Optimized Composite Geometric Protocol

• Optimized composite geometric protocol is a design strategy that decomposes a task into geometric segments with tunable parameters such as path, timing, and scaling.
• It bridges diverse fields—from nonequilibrium thermodynamics and quantum control to optical communications and convex optimization—by optimizing geometric primitives against domain-specific criteria.
• The approach improves performance and robustness through segmented optimization, though challenges remain in ensuring global optimality and balancing robustness with efficiency.

“Optimized composite geometric protocol” (Editor's term) denotes a family of constructions in which a task is realized by composing several geometrically defined segments, objects, or updates, and then optimizing free parameters—path, timing, labeling, scaling, or step coefficients—against a domain-specific criterion. In current arXiv literature, the phrase is most directly realized by concatenated geodesic strokes in nonequilibrium thermodynamics, composite and optimized geometric loops in quantum control, many-to-one constellation shaping and channel-aware scaling in optical communications, proximal extensions of optimized smooth methods in composite optimization, and finite-geometry announcement or search schemes in cryptography and secure computation (Chennakesavalu et al., 2022, Ding et al., 29 Sep 2025, Yankov et al., 2023, Bok et al., 30 Jun 2025, Cordón-Franco et al., 2013).

1. Cross-domain structure

The phrase is not a single standardized formalism. In some works the geometry is literal state-space or control-space geometry; in others it is the geometry of constellations, enclosing balls, affine subspaces, or wedge decompositions. This suggests a common pattern: a protocol is built from locally geometric primitives, and optimization selects either the path itself, the time allocation along the path, or auxiliary parameters that leave the nominal task unchanged while altering robustness, dissipation, throughput, or complexity.

Domain	Composite geometric form	Principal optimization goal
Nonequilibrium control	Concatenated Wasserstein or thermodynamic-metric geodesic segments (Chennakesavalu et al., 2022)	Minimize dissipation
Quantum control	Composite pulses, seven-segment and multi-loop orange-slice paths, DFS logical trajectories (Ding et al., 29 Sep 2025, Ding et al., 1 May 2026)	Maximize robustness and gate fidelity
Optical communications	Many-to-one geometric constellation shaping and channel-aware constellation scaling (Yankov et al., 2023, Sillekens et al., 2022)	Maximize AIR, GMI, MI, reach
Convex optimization	Geometric proximal-gradient balls and proximal extensions of optimized smooth methods (Chen et al., 2016, Bok et al., 30 Jun 2025)	Minimize objective gap or composite stationarity
Cryptography and secure geometry	Affine $k$-slicings and secure wedge-search decompositions (Cordón-Franco et al., 2013, 0705.4185)	Achieve informativity, $k$-safety, or lower round complexity

A further distinction concerns the status of the word “geometric.” In thermodynamics and geometric quantum control it refers to explicit geodesic structure or geometric phase. In communications it refers to constellation point geometry. In convex optimization it is literal in the geometric proximal-gradient method, but more indirect in PEP-based reduction frameworks and exact optimal composite methods, where the dominant language is proximal extension, fixed-step first-order methods, Lyapunov certificates, or reduction from smooth optimization (Chen et al., 2016, Jang et al., 2023).

2. Nonequilibrium control and thermodynamic geometry

In stochastic thermodynamics, the most explicit formulation of an optimized composite geometric protocol is the finite-time dissipation-minimization problem

$$\lambda_*=\arg\min_{\lambda:[0,t_{\rm f}]\to \mathbb{R}^k} \Delta \Sigma_{\rm tot}[\lambda]$$

with

$$\Delta \Sigma_{\rm tot}=\int_0^{t_{\rm f}} \beta(t)\int_{\Omega} v(x,t)^T v(x,t)\rho(x,t)\, dx\, dt,$$

subject to the continuity equation and endpoint distributions $\rho_A,\rho_B$. In the exact formulation this is the Benamou–Brenier dynamic optimal transport problem, so the optimal protocol is a Wasserstein-2 geodesic in distribution space. In the slow-driving limit the same objective reduces to

$$L[\lambda]=\int_0^{t_{\rm f}} \dot{\lambda}^T(t)\,\zeta(\lambda(t))\,\dot{\lambda}(t)\, dt,$$

where

$$\zeta(\lambda(t))=\beta\int_0^\infty \langle \delta f(x(s),t)\,\delta f^T(x(0),t)\rangle_{\lambda(t)}\, ds$$

is the thermodynamic friction tensor, so the optimal protocol becomes a geodesic in control-parameter space (Chennakesavalu et al., 2022).

This yields a two-layer construction. First one chooses the optimal path geometry: Wasserstein geodesics when the full density path is free, or thermodynamic-length geodesics on the equilibrium manifold in the near-equilibrium regime. Then one chooses the optimal time parameterization: constant Wasserstein speed in distribution space or constant thermodynamic speed in control space. The paper makes the composite aspect explicit in the Brownian Stirling engine and qubit engine, each divided into four intervals, $[0,\tau/4]$, $[\tau/4,\tau/2]$, $[\tau/2,3\tau/4]$, and $k$0, with each stroke treated as its own displacement-interpolation segment (Chennakesavalu et al., 2022).

The same work also sharpens an important limitation. If intermediate states are prescribed, a globally optimal protocol can be assembled from geodesics between those waypoints. If the intermediate states are not fixed, segmentwise geodesic optimality does not guarantee global optimality over the whole cycle. Another limitation is operational rather than geometric: lower dissipation does not imply higher efficiency in the qubit-engine example (Chennakesavalu et al., 2022).

3. Quantum control, geometric phases, and optimized composite loops

In quantum control, optimized composite geometric protocols are built from cyclic evolutions whose useful phase is geometric rather than dynamical. For composite pulses with proportional systematic error

$k$1

first-order cancellation

$k$2

implies vanishing total dynamical phase, so the resulting operation is a geometric quantum gate. This was established for one-qubit composite pulses compensating pulse length error and for two-qubit composite pulses compensating $k$3-coupling error. SCROFULOUS, time-symmetric BB1, Knill’s sequence, and Jones’ $k$4-coupling-compensating sequence are all shown to be geometric in this sense (Ichikawa et al., 2012). Closely related work further showed that geometric quantum gates with zero net dynamic phase are robust against control field strength errors, and that a Trotter–Suzuki-based composite rf pulse can be interpreted as a geometric quantum gate (Kondo et al., 2010).

The same literature also establishes a recurring misconception: geometricity and robustness are distinct objectives. The $k$5-$k$6-$k$7 pulse is geometric but not a Type A pulse-length-error-compensating composite pulse, a two-pulse $k$8-rotation can be geometric but not pulse-length-error robust, and CORPSE is robust but not geometric in the Aharonov–Anandan sense (Ichikawa et al., 2012, Kondo et al., 2010). In open-system settings the composite aspect can be moved into the environment rather than the control sequence: for a target spin coupled to an auxiliary spin and a bosonic bath, initial entanglement between the two spins increases the sturdiness of the geometric phase, especially for small $k$9, because the reduced Bloch-ball radius

$$\lambda_*=\arg\min_{\lambda:[0,t_{\rm f}]\to \mathbb{R}^k} \Delta \Sigma_{\rm tot}[\lambda]$$0

shrinks as concurrence increases (Villar et al., 2011).

More recent work makes the optimization layer explicit. A seven-segment composite nonadiabatic geometric gate keeps the target gate

$$\lambda_*=\arg\min_{\lambda:[0,t_{\rm f}]\to \mathbb{R}^k} \Delta \Sigma_{\rm tot}[\lambda]$$1

fixed while introducing a free path parameter $$\lambda_*=\arg\min_{\lambda:[0,t_{\rm f}]\to \mathbb{R}^k} \Delta \Sigma_{\rm tot}[\lambda]$$2, the angle between two orange-slice-type loops. Conventional composite nonadiabatic geometric quantum gates improve one error direction at the expense of another, whereas the optimized scheme chooses $$\lambda_*=\arg\min_{\lambda:[0,t_{\rm f}]\to \mathbb{R}^k} \Delta \Sigma_{\rm tot}[\lambda]$$3 numerically to increase robustness against both Rabi-frequency and detuning errors while preserving pulse-shape flexibility. For the $$\lambda_*=\arg\min_{\lambda:[0,t_{\rm f}]\to \mathbb{R}^k} \Delta \Sigma_{\rm tot}[\lambda]$$4 gate the reported optimum is $$\lambda_*=\arg\min_{\lambda:[0,t_{\rm f}]\to \mathbb{R}^k} \Delta \Sigma_{\rm tot}[\lambda]$$5 for X error, Z error, and both together; for the $$\lambda_*=\arg\min_{\lambda:[0,t_{\rm f}]\to \mathbb{R}^k} \Delta \Sigma_{\rm tot}[\lambda]$$6 gate the reported optima differ, with $$\lambda_*=\arg\min_{\lambda:[0,t_{\rm f}]\to \mathbb{R}^k} \Delta \Sigma_{\rm tot}[\lambda]$$7 against X error, $$\lambda_*=\arg\min_{\lambda:[0,t_{\rm f}]\to \mathbb{R}^k} \Delta \Sigma_{\rm tot}[\lambda]$$8 against Z error, and $$\lambda_*=\arg\min_{\lambda:[0,t_{\rm f}]\to \mathbb{R}^k} \Delta \Sigma_{\rm tot}[\lambda]$$9 against both. In superconducting transmons, the optimized single-qubit $$\Delta \Sigma_{\rm tot}=\int_0^{t_{\rm f}} \beta(t)\int_{\Omega} v(x,t)^T v(x,t)\rho(x,t)\, dx\, dt,$$0 gate has duration $$\Delta \Sigma_{\rm tot}=\int_0^{t_{\rm f}} \beta(t)\int_{\Omega} v(x,t)^T v(x,t)\rho(x,t)\, dx\, dt,$$1 ns, fidelity $$\Delta \Sigma_{\rm tot}=\int_0^{t_{\rm f}} \beta(t)\int_{\Omega} v(x,t)^T v(x,t)\rho(x,t)\, dx\, dt,$$2, fidelity $$\Delta \Sigma_{\rm tot}=\int_0^{t_{\rm f}} \beta(t)\int_{\Omega} v(x,t)^T v(x,t)\rho(x,t)\, dx\, dt,$$3 with $$\Delta \Sigma_{\rm tot}=\int_0^{t_{\rm f}} \beta(t)\int_{\Omega} v(x,t)^T v(x,t)\rho(x,t)\, dx\, dt,$$4, and fidelity $$\Delta \Sigma_{\rm tot}=\int_0^{t_{\rm f}} \beta(t)\int_{\Omega} v(x,t)^T v(x,t)\rho(x,t)\, dx\, dt,$$5 with $$\Delta \Sigma_{\rm tot}=\int_0^{t_{\rm f}} \beta(t)\int_{\Omega} v(x,t)^T v(x,t)\rho(x,t)\, dx\, dt,$$6; the optimized two-qubit controlled-phase gate has duration $$\Delta \Sigma_{\rm tot}=\int_0^{t_{\rm f}} \beta(t)\int_{\Omega} v(x,t)^T v(x,t)\rho(x,t)\, dx\, dt,$$7 ns, fidelity $$\Delta \Sigma_{\rm tot}=\int_0^{t_{\rm f}} \beta(t)\int_{\Omega} v(x,t)^T v(x,t)\rho(x,t)\, dx\, dt,$$8, robustness $$\Delta \Sigma_{\rm tot}=\int_0^{t_{\rm f}} \beta(t)\int_{\Omega} v(x,t)^T v(x,t)\rho(x,t)\, dx\, dt,$$9 at $\rho_A,\rho_B$0, and $\rho_A,\rho_B$1 at $\rho_A,\rho_B$2 (Ding et al., 29 Sep 2025).

A closely related superconducting logical construction combines nonadiabatic geometric control with decoherence-free subspace encoding and a multi-loop optimized composite trajectory. In the logical basis

$\rho_A,\rho_B$3

the target is a logical

$\rho_A,\rho_B$4

For the two-loop optimized composite geometric protocol, the additional free parameter $\rho_A,\rho_B$5 is chosen as $\rho_A,\rho_B$6 on Path 1 or $\rho_A,\rho_B$7 on Path 2, yielding fourth-order suppression of Rabi, detuning, and residual $\rho_A,\rho_B$8-crosstalk errors, while DFS encoding provides inherent suppression of collective dephasing. At $\rho_A,\rho_B$9 and $$L[\lambda]=\int_0^{t_{\rm f}} \dot{\lambda}^T(t)\,\zeta(\lambda(t))\,\dot{\lambda}(t)\, dt,$$0 MHz, the logical gate fidelity is approximately $$L[\lambda]=\int_0^{t_{\rm f}} \dot{\lambda}^T(t)\,\zeta(\lambda(t))\,\dot{\lambda}(t)\, dt,$$1, and for $$L[\lambda]=\int_0^{t_{\rm f}} \dot{\lambda}^T(t)\,\zeta(\lambda(t))\,\dot{\lambda}(t)\, dt,$$2 kHz the logical optimized composite geometric $$L[\lambda]=\int_0^{t_{\rm f}} \dot{\lambda}^T(t)\,\zeta(\lambda(t))\,\dot{\lambda}(t)\, dt,$$3 gate remains above $$L[\lambda]=\int_0^{t_{\rm f}} \dot{\lambda}^T(t)\,\zeta(\lambda(t))\,\dot{\lambda}(t)\, dt,$$4 over nearly the full tested error interval (Ding et al., 1 May 2026).

4. Communication and modulation protocols

In coherent optical communications, the composite geometric idea is realized by embedding rate adaptation, mapping structure, and constellation geometry into a single fixed-interface protocol. A BICM-compatible geometric constellation shaping scheme keeps the FEC engine and modulation order fixed, inserts $$L[\lambda]=\int_0^{t_{\rm f}} \dot{\lambda}^T(t)\,\zeta(\lambda(t))\,\dot{\lambda}(t)\, dt,$$5 dummy bits before mapping, and optimizes a many-to-one labeling and geometry so that labels differing only in dummy positions collapse onto the same or nearly the same constellation point. The net rate becomes

$$L[\lambda]=\int_0^{t_{\rm f}} \dot{\lambda}^T(t)\,\zeta(\lambda(t))\,\dot{\lambda}(t)\, dt,$$6

with

$$L[\lambda]=\int_0^{t_{\rm f}} \dot{\lambda}^T(t)\,\zeta(\lambda(t))\,\dot{\lambda}(t)\, dt,$$7

Backward compatibility is exact at $$L[\lambda]=\int_0^{t_{\rm f}} \dot{\lambda}^T(t)\,\zeta(\lambda(t))\,\dot{\lambda}(t)\, dt,$$8, where the scheme reduces to ordinary BICM. The positions and labeling are optimized with automatic differentiation, using an AIR objective that discards dummy-bit positions and thereby induces many-to-one geometry (Yankov et al., 2023).

The protocol-level consequence is near-continuous rate adaptation while keeping both $$L[\lambda]=\int_0^{t_{\rm f}} \dot{\lambda}^T(t)\,\zeta(\lambda(t))\,\dot{\lambda}(t)\, dt,$$9 and $$\zeta(\lambda(t))=\beta\int_0^\infty \langle \delta f(x(s),t)\,\delta f^T(x(0),t)\rangle_{\lambda(t)}\, ds$$0 fixed. The paper reports, for the DVB-S2 example with $$\zeta(\lambda(t))=\beta\int_0^\infty \langle \delta f(x(s),t)\,\delta f^T(x(0),t)\rangle_{\lambda(t)}\, ds$$1, $$\zeta(\lambda(t))=\beta\int_0^\infty \langle \delta f(x(s),t)\,\delta f^T(x(0),t)\rangle_{\lambda(t)}\, ds$$2, and $$\zeta(\lambda(t))=\beta\int_0^\infty \langle \delta f(x(s),t)\,\delta f^T(x(0),t)\rangle_{\lambda(t)}\, ds$$3, a rate step below $$\zeta(\lambda(t))=\beta\int_0^\infty \langle \delta f(x(s),t)\,\delta f^T(x(0),t)\rangle_{\lambda(t)}\, ds$$4 bits/symbol. In simulation, the proposed MTOM+GCS outperforms TH without shaping by up to $$\zeta(\lambda(t))=\beta\int_0^\infty \langle \delta f(x(s),t)\,\delta f^T(x(0),t)\rangle_{\lambda(t)}\, ds$$5 dB in AWGN, but PAS remains stronger in linear AWGN. In nonlinear fiber, PAS and the proposed GCS have similar performance, and experiments over up to $$\zeta(\lambda(t))=\beta\int_0^\infty \langle \delta f(x(s),t)\,\delta f^T(x(0),t)\rangle_{\lambda(t)}\, ds$$6 km show near-continuous zero-error flexible throughput, up to $$\zeta(\lambda(t))=\beta\int_0^\infty \langle \delta f(x(s),t)\,\delta f^T(x(0),t)\rangle_{\lambda(t)}\, ds$$7–$$\zeta(\lambda(t))=\beta\int_0^\infty \langle \delta f(x(s),t)\,\delta f^T(x(0),t)\rangle_{\lambda(t)}\, ds$$8 spans increased reach at the same net data rate, and up to $$\zeta(\lambda(t))=\beta\int_0^\infty \langle \delta f(x(s),t)\,\delta f^T(x(0),t)\rangle_{\lambda(t)}\, ds$$9 bits/2D symbol gain over conventional QAM; at $[0,\tau/4]$0 km, the reported example is $[0,\tau/4]$1 bits/2D symbol for MTOM-GCS versus $[0,\tau/4]$2 bits/2D symbol for conventional BICM forced back to 64QAM (Yankov et al., 2023).

A second line of work treats constellation design itself as a composite differentiable map. Constellation points $[0,\tau/4]$3 are first normalized by

$[0,\tau/4]$4

and for nonlinear fiber further transformed by a kurtosis-dependent scaling

$[0,\tau/4]$5

before MI or GMI is optimized with exact analytical gradients and a trust-region method using SR1 and Steihaug’s conjugate gradient. This reduced design time from days to minutes, enabled optimization up to $[0,\tau/4]$6 points in 2D and 4D, and produced GMI within $[0,\tau/4]$7 bit/2Dsymbol of AWGN capacity (Sillekens et al., 2022).

These communication results sharpen a recurring boundary condition. Geometric optimization is not channel-agnostic: the AWGN-optimal geometry, the nonlinear-fiber-optimal geometry, and the most implementation-friendly geometry need not coincide. In particular, PAS remains better in linear AWGN for finite constellations, whereas fixed-order MTOM-GCS becomes competitive in nonlinear fiber because it avoids a distribution matcher and reduces higher-order moment penalties (Yankov et al., 2023).

5. Composite optimization algorithms and reduction frameworks

In convex composite minimization,

$[0,\tau/4]$8

with $[0,\tau/4]$9 $[\tau/4,\tau/2]$0-strongly convex and $[\tau/4,\tau/2]$1-smooth and $[\tau/4,\tau/2]$2 convex proximable, the geometric proximal gradient method extends geometric descent to the nonsmooth composite setting by replacing $[\tau/4,\tau/2]$3 with the proximal gradient mapping

$[\tau/4,\tau/2]$4

and defining

$[\tau/4,\tau/2]$5

The optimizer is enclosed by balls centered at $[\tau/4,\tau/2]$6, and the method chooses iterates by a root-finding procedure for a strictly increasing scalar function $[\tau/4,\tau/2]$7, replacing exact line search. The enclosing-ball argument yields

$[\tau/4,\tau/2]$8

and with $[\tau/4,\tau/2]$9 the method attains the optimal accelerated linear factor

$[\tau/2,3\tau/4]$0

for strongly convex composite problems (Chen et al., 2016).

A different strand seeks exact worst-case optimality among fixed-step first-order composite methods. OptISTA solves

$[\tau/2,3\tau/4]$1

with horizon-dependent coefficients $[\tau/2,3\tau/4]$2 and $[\tau/2,3\tau/4]$3, and achieves

$[\tau/2,3\tau/4]$4

This improves the FISTA constant by a factor of $[\tau/2,3\tau/4]$5, and the paper proves a matching lower bound via a double-function semi-interpolated zero-chain construction, making the result exact for the stated oracle model (Jang et al., 2023).

A subsequent reduction framework shows that optimized methods for unconstrained smooth optimization can be converted systematically into optimized composite proximal methods by replacing each smooth gradient occurrence $[\tau/2,3\tau/4]$6 with $[\tau/2,3\tau/4]$7, $[\tau/2,3\tau/4]$8, and transporting the smooth PEP certificate into the composite setting. This yields stepsize acceleration for proximal gradient descent with exponent $[\tau/2,3\tau/4]$9, a convergence guarantee for proximal OGM faster than FISTA, and a proximal OGM-G method that improves the state-of-the-art rate for minimizing composite gradient norm (Bok et al., 30 Jun 2025). The geometric content here is partly explicit in GeoPG, but in the reduction framework it is mostly indirect, appearing through PEP certificates, rank-1 residual structure, and Laplacian or Schur-complement positivity rather than through an overt geometric-descent construction.

6. Finite geometry, cryptography, and secure computation

In combinatorial cryptography, a geometric protocol for the generalized Russian cards problem maps the deck to the affine space $k$00 and represents Alice’s hand as a $k$01-slicing, a union of $k$02 distinct parallel hyperplanes: $k$03 The protocol is parameterized by

$k$04

and Theorem 4.1 states that if

$k$05

then the protocol is informative and $k$06-safe. The same paper proves that for any fixed $k$07 there are infinitely many valid $k$08 with

$k$09

hence $k$10, and identifies the construction as the first solution guaranteeing $k$11-safety when Cath has more than one card (Cordón-Franco et al., 2013).

Secure computational geometry exhibits an analogous protocol architecture. For secure point inclusion, the classical geometric predicate is decomposed into simpler tests and then instantiated cryptographically. In a star-shaped polygon, inclusion is reduced to locating the wedge containing the query point and then checking one edge test. Determinants are rewritten as dot products,

$k$12

which are evaluated with secure scalar product and secure comparison. Protocol 4.1 gives $k$13 round complexity, whereas Protocol 4.2 hides the binary-search path by encrypted cyclic rotation and homomorphic commutative encryption, reducing the star-shaped case to $k$14 rounds. Protocol 5.1 extends the method to a more general polygonal domain with multiple disconnected nested components at $k$15 cost (0705.4185).

These finite-geometry constructions are not phase-geometric or geodesic, but they fit the same composite pattern: a global task is factored into geometrically structured local pieces, and efficiency or security comes from optimizing the composition rather than replacing geometry with a generic circuit.

7. Recurring design rules and limitations

Across domains, optimized composite geometric protocols repeatedly use extra degrees of freedom that do not alter the nominal task but do alter robustness or efficiency. In thermodynamics the critical freedom is time allocation along a fixed geodesic or the choice of prescribed waypoints; in optical modulation it is dummy-bit loading, many-to-one collapse, launch power, and channel-aware scaling; in composite optimization it is stepsize matrices, proximal embeddings, or enclosing-ball geometry; in quantum control it is the relative orientation of loops, such as the path parameter $k$16 or the inter-loop azimuthal offset $k$17 (Chennakesavalu et al., 2022, Yankov et al., 2023, Chen et al., 2016, Ding et al., 29 Sep 2025).

Several misconceptions are corrected explicitly by the literature. Geometricity does not automatically imply robustness, and robustness does not automatically imply geometricity, as shown by the geometric but non-robust $k$18-$k$19-$k$20 pulse, the geometric but pulse-length-error-nonrobust two-pulse $k$21-rotation, and CORPSE, which is robust but not geometric in the Aharonov–Anandan sense (Ichikawa et al., 2012, Kondo et al., 2010). Segmentwise geodesic optimization is not automatically globally optimal when intermediate states are free (Chennakesavalu et al., 2022). In communications, PAS remains better in linear AWGN even though fixed-order many-to-one GCS becomes competitive in nonlinear fiber (Yankov et al., 2023).

A further limitation is methodological. Exact optimality results in composite optimization are often horizon-dependent, require bespoke coefficient construction, or rely on computer-assisted PEP design; GeoPG requires a root-finding subproblem, OptISTA requires precomputation of horizon-dependent coefficients, and reduction-based composite acceleration depends on transferable certificate structure (Chen et al., 2016, Jang et al., 2023, Bok et al., 30 Jun 2025). Quantum-control variants similarly remain parameter-scan driven rather than closed-form in their strongest multi-error-robust forms, as in the numerical optimization of the path parameter $k$22 for composite nonadiabatic geometric gates (Ding et al., 29 Sep 2025).

This suggests that “optimized composite geometric protocol” is best understood not as a single standardized protocol class but as a design strategy: keep the outer task or interface fixed, decompose the implementation into geometrically structured components, and optimize internal free parameters so that the composed object outperforms the corresponding single-segment, unshaped, or non-optimized construction.