This is an old revision of the document!
Table of Contents
The Quantum Threat Against Cryptography is a Zero Gain Analog Computer
In 1996 Peter Shor released a computer science paper1) that challenged some widely held beliefs from that field with the aid of a theoretical computer based on quantum mechanics. I will refer to that computer as a “Shor's computer” here. That sparked interest in actually building such a computer. That interest has recently flared up to the point that fear of such a computer has infected public policy2). So the question of the feasibility of a Shor's computer is important.
I have heard that things like Shor's computer are now in the realm of mere engineering. This then is my response from the perspective of computer/electronic design. Unfortunately providing context is going to require a fair bit of exposition. Sorry about that. It now seems clear that Grover's algorithm is not a threat to cryptography3) so that leaves a Shor's computer as the only perceived threat. I will specifically be considering the threat from a Shor's computer against the particular bit of cryptography known as RSA2048. That does not mean that a Shor's computer intended to attack other forms of cryptography is feasible but I wanted a definite example.
You will not see the terms “quantum computer” or “qubit” outside of this paragraph as they beg the question I am addressing here by implying that such things could exist in some meaningful way. “Quantum computer” is a poor term anyway as it is used to distinguish from conventional computing. But conventional computing is based on the quantum field effect. You want to see a quantum computer? Just pull out your phone.
I confess that my understanding of the algorithm that would drive a Shor's computer is not complete. At some point in my research it became obvious that I was looking at what was probably the most unlikely proposal for a computing technology I had ever seen. The impracticality is nothing less than breathtaking. It did not seem to be worth my time to go further.
About Analog Electronic Computers
So what do I mean by an “analog computer”?
First a couple of basic electrical concepts:
- Electric pressure. The same idea as the pressure you measure in your bike tires. The unit of electric pressure is “volt” (abbreviated as “V”). The pressure itself is called voltage.
- Electric flow. When you run over a glass splinter with your bike, the flow of air from inside the tire to the environment is an analogy for electric flow. The unit of electric flow is “amp” (abbreviated as “A”). The flow itself is called current.
So electrical pressure (voltage) induces electric flow (current).
Let's consider the most basic analog computer operation: negation. We use voltage to indicate values. If we put in 6.32 volts we will expect -6.32 volts out. If we put in -8.37 volts we will expect 8.37 volts out. Here is my proposed circuit in schematic form (exact details not important):
The round device is a transistor. For this we can say that it converts a input voltage to an output current. The zigzag device is a resistor (“10k” means 10000). For this we can say it converts the current from the transistor back into a voltage that we can then output. This is kind of awkward, it would be better if the transistor controlled a voltage that we could output directly, but that is just how this type of transistor works. In electronics you have to design with the components that can actually exist, not the ones you wished you had. We will be using real devices (“BS138” is the name of a well known transistor for example).
The “+15V” represents a power supply of positive 15 volts as the “-15V” represents a power supply of negative 15 volts. We have negative voltages here because we defined a reference voltage of 0 volts halfway between the two power supplies. We call this 0V reference the “ground”.
I will be using an electronic simulator called ngspice4). I had to shift the input from the nominal +-10V to get any sort of meaningful output. The modified input range is from -15V to -5V. Here is the simulated result:
Simulated terrible analog negation
Well that isn't very good… The output is pinned at one extreme or the other. The problem here is something called “gain”. I calculated using the appropriate value off the BS138 datasheet that with the value of the resistor we used (10k) that the gain is -5000. So for every positive volt in we would expect 5000 negative volts out due to the gain. There are only +-15 volts available so we get that pinned at the extremes effect. We really want a gain of -1.
Obviously all we have to do is adjust the value of the resistor:
… which simulated as this:
simulated less terrible analog negation
The input range is still restricted to negative values but we are getting something that looks sort of OK for negation. The output lines are slightly curved (bad linearity) and the range is a bit offset and a bit too large but it seems that we have achieved some form of primitive analog computing.
Not so fast…
The worst case power absorbed by the transistor works out to 88 Watts. The worst case current through the transistor works out to 5.5 amps. This is what the transistor datasheet has to say about this:

BSS138 Absolute Maximum Ratings
The maximum current is 0.22 amp. That is considerably more than the 5.5 simulated amps. For that case, the datasheet says: “device functionality should not be assumed”. The device model for the simulator, which came from the manufacturer of the transistor, assumed that the current would not exceed 0.22 amps. So what the simulator produced was pure fantasy. It looked pretty definite on the plot, but that plot was just nonsense.
The maximum power as specified by the datasheet is 0.36 Watts. The simulated power is 88 Watts. In electronics the sound of the smoke detector is often an indication of a breach of a physical limit. The simulator did not take into account the fact that the transistor would quickly transmorph into smoke and flame.
This sort of error is very common when doing either computational or mathematical modelling … and yes, I am making a point here. Physical limits are important but can easily be overlooked.
Back in the vacuum tube (AKA valve) days it was common to add a resistor (“cathode resistor”) to move the operating curve to a place where things were more linear. It was known that the resistor would reduce the gain, but would improve the linearity more than expected. The effect was called “negative feedback”. It turns out that you can get more precision if you are willing to sacrifice gain. Well we are totally willing to sacrifice gain, we don't want any for our analog negator. So let's add a resistor:
… which simulates as:
Simulated better analog negation
I checked. Yes the transistor is working well within the physical limits. So this has the immediate advantage over the previous circuit of being potentially real.
We are still suffering from a restricted input range and we have about a volt offset. The output range looks dead on (10V) and the lines look quite straight (good linearity). Trading away gain with negative feedback seems to have provided definite benefits. It's possible to imagine that with a bit of care we could actually use this negation circuit for some sort of calculation.
We don't have to just use one transistor. Why not make a sort of ideal amplifier designed to work with negative feedback out of a bunch of components? Oh, and give it tons of gain… The result of that thinking produced the operational amplifier (an amplifier for doing arithmetic operations). That's a lot of syllables so everyone calls them “op amps”. Here is the schematic representation of the well known 741 op amp (details entirely unimportant):
In return for that complexity we get (among other things) a whole lot of gain (hundreds of thousands) that we can trade away for precision. When we trade off all that gain we also get much better resistance against external interference. All the negative feedback cancels out the interference. This is our negation using a 741 op amp:
Analog negation using an op amp
The symbol with the three parallel, unequal lines is the 0 volt ground reference as mentioned above. Here is the simulation output:
Simulated analog negation using an op amp
… which is visually perfect. I set 6.32 volts as the input for the simulation and got out -6.31692 which counts as correct. When I set -8.37 volts as the input I got out 8.37293 which also counts as correct. Using a state of the art op amp would likely improve the precision but keep in mind that in an analog computer, errors outside the feedback loop add up. So more operations mean more stages which means less precision. So 3 significant figures, perhaps 1 or 2 more, depending.
So after many decades of effort, the best we can do with this particular technology is 3 significant figures. Technologies usually have an ultimate capability that you can approach but not exceed.
I would like to specifically point out here that the progress that was made was mostly due to the use of gain and negative feedback.
About Digital Electronic Computers
Here we are talking about the sort of transistor based computing technology that runs most everything these days. Desktops, laptops, smart phones, … microwave ovens, …
Refer back to our first attempt at an analog computing function. That was where the output was pinned to one extreme or the other. Why don't we just go with that? Why bother with trying to make all those zillions of intermediate states be reliable, let's just have 2 states. The serious downside is that everything will have to be represented in terms of binary counts and we will need a whole lot of them. The 3 significant figures we were getting from a single channel of analog computing will now need 10 channels. Actually doing anything with these counts will require a large number of logical operations. So the electronics will have to be very cheap and very reliable.
Here is the circuit for an inverter:
All this does is switch one of the 2 states to the other state. The simplest possible logical operation. Contrast this with the circuit for the 741 op amp. We only have 2 possible states so we only need one positive power supply. So we will have to find some other way to deal with negative values. We don't care about intermediate values so we don't need as many volts from the power supply. The BSS84 is the reverse version of the BSS138. The fact that such a version can physically exist is why this is such a simple circuit. You can't create a reversed vacuum tube for instance. We got lucky.
For simulating this it makes more sense to use a square wave as the input. Oh, and faster too, we are going to have to do a lot more of these simple 2 state operations. Here is the result:
Since we are only concerned about 2 levels here this looks perfect. So let's break it:
Series inverters with interference
With 2 inverters in series we expect to get whatever we put in back out. The circle with the arrow in it is a source of current that interferes with the logic level. So we are trying to force the level to another bogus level that will cause the output to be different than the input.
I used a not very clever method of first determining the weakest transistor (BSS84), setting the input to a constant level (0V), and then manually increasing the level of triangular wave interference until I just barely got a failure that looked like this:
Simulated interference induced digital logic failure
Note that the interference tried to damage the transistor inputs by forcing the “mid” signal out of the normal range of the circuit but only got up to 4V and down to -0.7V. That is because of an intrinsic side effect ((body diode) of how this type of transistor is made. This protection wasn't planned, we just got lucky … and by “we” I mean everyone because this is a feature of contemporary digital logic.
I then copied the model files that represented the transistors and nerfed the transistors by reducing the gain by a factor of 10. Then I again determined what level of interference was required to cause a failure. The results:
| Gain | Interference (amps) |
|---|---|
| Normal | 0.118 |
| 1/10 | 0.018 |
So reducing the gain of the transistors by a factor of 10 caused the resistance against interference to reduce by a factor of 6.6. More gain means that the transistor has more capability to keep the level solidly pinned to the correct state. It's not obvious, but this is a form of negative feedback. The transistor resists any attempt to move the level out of the legitimate range by dynamically applying an opposing current.
Digital electronic computers are almost ridiculously robust and reliable. Vacuum tube computers were only unreliable because the tubes burned out. The processor in a contemporary smart phone has billions of transistors. But such processors rarely, if ever, make mistakes. You see error correction used for things like memory and communications, but never digital logic. This reliability comes from the fact that the transistors have gain.
Gain makes analog computers work. Gain makes digital computers work. We need a counterexample here:
Solid state diodes have been around forever. When my Father tickled the rock with a wire for his crystal radio he was effectively manufacturing his own diode. You can do digital logic with diodes. You know what no one has done with diodes? Created a digital computer that can do significant calculations. One important reason is because diodes do not have gain. Diodes have been used in general purpose digital computers but only in combination with either vacuum tubes or transistors to make up for the lack of gain.
You generally can't build a useful computer out of devices without gain. The complex mechanical computers of the mid 20th century had latches and clutches. Latches and clutches have gain. Relay computers existed in that same era. Relays have gain.
Shor's vs RSA2048
A Shor's computer critically depends on the properties of single atomic scale particles. When considering the feasibility of a Shor's computer, one aspect stands out. Single particles, and I can not emphasize this enough, do not and can not have gain. As a result, where conventional computing based on transistors is almost unimaginably robust and effective, computing based on single particles is almost unimaginably fragile and limited.
A Shor's computer depends on the properties of single particles, but the properties of those particles also work to prevent the construction of a Shor's computer in the first place.
A Shor's Computer is Made Out of Probability
While a electronic computer, either analog or digital, is made out of volts and amps, a Shor's computer is made out of the probability of the quantum state of individual particles. Computing would be a process of increasing the probability of the desired result at the expense of all the unwanted results. Here is what that looks like, with analog and digital computing for comparison:

Left: particle probability; middle: analog state; right: digital state5).
As with analog and digital computing, a Shor's computer would map data to a physical property. This property has an extra dimension and is often shown as a sphere. The two binary states are mapped as halves of the sphere. The extra dimension is the angle mapped to the 𝑥𝑦 plane (labelled as 𝜑 in the diagram). That angle is usually called “phase”. For binary operations, the phase is set and hopefully maintained at one of two opposite angles:
Binary phase convention (view looking down from the top of the sphere)
… but it is common to suggest that the phase could be used in a more analog way.
Superposition and Entanglement
Quantum superposition is why anyone cares about any of this. Superposition means that a single particle can be in an indeterminate probabilistic state that in some way reflects multiple states. Quantum entanglement links particles together. So the idea is that we could create a sort of digital quantum register where each bit corresponds to a single particle and the particles are linked together with entanglement. Then if you do operations on that register you are effectively doing operations on every possible state that register represents. That would be something like 22048 parallel operations for the case of RSA2048. So virtually infinite computation at a small fixed cost.
If you have had “The Talk”6) you will know that that interpretation is wrong or at least misleading. Instead you would be doing quantum interference in a way that could produce useful computation. You try to arrange things so that the value you want adds up and the values you don't want cancel out. This sort of cancellation puts this into the analog realm. Cancellation works perfectly in simulation/modelling but tends to do poorly in practice. Practical imprecision means that the cancellation is partial. I'm not a fan of the general idea.
What level of precision does the problem imply? The 2048 in RSA2048 is the number of bits that a Shor's computer would have to deal with. So one of the registers would be at least 2048 bits long. The register would be responsible for 22048 states. To be clear what is being suggested here I am going to present this number in full:
323170060713110073007148766886699519604441026697154840321303454275246551388
678908931972014115229134636887179609218980194941195591504909210950881523864
482831206308773673009960917501977503896521067960576383840675682767922186426
197561618380943384761704705816458520363050428875758915410658086075523991239
303855219143333896683424206849747865645694948561760353263220580778056593310
261927084603141502585928641771167259436037184618573575983511523016459044036
976132332872312271256847108202097251571017269313234696785425806566979350459
972683529986382155251663894373355436021354332296046453184786049521481935558
53611059596230656
That's a single 617 digit number. That's an unimaginably large number of states, it would only take something like 80 digits to represent every atom in the entire known universe. So suggesting that might be possible to pick one of two numbers out of all those states using some sort of reinforcement/cancellation scheme counts as an extraordinary claim. It has to be verified.
There is proof of mere entanglement in the 100 to 200 particle range. No superimposed calculation, just proof of some sort of entanglement. A Shor's computer would at the very least need thousands of superimposed and entangled particles that you could actually do something with. Actual physical demonstrations involve few particles, mostly 1 or 2.
The important point here is that the fundamental idea that makes Shor's work has not been validated, and it really needs to be validated. We can't directly extrapolate the behaviour of two particles to thousands of particles and expect to get a valid result.
A Shor's Computer is a Reversible Computer
A computer scientist/engineer reading about the virtually infinite amount of computing mentioned in the previous section might be thinking that it would be straight up impossible for purely thermodynamic reasons. There is a minimum amount of energy required for a computation. This would violate that particular physical limit to an insane degree.
If you believe a Shor's computer could work then you also believe that it is a completely reversible computer.
The dream of reversible computing, perhaps a faded dream at this point, has been around pretty much as long as there has been computation. This huge breakthrough is just assumed for the Shor's case. The critical superposition and entanglement would have to be preserved in such a reversible computer so a Shor's computer represents at least one fundamental breakthrough but might represent up to three.
Superposition and Entanglement are Extremely Fragile
Assume you have just constructed a completely reversible Shor's computer out of superimposed and entangled particles. If you want to do anything with it you had better hurry. It won't be around very long.
Superposition and entanglement come with a time limit called decoherence. The popular superconduction based scheme for example only lasts microseconds.
It's as if there was an industrial process that was made out of components that would spontaneously burst into flame. You would have to build your factory very quickly and get as much done as possible before your factory burned down. That is the basic idea behind a practical implementation of a Shor's computer. You build it as fast as you can with the hope that you can complete the last of it before the first part you built degrades enough to prevent the computation.
What sorts of things might prematurely destroy your Shor's computer? Pretty much anything:
- Electric fields.
- Magnetic fields.
- Heat.
- Other particles.
- Light (photon particles).
- Vibration.
Simply attempting to get the result of the calculation will break your Shor's computer. So the proposed method is to create the computer as fast as possible and then destroy the computer by taking the output as the final step, hoping that nothing went wrong in the meantime.
This extreme fragility is why single particles can not have gain. Any attempt to add energy will destroy the superposition and entanglement that represent the very data that we are interested in.
In terms of regular quantum physics this is not all that well understood. What is known is that if you try to build a Shor's computer you will get nothing but useless garbage out in the end. It is hoped that this is because of external interference which we could potentially control, but we don't really know.
Regeneration
A contemporary computer uses the gain of the transistors to regenerate the level at each stage. So based on the input level(s) a whole new level is created. Telegraph systems regenerated the signals using the gain of telegraph relays. It's not a new idea. Particle based computers such Shor's don't have access to such advanced 19th century technology due to the quantum no cloning principle. This turns out to be a huge limitation. It means that every intermediate state retains echos of everything that has happened to the particle since initialization. Like an analog system…
Shor's ... the Multiply/Modulus Step
The genius of the Shor's algorithm is that it manages to do something useful in the face of the extreme limitations of the underlying single particle based computing hardware. Here is a really brief description expressed as pseudo code:
Q[0] = 1
for i = 0 to {enough}
Q[i+1] = (Q[i] * A) mod P
endfor
… where P is the RSA public key we are attempting to break and A is a number. Taking into account some minor restrictions, A can be most any number. So we create a series of values by taking the last value, multiplying it by A and taking the result modulus P. That value becomes the new last value and we repeat. Let's use the smallest possible RSA key (P=15) and A=2 as an example which gives us this:
'1 2 4 8 1 2 4 8 1 2 4 8 1 2 4 8 1 2 4 8 …'
The significant aspect of this is that it repeats every 4 values. That periodicity can be then used as a hint to break the RSA key using conventional computing. Details not important.
Quantum Error Correction
The recent fear that a Shor's computer might break all the cryptography comes from the idea that quantum error correction will make such a computer feasible. That perhaps we can use redundancy to improve reliability. Let's first talk about conventional error correction:
The cell tower sends some data to your phone. Something goes wrong and it is modified during the journey. The error correction then uses the redundant data sent along with the data you actually care about to recreate what was originally sent. Consider that data at that exact point in time. For all practical purposes it is the data. All the effects of the unfortunate trip have been eliminated. That data can be sent along to the rest of the phone as if nothing untoward had happened. That is because we are error correcting a system with regeneration. Everything is recreated new at each stage of the process. Note that we did everything at the digital level. That is the great advantage here, we don't have to consider the possibly complex analog processes that created the error in order to remove the effects of that error.
That was boring old regular error correction. Quantum error correction is a radically different thing…
Shor's is inherently analog. So you *do* have to consider the analog processes. Shor's can't do regeneration. So we can't actually “correct” errors in the way that we can with gain based computing. Instead we have to somehow create a complete system with a lower chance of error at the end of processing. Oh, and we have to correct errors in phase as well. There have been many proposals for how to do this. Progress has been very slow. The point here is that applying error correction to a Shor's computer is much harder than to, say, a communications channel. The problem likely gets harder when you attempt to scale the system up.
Let's say we have manged to completely error correct the Shor's multiply/modulus step. Then we get to the next step of the process…
Shor's ... the Quantum Fourier Transform step
We have already discussed how Shor's generates a periodic series. Now we must discuss how Shor's detects the periodicity. The series is treated as a sort of a signal and a single particle method called the quantum Fourier transform (QFT) is used to detect the period. Our signal for our P=15 example looks like this8):
Output of the multiply/modulus step for P=15 and A=2
Human hearing is good at detecting the periodicity of a signal so we will use that as an analogy to the QFT. This is what the signal sounds like:
Output of the multiply/modulus step for P=15 and A=2 in audible form
You can clearly hear a dominant tone. That tone represents the periodicity. Let's listen to a larger RSA public key (P=1261, A=14):
Output of the multiply/modulus step for P=1261 and A=14 in audible form
… which looks like this:
Output of the multiply/modulus step for P=1261 and A=14
A lower frequency tone is clearly audible from what seems to be a complex waveform.
Part of the QFT involves “controlled phase rotations”. An attribute of those rotations is that the required precision for the rotation doubles with each bit. So for 2048 bits:
Required precision for 2048 QFT
Refer back to where I wrote out 22048 in full. This level of precision is obviously completely impossible. Because of the use of the QFT, Shor's counts as not just analog, but very analog.
There have been various proposals over the years for reducing the required precision9). Ultimately though, you need enough precision or time (there is a tradeoff) to get the data you are seeking into the process in the first place. These, and the original QFT idea, are based on mathematical modelling. This is very much the sort of thing that works in the model but not in reality. The mathematical model of our analog negator is 𝑦= -𝑥 which technically has infinite precision but the reality is much different. Verification is desperately required here but it seems that none of these proposals have been verified experimentally at any reasonable scale.
Let's imagine that we could get the precision requirement all the way from exponential to linear … an unlikely amount of improvement. So the required precision for RSA2048 would be 1 out of 2048. That's still 4 significant figures. That would seriously challenge the capability of electronic analog computing which has the advantage of a tremendous amount of gain traded for precision with negative feedback. But we are proposing to do this inherently analog operation with single particles which, again, have no gain to trade.
The particles used for the QFT step are not any more reliable than the ones we used for the multiply/modulus step. Even if we somehow resolved the precision problem, we would still be left with the unreliability problem. So error correction again?
The catch here is that the sort of error correction we have been talking about up to now depends on the fact that there are few possible states (2 for the case of binary). That error correction is a process of moving the degraded state to the most likely state. That doesn't work for something with lots of possible states like the QFT. You might try some redundancy. Do the process a bunch of times and average the outputs. But that wouldn't work for an analog electronic computer…
If you actually tried it, you would find that all the parallel analog computers were wrong the same way. All the components of the analog computer would be the same while existing in the same environment. You can't average out errors if the errors are all more or less the same. Our atomic scale particles very much exist in the same environment. This is another in the list of things that would have to be verified before we could consider that a Shor's computer might be real.
That list:
- That it is possible to create a useful computer using zero gain components.
- That we can get virtually infinite computing from the quantum realm with the superposition and entanglement of thousands of atomic scale particles.
- That it is possible to create a completely reversible computer that preserves superposition and entanglement.
- That we can do quantum error correction at a sufficient scale to make the multiply/modulus step of a Shor's computer work.
- That we can overcome the problem of precision in the quantum fourier transform step of a Shor's computer.
- That we can correct errors in the quantum fourier transform step of a Shor's computer.
In the face of all this, a functional Shor's computer would require a tremendous amount of luck. It is often not possible to prove a technology impossible when the basic principles that technology is based on are poorly understood as is the case here. It is sometimes possible to show that such a technology is extremely unlikely.
Some Arguments Directly Addressed
A Shor's computer has broken 15.
The idea that someone has broken the smallest possible RSA public key originally came from a paper called “Experimental realization of Shor's quantum factoring algorithm using nuclear magnetic resonance”10). That claim is fake11). It includes a “compilation” that reduced the problem to something trivially solvable with something completely unlike a Shor's computer (a single molecule consisting of 7 atoms). An actual Shor's computer to break 15 would at the very minimum require the interaction of thousands of particles12).
That paper may have inspired a whole series of fake claims. While the tortured logic can be a source of amusement13) that doesn't excuse the questionable ethics. There is an obligation to present the result of research so that it is not likely to be misunderstood. That is particularly true where the misunderstood result represents a danger to the general public. A public that would be under the false impression that a Shor's computer can exist and would only need to be scaled up to represent a threat to cryptography.
I can make a simple and definite statement here:
Any claims before November 2026 of breaking RSA with a Shor's computer are fake.
Shor's capability will sharply accelerate like conventional digital computing did in the past.
If I have achieved only one thing here I hope it is the understanding that a Shor's computer is nothing like a conventional digital electronic computer. Any comparison between the two is meaningless. A Shor's computer is more like an analog electronic computer. Analog computing over a period of decades reached a fairly mediocre level of performance that could not be exceeded.
Conventional computing had a noise problem just like Shor's.
I have seen this stated or implied in more than one place. I suppose it might be true in some limited technical sense. Basically when devices with gain became available then digital electronic computers suddenly became practical. Then it actually was just a matter of engineering. That's when the progress became rapid. So the comparison is misleading as single particles do not and can not have gain.
Shor's capability will sharply accelerate due to progress in quantum error correction.
The entire quantum threat against cryptography hinges on the feasibility of error correcting a Shor's computer enough that it might work. A recent and famous demonstration shows that it is possible to error correct a single particle enough to consider it a memory element14). That's it. Not even a single logical operation. No useful superposition. Obviously no useful entanglement either, that would require at least one more particle to entangle the first one with. It is hard to think that there is some sort of momentum here.
There is no reason to believe that any of the various ideas for error correction might work for a complete Shor's computer at this point. Again, there are two different problems to address at the same time, the multiply/modulus step and the quantum fourier transform step.
Consider this other thing that seemed impractical but later became practical.
There seems to be some sort of fallacy of inevitable progress at work here. That there is an inevitable march towards greater practicality. That has not been historically true even for ideas that seemed plausible. The vast majority of ideas are never realized. This argument seems to be an example of survivorship bias.
Rollout of post quantum encryption should be expedited due to the impending creation of a Shor's computer.
The idea behind post quantum (PQ) encryption is that we can come up with methods that a Shor's computer could not break. This seems to involve developing already understood methods that never became popular in day to day cryptography due to various disadvantages. One of these disadvantages is length. Everything tends to be significantly bigger15):
| Type | Public Key | Secret Key | Ciphertext | Signature | |
|---|---|---|---|---|---|
| cur | RSA2048 | 2048 | 2048 | 2048 | 2048 |
| cur | X25519/Ed25519 | 256 | 256 | 256 | 512 |
| PQ | ML-KEM-512 (Kyber) | 6400 | 13056 | 6144 | |
| PQ | ML-DSA-44 (Dilithium) | 10496 | 20480 | 19360 |
A comparison of lengths; cur: currently used methods; PQ: post quantum methods (units of bits)
RSA is known for being large. The leading post quantum methods are significantly larger than RSA. I have included the lengths of the popular X25519/Ed25519 methods for reference.
The other disadvantage has turned out to be security. One proposed PQ method turned out to be completely insecure after it had become a candidate for standardization16). Another turned out to be significantly less secure than expected after it had been listed as a national recommendation17). What seems to be happening is that due to the sense of crisis associated with the possibility of a Shor's computer, the cryptography researchers have not been given enough time to do their work.
So it seems that we now have both a definite security crisis and a definite length crisis prompted by a crisis with a very low chance of ever existing. The reaction to the PQ security crisis has been “hybrid PQ cryptography” which involves continuing to use current cryptographic methods in combination with PQ methods so that if the PQ method is broken the current method will still be effective. This strategy reflects the low probability of the Shor's threat relative to the threat of a new PQ method being found to be insecure. Using both methods at once would of course significantly increase the complexity of the system.
The reaction to the length crisis involves a whole lot of extra complexity18). All this extra complexity is very likely to increase attack surface and decrease security. So the cost of post quantum cryptography is fairly high: inefficiency due to the extra length and dual methods, less secure systems due to the added complexity.
We really need to step back, take a breath, and take a more measured approach here. There is no real crisis in the first place. Even if computation with single particles becomes a thing, chances are that what was learned along the way to that result will mean that some algorithm other than Shor's would end up being the threat against cryptography. That is normally how this sort of thing works, the algorithm has to take into account the limitations of the hardware. The current PQ cryptography strategy is based on an algorithm for hardware that doesn't exist.
The current post quantum cryptography approach seems ill advised.
Because of “harvest now, decrypt later” we should still consider the possibility of a Shor's computer an imminent crisis.
“Harvest now, decrypt later” expanded:
There might be entities out there collecting encrypted data, say, off the internet. Those entities can sit on their surreptitious archive of your data indefinitely. When a Shor's computer becomes viable they can then decrypt it all. The longer you wait the more data they will get.
This makes me think of Pascal's wager. It involves a decision based on an unknowable factor: if a Shor's computer is possible or not. The wager decision would be to implement PQ encryption or not. If you choose to transition to PQ cryptography and Shor's never comes to be, then you will incur the cost of the extra complexity and loss of efficiency indefinitely. If you don't transition to PQ cryptography before the age of Shor's, then you incur the risk that an attacker had the motivation and the means to create the surreptitious archive and that the old secrets contained in the archive could harm you in some way. The decision table looks something like this:
| Shor's will exist | Shor's will not exist | |
|---|---|---|
| Wager for Shor's | status quo | loss of efficiency and security indefinitely |
| Wager against Shor's | possible exposure of past secrets | status quo |
The cost of the loss of efficiency and security is practically unbounded. The cost due to the risk of exposed secrets is finite and for most people is quite low. If you believe as I do, that the chance of a working Shor's computer is very low, then the decision is simple: you would avoid a transition to PQ cryptography. You would only consider the transition if you thought the chance of a working Shor's computer was significant.
“Harvest now, decrypt later” only applies to encryption. It does not apply to authentication. But yet there are PQ authentication methods mandated by the various national regulatory agencies. That would only make sense if there were some definite indication that a Shor's computer could actually be built.

