Katikati ya utafiti huu kuna swali la msingi lakini lenye kina: Tunawezaje kusafirisha probability distribution moja kwenda nyingine kwa njia ya asili zaidi, inayowezekana zaidi au iliyopangwa zaidi? Swali hili linaweza kuonekana abstract kwa mara ya kwanza, lakini liko katika moyo wa maeneo mengi kama modern artificial intelligence, Bayesian statistics, diffusion models, optimal transport, stochastic control na data generation.

Tuanze na mfano wa kila siku. Fikiria tuna particles zilizotawanyika kama wingu. Mwanzoni particles hizi zina distribution fulani. Baada ya muda, tuchukulie kwamba tunaona particles hizo katika distribution nyingine. Ikiwa njia iliyotumika kati ya mwanzo na mwisho haijulikani moja kwa moja, swali linakuwa: Wingu hili la particles lingewezaje kutoka kwenye initial distribution kwenda final distribution kwa njia inayowezekana zaidi huku likibaki karibu iwezekanavyo na sheria ya random motion tunayojua?

Schrödinger bridge problem ilitokana hasa na swali hili. Erwin Schrödinger aliuliza: Ikiwa initial na final distributions za independent particles zinazofanya Brownian motion zinajulikana, ni stochastic process ipi inayoweza kuunganisha observations hizi mbili kwa uwezekano mkubwa zaidi? Kama initial distribution pekee ingejulikana, tungetarajia particles kusambaa kwa Brownian motion. Lakini final distribution inapowekwa pia, tunatafuta kati ya all possible random paths ile inayolingana vyema zaidi na final observation lakini pia inapotosha reference Brownian motion kwa kiwango kidogo zaidi.

Katika utafiti, Schrödinger bridge problem inaandikwa katika dynamic path-measure form kama:

\[ P^\star = \arg\min_{P \in \mathcal{P}(\Omega)} \left\{ KL(P \Vert W_\varepsilon) : P_0 = \rho_0, P_1 = \rho_1 \right\} \]

Katika formula hii P^★ ni optimal path measure inayotafutwa. Ω ni space ya all possible continuous paths. W_ε ni path law ya reference diffusion, mara nyingi Brownian motion. KL(P||W_ε) hupima ni kwa kiasi gani path measure P inatofautiana na reference process katika maana ya relative entropy. P₀ = ρ₀ inaonyesha initial distribution imewekwa, na P₁ = ρ₁ inaonyesha terminal distribution imewekwa.

Maana ya formula hii ni hii: Kati ya all stochastic processes zinazotimiza initial na final distribution constraints, chagua process iliyo karibu zaidi na Brownian motion katika information-theoretic distance. Kwa maneno mengine, mfumo unapounganisha distributions mbili, hauharibu reference random dynamics bila sababu.

Reference process inafafanuliwa katika utafiti kwa diffusion ifuatayo:

\[ dX_t = b_0(t,X_t)dt + \sqrt{\varepsilon}dW_t,\quad X_0 \sim m_0 \]

Hapa X_t ni state variable katika muda t. b₀(t,X_t) ni drift term ya reference process. Mara nyingi katika utafiti b₀ ≡ 0 huchukuliwa kwa Brownian motion. √ε ni diffusion coefficient inayodhibiti magnitude ya randomness. W_t ni Brownian motion. m₀ ni initial law.

Parameter ε ni muhimu sana. Inawakilisha noise level ya reference diffusion na pia regularization weight katika entropic optimal transport problem. ε ikiwa kubwa, transport inakuwa more spread, stochastic na smooth. ε ikiwa ndogo, process hukaribia deterministic optimal transport behavior. Katika uhusiano ambao utafiti unaweka na GBC, ε pia ni knob kuu inayodhibiti generative map iwe smooth kiasi gani au sharp kiasi gani.

Ingawa Schrödinger bridge problem inaonekana kama dynamic path problem, utafiti unakumbusha kwamba inaweza kupunguzwa kuwa static endpoint-matching problem. Path measure hutenganishwa kuwa endpoint coupling na conditional reference bridge law kati ya endpoints. Hivyo problem hubadilika kuwa kutafuta coupling kati ya initial na final distributions:

\[ \pi^\star = \arg\min_{\pi \in \Pi(\rho_0,\rho_1)} KL(\pi \Vert R_{01}) \]

Hapa π^★ ni optimal coupling inayogawanya initial na final points pamoja. Π(ρ₀,ρ₁) ni set ya all couplings zenye marginals ρ₀ na ρ₁. R₀₁ ni joint law ya initial na terminal points chini ya reference diffusion.

Transformation hii ni muhimu. Complexity yote ya dynamic path inajikusanya kwenye problem ya jinsi endpoints zinavyopaswa kuunganishwa. Optimal coupling ikijulikana, paths za kati zinaweza kusampled kwa reference Brownian bridges. Utafiti unaandika uhusiano huu kama:

\[ P^\star(\cdot) = \pi^\star(dx_0,dx_1) W_\varepsilon(\cdot \mid X_0=x_0, X_1=x_1) \]

Katika formula hii W_ε(·|X₀=x₀,X₁=x₁) ni reference Brownian bridge yenye fixed initial point x₀ na final point x₁. Yaani kwanza endpoints huchaguliwa kwa optimal coupling, kisha Brownian bridge inasampled kati ya points hizi mbili. Muundo huu ndiyo msingi wa numerical scheme ya utafiti.

Chini ya Brownian reference, static Schrödinger problem inakuwa entropic optimal transport problem:

\[ \pi^\star = \arg\min_{\pi \in \Pi(\rho_0,\rho_1)} \int \frac{\|x-y\|^2}{2}d\pi(x,y) - \varepsilon H(\pi) + const \]

Hapa term ya kwanza ni quadratic cost ya kusafirisha mass kutoka point x kwenda y. H(π) ni entropy ya coupling. Term εH(π) huzuia transport isiwe rigid na deterministic mno na hutoa regularized, smoother solution. Kwa hiyo problem inaweza kusomwa kama optimal transport na pia probabilistic diffusion.

Schrödinger factorization inatolewa kama:

\[ \pi^\star(dx,dy) = \phi(x)\psi(y)R_{01}(dx,dy) \]

Katika formula hii φ(x) na ψ(y) ni potentials kwenye initial na terminal endpoints. Potentials hizi zinaonyesha jinsi optimal coupling inavyoreweight reference coupling. Potentials hizo hutimiza Schrödinger system:

\[ \phi(x)\int p_1^\varepsilon(y|x)\psi(y)dy = \rho_0(x) \]

\[ \psi(y)\int p_1^\varepsilon(y|x)\phi(x)dx = \rho_1(y) \]

Hapa p₁^ε(y|x) ni transition density ya reference Brownian diffusion. Suluhisho la system hii kwenye discrete grid hupatikana kwa iterative proportional fitting algorithm inayojulikana kama Sinkhorn au IPF. Kwa classic form, kernel matrix ni:

\[ K_{ij} = \exp\left(-\frac{\|x_i-y_j\|^2}{2\varepsilon}\right) \]

Ikiwa initial na target discrete weights ni a na b, Sinkhorn iteration inategemea fixed-point idea:

\[ u \odot (Kv) = a,\quad v \odot (K^\top u)=b \]

Expression hii inaeleza row na column scalings zinazorekebisha coupling matrix mpaka marginals zake mbili zilingane na target. Jambo muhimu la utafiti ni kuunganisha classical solution hii na simulation-and-regression logic, bila kulazimika kutatua density au PDE moja kwa moja.

Stochastic-control form ya Schrödinger bridge pia ni mojawapo ya main connections za utafiti. Dynamic problem inaweza kuandikwa kama control problem:

\[ P^\star \leftrightarrow \min_u \frac{1}{2\varepsilon}E\int_0^1 \|u(t,X_t)\|^2dt \quad \text{s.t. } X_1 \sim \rho_1 \]

Hapa u(t,X_t) ni drift field ya controlled diffusion. Formula inaeleza kupunguza control energy inayohitajika kusafirisha system kutoka initial distribution kwenda target distribution. Coefficient 1/(2ε) inaunganisha noise level na control cost.

Hapa ndipo moja ya formulas muhimu zaidi ya utafiti inapojitokeza. Optimal drift field inaweza kuandikwa kama conditional expectation bila kuhesabu density au score function:

\[ u^\star(t,x) = E_{\pi^\star}\left[\frac{X_1-X_t}{1-t}\mid X_t=x\right],\quad 0 \leq t < 1 \]

Katika formula hii u^★(t,x) ni optimal drift katika state x na time t. (X₁-X_t)/(1-t) ni bridge velocity inayolenga kutoka current point kwenda terminal point. Conditional expectation hutoa average directing effect ya possible terminal targets kwa t na x ileile.

Practical meaning ya formula hii ni kubwa. Ikiwa una simulated trajectories, unajua X_t na terminal target X₁ kwa kila trajectory. Basi target velocity inaweza kuhesabiwa. Kisha regression model inaweza kujifunza ku-predict target velocity kutoka inputs t na X_t. Hivyo optimal drift field inaweza kujifunzwa bila density estimation, score learning au PDE solving.

Uhusiano na GBC, Generative Bayesian Computation, hujengwa hapa. GBC hutazama Bayesian posterior sampling kama transport-map learning problem. Katika classic Bayesian inference, posterior ni:

\[ p(\theta|y) \propto p(y|\theta)p(\theta) \]

Katika problems nyingi halisi ni vigumu kusample moja kwa moja kutoka distribution hii. GBC hujifunza map inayobadilisha simple random variable kuwa posterior sample:

\[ \theta = H(\tau,y),\quad \tau \sim U(0,1)^k \]

Hapa τ ni simple base random variable kama uniform au Gaussian. H ni learned map inayobadilisha τ kuwa posterior sample conditional on observed data y. Map hujifunzwa ili H(·,y) isukume base measure kwenda p(θ|y) posterior.

Distinctive feature ya GBC ni kwamba haihitaji ku-evaluate posterior density moja kwa moja. Badala yake, model husimulated:

\[ \theta^{(i)} \sim p(\theta),\quad y^{(i)} \sim p(y|\theta^{(i)}) \]

Kwenye simulation table hii, deep quantile network au generative network hujifunza conditional relationship kati ya θ na y. Observed y^★ inapopatikana, new τ values huingizwa kwenye network na posterior samples huzalishwa. Hii ni likelihood-free approach, yaani haihitaji direct likelihood evaluation.

Kwa upande wa Schrödinger bridge, kwa kuwa drift formula ni conditional expectation, inaungana naturally na regression logic ya GBC. Kwa hiyo utafiti unaposema “GBC solves Schrödinger bridge,” maana yake ni: drift katika IPF/Sinkhorn half-bridge steps ni conditional expectation ya simulated bridge velocities; expectation hii inaweza kujifunzwa kwa GBC-style regression.

Wazo hili linatolewa kama GBC–IPF algorithm. Main steps ni: kuchukua samplers za ρ₀ na ρ₁, kuunda time grid, na kuchagua initial drift kuwa zero kwa Brownian motion. Katika kila iteration, backward half-bridge inayolazimisha terminal marginal hujifunzwa kwanza, kisha forward half-bridge inayolazimisha initial marginal. Katika kila hatua, simulated velocities hu-regressed na network.

Katika backward half-bridge target velocity ina muundo:

\[ v = \frac{X_1-X_t}{1-t} \]

Katika forward half-bridge, target velocity kuelekea mwanzo kwa reverse time inaandikwa:

\[ v = \frac{X_0-X_t}{t} \]

Targets hizi zote hazihitaji density; zinategemea tu points kwenye simulated trajectory. Kwa hiyo utafiti unawasilisha method kama “simulation-only” Schrödinger bridge solver.

Utafiti pia unafanya distinction muhimu na score-based diffusion models. Score-based methods kawaida hujifunza quantities kama ∇logρ_t. Mbinu hapa hujifunza drift ambayo ni conditional expectation moja kwa moja. Hii inaweza kuwa practical zaidi katika baadhi ya hali kwa sababu target ni regression ya simulated velocities zinazoonekana.

Kwa upande mwingine, Schrödinger bridge hutoa geometric interpretation ya GBC. Transport map inayojifunzwa na GBC inaweza kuonekana kama time-one flow ya bridge kutoka base measure kwenda target posterior. Utafiti unaandika:

\[ (F_{0\to1}^{\varepsilon})_\#\rho_0 = \rho_1 \]

Hapa F_0→1^ε ni Schrödinger-bridge flow map. Symbol # inaonyesha pushforward operation; yaani samples kutoka ρ₀ zikipitishwa kwenye map zinatoa ρ₁. Kwa GBC, ρ₀ inaweza kuwa simple base distribution na ρ₁ target posterior.

Katika deterministic limit:

\[ \varepsilon \to 0 \]

Schrödinger bridge huconcentrate kwenye Monge–Brenier optimal transport map:

\[ T = \nabla \varphi,\quad T_\#\rho_0 = \rho_1 \]

Katika dimension moja, map hii ni monotone quantile map:

\[ T = Q_{\rho_1}\circ F_{\rho_0} \]

Hapa F_ρ0 ni CDF ya initial distribution na Q_ρ1 ni quantile function ya target distribution. Hii ndiyo map ambayo GBC hujifunza moja kwa moja katika dimension moja. Kwa hiyo utafiti unadai kwamba deterministic transport target ya GBC ni ε → 0 limit ya Schrödinger bridge.

Connection hii ni interesting pia kwa Bayesian inference. Ikiwa initial distribution ni prior na terminal potential inahusishwa na likelihood, posterior inaweza kutafsiriwa kama terminal marginal ya prior-to-posterior Schrödinger bridge. Hii inaruhusu Bayes rule kuonekana si tu kama static reweighting bali kama entropic na stochastic path kutoka prior kwenda posterior.

Hata hivyo, utafiti unaweka limitation muhimu hapa. Algorithm 1 inahitaji sampler kutoka target distribution ρ₁. Katika generative modeling examples hii si tatizo kwa sababu target distribution inajulikana wazi. Lakini kwa Bayesian posterior, sampling kutoka ρ₁ ndiyo problem yenyewe inayotakiwa kutatuliwa. Kwa hiyo prior-to-posterior bridge connection ni geometric na theoretical interpretation zaidi kuliko direct algorithmic solution. Katika practice, GBC hufanya amortized inversion kupitia joint-model simulation bila kudhani direct posterior sampler.

Numerical validation section ina examples tatu. Ya kwanza ni Gaussian-to-Gaussian Schrödinger bridge. Initial distribution ni:

\[ \rho_0 = N(0,0.7^2) \]

na target distribution ni:

\[ \rho_1 = N(3,1.1^2) \]

Kwa diffusion coefficient ε = 0.5, simulated bridge trajectories zinaanza kwenye tight Gaussian, zinaenea katikati chini ya reference noise, na mwisho zinaconcentrate tena kwenye wider terminal Gaussian. Left panel ya Figure 1 inaonyesha trajectories hizi na mean path; right panel inaonyesha kwamba marginals katika t = 0, t = 0.5 na t = 1 zinakubaliana na analytic Gauss–Markov marginals.

Katika Gaussian special case kuna closed-form coupling. Cross-covariance kwa scalar Gaussian bridge inatolewa kama:

\[ c = \frac{1}{2}\left(\sqrt{4\sigma_0^2\sigma_1^2+\varepsilon^2}-\varepsilon\right) \]

Optimal coupling covariance matrix ni:

\[ \Sigma^\star = \begin{pmatrix} \sigma_0^2 & c \\ c & \sigma_1^2 \end{pmatrix} \]

Hapa σ₀ na σ₁ ni standard deviations za initial na target Gaussian distributions. c ni cross-covariance kati ya initial na terminal points. ε inapopungua, c hukaribia comonotone-coupling value ya deterministic optimal transport. ε inapoongezeka, c hukaribia zero, ikimaanisha endpoints zinakuwa more independent.

Katika validation table, cross-covariance values zilizojifunzwa kwa Sinkhorn/IPF zinalinganishwa na closed-form solution. Kwa ε = 0.50 value ni 0.5596, kwa ε = 0.20 ni 0.6765, na kwa ε = 0.05 ni 0.7454. Katika hali zote tatu absolute error ni chini ya 10^-4. Hii inaonyesha kwamba kwenye one-dimensional Gaussian test numerical method inakadiria analytic solution kwa accuracy ya juu sana.

Left panel ya Figure 2 inaonyesha Sinkhorn/IPF marginal violation ikipungua geometrically na iteration count. Lakini ε inapopungua convergence inakuwa polepole. Sababu ni kwamba kernel inakuwa sharper na entropic transport problem inakuwa numerically stiffer. Right panel inaonyesha learned coupling cross-covariance points karibu zikiangukia closed-form curve.

Example ya pili ni transition kutoka Gaussian kwenda bimodal mixture target:

\[ \rho_0 = N(0,1) \]

\[ \rho_1 = \frac{1}{2}N(-3,0.55^2)+\frac{1}{2}N(3,0.55^2) \]

Example hii ni interesting zaidi kwa generative modeling kwa sababu simple unimodal noise distribution inasafirishwa kwenda structured target distribution yenye modes mbili. Kwa ε = 0.15, bridge trajectories zinaanza kutoka single Gaussian cloud na kugawanyika kuwa terminal modes mbili kwa muda. Figure 3 inaonyesha hili wazi: left panel trajectories zinagawanyika katika branches mbili; right panel inaonyesha endpoint marginals katika t = 0 na t = 1 kwa simulated histograms na analytic density curves.

Point muhimu hapa ni multimodal target. Multimodal distributions ni ngumu katika generative models kwa sababu model inapaswa kugawanya mass kuelekea modes mbili au zaidi bila kuiangusha kwenye one averaged region. Schrödinger bridge hufanya hii kama entropic na stochastic transport. Utafiti unawasilisha ε = 0.15 kama balance inayoweka paths smooth vya kutosha huku ikihifadhi mode separation.

Example ya tatu ni kujifunza drift field kwa simulation katika multimodal target. Target velocity:

\[ \frac{X_1-X_t}{1-t} \]

inahesabiwa kutoka simulated trajectories na regressed kwenye X_t. Degree-5 polynomial basis na ridge regularization hutumiwa. Kwa more complex au high-dimensional targets, deep quantile network inapendekezwa badala yake.

Figure 4 inaonyesha learned drift field. Hii ni mojawapo ya figures zenye maelezo zaidi. Drift inafanya kama branching separator karibu na x = 0. Kwenye regions juu ya origin, drift ni positive na inaelekeza kuelekea upper terminal mode. Kwenye regions chini ya origin, drift ni negative na inasukuma kuelekea lower terminal mode. Karibu na x = 0 kuna unstable ridge au saddle line inayotenganisha basins mbili.

Branching drift field hii inaonyesha kile generative model inapaswa kujifunza. Ili kubadilisha simple initial noise kuwa bimodal target, model inapaswa kugawanya mass kwa state-dependent way, si randomly tu. Utafiti unaonyesha kwamba directing field hii inaweza kupatikana kupitia simulated trajectories na regression pekee.

Scientific importance ya utafiti inaweza kutathminiwa katika levels kadhaa. Kwanza, inaweka conceptual unity kati ya Schrödinger bridge na GBC. Upande mmoja kuna entropic optimal transport na stochastic control, upande mwingine simulation-based Bayesian transport. Utafiti unaonyesha maeneo haya mawili yanaweza kuunganishwa kupitia conditional expectation na transport idea ileile.

Pili, unasisitiza possibility ya kujifunza Schrödinger bridge bila density au PDE solving. Hii ni muhimu kwa modern generative models. Katika real data distributions, densities mara nyingi hazijulikani; kuna samples au simulators tu. Ikiwa bridge drift inaweza kujifunzwa kama conditional expectation, more flexible methods zinaweza kutengenezwa kwa high-dimensional na implicit models.

Tatu, unatoa geometric perspective mpya kwa Bayesian inference. Posterior inaweza kuonekana si tu kama product ya prior na likelihood bali pia kama terminal distribution ya entropic transport path kutoka prior kwenda posterior. Hii inaweza kuwa useful hasa katika annealed sampling, diffusion-based posterior sampling na amortized inference.

Kwa daily-life relevance, utafiti huu hautoi direct device au software product; lakini unaunganisha jinsi AI systems zinavyogenerate data na kufanya inference chini ya uncertainty na stronger mathematical foundations. Generative models nyingi za leo hubadilisha simple noise kuwa complex data. Utafiti unaonyesha transformation hii inaweza kueleweka kwa lugha ya entropic optimal transport na Schrödinger bridges.

Kwa mtazamo wa historia, umuhimu ni kuunganisha problem ya particle-cloud iliyotolewa na Schrödinger katika miaka ya 1930 na modern diffusion models na generative Bayesian computation. Kwa sasa, unasaidia kusoma score-based diffusion, flow matching, Sinkhorn na GBC, ambazo zinaonekana tofauti, kama sehemu ya mathematical transport family moja. Kwa siku zijazo, unaweza kutoa msingi wa simulation-only algorithms kwa high-dimensional posterior sampling na multimodal generative modeling.

Nguvu za utafiti ni pamoja na explicit mathematical mapping, validation kwa closed-form Gaussian example, demonstration ya branching drift field kwenye multimodal target, na clear description ya simulation-regression framework isiyohitaji density/PDE. Pia inaonyesha wazi kwamba ε si technical regularizer tu bali meaningful control knob kati ya deterministic transport na stochastic bridge.

Vikwazo pia ni wazi. Numerical examples zote ni one-dimensional. Discrete Sinkhorn grid na polynomial drift basis zilizotumika hazi-scale kwa high dimensions. Waandishi wanasema hili wazi na wanapendekeza log-domain Sinkhorn stabilization na deep networks kwa high dimensions. Pia, direct target-sampling assumption ya ρ1 kwa Bayesian posterior inaweza kuwa algorithmically circular; utafiti unatenganisha theoretical bridge interpretation na practical GBC training.

Pia lazima kutenganishwe kile utafiti hausisemi. Haujasolve Bayesian problems zote za high dimension. Hauwasilishi full scaled implementation kwa deep networks. Hautoi comprehensive real-data benchmarks. One-dimensional Gaussian na bimodal examples ni demonstrations za logic, si general success guarantee. Main contribution ni kufafanua mathematical connection kati ya Schrödinger bridges na generative Bayesian computation na kuonyesha jinsi connection hii inaweza kutafsiriwa kuwa simulation-based algorithm.

Mbinu na Matokeo ya Utafiti

Mbinu ya utafiti ina-formulate Schrödinger bridge problem katika levels tatu: dynamic path-measure problem, static coupling problem na stochastic-control problem. Kisha inaunganisha structure hii na simulation-based transport-map learning ya GBC.

1. Dynamic Schrödinger bridge problem

Reference diffusion:

\[ dX_t = b_0(t,X_t)dt+\sqrt{\varepsilon}dW_t,\quad X_0\sim m_0 \]

Dynamic problem:

\[ P^\star = \arg\min_{P \in \mathcal{P}(\Omega)} \left\{ KL(P \Vert W_\varepsilon) : P_0=\rho_0,\; P_1=\rho_1 \right\} \]

Problem hii hupata path measure iliyo karibu zaidi na reference Brownian motion huku initial na terminal marginals zikibaki fixed.

2. Static coupling problem

Dynamic problem hupunguzwa kuwa optimal coupling kati ya endpoints:

\[ \pi^\star = \arg\min_{\pi \in \Pi(\rho_0,\rho_1)} KL(\pi \Vert R_{01}) \]

Optimal dynamic bridge hujengwa upya kama:

\[ P^\star(\cdot) = \pi^\star(dx_0,dx_1)W_\varepsilon(\cdot|X_0=x_0,X_1=x_1) \]

3. Entropic optimal transport form

Chini ya Brownian reference, problem inakuwa:

\[ \pi^\star = \arg\min_{\pi \in \Pi(\rho_0,\rho_1)} \int \frac{\|x-y\|^2}{2}d\pi(x,y)-\varepsilon H(\pi)+const \]

Hapa ε hudhibiti coupling iwe distributed na entropic kiasi gani. ε → 0 inapofanyika solution hukaribia deterministic optimal transport; ε ikiongezeka coupling inakuwa more diffuse.

4. Schrödinger system na Sinkhorn/IPF

Schrödinger factorization:

\[ \pi^\star(dx,dy)=\phi(x)\psi(y)R_{01}(dx,dy) \]

Potential system:

\[ \phi(x)\int p_1^\varepsilon(y|x)\psi(y)dy=\rho_0(x) \]

\[ \psi(y)\int p_1^\varepsilon(y|x)\phi(x)dx=\rho_1(y) \]

Discrete-grid kernel:

\[ K_{ij}=\exp\left(-\frac{\|x_i-y_j\|^2}{2\varepsilon}\right) \]

Sinkhorn/IPF fixed point:

\[ u\odot(Kv)=a,\quad v\odot(K^\top u)=b \]

Structure hii hutumika katika static-coupling validations za utafiti.

5. Stochastic control na drift formula

Control problem:

\[ \min_u \frac{1}{2\varepsilon}E\int_0^1\|u(t,X_t)\|^2dt\quad \text{s.t. } X_1\sim \rho_1 \]

Conditional-expectation representation ya optimal drift:

\[ u^\star(t,x)=E_{\pi^\star}\left[\frac{X_1-X_t}{1-t}\mid X_t=x\right] \]

6. Transport map ya GBC

GBC huandika posterior sampling kwa map:

\[ \theta = H(\tau,y),\quad \tau\sim U(0,1)^k \]

Simulation table:

\[ \theta^{(i)}\sim p(\theta),\quad y^{(i)}\sim p(y|\theta^{(i)}) \]

Structure hii huwezesha conditional transport map kujifunzwa kwa simulation bila posterior-density evaluation.

7. Kiini cha GBC–IPF algorithm

8. Schrödinger-bridge interpretation ya GBC map

Bridge flow map:

\[ (F_{0\to1}^{\varepsilon})_\#\rho_0=\rho_1 \]

Deterministic limit:

\[ \varepsilon\to0 \]

Monge–Brenier map:

\[ T=\nabla\varphi,\quad T_\#\rho_0=\rho_1 \]

One-dimensional quantile map:

\[ T=Q_{\rho_1}\circ F_{\rho_0} \]

Matokeo yanaonyesha deterministic transport map inayojifunzwa na GBC inaweza kutafsiriwa kama zero-noise limit ya Schrödinger bridge.

9. Gaussian bridge validation

Distributions zilizotumika:

\[ \rho_0=N(0,0.7^2),\quad \rho_1=N(3,1.1^2) \]

Closed-form cross-covariance:

\[ c=\frac{1}{2}\left(\sqrt{4\sigma_0^2\sigma_1^2+\varepsilon^2}-\varepsilon\right) \]

Validation hii inaonyesha katika one-dimensional Gaussian case, method inakadiria analytic coupling kwa accuracy ya juu sana.

10. Multimodal target example

Initial distribution:

\[ \rho_0=N(0,1) \]

Target distribution:

\[ \rho_1=\frac{1}{2}N(-3,0.55^2)+\frac{1}{2}N(3,0.55^2) \]

Diffusion parameter:

\[ \varepsilon=0.15 \]

11. Kujifunza drift field kwa simulation

Regression target:

\[ \frac{X_1-X_t}{1-t} \]

12. Maana ya kisayansi ya figures

• Gaussian bridge figure: Inaonyesha sample trajectories kutoka ρ₀ = N(0,0.7²) kwenda ρ₁ = N(3,1.1²) na marginals katika t = 0, 0.5, 1. Bridge inatoa smooth Gauss–Markov evolution kutoka mwanzo kwenda target.
• Sinkhorn/IPF convergence figure: Inaonyesha marginal violation ikipungua geometrically kwa iterations, lakini convergence inapungua ε ikipungua. Figure hiyo hiyo inaonyesha learned cross-covariance ikilingana na closed-form solution.
• Multimodal bridge figure: Inaonyesha unimodal Gaussian initial distribution ikigawanyika kwenda bimodal mixture target. Hii inawakilisha generative-modeling problem ya kuunda structured data kutoka simple noise.
• Drift-field figure: Inaonyesha learned drift ikigawanyika karibu na x = 0, ikielekeza juu kwenye positive mode na chini kwenye negative mode. Field hii ndiyo directing structure ambayo generative sampler inapaswa kujifunza kwa multimodal target.

13. Main findings

• Schrödinger-bridge drift inaweza kuandikwa kama conditional expectation.
• Conditional expectation hii inaweza kujifunzwa kwa regression kutoka simulated trajectories.
• Simulation-based transport-map learning ya GBC ni natural tool ya Schrödinger-bridge solution.
• Deterministic quantile/transport map ya GBC inaweza kutafsiriwa kama ε → 0 limit ya Schrödinger bridge.
• Katika Gaussian bridge example, Sinkhorn/IPF results zinaendana na closed-form cross-covariance kwa accuracy bora kuliko four significant digits.
• Katika Gaussian-to-bimodal example, method inaweza kujifunza branching drift field inayogawanya mass kwenda modes mbili kwa simulation.

14. Nguvu na vikwazo

Chanzo na Dokezo la Mbinu

Makala hii imeandaliwa kwa kutegemea utafiti “Generative Bayesian Computation for Schrödinger Bridges” ulioandaliwa na Nicholas G. Polson na Vadim Sokolov. Utafiti unatoa affiliation ya Polson kama University of Chicago Booth School of Business, na ya Sokolov kama George Mason University Department of Systems Engineering and Operations Research. Maandishi yana taarifa “First draft: May 2026” na “This version: June 1, 2026”.

Kwa kuzingatia aina ya chanzo, muundo na tarehe, kazi hii inapaswa kutathminiwa kama academic research note / preprint ya theoretical na numerical study. Kwa kuwa hakuna peer-reviewed journal acceptance, DOI, conference acceptance au open-review information inayoweza kuthibitishwa kutoka kwenye maandishi, inafaa kutumia maelezo utafiti ambao peer-review yake haiwezi kuthibitishwa kutoka kwenye maandishi.

Katika kuandaa maudhui haya, Schrödinger bridge formulations, entropic optimal transport relation, Schrödinger system, Sinkhorn/IPF iteration, stochastic-control interpretation, conditional-expectation drift formula, GBC transport map, GBC–IPF algorithm, Gaussian-bridge closed-form validation, multimodal-target example, drift-field figure na limitations katika discussion zimetumika kama msingi. Hakuna madai ambayo hayapo katika maandishi—kama high-dimensional application success, real-data benchmark result, completed deep-network implementation, universal posterior-sampling solution au peer-reviewed publication acceptance—yaliyoongezwa.

Utafiti unapendekeza strong theoretical connection kati ya Schrödinger bridges na generative Bayesian computation, lakini numerical demonstrations zimewekewa mipaka ya one-dimensional examples. Kwa hiyo matokeo hayapaswi kusomwa kama direct success guarantee kwa high-dimensional generative modeling au complex Bayesian posterior, bali kama mathematical na algorithmic foundation kwa maeneo haya.