Sign in

Rasmus Pagh

@rasmuspagh.net
654 followers 330 following 213 posts

Professor of computer science at University of Copenhagen. Interested in random things & their application (especially to algorithms and privacy). rasmuspagh.net

PostsRepliesMedia
Rasmus Pagh @rasmuspagh.net · 29/09/2026
Very interesting report about AI and theoretical computer science, based on a recent meeting at @simonsinstitute.bsky.social simons.berkeley.edu/ai-tcs-worki... These are times of change, but I am optimistic that there will always be a valuable place for humans who seek to understand.
simons.berkeley.edu
AI + TCS Working Group
0123
Rasmus Pagh @rasmuspagh.net · 17/09/2026
An intense 1st EuroTDP workshop has concluded, with many great talks, interesting posters and more colleagues to talk to than time would allow. Looking forward to the next time we can gather the privacy community in Europe!
Participants entering the EuroTDP venue
080
Rasmus Pagh @rasmuspagh.net · 16/09/2026
Th
000
Rasmus Pagh @rasmuspagh.net · 16/09/2026
Since a BLT is a sandwich, and our data structure is filled to the limit with contents, we found it suitable to name it the "Stuffed IBLT". It is always good to see such long-running projects come to a conclusion, and the (tasty?) paper is now out on arXiv: arxiv.org/abs/2609.17487
arxiv.org
Stuffed IBLTs: Optimal Linear Multiset Sketches
A \emph{linear sketch} is a randomized linear mapping of a vector $v$ to a lower dimensional sketch vector, designed to preserve relevant information about $v$. We consider sketches of vectors $v \in ...
130
Rasmus Pagh @rasmuspagh.net · 16/09/2026
Together with Jonas Klausen who joined the project in 2023 we figured out how to not only reduce the space usage of IBLTs to almost optimal while retaining efficiency, but also make them more reliable by substantially decreasing error probability and how to do it with efficient hash functions.
110
Rasmus Pagh @rasmuspagh.net · 16/09/2026
Back in 2022, Stefan Walzer visited Copenhagen and we talked about getting "best of both worlds" results with near-optimal space efficiency and efficient operations. Our first thought was to combine the known techniques to a form of bucketed IBLTs. But it turns out there is a better way...
110
Rasmus Pagh @rasmuspagh.net · 16/09/2026
There exist other linear methods for set reconciliation that allow recovery of multisets all the way up to the information-theoretic barrier (the number of entries needed to store a set of elements and their counts). However, unlike IBLTs these methods have slow update and recovery times.
110
Rasmus Pagh @rasmuspagh.net · 16/09/2026
IBLTs can be used for *set reconciliation*, where we must compute the symmetric difference A∆B of multisets A and B held by different servers. If A∆B is small enough, IBLT(A)-IBLT(B) can be used to recover its elements. Servers need only communicate the IBLTs, independent of the sizes of A and B.
110
Rasmus Pagh @rasmuspagh.net · 16/09/2026
An interesting property is that while an IBLT vector has fixed dimension d, it can be used to represent sets of arbitrary size. However, it is only possible to recover A from IBLT(A) if A is sufficiently small (at most of size some constant fraction of d).
110
Rasmus Pagh @rasmuspagh.net · 16/09/2026
An IBLT (Goodrich & Mitzenmacher 2011) is an integer vector encoding a multiset in a linear way: IBLT(A)+IBLT(B)=IBLT(A∪B). Subtracting, IBLT(A)-IBLT(B) is an IBLT where the multiplicity of x is the difference between its multiplicity in A and its multiplicity in B (negative multiplicities are ok).
130
Rasmus Pagh @rasmuspagh.net · 14/09/2026
An addition to make the result more clear: The polynomial coefficients are known ahead of time and it is allowed to do preprocessing, not counted in the number of multiplications. The new bound is the number of multiplications needed after receiving an input x.
030
Rasmus Pagh @rasmuspagh.net · 14/09/2026
I have known about this result for about 3 years, but since Thomas and Jakob are working in industry it took some time to get the paper written. A nice contrast to the rush to publish that we see too often, and as a bonus the paper is (partly) Lean formalized. Congratulations on this achievement!
141
Rasmus Pagh @rasmuspagh.net · 14/09/2026
In a new result published on arXiv, former BARC PhD students @thomasahle.bsky.social and Jakob Bæk Tejs Houen make great progress on this question: floor(n/2)+1 multiplications suffice for any n. Great to see progress on a classic problem with such a clean result! arxiv.org/abs/2609.06022
arxiv.org
160
Rasmus Pagh @rasmuspagh.net · 14/09/2026
A 1972 result by Rabin and Winograd showed that fewer multiplications suffice for large enough n. Specifically, n/2+O(log n) multiplications suffice for polynomials over the reals (as well as some other fields). But the O(log n) term makes the method impractical for low-degree polynomials.
120
Rasmus Pagh @rasmuspagh.net · 14/09/2026
Horner's rule is one of the first nontrivial algorithms many students see: Rather than using a quadratic number of multiplications to evaluate a degree-n polynomial term-by-term, it suffices to use n-1 multiplications. This optimization is important, e.g., for evaluating polynomial hash functions.
181
Reposted by Rasmus Pagh
Claudio Orlandi @claudiorlandi.bsky.social · 07/09/2026
Registration for the PPML workshop in Aarhus November 2-4 is now open. The deadline for talk submissions is also this week (September 10). Feel free to reshare! p1dpml.github.io/workshops/wo...
p1dpml.github.io
Workshop 2 | Data Privacy in Machine Learning
042
Rasmus Pagh @rasmuspagh.net · 07/09/2026
The deadline for talk submissions is **September 10**. Feel free to submit also contributions that are not explicitly about distributed privacy, but relevant to the broader area.
000
Rasmus Pagh @rasmuspagh.net · 07/09/2026
Upcoming workshop in Aarhus on Privacy-preserving Machine Learning and Privacy in Distributed Settings in early November. Registration: p1dpml.github.io/workshops/wo... This is a great opportunity to learn about new developments in the privacy/cryptography intersection!
p1dpml.github.io
Workshop 2 | Data Privacy in Machine Learning
110
Rasmus Pagh @rasmuspagh.net · 03/09/2026
Congratulations to Susanna Rezende, Or Zamir, and other ERC Starting Grant recipients!
050
Reposted by Rasmus Pagh
Foundations of Responsible Computing @forcconf.bsky.social · 17/08/2026
Recordings from FORC 2026 are now available! Please check them out. Also, subscribe to FORC's new YouTube channel while you're at it! www.youtube.com/playlist?lis...
youtube.com
FORC 2026 - YouTube
Talk recordings from FORC 2026
056
Reposted by Rasmus Pagh
Claudio Orlandi @claudiorlandi.bsky.social · 12/08/2026
Join us in Aarhus, Nov 2–4, for the Workshop on Privacy-Preserving Machine Learning and Privacy in Distributed Settings! We invite talks on relevant topics including federated learning, differential privacy, MPC, attacks, and more. Deadline: Sept 10 p1dpml.github.io/workshops/wo...
p1dpml.github.io
Workshop 2 | Data Privacy in Machine Learning
074
Rasmus Pagh @rasmuspagh.net · 12/08/2026
The organizers (@hannah-keller.bsky.social @claudiorlandi.bsky.social @amartyasanyal.bsky.social and myself) look forward to reading your (non-archival) submissions! Submission Deadline: September 10, 2026.
000
Rasmus Pagh @rasmuspagh.net · 12/08/2026
We are excited to invite submissions for our upcoming Workshop on Privacy-preserving Machine Learning and Privacy in Distributed Settings! The workshop will be held November 2-4 at Aarhus University in Aarhus, Denmark and is part of the @aicentre.dk Program on Data Privacy in Machine Learning.
p1dpml.github.io
Workshop 2 | Data Privacy in Machine Learning
192
Reposted by Rasmus Pagh
Ted @desfontain.es · 16/07/2026
Over the past year, I've had the honor to work with the European Commission to help them understand Google's (terrible) approach to sharing anonymised search data with competitors under the Digital Markets Act, and design an alternative solution that preserves more utility. [1/3]
195
Rasmus Pagh @rasmuspagh.net · 12/06/2026
Talks were not recorded, but maybe the speakers could be convinced to share their slides
120
Rasmus Pagh @rasmuspagh.net · 12/06/2026
bsky.app/profile/rasm...
010
Rasmus Pagh @rasmuspagh.net · 12/06/2026
Great to see participants from all over the world! The contributed talks have been very high quality, ranging from using formal methods to verify DP implementations, to new ways of reasoning about approximation algorithms under continual observation, to privacy in distributed ML and analytics.
110
Rasmus Pagh @rasmuspagh.net · 12/06/2026
For a quick overview I can recommend dl.acm.org/doi/pdf/10.1... For more in-depth treatment see dimacs.rutgers.edu/~graham/ssbd... (which is also a great book for teaching). For applications in linear algebra see arxiv.org/abs/1411.4357
dl.acm.org
010
Rasmus Pagh @rasmuspagh.net · 12/06/2026
(I just realized that in fact the first paper co-authored with Sia came out a few weeks back, but this paper was our first project together so it is first in this sense.)
000
Rasmus Pagh @rasmuspagh.net · 12/06/2026
Finally, we present two applications based on Private CountSketch: An improved range counting data structure, and private sketches for join size (aka. F2) estimation.
110
Rasmus Pagh @rasmuspagh.net · 12/06/2026
The paper also contains a new analysis of the binary tree mechanism with Gaussian noise, which turns out to have a better utility-privacy trade-off than some later refinements such as the smooth binary mechanism.
110
Rasmus Pagh @rasmuspagh.net · 12/06/2026
The idea is to simulate the distribution of the binary tree mechanism without generating all noise values. Only when accessing a sketch entry do we generate the noise, and with a suitable data structure of size O(log T) it turns out that this is possible in constant time!
110
Rasmus Pagh @rasmuspagh.net · 12/06/2026
The first paper co-authored with my student Sia Sejer is out! We show how to do continual observation of sketches (and other data structures) with only a constant-factor time overhead relative to the non-private versions.
240
Rasmus Pagh @rasmuspagh.net · 12/06/2026
Great invited talks by @ahonkela.bsky.social and @grahamrc.bsky.social on the Data Privacy in Machine Learning workshop's first day. Looking forward to day 2 which will focus on unlearning!
143
Rasmus Pagh @rasmuspagh.net · 01/06/2026
The program for our upcoming Workshop on Differential Privacy and Unlearning is now up on p1dpml.github.io/workshops/wo... Registration deadline is Sunday June 7.
p1dpml.github.io
Differential Privacy and Unlearning in Machine Learning | Data Privacy in Machine Learning
062
Rasmus Pagh @rasmuspagh.net · 11/05/2026
June 13-14, right after the workshop at nearby DTU, there will be a summer school on on strings & privacy featuring Teresa Steiner and Solon Pissis, ahead of CPM 2026. cpm2026.compute.dtu.dk/p/summer_sch...
cpm2026.compute.dtu.dk
- CPM 2026
000
Rasmus Pagh @rasmuspagh.net · 11/05/2026
The workshop is organized by @amartyasanyal.bsky.social, @claudiorlandi.bsky.social and myself. See the call for contributions as well as registration information on the workshop web page p1dpml.github.io/workshops/wo...
p1dpml.github.io
Differential Privacy and Unlearning in Machine Learning | Data Privacy in Machine Learning
120
Rasmus Pagh @rasmuspagh.net · 11/05/2026
Join us June 11-12 for a workshop on Differential Privacy and Unlearning in Machine Learning at University of Copenhagen! The workshop will feature tutorials, three great invited speakers (@grahamrc.bsky.social, @ahonkela.bsky.social and @koloskova.bsky.social), as well as contributed talks.
164
Rasmus Pagh @rasmuspagh.net · 11/05/2026
Great question! If I am not mistaken we could only hope to sample from discrete distributions using a Bernoulli Factory, so if this is the case there is no hope to get Laplace (unlike discrete Laplace which is easy to get). In any case, we are not aware of any link to Bernoulli Factories.
010
Rasmus Pagh @rasmuspagh.net · 08/05/2026
Our paper, also containing a bunch of other results not mentioned above, is now on arXiv: arxiv.org/abs/2605.06502
arxiv.org
Privacy by Postprocessing the Discrete Laplace Mechanism
We show that an "old dog", the classical discrete Laplace (aka.~geometric) mechanism, can "perform new tricks": 1. It can be post-processed to yield a simple, unbiased estimator of any subexponentia...
140
Rasmus Pagh @rasmuspagh.net · 08/05/2026
We also found out that it is possible to post-process the output of the discrete Laplace mechanism to exactly recreate the Laplace mechanism, using additive noise. So for integer data, adding discrete Laplace noise is strictly more general that Laplace noise (and the variance is better).
Illustration of the transformation from discrete Laplace mechanism to Laplace mechanism
130
Rasmus Pagh @rasmuspagh.net · 08/05/2026
The estimator works under minimal assumptions on f, which e.g. does not have to be continuous. The flip side is that the variance can sometimes be much higher than for the naive biased estimator that simply evaluates f(x*). But in settings where error is dominated by bias, error becomes better.
110
Rasmus Pagh @rasmuspagh.net · 08/05/2026
Last year Quentin Hillebrand, Jacob Imola, Sia Sejer and myself set out to study what happens if the noise is *discrete* Laplace, which is natural for integer vectors such as histograms. We found that a clean, closed-form, unbiased estimator for f(x) exists, even in the multivariate setting.
120
Rasmus Pagh @rasmuspagh.net · 08/05/2026
Back in 2023, Hillebrand, Suppakitpaisarn, and Shibuya had shown such a result for polynomials and applied it to local differential privacy algorithms for graph problems. Unlike Calmon et al., their method works for multivariate functions. dl.acm.org/doi/10.1145/...
dl.acm.org
Unbiased Locally Private Estimator for Polynomials of Laplacian Variables | Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining
110
Rasmus Pagh @rasmuspagh.net · 08/05/2026
Last year Calmon, Du, Dwork, and Finley presented a remarkable result: If x is released with the Laplace noise, giving a noisy value x*, it is possible to estimate f(x) in an unbiased way for any f that is twice differentiable and does not have too high growth rate. drops.dagstuhl.de/entities/doc...
drops.dagstuhl.de
Debiasing Functions of Private Statistics in Postprocessing
1101
Rasmus Pagh @rasmuspagh.net · 01/05/2026
More here: www.compute.dtu.dk/news/dtu-res...
compute.dtu.dk
DTU research receives SIGMOD Research Highlight Award 2026 for groundbreaking work in private data processing
Two DTU Compute professors, Inge Li Gørtz and Philip Bille, have received an international award as co-authors of a research paper recognised for its potential to significantly impact database and dat...
010
Rasmus Pagh @rasmuspagh.net · 29/04/2026
This is such a fun idea: An LLM trained only on data from before 1930. You can try it out here: talkie-lm.com/chat
User: Didn't Babbage try to mechanize computers?

Talkie-1930: Mr. Babbage did make an attempt to construct a calculating machine, which should perform arithmetical operations; but the project was ultimately abandoned. Computers are, consequently, still necessary.User: Would it be feasible to continue Babbage's project?

Talkie-1930: It would be possible to continue Mr. Babbage's project, but very improbable that it would ever be made practically useful. The great expense of such a machine would be an insuperable obstacle to its general adoption; and the probable errors to which all calculations must be liable, when performed by machinery, instead of by intelligent beings, would render the results unreliable.
6566
Rasmus Pagh @rasmuspagh.net · 28/04/2026
Interested in strings and privacy? It is worth your time! Don’t take my word for it, read the technical perspective by Grigorios Loukides dl.acm.org/doi/10.1145/...
dl.acm.org
Technical perspective on 'Differentially Private Substring andDocument Counting' | ACM SIGMOD Record
Differential privacy has become the de facto privacy standard, as it is widely employed by various organizations. However, only a few research works have focused on strings (e.g., [1, 2, 5]). This is ...
100
Rasmus Pagh @rasmuspagh.net · 28/04/2026
The privacy bug is spreading! Good colleagues from Technical University of Denmark and University of Southern Denmark featured in SIGMOD Record with a very nice paper on differentially private string data structures. dl.acm.org/doi/10.1145/...
dl.acm.org
A Differentially Private Data Structure for Substring and Document Counting | ACM SIGMOD Record
For databases consisting of many text documents, one of the most fundamental data analysis tasks is counting (i) how often a pattern appears as a substring in the database (substring counting) and (ii...
150
Rasmus Pagh @rasmuspagh.net · 27/04/2026
Congratulations to BARC alumnus Vincent Cohen-Addad and to @gautamkamath.com, very well deserved!
160