To get infinite bandwidth on a ratio, you need a Guanella TLT.

The principle is using TLs in series-parallel combinations at each port, so that the signal sent down each TL is exactly the right fraction of the total, in exactly the right impedance, so that the waves are put back together exactly in phase, no delay shift. Ultimate bandwidth is limited by the cross section of the TLs, the routing in the port area, and crosstalk between TLs (because, of course, no TL is ideal with infinite common mode impedance).

To get maximum (note, it will always be finite) bandwidth on an isolating transformer, you basically want 1:1. Other ratios are effective, but you will have some compromise between a Guanella style construction, a Ruthroff type (where fewer TLs are used and the delays are unmatched), and a conventional (not-TL-driven) design.

For example, consider the 1:N transmission line transformer, where a signal line is routed along a reference plane multiple times. The plane forms a slotted loop (it might be a cylindrical shell made of copper sheet or foil, or a planar circle on a PCB, say), while the signal forms a spiral or helix (or equivalent). The slot in the plane, in turn, connects to a parallel plate transmission line.

A cylindrical example looks like thus:

The "signal" line here is 1/8" (3.1mm) copper tubing (with sleeving), bifilar, inside a sheet metal surround. The transmission line impedance I would guess is on the order of 50 ohms (per signal line). The output line should be around 9 ohms.

The delay through the signal line is N turn lengths, while the delay through the plane is just one turn length. If a step change in voltage is applied to the plane, that voltage is induced on the signal line, simultaneously, at each point the line crosses the plane split. When this propagates out to the terminals, the result is a series of N steps (of complementary polarity from each end of the line, with respect to the reference plane).

The signal is always with respect to a reference plane, so it's balanced, and each end of the winding has that same reference. This is good when you need a balanced signal, but when you need isolation, you need imbalance -- the delay of the signal line corresponds to the equivalent isolation capacitance and leakage inductance in the transformer.

Any construction that is similar to this, will get you pretty good bandwidth.

You can apply the same theory to conventional (multilayer) windings, as well: in that case, each layer of turns is the reference plane for the layers nearby, and so on, so the characteristic impedance of a multilayer (per segment) winding can become quite high indeed (to first approximation, the impedance is the impedance of how many turns tall+wide it is). Predicting what happens in cutoff, in windings like this, who knows -- but estimating where the first cutoff is, isn't so bad!

Incidentally, the characteristic impedance of a common mode choke like this,

https://www.digikey.com/product-detail/en/taiyo-yuden/TLF9UA102W0R8K1/587-2788-ND/2573875is around 600 ohms. There's a few layers in there, and the two windings are rather poorly coupled to each other! The resulting bandwidth is even lower than you'd expect given the wire length (which probably isn't much on a small one like this, though I haven't counted). The isolation capacitance is quite low, though.

Tim