models and worst-case approximation bounds

The review of maximal information divergence from neural-network statistical models considers the independence model $M$07, naive Bayes mixtures $M$08, restricted Boltzmann machines with $M$09 visible and $M$10 hidden units, deep belief networks, and comparison classes of exponential families such as partition models and hierarchical/log-linear models (Montufar et al., 2013).

The main technical strategy is geometric and model-theoretic. If a complicated model $M$11 contains a tractable submodel $M$12, then

$M$13

This allows upper bounds for hard hidden-variable models to be obtained from simpler exponential submodels. A second ingredient is a decomposition theorem for mixtures supported on disjoint blocks: if $M$14 partitions $M$15 and $M$16, then the $M$17-projection onto the mixture $M$18 decomposes blockwise as

$M$19

This enables exact or near-exact reduction of union models to tractable block components (Montufar et al., 2013).

Several closed-form or nearly closed-form expressions are available. For the partition model $M$20 with coarseness $M$21,

$M$22

Its maximizers have support intersecting each block in at most one point, and exactly one point in a block of maximal size. For an exponential family of dimension $M$23,

$M$24

and if equality holds, then $M$25 must be a homogeneous partition model. For the $M$26-ary independence model,

$M$27

while in general

$M$28

The maximizers in the $M$29-ary case are uniform distributions on $M$30-ary codes of size $M$31 and minimum distance $M$32 (Montufar et al., 2013).

For hidden-variable neural-network models, the same submodel method yields explicit worst-case bounds. For the naive Bayes model $M$33, a general bound is

$M$34

under the condition $M$35. In the binary case,

$M$36

For restricted Boltzmann machines with hidden state-space sizes $M$37, if

$M$38

then

$M$39

and for binary units with $M$40,

$M$41

These formulas subsume the naive Bayes and independence cases as special cases at $M$42 and $M$43 (Montufar et al., 2013).

For deep belief networks with finite-valued units, the paper gives a new result for deep and narrow architectures. If the network has $M$44 layers of width $M$45, unit cardinalities $M$46, and if

$M$47

with $M$48 chosen so that

$M$49

then

$M$50

In the binary case, if the network has

$M$51

layers of size

$M$52

then

$M$53

The paper notes that the binary case follows prior work, whereas the non-binary case is new. Within this approximation-theoretic framework, universal approximation corresponds to vanishing maximal divergence, and deepening the network can improve approximation power, but narrow DBNs may still fail to be universal approximators (Montufar et al., 2013).

5. Maximal quantum $M$54-divergence and reverse-test formulations

In finite dimensions, a reverse test of a pair of positive operators $M$55 is a triple $M$56 where $M$57 is a positive trace-preserving map from classical probability measures to operators, $M$58 are probability distributions on a finite set $M$59, and

$M$60

The maximal quantum $M$61-divergence is defined by the optimization problem

$M$62

Among all quantum divergences $M$63 satisfying data processing under CPTP maps and agreement with the classical $M$64-divergence on commuting states, $M$65 is the largest one: any such $M$66 obeys

$M$67

The same paper gives the dual characterization

$M$68

showing that $M$69 is a pointwise supremum of linear functionals (Matsumoto, 2013).

When $M$70 is proper, lower semicontinuous, operator convex, $M$71, and $M$72, a closed formula is available: $M$73 If $M$74, the correction term vanishes and the expression reduces to

$M$75

The same work establishes joint convexity, positive homogeneity, direct-sum additivity, and lower semicontinuity. It also proves that for operator convex $M$76, the maximum $M$77-divergence of the outcome distributions of a measurement is strictly less than $M$78, while this fails in general without operator convexity; the counterexample $M$79 corresponds to total variation distance (Matsumoto, 2013).

In the von Neumann algebra setting, maximal $M$80-divergence is developed for normal positive functionals using Haagerup’s $M$81-space. If $M$82, then there exists a unique operator $M$83 such that

$M$84

and for operator convex $M$85,

$M$86

For arbitrary $M$87, the definition is extended by approximation. The central reverse-test theorem states

$M$88

and the minimum is attained. The theory further yields a general integral expression, monotonicity under unital positive normal maps, joint convexity, joint lower semicontinuity in the norm topology, and martingale convergence

$M$89

for increasing nets of von Neumann subalgebras (Hiai, 2018).

This operator-algebraic theory also clarifies the meaning of “maximal.” The maximal divergence dominates the standard one, coincides with it in the commuting case, and for $M$90 yields the Belavkin–Staszewski relative entropy. In particular,

$M$91

with equality when $M$92 and $M$93 commute (Hiai, 2018).

6. Efficient measurements, many-copy regularization, and algorithmic computation

A recent quantum-information formulation defines computational max-divergence relative to a family of polynomially generated sets of effect operators

$M$94

where each $M$95 consists of effects implementable by at most $M$96 elementary gates for some polynomial $M$97. The family is required to be closed under complements and tensor products, and generates a cone

$M$98

which is proved to be a proper cone: convex, closed, pointed, and solid. The induced partial order is defined by $M$99 (Yángüez et al., 25 Sep 2025).

Using the general $f$00-max-divergence formalism, the computational max-divergence is defined for $f$01 with $f$02 and $f$03 by

$f$04

The equality with the dual expression follows from standard conic duality for proper cones. Operationally, this quantity measures the strongest bias obtainable in state discrimination when restricted to efficient binary tests. The paper also stresses that the converse to the ordinary positive-semidefinite implication need not hold: $f$05 but not vice versa, so the computational order only captures positivity as seen through efficient measurements (Yángüez et al., 25 Sep 2025).

The same work defines the computational measured Rényi divergence over the set $f$06 of efficient binary POVMs and proves the central equivalence

$f$07

Thus the computational max-divergence is exactly the $f$08 limit of the computational measured Rényi divergence, mirroring the unconstrained identity for unrestricted measurements. The paper also proves super-additivity under a strictly super-additive polynomial bound, approximate sub-additivity under finite preparation cost $f$09, and the lower bound

$f$10

where $f$11 is the computational trace distance (Yángüez et al., 25 Sep 2025).

For many-copy regimes, regularized versions are introduced. The regularized computational max-divergence satisfies

$f$12

with the equality of limit and supremum obtained from super-additivity and Fekete’s lemma. The measured regularizations obey the ordering

$f$13

and a computational Stein bound gives the regularized computational measured relative entropy an operational role as an upper bound on the achievable type-II error exponent under efficient binary measurements (Yángüez et al., 25 Sep 2025).

The same framework induces resource measures. For a free-state set $f$14,

$f$15

The measured resource divergence is monotone under efficient free operations and faithful if $f$16 is closed. An asymptotic continuity bound is proved when $f$17 is closed, convex, bounded, and contains a full-rank state. For entanglement, the computational measured relative entropy of entanglement

$f$18

acts as an entanglement monotone under efficient LOCC, and if the Bell projector $f$19 is efficiently generated, then

$f$20

The paper also constructs explicit separations using pseudoentangled states, showing that computational entanglement quantities can be negligible while the corresponding information-theoretic quantities remain maximal (Yángüez et al., 25 Sep 2025).

Related algorithmic work addresses the direct computation of maximal quantum $f$21-divergence. Using the block-encoding + QSVT paradigm, one can compute

$f$22

for operator-convex $f$23, under purification access or sample access. The method constructs block-encodings of $f$24 and $f$25, forms

$f$26

approximates functions such as $f$27 either by direct QSVT polynomial approximation or by Stieltjes/resolvent representations, and then estimates the trace. For general operator-convex $f$28, the Löwner–Kraus representation

$f$29

reduces computation to resolvent kernels implemented with the same block-encoding and inversion primitives (Dinh et al., 13 Nov 2025).

Across these lines of work, computational max-divergence is a unifying extremal principle with several concrete realizations: worst-case KL approximation error for statistical models, reverse-test maximality for quantum $f$30-divergences, and efficient-observer distinguishability under computational constraints. The common structure is an outer maximization over admissible targets or tests together with an inner projection, comparison, or conic-order minimization; the differences lie in which model class, geometry, and operational restrictions are built into that extremal problem.