WorksPhysics and Astronomy
A non-equilibrium thermodynamics formulation of shear set-points in adaptively remodeling vascular networks
Abstract
Blood vessels remodel in response to wall shear stress while the network operates under a metabolic maintenance budget. For a power-law maintenance cost, minimizing dissipation at fixed budget yields a local shear set-point, and a purely local adaptation rule drives the network towards this optimum. Here we give this result a non-equilibrium thermodynamic formulation. Blood flow is an exactly linear Onsager system, and the slow remodeling of vessel radii is a gradient flow generated by a Rayleigh dissipation functional. This identifies the structural mobility, which scales inversely with the maintenance cost of each vessel and can be measured from radius fluctuations. Enforcing the budget exactly requires a time-dependent metabolic multiplier, which we obtain in closed form; with a fixed multiplier the budget is violated by up to one third during transients. The fluid displaced by remodeling vessels couples flow and structure through a symmetric, non-diagonal mobility, so that Onsager reciprocity becomes a testable statement: neighbouring vessels should show anticorrelated radius fluctuations. With noise, the structural sector obeys a second law and an integral fluctuation theorem. Non-reciprocal signaling between vessels, such as upstream conducted responses, then breaks the time-reversal symmetry of radius fluctuations and produces housekeeping entropy, detectable from time series of realistic length. Finally, we construct a partition function over network topologies with the number of loops as order parameter. Numerical tests show that loop counts do not collapse onto a single curve against an effective temperature, and that they depend strongly on the detection threshold.
Introduction
Vascular networks remodel locally (i.e., vessels dilate or constrict in response to wall shear stress), yet the efficiency of blood transport is a property of the network as a whole. Reconciling this local sensing with a global energetic goal is a long-standing question in the physiology of adaptive flow networks, and it applies equally to leaf venation and to engineered transport systems [1, 2, 3, 4]. A recent variational treatment resolved this tension for a power-law maintenance cost: minimizing dissipation under a fixed maintenance budget yields an exact local shear set-point that depends jointly on vessel radius and length, recovers Murray’s law as a special case, and is realized dynamically by a purely local adaptation rule that performs descent on the associated global transport functional [5, 6, 7, 8, 9, 10]. Simulation of that rule further showed that developmental history and fluctuating demand shape network topology: staged growth with regrowth of pruned connections systematically outperforms adaptation from a dense initial state, and demand fluctuations stabilize redundant loops without increasing the maintenance budget [11, 12, 13, 14].
That analysis is thermodynamic in content but not in form: the local adaptation rule and its associated Lyapunov inequality are asserted rather than derived from a variational dissipation principle, no phenomenological (Onsager) coefficients are identified, and the topological degeneracy and loop-selection phenomena are documented numerically without a statistical-mechanical framework that would make them predictive beyond the simulated parameter range. The purpose of the present paper is to supply that framework. We show that the vascular remodeling problem is a two-time-scale system whose fast sector is exactly linear and whose slow sector is a gradient flow that reduces to linear response near the set-point, in the sense of Onsager and Prigogine [15, 16, 17, 18], that its local adaptation rule can be written as a gradient flow generated by a Rayleigh dissipation functional together with the maintenance constraint, and that its fluctuation-driven topological transitions admit a partition-function description formally analogous to that of other adaptive network models [12, 19].
The present work builds on several lines of research. Flux-driven adaptation rules were introduced for Physarum [20, 21] and for microvascular networks, where the models of Pries, Secomb and coworkers already combine shear, pressure, metabolic and conducted (non-local) stimuli [22, 23]; convergence of such dynamics to optimal transport and the existence of Lyapunov functionals have been proved or characterized in detail [24, 25, 26], and gradient flows that respect constraints exactly have been constructed for microvascular optimization [27]. The shear set-point itself has direct experimental support [28], and set-points cued to shear peaks have been shown to stabilize loopy capillary beds [29]. Time-dependent loads act through an averaged dynamics with an explicit criterion for tree topologies [30]; topology selection under noise has been formulated in statistical-mechanical terms [31, 32, 33], and a curve collapse of fluctuation-induced topology transitions has been reported for coupled adaptive networks [34]. Global conservation of fluid volume has been identified as the parameter that integrates network architecture into vein fate late in Physarum [35]. On the thermodynamic side, network thermodynamics provides the classical Onsager structure on graphs [36, 37], and local learning rules in physical networks are a related instance of local rules descending global functionals [38]. Finally, generalizations of Murray’s law with non-cubic cost exponents [39, 40, 41, 42] and empirical scalings of wall shear stress with radius [43] bear directly on the radius dependence of Eq. (3).
Gradient-flow representations of local adaptation rules have been discussed before [18, 19, 25, 26]. What is new here is: (i) the structural Onsager coefficient and its testable scaling with the maintenance cost; (ii) the exact enforcement of the metabolic budget through a closed-form Gaussian multiplier; (iii) a hydraulic coupling between vessels that makes Onsager reciprocity testable; and (iv) a second law for the structural sector, which turns reciprocity into a time-reversal test for active signaling between vessels.
The paper is organized as follows. Section 2 recalls the network model and the central optimality result. Section 3 establishes the fast hydraulic sector as an exact linear Onsager system. Section 4 writes the slow structural sector as a Rayleighian gradient flow, obtains the structural Onsager coefficient in closed form, recovers the reported Lyapunov inequality as a special case of the bilinear form , and derives the hydraulic–structural coupling induced by the displaced volume.
Section 5 examines the metabolic multiplier and shows that its quenched treatment is a canonical-ensemble approximation to an exact micro-canonical (Gaussian-constrained) dynamics, which we derive in closed form. Section 6 extends the theory to fluctuating demand via a Langevin equation and constructs the associated partition function over network topologies, from which the loop count emerges as a thermal order parameter; it also formulates the second law and the fluctuation theorem of the structural sector and tests both predictions numerically. Section 7 discusses the regime of validity of the classical minimum-entropy-production theorem in this system, collects the falsifiable predictions of the theory, and concludes. Appendix A collects the derivations step by step, and Appendix B describes the numerical methods and parameters.
Network model and the local optimality condition
We adopt the network model and notation of Ref. [10]. Consider a connected graph with sink nodes, a single source, and nodes in total; each sink withdraws one unit of flow. An edge has radius , length , Poiseuille conductance (in reduced units absorbing viscosity and geometric prefactors), signed flow , and wall shear stress . Writing for the oriented incidence vector of edge and for the weighted graph Laplacian, node pressures are , where is the injection vector and the Moore–Penrose pseudoinverse. The total dissipated power is
and the maintenance cost, for vessel mass raised to a power , is
and we denote by the maintenance cost of edge , so that . The budget is fixed, . Minimizing at fixed by a Lagrange multiplier yields, on the active support of a tree, the shear set-point
which reduces to Murray’s uniform-shear condition at . Equivalently, the stationarity condition reads . Writing and the normalized shear , the reported local adaptation rule is
under which the composite functional
obeys
Equations (4)–(6) are the starting point of the present analysis: we now show that they are not independent postulates but admit an Onsager–Rayleigh gradient-flow representation once the network is separated into its fast and slow sectors.
The fast hydraulic sector as an exact linear Onsager system
For fixed geometry , Poiseuille’s law is already a linear force–flux relation, with flux , thermodynamic force (the pressure drop across edge ), and phenomenological coefficient . The corresponding entropy-production (dissipation) rate is the bilinear Onsager form
The dissipation rate is thus the total power of Eq. (1). (Strictly, the entropy production is , with the temperature of the fluid; in reduced units is absorbed and we identify both.) Because is a sum of symmetric outer products, the Onsager reciprocity relations for this sector are satisfied identically, with no additional symmetry assumption required. Moreover, for fixed and fixed sink demands, is a strictly convex functional of the flow field subject to Kirchhoff’s current law, so the physically realized flow is exactly its unique minimizer: this is Thomson’s principle, the linear-network counterpart of the minimum-entropy-production theorem of Prigogine [17]. A useful corollary, used in § 4, follows from the envelope theorem: since the flow minimizes at fixed demands, , with no contribution from the implicit dependence of the flows on . The reduced dissipation appearing in , Eq. (5), is therefore the exact result of adiabatically eliminating the fast sector: introducing and writing , with the hydraulic co-content, the limit collapses to the algebraic constraint used throughout § 2. The physiological separation of time scales between flow relaxation (seconds) and structural remodeling (hours to days) justifies this elimination a priori.
The slow structural sector: a Rayleighian formulation of the adaptation rule
The Rayleigh dissipation functional
We now show that Eq. (4) has the structure of a gradient flow generated by an Onsager variational principle [15, 16, 44, 45]. Define the Rayleighian [46]
in which the first term is the instantaneous rate of change of the free-energy-like functional and the second is a quadratic dissipation function penalizing the rate of structural change, with an as-yet-unspecified structural mobility. The physical trajectory extremizes over at fixed ; stationarity of Eq. (8), , gives the general force–flux law
The structural Onsager coefficient
Because at fixed , . Using and the set-point relation (from Eqs. (2)–(3)),
so that the thermodynamic force conjugate to is
Matching Eq. (9), with the force of Eq. (11), to the reported rule, Eq. (4), fixes the structural Onsager coefficient uniquely:
Two remarks qualify this result. First, the argument runs from the rule to the mobility: given Eq. (4), is fixed uniquely, but conversely any rule with admits a representation of the form (9) with a positive mobility. The physical content of Eq. (12) is therefore not the rule itself, but the identification of its mobility, which carries testable information. Second, depends on , and hence on itself, so the structural force–flux relation is nonlinear; it reduces to linear Onsager response, with the state-independent coefficient , only in the neighbourhood of the set-point .
Three consequences follow immediately. First, for all admissible states, since , , , and ; this positivity is inherited from the positivity of the maintenance cost and the multiplier, not imposed separately, and it is what guarantees the Lyapunov property below. Second, : vessels with larger local maintenance cost are structurally stiffer, i.e. their radius responds more slowly per unit thermodynamic force. This is a falsifiable prediction not present in the original formulation: fitted structural mobilities should scale inversely with local maintenance cost across a vascular bed.
Third, enters exactly as an intensive potential conjugate to the extensive budget , structurally analogous to a chemical potential in a grand-canonical ensemble.
Figure 1A shows how the second consequence can be tested. The fluctuation–dissipation relation of § 6 links the mobility to the short-time diffusivity of the logarithmic radius, . The prediction can therefore be checked from radius time series, without knowing or . In Langevin simulations around the optimal tree of a jittered lattice, the fitted exponents are , and for , 1 and 1.5. A fit against alone gives a biased exponent ( instead of at ), because also depends on the vessel length, which must therefore be included. In simulation this is a consistency check, since the rule defines . Its empirical content is that is the same for all vessels, so that intravital measurements of radius fluctuations across a vascular bed should give a log–log slope of against , and of against at fixed .
Dissipation function and Lyapunov stability
which reproduces Eq. (6) term for term. In the present derivation, however, non-negativity follows directly from the bilinear Onsager form together with , rather than from a separate convexity argument. Equality in Eq. (13) holds only at for every active edge, i.e. the shear set-point of Eq. (3) is exactly the zero-dissipation manifold of the slow sector. We stress that has the bilinear Onsager form but is not a thermodynamic entropy production in the strict sense: is itself a rate of dissipation, so has units of power per unit time. Its non-negativity is the structural-sector analogue of the second law, namely a Lyapunov property of the minimum-dissipation principle, in the spirit of Prigogine’s theorem and of the Glansdorff–Prigogine evolution criterion [47]. A dimensionally consistent second law, in which is converted into an energy and the configurational entropy of the structural state is included, is formulated in § 6.2.
Reciprocity
The structural Onsager matrix implied by Eq. (9) is diagonal, , reflecting the assumption of purely local stress sensing; Onsager reciprocity, , then holds trivially. A genuine test of this assumption would require detecting cross-couplings with (i.e. remodeling at one vessel driven by shear elsewhere via long-range mechanotransductive signaling) which the present rule excludes by construction and which we leave for future work as a natural extension of the theory. Part of the cross-couplings, however, is not an additional assumption: it follows from hydraulics alone, as shown next.
Hydraulic–structural coupling through the displaced volume
The separation into sectors of Sections 3–4 neglects the fact that a remodeling vessel changes its lumen volume, (with absorbing geometric prefactors), and must therefore exchange fluid with the network. Mass conservation at the nodes reads , where assigns half of the volume change of each vessel to each of its end nodes. Treating the node pressures as Lagrange multipliers of this constraint, the joint Rayleighian of both sectors is
where is the structural potential expressed as an energy, with a metabolic time scale converting the power into an energy, and is the bare structural mobility. The dissipation thus appears twice: as the physical dissipation of the hydraulic sector, first term, and, evaluated on the static flow, as part of the biological target that drives remodeling; the second role is phenomenological and is what makes the structural drive active. Stationarity with respect to returns Poiseuille’s law, and stationarity with respect to gives
with . The same coefficient that turns wall motion into a volume source in the continuity equation turns the mean luminal pressure into a force on the wall: this is Onsager reciprocity between the two sectors, obtained from the variational structure rather than imposed. We assume that the static transmural load is balanced by wall tension (Laplace’s law) and is absorbed into the reference state, so only the pressure generated by the wall motion itself remains, , where the subscript denotes removal of the source node, whose pressure is fixed, and . Eliminating it,
so that the effective structural Onsager matrix is . The assumption is not innocuous: the static mean pressure depends on the whole configuration , and the force is in general not a gradient, so that, if it is not fully balanced by the wall, it adds a non-conservative drive that breaks detailed balance (§ 6.2).
Four properties follow. (i) is symmetric and positive definite, because is; reciprocity, , now holds for a non-diagonal matrix and is therefore a non-trivial, testable statement. (ii) To first order in the coupling, for every pair of vessels in the same branch, because has positive entries on each connected component (Appendix § 8.8). Physically, a dilating vessel draws fluid from its neighbourhood, lowers the local pressure and drives the contraction of hydraulically connected vessels. Numerically, the sign persists up to in all 72 cases tested (12 networks, 3 values of ) and breaks down only at strong coupling, . The coupling decays with distance along the tree and vanishes between branches that meet only at the pressure-fixed source (Figure 1C). (iii) The coupling strength is set by , the ratio between the hydraulic drag time of wall motion and the remodeling time, of the same order as the small parameter of Section 3. To first order, is the leading correction to the adiabatic elimination: it is small in mature vasculature (seconds against hours to days), but not in systems in which flow and adaptation evolve on comparable time scales, such as Physarum or rapidly growing plexuses. (iv) The uncoupled theory only fixes the product of Eq. (12); the coupling makes and separately identifiable.
Figure 1. Structural mobilities obtained from fluctuations, in Langevin simulations around the optimal tree of a jittered lattice with the source at a corner. (A) Short-time diffusivity of divided by , which equals the mobility , against the maintenance cost of each vessel, without hydraulic coupling. Lines: Eq. (12) at . (B) Cross-mobilities measured from the covariance of radius increments against the prediction , Eq. (16), for and ; both are normalized by , and the line marks equality. (C) Normalized cross-mobility against the distance between vessels along the tree, for three coupling strengths. The shaded band contains pairs in branches that meet only at the source, where the coupling vanishes.
Multiplying Eq. (16) by gives the power balance of the coupled system at fixed (with the Gaussian multiplier of Section 5, is replaced by ),
where are the flows induced by wall motion. Part of the decrease of the structural potential is therefore dissipated viscously in the lumen, and the entropy production of the slow sector has a genuinely hydraulic component. Along relaxations towards the optimal tree the balance is satisfied numerically to 0.2% (the finite-difference error), with a hydraulic share of 0.8%, 20% and 53% for , 0.1 and 1. The cross-mobilities can be measured from the covariance of radius increments, (negative for neighbouring vessels, whose increments are therefore anticorrelated). In simulation, the measured and predicted values agree with a correlation of 0.993 (Figure 1B).
The metabolic multiplier: quenched versus dynamically constrained dynamics
Equation (13) treats as a constant. Differentiating the budget,
which vanishes only at the fixed point, not along the trajectory: a constant enforces the budget constraint only asymptotically, and only if is set to its optimal value ; for any other value the dynamics converges to the optimum of a different budget in a canonical-ensemble sense, analogous to running constrained dynamics with a frozen Lagrange multiplier rather than a holonomic constraint force solved self-consistently at every instant (as in the SHAKE/RATTLE algorithms or a Gaussian isokinetic thermostat).
We now derive the exact constrained dynamics. Define the -independent gradients
so that and the dynamics reads , with given by Eq. (12) evaluated at . Introducing, with the gradients of Eq. (19), the local, -independent shear ratio , the normalized shear is , and the rule (4) reads . Requiring that Eq. (18) vanish at all times, , then gives the multiplier in explicit form,
which is the Gaussian-constraint multiplier for this network: is the cost-weighted mean of the local shear ratios, determined instantaneously by the current flow field. Substituting back,
which is a projected gradient flow: the multiplier is chosen at every instant so that the velocity is tangent to the constraint manifold, , i.e. orthogonal in the metric to the Riemannian gradient of the constraint, exactly as in constrained steepest descent [27]. Because , the time dependence of does not enter the rate of change of on the constraint manifold, and itself is the natural Lyapunov functional, with dissipation function
vanishes exactly only at the true fixed point for every edge simultaneously (cf. Eq. (3)) and is free of the residual budget drift implicit in Eq. (13).
We note that a numerical implementation of Eq. (21) should conserve to machine precision at every integration step, in contrast with the empirical budget tolerance reported under quenched- integration [10]; this provides a direct numerical consistency test of the present formalism. Since the constraint is nonlinear, a first-order integrator with evaluated at the beginning of each step leaves a drift of order per step; exact conservation is restored by a SHAKE-like retraction, which here has a closed form: because each , a uniform shift after each step restores exactly (Appendix § 8.§ 6).
Table 1 compares integration schemes on a jittered lattice (60 edges, floor radius ) from the same dense initial state, integrated up to . Three conclusions follow. First, the quenched dynamics, even with set to its optimal value , violates the budget by 13–33% during the transient, independently of ; it returns to only if the final topology coincides with the one for which was computed, and for and the dynamics selects a different tree and a residual violation of persists. Second, explicit Euler integration of the Gaussian dynamics, Eq. (21), drifts by and does not self-correct, so it converges to the optimum of a slightly different budget. Third, machine precision is reached only with the retraction; the retraction also works with quenched , so exact conservation is a property of the integrator, whereas the Gaussian multiplier is what keeps the continuous dynamics on the constraint manifold. The large transient violations under quenched also show that the two treatments differ most during transients, as discussed next.
The quenched treatment of § 4 and the exact treatment derived here correspond, respectively, to a canonical ensemble (fixed , fluctuating ) and a micro-canonical ensemble (fixed , dynamical ) for the structural sector; the two are expected to coincide near the fixed point and in the thermodynamic limit, but may differ measurably during rapid growth transients, precisely where staged-growth protocols operate.
Fluctuating demand: a Langevin theory and the partition function over topologies
Effective temperature from fluctuating demand
Let each sink demand have unit mean and independent fluctuations of variance , with second moment and expected dissipation , as in Ref. [10]. Two regimes must be distinguished. If demand fluctuates on the hydraulic time scale , the adiabatic elimination of Section 3 replaces by in , so that fluctuations modify the energy landscape rather than acting as noise, . This is the mechanism of Refs. [11, 12], in which loops are already the zero-temperature optimum of the averaged functional. Only the component of the demand fluctuations with time scales comparable to or longer than acts as structural noise. In what follows is understood as , and accounts for the slow component.
We extend the Rayleighian dynamics of Section 4 to a Langevin equation consistent with the fluctuation–dissipation theorem relative to the mobility already derived in Eq. (12):
where the effective structural temperature is sourced by the demand variance through the stochastic dependence of on . At , and Eq. (23) reduces to the deterministic descent of Section 4; for , thermal noise competes with the energetic drive toward the minimum of , which for need not be a tree. We emphasize that Eq. (23) neglects multiplicative-noise (spurious-drift) corrections arising from the -dependence of ; these are subleading provided is small compared with the local curvature scale of , and we flag this as a controlled approximation rather than an exact statement. Alternatively, adding the Itô drift to Eq. (23) restores detailed balance with respect to exactly, since the mobility matrix is diagonal. With the coupling of Section 4.5, is replaced by in both the drift and the noise covariance, by (so that has units of energy), and the Itô drift by .
Second law and fluctuation theorem for the structural sector
The Langevin extension provides the structural sector with a second law in the strict sense of stochastic thermodynamics [48, 49]. The probability density of the structural state obeys the Fokker–Planck equation
Defining the structural free energy , one finds
which is a second law for the structural sector: the total entropy production is the sum of the change of configurational entropy and of the entropy delivered to the medium by the descent of , and it is non-negative because is positive definite. Along single trajectories the total entropy production satisfies the integral fluctuation theorem [48], of which Eq. (25) is the average consequence. In the limit the density concentrates on the deterministic trajectory, , and Eq. (25) reduces to , i.e. to Eq. (13) generalized by the coupling, Eq. (17). The stationary state, , is , with detailed balance, because is symmetric and is a gradient; for the micro-canonical treatment the same construction holds with projected onto the manifold , , which is symmetric and positive semi-definite with null direction normal to the manifold. The structure is that of a dissipative (metric) GENERIC system [50].
Two remarks delimit this result. First, it is the second law of the effective description: measures structural fluctuations, not the temperature of the fluid, and is a potential sustained by metabolism rather than an equilibrium free energy, so the full physical entropy production also contains the metabolic dissipation that sustains the drive, which bounds Eq. (25) from above. Second, the result provides a sharp test of the reciprocal structure: if long-range mechanotransductive couplings (§ 4.4) were non-reciprocal, , or non-gradient, detailed balance would fail, the stationary current would not vanish, and a housekeeping entropy production [51] would appear. Time-reversal asymmetry of radius fluctuations in pairs of vessels would then be a direct signature of active signaling beyond hydraulics.
Figure 2 quantifies this test for a minimal model of upstream conduction in the spirit of Refs. [22, 23]: each vessel also responds to the deviations from the set-point of the vessels it feeds,
Figure 2. Time-reversal test of reciprocity, for the linearized dynamics around the optimal tree of the lattice (, conduction range edges). (A) Housekeeping entropy production, multiplied by the slowest relaxation time, against the strength of the conducted signal, Eq. (26), without () and with () hydraulic coupling; the dotted line shows the scaling . For it vanishes to machine precision in both cases. (B) Lagged cross-correlations of a parent–child pair in both time orders. With hydraulic coupling alone (, ; grey) the two orders coincide; with conduction (, ; red) they differ. (C) Probability of detecting at a 5% false-positive rate against observation time, for two signal strengths and three levels of measurement noise , given in units of the intrinsic fluctuation amplitude; the dotted line marks the 5% false-positive level.
where is the distance along the tree and the noise obeys the fluctuation–dissipation relation with respect to the passive mobility only. Because is not symmetric, the mobility is not either. Linearizing around the optimal tree gives an Ornstein–Uhlenbeck process whose stationary entropy production is entirely housekeeping (Appendix § 8.11). It vanishes to machine precision for , with or without hydraulic coupling, and grows as (fitted exponent 1.95; Figure 2A). The cross-correlations of a parent–child pair become asymmetric under time reversal (Figure 2B), whereas the hydraulic coupling alone generates correlations that are symmetric. A statistic that sums the asymmetries of all parent–child pairs, weighted by their predicted sign, detects with power above 0.8 after 30 relaxation times without measurement noise, and after 100 relaxation times when the measurement noise equals the intrinsic fluctuation amplitude; requires about 300 relaxation times (Figure 2C). In mature vasculature, where structural relaxation takes days, the test therefore calls for imaging campaigns of months; in Physarum, where veins respond to shear within minutes [35], hours of imaging suffice.
Partition function over discrete supports
Let denote a connected spanning sub-network (support) of the underlying spatial graph, so that every sink is supplied and below is well defined, and define the coarse-grained free energy of a support as the optimal dissipation attained on it,
which coincides with the numerical benchmark of Ref. [10] up to the fixed rescaling by and the additive constant . Because throughout , the cost is concave in and, at , the constrained minimum on a support containing loops is generically not interior: it coincides with that of one of its spanning subtrees, with some radii driven to zero. Eq. (27) must therefore be understood with a minimal radius, , supports whose optimum degenerates to a subtree must be identified with it to avoid double counting, and interior minima on loopy supports exist only for with . For a Langevin process of the form (23) with mobility approximately constant over the relevant fluctuations, the stationary distribution over configurations satisfies detailed balance with respect to .
Coarse-graining over the continuous radii within each support by a Laplace (saddle-point) approximation yields an effective partition function over topologies,
where is the number of continuous degrees of freedom on . The ensemble must be chosen consistently: in the micro-canonical form, Eq. (27) is minimized on , the Hessian is restricted to the tangent space of the constraint and ; in the canonical form, is fixed, is the unconstrained minimum of on (whose budget then differs from support to support) and . Since , the prefactor contains : an explicitly temperature-dependent entropic weight that favours loops as grows and must be kept in Eq. (29). The entropic prefactor is the network analogue of a vibrational entropy, obtained from the Gaussian fluctuations of around its minimum on $S ({xref}Appendix A.10`). The Laplace approximation requires the Hessian restricted to to be positive definite and to be small compared with its smallest eigenvalue.
The loop count as a thermal order parameter
With the number of independent loops of a support, its thermal average is
As , is dominated by the lowest- support, which is generically a tree for concave transport costs [7, 8, 9], recovering ; as , all supports are weighted by alone, which now grows as , and since loopy connected subgraphs vastly outnumber spanning trees combinatorially, saturates at a -independent value, which in the tests of Figure 3 is set by the detection threshold and lies at about half the maximum loop count; this limit lies outside the validity of the Laplace approximation and should be read as a qualitative extrapolation, to be checked by direct enumeration on small graphs. This reinterprets the previously reported monotonic growth of surviving loops with demand fluctuation [10] as an order–disorder crossover, which can sharpen into a genuine transition only in the limit of large graphs [13], with playing the role of the order parameter and that of the control parameter.
Figure 3. Test of the data-collapse hypothesis by Metropolis sampling on a jittered lattice; the dotted line marks the maximum number of independent loops (16). (A) Mean loop count against for three cost exponents. Filled symbols mark temperatures at which the equipartition calibration agrees with within 35%. (B) The same data against . (C) Mean loop count at for three detection thresholds .
This picture suggests a specific test that goes beyond fitting against for each cost exponent : if is calibrated independently from the stationary variance of structural fluctuations in simulation, , with the Hessian at the minimum on (equipartition; for a single decoupled edge, with ) at a reference edge, then plotted against this calibrated should collapse the family of curves obtained for different values of onto a single master curve, to the extent that the -dependence of the free-energy differences and of is captured by the single scale ; systematic deviations from collapse would then measure the residual -dependence of the landscape itself. This hypothesis can be tested on existing simulation data, provided the stationary radius fluctuations were recorded.
We carried out this test by Metropolis sampling of on a jittered lattice (40 edges, at most 16 independent loops) for , 1 and 1.5, counting as active the vessels with (). The results (Figure 3) do not support the collapse, in contrast with the curve collapse reported for fluctuation-driven transitions in coupled Hu–Cai networks [34], where the control parameter is the fluctuation amplitude itself rather than a calibrated temperature. (i) The equipartition calibration is valid only in the harmonic regime, up to ; beyond it the reference edge leaves its well and the calibrated temperature diverges. (ii) In the crossover region the curves do not collapse, neither against nor against : at one finds , 1.8 and 4.2 for , 1 and 1.5. The -dependence is dominated by the cost of small vessels, , which makes the thermal activation of loops cheaper at larger and is not absorbed by a single temperature scale. (iii) At high the curves merge on a -independent plateau , whose value, like everywhere, depends strongly on . A meaningful order parameter therefore requires either a threshold-free definition, for instance a conductance-weighted measure of redundancy, or an explicit threshold matched to the resolution of the imaging. These tests use purely thermal noise (); demand fluctuations enter through and must be tested separately.
Discussion and conclusion
The classical minimum-entropy-production theorem guarantees a unique stationary state only near equilibrium, in the strictly linear regime, and on a fixed support. Both conditions hold for the fast hydraulic sector (§ 3 and Eq. (7)), where the physical flow field is exactly the unique minimizer of dissipation on a fixed conductance network. Neither condition holds for the slow structural sector once pruning and growth are allowed to change the admissible support: is strictly convex in only on a fixed tree, and the map from initial to final support is not unique. This is the thermodynamic origin of the previously reported topological degeneracy and developmental-history dependence [10]; it is an expected consequence of leaving the domain of validity of the classical uniqueness theorem once the constitutive support itself is allowed to vary, rather than a numerical artifact of the search procedure.
In summary, we have shown that: (i) the hydraulic sector is an exact linear Onsager system with automatic reciprocity and an associated minimum-dissipation theorem; (ii) the structural adaptation rule and its Lyapunov inequality admit a gradient-flow representation through a Rayleighian variational principle, which identifies the structural Onsager coefficient in closed form and predicts an inverse scaling of structural mobility with local maintenance cost; (iii) exact enforcement of the metabolic budget constraint requires a Gaussian-constraint correction to the multiplier, Eq. (20), which takes an explicit closed form, distinguishing a canonical from a micro-canonical treatment of the structural dynamics; (iv) fluctuating demand admits a Langevin extension whose stationary measure defines a partition function over discrete network topologies, reframing loop selection as an order–disorder crossover, although the data-collapse hypothesis suggested by this picture is not supported by the numerical test of § 6.4; and (v) the fluid displaced by wall motion couples both sectors through a symmetric, non-diagonal effective mobility, and the Langevin extension satisfies a second law and an integral fluctuation theorem for the structural sector. Together these results recast a variational optimization result as a predictive non-equilibrium thermodynamic theory, with falsifiable consequences, namely the mobility–cost scaling, the exact-conservation numerical test, the reciprocal, negative cross-mobilities of hydraulically connected vessels, and the time-reversal symmetry of their fluctuations, that go beyond what is accessible from the original formulation.
Natural extensions include relaxing the diagonal (purely local) structural Onsager matrix to admit long-range mechanotransductive couplings beyond the hydraulic coupling of § 4.5, testing Onsager reciprocity for such couplings directly against vascular data, and carrying out the small-network enumeration outlined in § 6 to validate the partition-function description quantitatively, with a threshold-free order parameter for loop selection.
Derivations
Notation
Table A.1 lists the symbols used in the main text.
| Symbol | Meaning |
| , , | radius, length, logarithmic radius of edge |
| , , | conductance, flow, pressure drop () |
| , , | wall shear stress, set-point, |
| , | cost exponent |
| , | length factor, maintenance cost of edge |
| , , | total cost, budget, metabolic multiplier |
| , | dissipation, composite functional |
| , , | rate constant, structural mobility, force |
| , , | metabolic time, structural potential, bare mobility |
| , , | lumen volume, hydraulic friction, effective mobility |
| coupling strength | |
| , , | structural temperature, conducted-signal strength and range |
Table A.1. Notation (reduced units).
Shear set-point
This subsection derives Eq. (3). With the Lagrangian and (Appendix § 8.§ 3), stationarity gives . Since ,
which is Eq. (3). For an arbitrary state the same algebra defines through .
Envelope theorem
At fixed demands the flow minimizes subject to Kirchhoff’s law. The derivative of the minimum with respect to equals the partial derivative of the Lagrangian at the optimum, ; the implicit dependence of the flows on does not contribute.
Structural force and Lyapunov identity
This subsection derives Eqs. (6), (10) and (13). With and the results of Appendices § 8.§ 2 and § 8.§ 3,
With and the rule (4), .
Gaussian multiplier
This subsection derives Eqs. (20)–(22). From Appendix § 8.§ 2, , with independent of . Since , the constraint reads , whose solution is Eq. (20). Along the constrained flow, , because and .
Retraction
Since , a uniform shift of the free edges multiplies their cost by . With and the costs of free and clamped (floor) edges, restores exactly.
Coupled sectors
This subsection derives Eqs. (15)–(17). Stationarity of Eq. (14) with respect to gives (); inserted in the constraint, . The static part, , is the pressure of § 2; the dynamic part, with the source pressure fixed, is . Stationarity with respect to gives Eq. (15); keeping the dynamic pressure only, , i.e. Eq. (16). The flows induced by wall motion, , dissipate , which gives Eq. (17).
Sign of the cross-mobilities
Writing , the off-diagonal elements are to first order. is the Laplacian with the source grounded; on each connected component of the graph without the source it is an irreducible M-matrix, whose inverse has strictly positive entries, and it is block-diagonal across components. Hence for in the same branch, and otherwise.
Second law
This subsection derives Eq. (25). With Eq. (24),
where the second step integrates by parts (the term vanishes) and the third uses .
Laplace prefactor
This subsection derives the prefactor of Eq. (28). Expanding around the minimum on , .
Ornstein–Uhlenbeck entropy production
For with , the stationary covariance solves , and the stationary current is with . The stationary entropy production, which is entirely housekeeping, is
For Eq. (26), and , with the Hessian of . At , solves the Lyapunov equation and , so that and : detailed balance holds for any symmetric . The lagged correlations are ; their antisymmetric part vanishes identically under detailed balance.
Numerical methods
Networks
Square lattices of nodes (, 60 edges; , 40 edges) with node positions perturbed uniformly by up to 0.2 lattice units, so that lengths vary; the source is a corner node, every other node withdraws one unit of flow, and pressures are obtained from the grounded Laplacian. A floor radius prevents singular conductances; edges at the floor with negative velocity are clamped.
Deterministic dynamics
This subsection describes the runs of Table 1. The rule (4) with is integrated by explicit Euler up to from a dense initial state, . The multiplier is either fixed at the value of a reference run (, Gaussian multiplier with retraction) or given by Eq. (20); the retraction is that of Appendix § 8.§ 6.
| Fixed max | Fixed final | Gaussian Euler | Gaussian with retraction | ||
| 0.5 | 0.05 | ||||
| 0.5 | 0.02 | ||||
| 0.5 | 0.005 | ||||
| 1.0 | 0.05 | ||||
| 1.0 | 0.02 | ||||
| 1.0 | 0.005 | ||||
| 1.5 | 0.05 | ||||
| 1.5 | 0.02 | ||||
| 1.5 | 0.005 |
Table 1. Maximum relative budget violation, , for three integration schemes, and final violation for the scheme with a fixed multiplier. Fixed: throughout. Gaussian: multiplier of Eq. (20) with explicit Euler steps, without and with the retraction of Appendix § 8.§ 6. Values below are at machine precision.
Langevin simulations
For Figure 1 we use Euler–Maruyama integration around the optimal tree with , steps, (with the mean edge cost) and fixed ; mobilities are estimated from the covariance of one-step increments, . The coupling strength is set by rescaling so that the median of equals .
Monte Carlo sampling
For Figure 3 we use Metropolis sampling of with single-edge Gaussian proposals of width 0.35 in , reflected at the floor; is updated with the Sherman–Morrison formula and recomputed every 50 sweeps; 1500 sweeps per temperature, the first 30% discarded. Loops are counted as the cycle rank of the subgraph of edges with .
Reciprocity test
For Figure 2, the linearized dynamics of Appendix § 8.11 is simulated exactly in discrete time, with , sampled at with the smallest real part of the eigenvalues of . Measurement noise of standard deviation is added independently to each sample. The test statistic is over parent–child pairs, with the normalized antisymmetric part of the lagged correlation at and its theoretical value; the threshold is the 95% quantile of the statistic for , and the power is estimated from 300 independent realizations per point.
Code availability
All calculations use Python (NumPy, SciPy); the scripts reproduce every figure and table, and are available at https://github.com/Danpc11/setpoint-thermo.
Acknowledgements
This work was supported by DGAPA-PAPIIT under Grant No. IN234029 and by SECIHTI under Grant No. CBF-2025-G-789.
References
- [1]P. F. Davies, “Flow-mediated endothelial mechanotransduction,” Physiol. Rev., vol. 75, pp. 519–560, 1995.pubmed.ncbi.nlm.nih.gov/7624393
- [2]A. Kamiya and T. Togawa, “Adaptive regulation of wall shear stress to flow change in the canine carotid artery,” Am. J. Physiol., vol. 239, pp. H14–H21, 1980.pubmed.ncbi.nlm.nih.gov/7396013
- [3]A. R. Pries, T. W. Secomb, and P. Gaehtgens, “Design principles of vascular beds,” Circ. Res., vol. 77, pp. 1017–1023, 1995.pubmed.ncbi.nlm.nih.gov/7554136
- [4]R. S. Reneman, T. Arts, and A. P. G. Hoeks, “Wall shear stress—An important determinant of endothelial cell function and structure—In the arterial system in vivo. Discrepancies with theory,” J. Vasc. Res., vol. 43, pp. 251–269, 2006.pubmed.ncbi.nlm.nih.gov/16491020
- [5]C. D. Murray, “The physiological principle of minimum work: I. The vascular system and the cost of blood volume,” Proc. Natl. Acad. Sci. USA, vol. 12, pp. 207–214, 1926.pubmed.ncbi.nlm.nih.gov/16576980
- [6]G. B. West, J. H. Brown, and B. J. Enquist, “A general model for the origin of allometric scaling laws in biology,” Science, vol. 276, pp. 122–126, 1997.pubmed.ncbi.nlm.nih.gov/9082983
- [7]J. R. Banavar, F. Colaiori, A. Flammini, A. Maritan, and A. Rinaldo, “Topology of the fittest transportation network,” Phys. Rev. Lett., vol. 84, pp. 4745–4748, 2000.pubmed.ncbi.nlm.nih.gov/10990786
- [8]S. Bohn and M. O. Magnasco, “Structure, scaling, and phase transition in the optimal transport network,” Phys. Rev. Lett., vol. 98, Art. no. 088702, 2007.pubmed.ncbi.nlm.nih.gov/17359138
- [9]M. Durand, “Structure of optimal transport networks subject to a global constraint,” Phys. Rev. Lett., vol. 98, Art. no. 088701, 2007.DOI
- [10]G. Vázquez-Victorio, E. Hernández-Lemus, A. Cervera, L. Gómez-Romero, H. Tovar, and D. Pérez-Calixto, “Metabolic cost sets the shear stress profile of optimal vascular trees,” submitted for publication, 2026.
- [11]F. Corson, “Fluctuations and redundancy in optimal transport networks,” Phys. Rev. Lett., vol. 104, Art. no. 048703, 2010.pubmed.ncbi.nlm.nih.gov/20366745
- [12]E. Katifori, G. J. Szöllősi, and M. O. Magnasco, “Damage and fluctuations induce loops in optimal transport networks,” Phys. Rev. Lett., vol. 104, Art. no. 048704, 2010.pubmed.ncbi.nlm.nih.gov/20366746
- [13]F. Kaiser, H. Ronellenfitsch, and D. Witthaut, “Discontinuous transition to loop formation in optimal supply networks,” Nat. Commun., vol. 11, Art. no. 5796, 2020.DOI
- [14]E. A. Martens and K. Klemm, “Transitions from trees to cycles in adaptive flow networks,” Front. Phys., vol. 5, Art. no. 62, 2017.DOI
- [15]L. Onsager, “Reciprocal relations in irreversible processes. I,” Phys. Rev., vol. 37, pp. 405–426, 1931.DOI
- [16]L. Onsager, “Reciprocal relations in irreversible processes. II,” Phys. Rev., vol. 38, pp. 2265–2279, 1931.DOI
- [17]I. Prigogine, Introduction to Thermodynamics of Irreversible Processes. Springfield, IL, USA: Charles C. Thomas, 1955.
- [18]D. Hu and D. Cai, “Adaptation and optimization of biological transport networks,” Phys. Rev. Lett., vol. 111, Art. no. 138701, 2013.pubmed.ncbi.nlm.nih.gov/24116821
- [19]H. Ronellenfitsch and E. Katifori, “Global optimization, local adaptation, and the role of growth in distribution networks,” Phys. Rev. Lett., vol. 117, Art. no. 138301, 2016.DOI
- [20]A. Tero, R. Kobayashi, and T. Nakagaki, “A mathematical model for adaptive transport network in path finding by true slime mold,” J. Theor. Biol., vol. 244, pp. 553–564, 2007.
- [21]A. Tero et al., “Rules for biologically inspired adaptive network design,” Science, vol. 327, pp. 439–442, 2010.DOI
- [22]A. R. Pries, T. W. Secomb, and P. Gaehtgens, “Structural adaptation and stability of microvascular networks: Theory and simulations,” Am. J. Physiol., vol. 275, pp. H349–H360, 1998.DOI
- [23]A. R. Pries, B. Reglin, and T. W. Secomb, “Structural adaptation of microvascular networks: Functional roles of adaptive responses,” Am. J. Physiol., vol. 281, pp. H1015–H1025, 2001.pubmed.ncbi.nlm.nih.gov/11514266
- [24]V. Bonifaci, K. Mehlhorn, and G. Varma, “Physarum can compute shortest paths,” J. Theor. Biol., vol. 309, pp. 121–133, 2012.
- [25]E. Facca, F. Cardin, and M. Putti, “Towards a stationary Monge–Kantorovich dynamics: The Physarum polycephalum experience,” SIAM J. Appl. Math., vol. 78, pp. 651–676, 2018.DOI
- [26]E. Facca, F. Cardin, and M. Putti, “Branching structures emerging from a continuous optimal transport model,” J. Comput. Phys., vol. 447, Art. no. 110700, 2021.arxiv.org/abs/1811.12691
- [27]S.-S. Chang and M. Roper, “Microvascular networks with uniform flow,” J. Theor. Biol., vol. 462, pp. 48–64, 2019.DOI
- [28]N. Baeyens et al., “Vascular remodeling is governed by a VEGFR3-dependent fluid shear stress set point,” eLife, vol. 4, Art. no. e04645, 2015.DOI
- [29]Y. Qi et al., “Hemodynamic regulation allows stable growth of microvascular networks,” Proc. Natl. Acad. Sci. USA, vol. 121, Art. no. e2310993121, 2024.DOI
- [30]A. Lonardi, E. Facca, M. Putti, and C. De Bacco, “Infrastructure adaptation and emergence of loops in network routing with time-dependent loads,” Phys. Rev. E, vol. 107, Art. no. 024302, 2023.DOI
- [31]G. Palla, I. Derényi, I. Farkas, and T. Vicsek, “Statistical mechanics of topological phase transitions in networks,” Phys. Rev. E, vol. 69, Art. no. 046117, 2004.pubmed.ncbi.nlm.nih.gov/15169079
- [32]F. G. Woodhouse, A. Forrow, J. B. Fawcett, and J. Dunkel, “Stochastic cycle selection in active flow networks,” Proc. Natl. Acad. Sci. USA, vol. 113, pp. 8200–8205, 2016.DOI
- [33]P. Koehl, M. Delarue, and H. Orland, “Optimal transport at finite temperature,” Phys. Rev. E, vol. 100, Art. no. 013310, 2019.DOI
- [34]F. Kramer and C. D. Modes, “How to pare a pair: Topology control and pruning in intertwined complex networks,” Phys. Rev. Res., vol. 2, Art. no. 043171, 2020.DOI
- [35]S. Marbach, N. Ziethen, L. Bastin, F. K. Bäuerle, and K. Alim, “Vein fate determined by flow-based but time-delayed integration of network architecture,” eLife, vol. 12, Art. no. e78100, 2023.DOI
- [36]G. Oster, A. Perelson, and A. Katchalsky, “Network thermodynamics,” Nature, vol. 234, pp. 393–399, 1971.
- [37]G. F. Oster, A. S. Perelson, and A. Katchalsky, “Network thermodynamics: Dynamic modelling of biophysical systems,” Q. Rev. Biophys., vol. 6, pp. 1–134, 1973.pubmed.ncbi.nlm.nih.gov/4576440
- [38]M. Stern, D. Hexner, J. W. Rocks, and A. J. Liu, “Supervised learning in physical networks: From machine learning to learning machines,” Phys. Rev. X, vol. 11, Art. no. 021045, 2021.DOI
- [39]H. B. M. Uy(l)ings, “Optimization of diameters and bifurcation angles in lung and vascular tree structures,” Bull. Math. Biol., vol. 39, pp. 509–520, 1977.
- [40]T. F. Sherman, “On connecting large vessels to small: The meaning of Murray’s law,” J. Gen. Physiol., vol. 78, pp. 431–453, 1981.pubmed.ncbi.nlm.nih.gov/7288393
- [41]Y. Huo and G. S. Kassab, “Intraspecific scaling laws of vascular trees,” J. R. Soc. Interface, vol. 9, pp. 190–200, 2012.DOI
- [42]R. Marchesi, “Beyond Murray’s law: Non-universal branching exponents from vessel-wall metabolic costs,” 2026, arXiv:2603.13687.arxiv.org/abs/2603.13687
- [43]R. S. Seymour, Q. Hu, E. P. Snelling, and C. R. White, “Interspecific scaling of blood flow rates and arterial sizes in mammals,” J. Exp. Biol., vol. 222, Art. no. jeb199554, 2019.DOI
- [44]M. Doi, “Onsager’s variational principle in soft matter,” J. Phys.: Condens. Matter, vol. 23, Art. no. 284118, 2011.DOI
- [45]M. Doi, Soft Matter Physics. Oxford, U.K.: Oxford Univ. Press, 2013.
- [46]A. N. Beris and B. J. Edwards, “Dissipation in nonequilibrium thermodynamics and its connection to the Rayleighian functional,” Phys. Fluids, vol. 36, Art. no. 013102, 2024.DOI
- [47]P. Glansdorff and I. Prigogine, Thermodynamic Theory of Structure, Stability and Fluctuations. London, U.K.: Wiley, 1971.
- [48]U. Seifert, “Entropy production along a stochastic trajectory and an integral fluctuation theorem,” Phys. Rev. Lett., vol. 95, Art. no. 040602, 2005.DOI
- [49]U. Seifert, “Stochastic thermodynamics, fluctuation theorems and molecular machines,” Rep. Prog. Phys., vol. 75, Art. no. 126001, 2012.DOI
- [50]M. Grmela and H. C. Öttinger, “Dynamics and thermodynamics of complex fluids. I. Development of a general formalism,” Phys. Rev. E, vol. 56, pp. 6620–6632, 1997.DOI
- [51]T. Hatano and S.-i. Sasa, “Steady-state thermodynamics of Langevin systems,” Phys. Rev. Lett., vol. 86, pp. 3463–3466, 2001.pubmed.ncbi.nlm.nih.gov/11327999


