Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Efficient Spectral Control of Partially Observed Linear Dynamical Systems
Authors: Anand Brahmbhatt, Gon Buzaglo, Sofiia Druchyna, Elad Hazan
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | C Experiments We present a series of synthetic experiments designed to evaluate the performance of the DSC, as specified in Algorithm 1. We compare DSC against the gradient response controller (GRC) [30] and the linear quadratic gaussian controller (LQG) [8]. We analyze the performance of these three controllers under the following system settings: a linear signal, where the initial state x0 is randomly sampled from the Gaussian distribution; and a signal with the Re LU state transition. For each signal type, we evaluate the performance of controllers under two types of perturbations: (i) Gaussian noise, and (ii) sinusoidal disturbances; and provide 95% confidence intervals for each setting. The empirical results presented in Figures 3a demonstrate that the DSC controller outperforms the GRC controller, while LQG remains optimal for the setting with Gaussian perturbations. Furthermore, under sinusoidal perturbations, DSC maintains a performance advantage over both LQG and GRC, as illustrated in Figure 3b. Similar conclusions on the performance of the DSC controller hold for the experiments with Re LU state transition as shown in Figures 3c and 3d. The confidence intervals further substantiate that the performance gains of DSC over GRC are consistent and statistically robust across random system initializations. |
| Researcher Affiliation | Collaboration | Anand Brahmbhatt1 Gon Buzaglo1 Sofiia Druchyna1 Elad Hazan1,2 1Computer Science Department, Princeton University 2Google Deep Mind Princeton |
| Pseudocode | Yes | Algorithm 1 Double Spectral Control Algorithm 1: Input: Horizon T, number lifting filters h, number of learning filters h, memories m, m, step size η, convex constrains set K R( h+1) n (h+2)p. 2: Compute {(σj, ϕj)}h j=1 and {(λj, φj)} h j=1, the top eigenpairs of a matrices whose i, j th entry is (1 γ)i+j 1 i+j 1 of dimensions m and m respectively. 3: Initialize M 0 i Rn (h+2)p for all i {0, . . . , h}, and z0 = 0 Rd. 4: for t = 0, . . . , T 1 do 5: Perform spectral lifting by computing ynat t = h ynat t σ1/4 0 Y nat t:t mϕ0 . . . σ1/4 h Y nat t:t mϕh i R(h+2)p . 6: Define Y nat t:t m = ynat t . . . ynat t m and compute control ut = M t 0 ynat t + i=1 λ1/4 i M t i Y nat t:t mφi 7: Update zt+1 = Azt + But 8: Observe yt+1 and record ynat t+1 = yt+1 Czt+1. 9: Set M t+1 = ΠK [M t η Mℓt (M t)] 10: end for 11: return M T |
| Open Source Code | No | Our code repository will be open sourced and a link will be provided in the final version. A link to the code will be provided in the final version. However, we omit the Git Hub repository with all of our code at the moment in order not to violate the double blind policy and keep the paper anonymous. We release the code for our simulations. |
| Open Datasets | No | Our research does not involve human participants, and all experiments are simulation-based without the use of any external datasets. |
| Dataset Splits | No | The paper describes using |
| Hardware Specification | No | We will completely describe the necessary computing resources in the supplementary materials. |
| Software Dependencies | No | The paper does not explicitly mention any specific software dependencies or their version numbers in the main text. |
| Experiment Setup | Yes | For the described settings, we use h = h = 5 filters and m = m = 10 memory for both controllers. Each figure reports the average quadratic loss computed over a sliding window, with its size set to 10% of the sequence length. Hyperparameters are selected to reflect representative and stable performance, though the experimental setup generalizes to higher-dimensional systems and alternative perturbation models. |